Integration of artificial intelligence and wearable technology in the management of diabetes and prediabetes

Raphael A Fraser; Rebekah J Walker; Jennifer A Campbell; Obinna Ekwunife; Leonard E Egede

doi:10.1038/s41746-025-02036-9

. 2025 Nov 18;8:687. doi: 10.1038/s41746-025-02036-9

Integration of artificial intelligence and wearable technology in the management of diabetes and prediabetes

Raphael A Fraser ^1,^✉, Rebekah J Walker ¹, Jennifer A Campbell ¹, Obinna Ekwunife ¹, Leonard E Egede ¹

PMCID: PMC12627454 PMID: 41254217

Abstract

Artificial intelligence and wearable technology are increasingly used in healthcare and hold significant potential for improving the management of diabetes. Wearable devices enable continuous monitoring and real-time data collection, supporting AI-driven personalized interventions. This systematic review evaluated peer-reviewed studies that examined the integration of AI and wearable technology in diabetes management, with a focus on clinical and self-management outcomes. Sixty studies were included following a review of over 5000 records. AI models paired with wearable devices showed promise in glycemic monitoring, adaptive insulin management, and predicting diabetes-related events. Continuous glucose monitors and other wearables also enhanced self-management and informed clinical decision-making. However, key challenges persist, including limited demographic diversity, variable data quality, a lack of standardized benchmarks for evaluating AI performance, and limited interpretability of complex models. Future research should prioritize improving model transparency, addressing demographic disparities, and establishing clear benchmarks to support equitable and effective implementation in diabetes care.

Subject terms: Biotechnology, Metabolic disorders, Diabetes

Introduction

Diabetes mellitus, a chronic disease characterized by hyperglycemia, altered metabolism, and complications affecting both microvascular and macrovascular systems, represents a significant and escalating public health challenge^1,2. Current estimates place global prevalence at over 800 million people, with projections suggesting this figure could rise to 1.3 billion by 2050^3–5. More than 90% of these cases are attributed to type 2 diabetes (T2D), emphasizing the urgent need for innovative, effective management strategies that can meet the diverse and evolving needs of this population^4,6.

The emergence of artificial intelligence (AI) and wearable technology has revolutionized healthcare, offering innovative solutions for diabetes management⁷. AI, broadly defined as the ability of computer systems to perform tasks typically requiring human intelligence, is playing an important role in healthcare⁸. AI-powered systems can analyze vast amounts of data, identify patterns, and make predictions, enabling them to support clinical decision-making⁹, predict outcomes, and personalize treatment approaches¹⁰. This ability to process and interpret complex datasets is particularly valuable in diabetes management, where individual patient needs and responses to treatment vary widely⁹.

Wearable devices, such as continuous glucose monitors (CGMs), smartwatches, and other sensors, are becoming increasingly important in capturing real-time data for individuals living with diabetes⁸. These devices provide continuous monitoring of physiological parameters⁸, allowing individuals to gain insights into their interstitial glucose levels and make informed decisions about their lifestyle choices¹¹. CGMs have revolutionized diabetes management for individuals with type 1 diabetes and are increasingly being adopted by people with type 2 diabetes as an effective tool for real-time glucose monitoring.^12–15. They provide real-time interstitial glucose readings, enabling individuals to adjust insulin doses, dietary intake, physical activity, and other lifestyle factors to prevent hypoglycemia and hyperglycemia¹⁶.

While the use of wearable technology in diabetes management is expanding¹⁷, and numerous studies have explored the use of AI and wearable devices in healthcare^9,18–20, the integration of AI with these devices, particularly for T2D, remains relatively unexplored^8,9. This gap in research is significant because AI has the potential to significantly enhance the effectiveness of wearable devices^8,13. By analyzing data from wearable sensors, AI algorithms can provide personalized insights, predict interstitial glucose fluctuations, and even suggest dietary and lifestyle adjustments^11,21. AI-powered systems can also be used to automate insulin delivery, reducing the burden on individuals with diabetes and improving effectiveness of treatment plans^10,22.

While prior reviews have examined either AI applications in diabetes care or the role of wearable technologies independently, this review is, to our knowledge, one of the first to systematically evaluate the intersection of artificial intelligence and wearable technology specifically T2D management. Our review uniquely focuses on AI models that operate on physiological data collected from wearable devices—such as CGMs, fitness trackers, and smartwatches—to improve glycemic prediction, clinical decision support, and self-management outcomes. We also employed a novel synthesis framework to extract and analyze key dimensions across included studies: study population characteristics (including demographic diversity and metabolic profile), AI model type and interpretability, sensor modality and fusion, and clinical endpoints. This enables a more granular and equity-focused evaluation than prior reviews. This review focuses specifically on empirical studies where artificial intelligence models were applied to physiological data collected from wearable devices, such as continuous glucose monitors and smartwatches, for clinical or self-management outcomes in individuals with T2D. Studies that involved mobile health tools or digital coaching platforms without a direct AI modeling component were not included.

This systematic review aims to assess the current state of research on the integration of AI and wearable technology in T2D management, highlighting potential benefits and challenges while identifying future directions for research and development.

Results

Study selection

Figure 1 illustrates publication by year, indicating a steady increase in AI-related manuscripts primarily focused on T2D over the past decade, with significant growth beginning in 2022. Early years featured sparse publications, predominantly centered on interstitial glucose prediction, while later years showed diversification into insulin management and classification tasks. In subsequent years, the scope expanded significantly, with interstitial glucose prediction remaining dominant, accompanied by growing contributions from classification tasks and insulin management. Other objectives included detecting and classifying physical activity, evaluating diabetic retinopathy using CGM data, estimating stress levels based on several physiological parameters, and assessing the impact of CGM sensor location on glucose forecasting errors. Overall, there was a rapid growth and diversification of AI applications in diabetes research, reflecting a shift toward more balanced and varied study objectives in recent years.

Fig. 1 — Trends in study objective across publication years.

Study characteristics

Studies included in the review demonstrated a range of wearable technologies and diverse applications of AI models for managing T2D. Key attributes such as study design, geographic location, population demographics, wearable device types, AI architectures, model performance, and interpretability measures were summarized in Tables 1, 2. Sixty-seven percent (40 of 60) of the studies were observational or experimental in nature, with 20% (12 of 60) consisting of prospective observational studies focused on real-time data from wearable devices in naturalistic settings. Thirteen percent (8 of 60) of studies employed randomized controlled trials or non-randomized experimental designs to evaluate interventions involving CGMs and wearable activity trackers.

Table 1.

Summary of study characteristics

Ref. #	Author	Title	Study objective	Mean age or age range	Percent female or Sex Distribution	Study population size	Glycemic Profile	Metabolic Status	Race/Ethnicity	Reported Comorbidities	Country of Data Source	Data Source	Dataset Type	Study Duration	Study Setting
¹¹	Anjum, 2024	Optimizing type 2 diabetes management: AI-enhanced time series analysis of continuous glucose monitoring data for personalized dietary intervention	To forecast interstitial glucose values using AI models and provide personalized dietary interventions to manage T2D.	40, 45 (two patients reported)	NR	8 T2D patients, 10,160 CGM readings total	Mean glucose: 8.295 mmol/L; Min: 3.4, Max: 19.9; SD: 2.584	T2D	NR	NR	UK	CGM data collected from FreeStyle Libre system	Private	Mar 2021 - Dec 2021 ( ~ 9 months)	Real-world
¹⁶	Vettoretti, 2020	Advanced Diabetes Management Using Artificial Intelligence and Continuous Glucose Monitoring Sensors	To review AI methods integrated with CGM for DSS in T1D: bolus calculation, BC parameter tuning, and glucose prediction	NR	NR	10 clinical participants; 20–100 simulated subjects	NR	T1D	NR	NR	Italy	ABC4D; UVa/Padova T1D Simulator	Private	NR	Mixed
²¹	Ribeiro 2024	A Novel AI Approach for Assessing Stress Levels in Patients with Type 2 Diabetes Mellitus Based on the Acquisition of Physiological Parameters Acquired during Daily Life	Develop an AI-based wearable system for stress classification in T2D patients using physiological indicators.	Mean (SD) Min-Max: Years 42 (17) 12-75	47% F	128	HbA1c: 7.9 (5-12) BMI (kg/m2): 35 (10-67), BP: 62% High, 30% Normal, 8% Other FBS: 163 (80-280) FH T2D: 40% Yes Smoke: 62% Yes	T2D	NR	Hypertension	N/A	Public dataset of 128 diabetic patients [Kaggle]	Public; in silico	N/A	N/A
⁵⁷	Zhu, 2024	Population-Specific Glucose Prediction in Diabetes Care With Transformer-Based Deep Learning on the Edge	To develop a low-power, population-specific, multi-horizon glucose prediction model deployable on wearable devices	NR (varies by dataset; age used as input)	NR	124 total (24 T1D, 100 T2D)	RMSE, MAE, MAPE, gRMSE reported; example RMSE: 14.7-23.5 mg/dL	Mixed (T1D + T2D)	NR	NR	UK, China	OhioT1DM, ShanghaiDM	Public	8 weeks (OhioT1DM), 14 days (ShanghaiDM)	Real-world
⁵⁸	Zhu 2024	Multi-Horizon Glucose Prediction Across Populations with Deep Domain Generalization	Develop a generalizable, multi-horizon glucose prediction model for diverse populations using CGM data	Mean (SD): Years REPLACE-BG: 44 (14), OhioT1DM: 50 (10), ABC4D: 47 (17), GVAS: 74 (12)	REPLACE-BG: 112 (50%) F, OhioT1DM: 5 (42%) F, ABC4D: 12 (55%) F, GVAS: 18 (40%) F	REPLACE-BG: 226; OhioT1DM: 12; ABC4D: 22; GVAS: 45	Mean (SD): mg/dL REPLACE-BG: 160 (26), OhioT1DM: 162 (20), ABC4D: 178 (18), GVAS: 140 (48)	T1D; T2D	REPLACE-BG: ~90% White, OhioT1DM: NR, ABC4D: NR, GVAS: NR	GVAS: Stroke	REPLACE-BG: US, OhioT1DM: US, ABC4D: UK, GVAS: UK	REPLACE-BG, OhioT1DM (public); ABC4D, GVAS (proprietary)	Public: OhioT1DM; Private: REPLACE-BG, ABC4D, GVAS; all in vivo	3-day CGM segments per participant used for modeling; dataset durations vary (72 h to 26 wks)	REPLACE-BG: Controlled, OhioT1DM: Controlled, ABC4D: Controlled, GVAS: Real-world
⁵⁹	Warren, 2024	A Scalable Application of Artificial Intelligence-Driven Insulin Titration Program…	To assess the scalability and effectiveness of AI-driven insulin titration (d-Nav) in real-world T2D management	67.1 ± 11.5 years	52.8% male	600 patients	Baseline HbA1c: 8.6% ± 2.1%; reduced to 7.3% ± 1.2%; TIR improved from 47.7% to 65.4% in CGM subgroup; hypoglycemia <54 mg/dL: 0.4–0.6/month	T2D	46.2% non-Caucasian; 37.2% African American; 53.8% Caucasian	66.3% hypertension, 64.8% dyslipidemia, 36% CVD, 20.5% nephropathy	USA	Physicians East, North Carolina	Private	Oct 2022-Sep 2023	Clinical
⁶⁰	Wang 2023	A novel hypoglycemia alarm framework for type 2 diabetes with high glycemic variability	Develop a hypoglycemia early alarm framework for T2D patients with high glycemic variability using dynamic multi-scale features	NR	NR	204	Mean (Range): mg/dL Train: 158 (40-382) Test: 164 (40-400) CV of 0.36 or greater	T2D	NR	NR	China	Shanghai Sixth People’s Hospital	Private; in vivo	36 hours; 3–5 consecutive days of CGM data per participant	Hospital-based (real-world)
⁶¹	Tucker, 2024	Neural Networks With Gated Recurrent Units Reduce Glucose Forecasting Error Due to Changes in Sensor Location	To examine whether GRU neural networks can reduce glucose forecast error caused by changes in CGM sensor location	Mean 60.4 years (range 45–79)	6 females, 7 males	13 participants	RMSE, MARD, glycemic standard deviation	T2D	NR	NR	USA	Clinical study, Univ. of Minnesota (NCT03481530)	In vivo	15 weeks	Real-world
⁶²	Tao 2023	A Double Deep Latent Autoencoder for Diabetic Retinopathy Diagnose Based on Continuous Glucose Sensors	To diagnose diabetic retinopathy using CGM data with a deep latent autoencoder framework	NR	NR	26 individuals	NR	T2D	NR	NR	Iran	Sina Hospital, Tehran University	Private	NR	Clinical
⁶³	Shuzan 2024	QU-GM: An IoT Based Glucose Monitoring System From Photoplethysmography, interstitial Pressure, and Demographic Data Using Machine Learning	Develop a non-invasive, real-time wearable system for glucose estimation and severity classification	Median (Min-Max) 43 (13–87) years	42% F	139	Median (Min-Max) Glucose (mg/dL): 118 (66-600), BMI (kg/m2): 27 (17-46), SBP: 135 (90-234) DBP: 93 (60-169) Pulse Rate: 79 (53-128)	Diabetes 50%, Healthy 50%	NR	NR	Qatar	Qatar Diabetic Society; IRB-HMC-2021-011	Private; in vivo	Single visit (spot measurement)	Controlled/laboratory setting
⁶⁴	Seo, 2024	Generative Adversarial Network-Based Data Augmentation for Improving Hypoglycemia Prediction	To improve hypoglycemia prediction performance using GAN-based synthetic data augmentation in training data	NR	NR	10 patients (over 86,000 data points)	Baseline accuracy improved by 4-6% post-GAN augmentation	T1D	NR	NR	China	Jiangsu Province Hospital (NR dataset)	Private	10 days per patient	Clinical
³⁰	Ramazi, 2021	Predicting progression patterns of type 2 diabetes using multi-sensor measurements	To predict future HbA1c, HDL, LDL, and triglyceride levels in T2D patients using wearable and static data	33–78 years	19 female, 31 male	54	CGM-derived RMSE metrics	T2D	NR	NR	USA	Christiana Care Health System (Delaware)	Private	1 year (7-day wear + 1-yr follow-up)	Real-world clinical
⁶⁵	Yin, 2017	A Health Decision Support System for Disease Diagnosis Based on Wearable Medical Sensors and Machine Learning Ensembles	To develop a hierarchical, closed-loop health decision support system (HDSS) combining wearable sensors and CDSSs to enable disease diagnosis and tracking	NR	NR	6 diseases analyzed (datasets ranged from 120–3163 records)	Included for T2D only	Mixed	NR	Multiple (e.g., T2D, arrhythmia, hypothyroid)	USA	Public UCI Machine Learning Repository, literature-based private datasets	Public	Cross-sectional	Real-world and simulated tiers
⁶⁶	Kim 2024	Impartial feature selection using multi-agent reinforcement learning for adverse glycemic event prediction	Develop a deep learning model for predicting adverse glycemic events in hospitalized T2D patients using CGM, EMR, and MARL-based feature selection	Range: 20–90 Years	39% F	102	CGM: 40-400 mg/dL	T2D	NR	NR	South Korea	Soonchunhyang University Cheonan Hospital, IEEE Dataport	Private; in vivo	CGM: 7-10 days per participant EMR: April 2019 to January 2022	Real-world (hospital-based)
⁶⁷	Kashif 2024	GLSTM: On Using LSTM for Glucose Level Prediction	To predict glucose levels in individuals with prediabetes using a personalized Long Short-Term Memory (LSTM) model trained on multimodal wearable sensor data	35–65 years	100% F	11		Prediabetes	NR	NR	US	BIG IDEAs Lab dataset (PhysioNet)	Public; in vivo	8–10 days per participant	Real-world
⁶⁸	Kannenberg, 2024	Personalized Lifestyle Therapy for Type 2 Diabetes Through a Predictive Algorithm-Driven Digital Therapeutic	To evaluate the clinical impact of a CGM-AI-powered digital therapeutic on glycemic management and weight in non-insulin-treated T2D patients	31-78 years (mean 55.5)	58% female (n = 69), 38% male (n = 45)	118 participants	HbA1c: mean 7.46% (SD 0.93), range 6.5–11.0%; SDs for glucose/BMI included	T2D	NR	64% had comorbidities; HTN, dyslipidemia, mental & CV disorders	Germany	Direct-to-patient recruitment	Real-world	6 months (2–3-month phases)	Free-living
⁶⁹	Hotta 2024	Optimizing postprandial glucose prediction through integration of diet and exercise	To improve postprandial glucose prediction in GDM patients by integrating diet and exercise data using Bayesian transfer learning informed by randomized controlled trial data.	18–45 years	100% F	68	NR	Gestational diabetes	NR	The study included only GDM patients without major comorbidities	Finland	eMoM GDM trial & separate healthy-subject RCT data	Private; in vivo	GDM data: 3-day sessions collected monthly during pregnancy Healthy controls: 6-day data collection per participant	Real-world
⁷⁰	Metwally, 2024	Predicting Type 2 Diabetes Metabolic Phenotypes Using Continuous Glucose Monitoring and a Machine Learning Framework	To identify distinct metabolic subphenotypes in individuals with normoglycemia or prediabetes using OGTT-derived CGM data and machine learning to inform precision prevention of T2D.	Mean ~55 years	50% F	56 total (32 training, 24 validation, 29 CGM)	Normoglycemia or prediabetes	Pre-diabetes and early T2D	74% Caucasian, 27% Asian	NR	USA	Stanford CTRU & at-home OGTT	In vivo human trial	Single 10-day CGM session	Clinical + At-home
⁷¹	Goncharov 2024	Insertable Glucose Sensor Using a Compact and Cost-Effective Phosphorescence Lifetime Imager and Multi-Parametric Data-Driven Calibration	Demonstrate feasibility of a low-cost, compact phosphorescence lifetime imager integrated into an insertable glucose sensor system	N/A	N/A	no human/animal participants. The glucose sensor was tested in phosphate-buffered saline (PBS) glucose solutions ranging from 0 to 30 mM.	N/A	N/A	N/A	N/A	Finland	Sensor and animal lab data	Private; in vitro	Short-term sensor evaluation (hours)	Controlled laboratory; experimental study
⁷²	Dénes-Fazakas 2024	Physical Activity Detection for Diabetes Mellitus Patients Using Recurrent Neural Networks	Detect physical activity in T1DM patients using CGM, heart rate, and step data with RNNs	Mean (SD): years 50 (10)	OhioT1DM: 5 (42%) F	12	Mean (SD): mg/dL 162 (20)	T1D	NR	NR	US	Ohio T1DM dataset	Public; in vivo	8 weeks per participant	Controlled
⁷³	Dave, 2024	Detection of Hypoglycemia and Hyperglycemia Using Noninvasive Wearable Sensors: Electrocardiograms and Accelerometry	To detect hypoglycemia and hyperglycemia using noninvasive ECG and accelerometer data as alternatives to CGMs	27-58 years, mean 42.6	3 men, 2 women	5 participants	Sensitivity & specificity: 76% for hypoglycemia, 79% for hyperglycemia with fusion model	Healthy	NR	NR	UK (data origin), USA (analysis)	Warwick University (UK), Texas A&M (USA)	Secondary analysis	14 days	Free-living
⁷⁴	Chowdhury, 2024	Multi-modal Approach to Estimate interstitial Glucose Using Multi-Stream and Cross-Modality Attention	To improve non-invasive interstitial glucose estimation by integrating CGM, ECG, and PPG using multi-stream transformer networks	NR	NR	12 subjects (randomly selected from 22)	RMSE (CGM only): 14.28; RMSE (multi-modal): 12.81	T2D	NR	NR	Singapore	Private dataset from NTU	Private	NR	Lab-based
⁷⁵	Chikwetu, 2024	Carbohydrate Content Classification Using Postprandial Heart Rate Responses from Non-Invasive Wearables	Classify carbohydrate meal content (low, medium, high) using heart rate response from wearables	Mean (SD): years 27.6 (4.2)	33% F	9	N/A	Healthy	NR	NR	US	Empatica E4 device, collected at Duke University	Private; in vivo	9 days per participant	Controlled meals in real-world setting
⁷⁶	Bonet, 2024	Smart Algorithms for Efficient Insulin Therapy Initiation in Individuals With Type 2 Diabetes: An in Silico Study	To develop and evaluate new AI-driven algorithms for faster and safer basal insulin titration in insulin-naive T2D individuals	51 years (median, range 37–65)	102 women out of 300 virtual subjects	300 in silico subjects	TIR, TAR, TBR1 ( < 70 mg/dL), TBR2 ( < 54 mg/dL), OID (U), insulin dosing frequency	T2D	NR	NR	Italy	Padova T2D simulator	Synthetic	12 months	Simulated
⁷⁷	Beolet 2024	End-to-end offline reinforcement learning for glycemia control	Develop personalized offline RL models for basal insulin control using real-world CGM data from a commercial closed-loop system	Mean (SD): years 44.8 (13.0)		150	Mean (SD) Glucose: 160.8 mg/dL; TIR: 68.4%; TBR: 1.3%; TAR: 30.3%; CV: 33% Weight: 77.1 (17.5)	T1D			France	Real-life usage data from Diabeloop DBLG1 closed-loop system	Private; in vivo	9 months per patient on average
⁷⁸	Allam, 2024	Using nonlinear auto-regressive with exogenous input neural network in interstitial glucose prediction	To improve long-horizon interstitial glucose prediction using a modified RNN architecture.	3–18 years	0% F	9 patients; 4916 samples total	Range: 3.5-15.5 mmol/L; RMSEs range from 0.74 to 1.34 mmol/L; FIT from 75.2% to 84.7%	T1D	NR	NR	NR (Data from DirecNet, likely USA)	Diabetes Research in Children Network (DirecNet)	Public	~2 days per patient	Real-world
⁷⁹	Zanelli, 2023	Type 2 Diabetes Detection With Light CNN From Single Raw PPG Wave	To detect type 2 diabetes using a lightweight CNN applied to raw PPG pulses, with and without transfer learning.	50–70 years (middle-aged subjects)	Reported as input but specific distribution NR	100 subjects in DB_DT2 (15% with T2D); additional shape and HT datasets for pretraining	NR (diabetes based on clinical records, not glucose values)	T2D	NR	Excluded HT and other overlapping conditions	France (University Hospital of Nice)	DB_DT2 dataset; also DB_shape and DB_HT; collected using pOpm©tre device	Mixed	NR (single pulse acquisition from PPG; sampling rate = 1 kHz)	Controlled setting (non-clinical + clinical PPG acquisition)
⁸⁰	Yang, 2023	Glucose trend prediction model based on improved wavelet transform and gated recurrent unit	To improve CGM-based glucose trend prediction using an enhanced wavelet denoising algorithm combined with GRU.	56.1 ± 8.3 years	54 male / 26 female	80 patients	NR (but includes performance metrics: RMSE = 0.5537 mmol/L, MAPE = 2.21%, R² = 0.989)	T2D	NR	NR	China	Silicon-based CGM device (not named); 5-min interval data	Private	7 consecutive days per patient	Real-world
⁸¹	Tao, 2023	A deep learning nomogram of continuous glucose monitoring data for the risk prediction of diabetic retinopathy in type 2 diabetes	To predict the risk of diabetic retinopathy (DR) in T2D patients using a deep learning nomogram based on CGM data.	Mean age ~58.5 years across cohorts	~38% female	788 patients (Training: 494, Testing: 294)	TIR, SDBG, GRADE, CV, MBG, MAGE, LI, ADRR, etc.	T2D	NR	Excluded patients with malignancy, mental illness, or acute conditions	China	Shanghai Jiao Tong University Affiliated Sixth Peoples Hospital (iPro2 CGM device)	Private	3 days of CGM per patient	Hospital
⁸²	Site, 2023	Machine-learning-based diabetes prediction using multi-sensor data	To compare diabetes prediction performance using single vs. multi-sensor combinations from wearable data (glucose, ECG, accelerometer, breathing).	NR	NR	29 participants (20 non-diabetic, 9 diabetic)	NR (diabetes diagnosis status used, not glucose levels directly)	Type 2 Diabetes (T2D) and healthy controls	NR	NR	Finland (Tampere University)	D1NAMO dataset (Zephyr BioHarness 3 device)	Private	4 days of continuous data per subject	Free-living
⁸³	Nazha, 2023	Portable Infrared-Based Glucometer Reinforced with Fuzzy Logic	To develop a non-invasive portable glucometer that estimates glucose using infrared sensors and fuzzy logic, incorporating both finger and tear measurements.	16 to 65 years (as shown in participant table)	NR (Sex shown for each participant in table, no summary reported)	30 participants (15 with diabetes, 15 healthy controls)	Estimated glucose via voltage correlation; reported range 96-178 mg/dL; Error <3%	Mixed	NR	NR	Syria (Tartous University) and Germany (Otto Von Guericke University)	Study-specific device; not based on existing dataset	Private	Single-visit, with 12 measurements per participant pre- and post-meal	Free-living
⁸⁴	Lin, 2023	Prediction of interstitial Glucose Concentration Based on OptiScanner and XGBoost in ICU	To improve ICU interstitial glucose prediction using mid-IR spectral data collected from OptiScanner and XGBoost, accounting for hetastarch treatment effects.	NR	NR	1,021 ICU patients (Training: 633, Test: 388)	Reference glucose measured using YSI STAT 2300; RMSE used as evaluation metric	Unspecified	NR	Includes subgroup treated with hetastarch for hypovolemia	Taiwan	OptiScanner mid-IR spectrometry (data not publicly available)	Private	NR (bedside OptiScanner data acquisition)	Hospital
⁸⁵	Lee, 2023	An Integrated Digital Health Care Platform for Diabetes Management With AI-Based Dietary Management	To evaluate whether an integrated AI-based dietary management platform improves glycemic management and weight in T2D patients	Mean 56.1 years	66.3% male (181/273)	294	HbA1c (change from baseline)	T2D	Korean	Not specified	South Korea	3 university hospitals in Seoul	RCT	48 weeks	Outpatient
⁸⁶	Lee, 2023	Glucose Transformer: Forecasting Glucose Level and Events of Hyperglycemia and Hypoglycemia	To develop a deep learning model using the Transformer encoder to predict glucose levels and classify events of hypoglycemia/hyperglycemia in hospitalized T2D patients.	T1D dataset: mean age 44 years; T2D: age 20-90 years	NR	104 T2D patients; 226 T1D patients for transfer learning	T2D group: higher distribution and larger SD than T1D; glucose values ranged from 40 to 400 mg/dL	Type 2 Diabetes (T2D); transfer learning data from T1D patients	NR	NR	South Korea	Soonchunhyang University Cheonan Hospital (T2D); OhioT1DM dataset (T1D)	Public	3 to 7 days per patient	Hospital
⁸⁷	Kistkins, 2023	Comparative Analysis of Predictive Interstitial Glucose Level Classification Models	Compare the efficacy of ARIMA, logistic regression, and LSTM models in classifying glucose states (hypo-, eu-, hyperglycemia) 15 min and 1 h in advance.	Mean age: 47 ± 11 years	6 female, 5 male	11 T1D patients (real data); 30 virtual patients (in silico: 10 adults, 10 adolescents, 10 children)	Mean HbA1c: 57 ± 8 mmol/mol; variability discussed, not numerically detailed	T1D	NR	NR	Austria	COVAC-DM clinical study; UVA/Padova T1D Simulator (via simglucose v0.2.1)	Real-world	Two CGM phases per participant pre- and post-vaccination; virtual: 10 days	Clinical cohort and simulation study
⁸⁸	Arbi, 2023	interstitial glucose estimation based on ECG signal	To develop a method to estimate interstitial glucose concentration (BGC) using ECG signal features instead of invasive CGMs.	NR	NR	3 T1D patients from D1NAMO dataset	NR (BGC range evaluated per patient, model R² up to 98%)	T1D	NR	NR	Algeria	D1NAMO dataset (real-world, open access); PhysioNet QT dataset (for training ECG segmentation)	Public	8 days of ECG and BGC measurements per patient	Free-living
²³	Avram, 2020	A digital biomarker of diabetes from smartphone-based vascular signals	To detect prevalent diabetes using smartphone-based photoplethysmography and deep learning	Mean age ~45–55 yrs	~59% male with diabetes in primary cohort	53,870 (primary cohort)	HbA1c linear assoc: Î²=2.28 per SD DNN score; DNN AUC: 0.766 (primary), 0.740 (contemp), 0.682 (clinic)	T2D	NH White, Black, Asian, Hispanic, Multi-ethnic	Hypertension, hypercholesterolemia, CAD, CHF, PVD, stroke, sleep apnea	USA (and Canada for clinic)	Health eHeart + 3 clinical sites	Mixed	2014-2019 (PPG + HbA1c matched within 180 days)	Remote
⁸⁹	Andellini, 2023	Artificial intelligence for non-invasive glycaemic-events detection via ECG in a pediatric population	To validate an AI-based algorithm for detecting glycaemic events (hypo/hyperglycemia) using ECG signals in T1D children	<18 years	NR	64 pediatric T1D patients	HbA1c, mean glucose, glycaemic variability, time in range, event frequency	T1D	NR	Excluded: Celiac disease, cardiovascular disease, arrhythmias, pregnancy	Italy	Bambino Ges¹ Childrens Hospital	Observational	3 days per participant	Real-world
⁹⁰	Alvarado, 2023	Combining wavelet transform with convolutional neural networks for hypoglycemia events prediction from CGM data	To detect and predict hypoglycemia events up to 24 hours in advance using deep learning and wavelet transforms	40.16 ± 11.86 years	50% female	20 T1D participants	Time in range, % in hypo/hyper ranges, glycemic variability, interquartile range plots	T1D	NR	NR	Spain	Hospital Prncipe de Asturias, Alcal¡ de Henares	Real-world	~15 days/patient	Clinical trial
⁹¹	Ahmed, 2023	Performance of artificial intelligence models in estimating interstitial glucose level among diabetic patients using non-invasive wearable device data	To evaluate AI model performance in estimating interstitial glucose levels using non-invasive wearable device data	9–77 years	5 females, 8 males	13	RMSE: 0.099-0.197, MAE: 0.097-0.112	Mixed	NR	NR	Qatar	Open-source dataset (Riversong Smart Band, Freestyle LibrePro CGM)	Public	3 months/patient (June 2020-Dec 2021)	Wearable study
⁹²	Zale, 2022	Machine Learning Models for Inpatient Glucose Prediction	Review clinical evidence on machine learning models for predicting glucose trends in hospitalized patients and assess their predictive performance and clinical applicability.	NR	NR	Ranged from N = 20 to >100,000 across reviewed models; Zale 2022 model used data from 5 hospitals totaling over 100,000 patients	Outcomes included hypoglycemia (<70 mg/dL, <54 mg/dL), hyperglycemia (>180 mg/dL), and glucose trajectory; CGM studies reported mean squared errors (e.g., RMSE = 21.5 mg/dL)	Mixed	NR	CKD (severe), use of sulfonylureas or insulin, Charlson Comorbidity Index considered in some models	USA, multiple hospitals including Johns Hopkins; international studies also cited	EHRs from multiple hospitals (including Johns Hopkins), CGM data, MIMIC-III ICU database	Real-world	NR for most models; prediction horizons discussed rather than total study periods	Hospital
⁹³	Xiong, 2022	Identifying daily activities of patient work for type 2 diabetes and co-morbidities: a deep learning and wearable camera approach	To classify self-management activities of people with type 2 diabetes and comorbidities using wearable camera data and deep learning.	Median age: 72 years	10 females, 16 males	26 participants	NR	Type 2 Diabetes with comorbidities	NR	Yes; multiple chronic conditions	Australia	Wearable camera data from Macquarie University study	Private	1 day of wearable camera footage ( ~ 16 hours per participant)	Free-living
⁹⁴	Nemat, 2022	interstitial Glucose Level Prediction: Advanced Deep-Ensemble Learning Approach	To improve interstitial glucose prediction in Type 1 diabetes using deep-ensemble learning models and novel meta-learning techniques	20-80 years	5 females, 7 males	12 participants	RMSE, MAE, MCC, Surveillance Error	T1D	NR	NR	UK	Ohio T1DM dataset	Public	8 weeks per participant	Real-world
⁹⁵	Malerbi, 2022	Diabetic Retinopathy Screening Using Artificial Intelligence and Handheld Smartphone-Based Retinal Camera	To evaluate the diagnostic accuracy of a DL algorithm and a handheld retinal camera in detecting diabetic retinopathy in a real-world, underserved setting.	Mean age 60.8 ± 11.4 years	64.9% female	824 enrolled; 679 with gradable images	NR	T2D	NR	Hypertension (68.4%), smoking (48.4%)	Brazil	Itabuna Diabetes Campaign, Bahia State, Brazil	Private	Single-day campaign (Nov 23, 2019)	Other
⁹⁶	Lobo, 2022	A Data-Driven Approach to Classifying Daily Continuous Glucose Monitoring (CGM) Time Series	To identify a finite set of representative daily CGM profiles (motifs) for use in classification, modeling, and automated systems.	NR	NR	9741 (training), 14,175 (validation), 42,595 (testing); total = 66,511 profiles from 491 patients	Mean BG, SD, CV, TIR, TAR, TBR, LBGI, HBGI	Mixed (T1D + T2D)	NR	NR	USA	DCLP1, DCLP3, DIA1, DIA2, DSS1, NTLT	Private	3 to 6 months per study	Clinical trial
⁹⁷	Lim, 2022	Multi-Task Disentangled Autoencoder for Time-Series Data in Glucose Dynamics	To develop a multi-task disentangled VAE (MD-VAE) for glucose forecasting, event detection, and temporal clustering	NR	NR	T1DMS: 20 virtual patients (20 days); DCLP3: 112 real patients (6 months CGM)	MAE: 12.02-20.02 mg/dL; RMSE: 16.56-28.84 mg/dL; MARD: 8.01-13.25%; TIR, TAR, TBR, HBGI, LBGI, CV, IPI	T1D	NR	NR	South Korea	T1DMS simulator (FDA-approved), DCLP3 real-world clinical trial (Control-IQ, Dexcom G6)	Real-world	20 days (sim); 6 months (real patients)	In silico
²⁹	Kim, 2022	Identification of Healthy and Unhealthy Lifestyles by a Wearable Activity Tracker in Type 2 Diabetes	To classify healthy vs. unhealthy lifestyle patterns in T2D using data from wearable activity trackers and unsupervised clustering	29–60 years	58.3% male	24 patients	ÎHbA1c after 3 and 9 months; 0.33% lower HbA1c in healthy group at 9 months (not statistically significant)	T2D	NR	NR	South Korea	Fitbit Charge 2 tracker, Korea University Anam Hospital	Prospective observational study	9 months total	Outpatient
⁹⁸	Aloraynan, 2022	A Single Wavelength Mid-Infrared Photoacoustic Spectroscopy for Noninvasive Glucose Detection Using Machine Learning	To develop a noninvasive glucose detection system using single-wavelength MIR photoacoustic spectroscopy with ML classification	NR	NR	In vitro only - 3 samples per class across 10 classes; 40,200 measurements total	Detection resolution of ±25 mg/dL; Clarkes EGA: 96.1% zone A, 3.9% zone B (after preprocessing)	Non-human	NR	NR	Canada	Artificial gelatin-based skin phantoms prepared with glucose concentrations from 75-300 mg/dL	Synthetic	3 days of repeated measurements	Lab-based
²⁴	Yin, 2021	DiabDeep: Pervasive Diabetes Diagnosis Based on Wearable Medical Sensors and Efficient Neural Networks	Develop an efficient, wearable-sensor-based diabetes diagnosis system for edge and server environments	NR	NR	52 participants (14 T1D, 13 T2D, 25 healthy)	NR	Mixed	NR	None reported	USA	Empatica E4 + Samsung S4 smartphone	Real-world	1–1.5 hours/participant	Real-world daily use
⁹⁹	Garcia-Tirado, 2021	Advanced hybrid artificial pancreas system improves on unannounced meal response - In silico comparison to currently available system	To evaluate a fully automated, adaptive hybrid artificial pancreas (RocketAP) with a novel bolus priming system for managing unannounced meals in type 1 diabetes using in silico simulations	Simulated adult cohort	Simulated cohort, not reported	100 virtual adults (UVA/Padova simulator)	TIR, TTR, %time <70 mg/dL, %time >180 mg/dL, LBGI, HBGI, SD-glucose, %CV-glucose, Mean BG; model RMSE (identification: 7.67, validation: 7.73 mg/dL)	T1D	NR	NR	USA	UVA/Padova Type 1 Diabetes Simulator	Synthetic	14 days of training data + test simulations	In silico
²⁵	Deng, 2021	Deep transfer learning and data augmentation improve glucose levels prediction in type 2 diabetes patients	To improve short-term prediction of glucose levels in T2D using transfer learning and data augmentation on imbalanced datasets	65.1 ± 8.8 years	52.5% female	40 patients with T2D	BG range 40-400 mg/dL, mean 130.6 mg/dL; hypoglycemia <80, hyperglycemia >180 mg/dL	T2D	NR	NR	USA	Beth Israel Deaconess Medical Center (BIDMC)	Private	Up to ~170 hours per participant	Outpatient
¹⁰⁰	Bent, 2021	Engineering digital biomarkers of interstitial glucose from noninvasive smartwatches	To develop digital biomarkers from noninvasive wearable data and food logs to predict personalized interstitial glucose excursions and real-time glucose values.	35–65 years	NR	16 participants	Mean PersNorm=112.4 mg/dL, PersHigh=149.9, PersLow=90.8; SDs reported	Prediabetes	NR	Excluded: cancer, COPD, CVD, food allergies, antidiabetic meds	USA	Duke University; study-specific collection	Private	8–10 days per participant	Outpatient
¹⁰¹	Baig, 2021	Early Detection of Prediabetes and T2D Using Wearable Sensors and IoT-Based Monitoring Applications	To develop and evaluate an AI-powered early detection system for prediabetes and T2D using wearable technology and IoT-based monitoring	55–62 years	1 male, 1 female in testing; unclear for full training set	36 participants (training), 2 participants (real-time testing)	HbA1c (avg): 87.5 vs 64 mmol/mol; HR, ventilation, cadence, activity ranges; 91% accuracy; sensitivity 94%; specificity 90%; predictability 72%; Kappa = 0.75	T2D, Prediabetes	NR	NR	New Zealand	Hexoskin vest; PhysioNet; manual participant data	Mixed	2 years (follow-up); 10 months data collection	Real-world
²⁷	Zhu, 2020	An Insulin Bolus Advisor for Type 1 Diabetes Using Deep Reinforcement Learning	To optimize mealtime insulin dosing in T1D using deep reinforcement learning bolus advisor with CGM data	NR	NR	20 virtual subjects (10 adults, 10 adolescents)	TIR, TBR, TAR, mean BG, CV, LBGI, HBGI, CVGA	T1D	NR	NR	UK	UVA/Padova T1D simulator	Simulated	6-month training, 3-month testing	In silico
¹⁰²	Xie, 2020	Benchmarking Machine Learning Algorithms on interstitial Glucose Prediction for Type I Diabetes	To compare ML algorithms and classical ARX models for BG prediction in T1D using the same dataset	40–60 years	4 females, 2 males	6 T1D patients	RMSE, Temporal Gain (TG), Normalized ESODn	T1D	NR	NR	USA	OhioT1DM dataset	Public	8 weeks	Real-world
²⁶	He, 2020	interstitial glucose concentration prediction based on kernel canonical correlation analysis with particle swarm optimization and error compensation	To improve short-term interstitial glucose prediction using nonlinear modeling and personalized error compensation	NR	NR	10 patients	RMSE at 5–30 min: 8.01 to 16.40 mg/dL; R² range: 0.95 to 0.98; EC-CCA reduced RMSE by 33.45% over CCA	T1D	NR	NR	China	DirecNet dataset (USA origin)	Public	~4 days per subject	Clinical (pediatric)
²⁸	Sun, 2019	A Dual Mode Adaptive Basal-Bolus Advisor Based on Reinforcement Learning	To develop a personalized, dual-mode basal-bolus insulin advisor (ABBA) using reinforcement learning, adaptable to CGM or SMBG input	NR	NR	100 virtual adults (simulated)	% time in range, % hypo/hyper, LBGI, HBGI, MAGE, total daily insulin dose	T1D	NR	NR	Switzerland	UVA/Padova FDA-accepted T1DM simulator	Synthetic	3 months + 1 week (98 days)	In silico
¹⁰⁴	Samadi, 2018	Automatic Detection and Estimation of Unannounced Meals for Multivariable Artificial Pancreas System	To develop a fuzzy-logic-based system to automatically detect unannounced meals and estimate CARB intake in T1D patients	18–35 years	NR	11 subjects	Time to detection, detection rate, estimation error, false positives, insulin dose change	T1D	NR	NR	USA	University of Chicago clinical trial data	Real-world	~60 hours per subject	In-clinic (closed loop)
¹⁰⁵	Zecchin, 2015	Jump Neural Network for Real-Time Prediction of Glucose Concentration	To predict short-term (30 min) future glucose levels in T1D patients using a jump neural network incorporating CGM and meal information	NR	NR	20 T1D individuals (10 for training, 10 for testing)	T1D	T1D	NR	NR	Italy	DIAdvisor„¢ project CGM dataset (Dexcom SEVEN PLUS)	Public	2–3 days per subject	Free-living
¹⁰⁶	Zecchin, 2014	Jump neural network for online short-time prediction of interstitial glucose…	To evaluate a novel jump neural network architecture for short-term glucose forecasting after meals	NR	NR	11 T1D subjects (37 meal tests)	Reported mean RMSE across horizons: ~20-30 mg/dL	T1D	NR	NR	Italy	Padova Hospital, Italy	Private	3-day sessions post meal	Clinical

Open in a new tab

AI Artificial Intelligence, AUC Area Under the Curve, BMI Body Mass Index, CARB Carbohydrate, CGM Continuous Glucose Monitoring, ECG Electrocardiogram, EMR Electronic Medical Record, HbA1c Hemoglobin A1c, MAE Mean Absolute Error, MAPE Mean Absolute Percentage Error, MARD Mean Absolute Relative Difference, NR Not Reported, OGTT Oral Glucose Tolerance Test, PPG Photoplethysmography, R²: Coefficient of Determination, RMSE Root Mean Square Error, ROC Receiver Operating Characteristic, T2D Type 2 Diabetes Mellitus, TAR Time Above Range, TBR Time Below Range, TIR Time in Range.

Table 2.

Summary of studies integrating artificial intelligence and wearable technology for diabetes management

Ref. #	Author	Title	Study objective	AI/ML Model	Model input	Model Output or Outcome	Prediction Horizon	CGM Device Used	Other Modalities Used	Were Modalities Combined or Used Separately?	Did Additional Modalities Improve Prediction?	Model Performance	Interpretability Addressed?
¹¹	Anjum, 2024	Optimizing type 2 diabetes management: AI-enhanced time series analysis of continuous glucose monitoring data for personalized dietary intervention	To forecast interstitial glucose values using AI models and provide personalized dietary interventions to manage T2D.	XGBoost, SARIMA, Prophet	CGM glucose readings (timestamped), date/time features (day, hour, etc.)	interstitial glucose level predictions (forecasted CGM values)	2 months	FreeStyle Libre (with smartphone and app)	Meal photographs via FoodLens app	Combined	Yes	XGBoost: MAE = 0.32, MSE = 0.31, RMSE = 0.46, R2 = 0.96, MAPE = 0.04	No
¹⁶	Vettoretti, 2020	Advanced Diabetes Management Using Artificial Intelligence and Continuous Glucose Monitoring Sensors	To review AI methods integrated with CGM for DSS in T1D: bolus calculation, BC parameter tuning, and glucose prediction	Neural networks, linear regression, LASSO, random forest, SVR, gradient boosting, reinforcement learning, case-based reasoning, grammatical evolution	CGM, insulin, carbs, physical activity, time of day, IOB, body weight, ROC	Insulin dose; glucose level; hypoglycemic event	Typically 30 min (up to 6 hours for classification)	Not specified (generic CGM review)	Insulin pump, smart pens, activity trackers, apps	Combined	Yes	NN RMSE ~ 19-20 mg/dL; TG up to 7 min; XGBoost improved euglycemia time 67% vs 62%; SVR + grammar evolution used for 60-min ahead; DSSs improved therapy personalization & glycemic management	Yes
²¹	Ribeiro 2024	A Novel AI Approach for Assessing Stress Levels in Patients with Type 2 Diabetes Mellitus Based on the Acquisition of Physiological Parameters Acquired during Daily Life	Develop an AI-based wearable system for stress classification in T2D patients using physiological indicators.	Fuzzy Logic Model	PPG, BP, GSR, body temp, SpO2, HR, RR, glucose, demographics	Stress level (Non-temporal): Calm, Normal, Stressed	Real-time; up to 4 hours wearable battery life; additional 5-day data for subgroup	None	Yes	Separately	N/A	Metric: Calm, Normal, Stressed Sens.: 0.94, 0.82, 0.93 Spec.: 0.97, 0.94, 0.94 Prec.: 0.94, 0.88, 0.88 Acc.: 0.96, 0.90, 0.94 F1: 0.94, 0.85, 0.90	Interpretability is achieved through transparent, rule-based fuzzy logic using clinically grounded physiological inputs
⁵⁷	Zhu, 2024	Population-Specific Glucose Prediction in Diabetes Care With Transformer-Based Deep Learning on the Edge	To develop a low-power, population-specific, multi-horizon glucose prediction model deployable on wearable devices	Temporal Fusion Transformer (TFT), compared with LSTM, DRNN, N-BEATS, N-HiTS, SVR, XGBoost, Linear Regression	CGM, timestamp, age, sex; BMI and diabetes type for ShanghaiDM	Future interstitial glucose levels	30 and 60 minutes	Medtronic Enlite (OhioT1DM), Abbott FreeStyle Libre (ShanghaiDM)	Demographics (static data)	Combined	Yes	RMSE ‰¤ 14.7 (30-min) and ‰¤ 23.5 (60-min) on unseen cohorts	Yes
⁵⁸	Zhu 2024	Multi-Horizon Glucose Prediction Across Populations with Deep Domain Generalization	Develop a generalizable, multi-horizon glucose prediction model for diverse populations using CGM data	Sparse Transformer with meta-learning	CGM data, timestamp encoded	Predict glucose levels	30, 60, 90, and 120 minutes	REPLACE-BG: Dexcom G4, OhioT1DM: Dexcom G4, ABC4D: Dexcom G5, GVAS: Dexcom G6/Freestyle Libre	N/A	N/A	N/A	RMSE at 30 mins: mg/dL REPLACE-BG: 19.2, OhioT1DM: 18.9, ABC4D: 22.9, GVAS: 15.9	SHAP
⁵⁹	Warren, 2024	A Scalable Application of Artificial Intelligence-Driven Insulin Titration Program…	To assess the scalability and effectiveness of AI-driven insulin titration (d-Nav) in real-world T2D management	AI-based insulin titration (d-Nav app)	Self-monitored glucose or CGM, insulin regimen, patient response	Insulin dose recommendations	Weekly adjustments	35.8% used CGM	GLP-1RA, GIP/GLP-1RA, SGLT2i	Combined	Yes	HbA1c reduced 1.3%; severe hypoglycemia 1.7/100 pt-yrs; TIR improved to 64.2%	No
⁶⁰	Wang 2023	A novel hypoglycemia alarm framework for type 2 diabetes with high glycemic variability	Develop a hypoglycemia early alarm framework for T2D patients with high glycemic variability using dynamic multi-scale features	SVR for prediction + Random Forest with Relief-SVM-RFE	CGM data, derived statistical characteristics (e.g. range, std deviation, CV, mean, median, mode) and clinical parameters (ADRR, LBGI, etc.)	Hypoglycemia (BG < 70 mg/dL) occurrence within 30 minutes	30 minutes	Medtronic iPro 2	N/A	N/A	N/A	ACCURACY, SENSITIVITY, SPECIFICITY, PPV, NPV, F1	Feature selection via Relief algo (ranks feature importance) and SVM-RFE (recursive elimination of less relevant features)
⁶¹	Tucker, 2024	Neural Networks With Gated Recurrent Units Reduce Glucose Forecasting Error Due to Changes in Sensor Location	To examine whether GRU neural networks can reduce glucose forecast error caused by changes in CGM sensor location	Feedforward NN and GRU NN	CGM glucose time series (3 prior readings with bias)	Future glucose value	15, 30, 45, 60 minutes	FreeStyle Libre Pro	None			No significant location-induced error with GRU NNs; RMSE and MARD reported in tables 3–4	No
⁶²	Tao 2023	A Double Deep Latent Autoencoder for Diabetic Retinopathy Diagnose Based on Continuous Glucose Sensors	To diagnose diabetic retinopathy using CGM data with a deep latent autoencoder framework	Double Deep Latent Autoencoder (DDL-AE)	CGM data from 26 individuals with 71,663 glucose points	Retinopathy classification	Non-temporal	Unspecified	None			Accuracy = 91.1%, Precision = 94.3%, Recall = 86.8%, F1 = 90.3%	No
⁶³	Shuzan 2024	QU-GM: An IoT Based Glucose Monitoring System From Photoplethysmography, interstitial Pressure, and Demographic Data Using Machine Learning	Develop a non-invasive, real-time wearable system for glucose estimation and severity classification	For regression: Bagged Ensemble Trees (BET); For classification: K-Nearest Neighbor (KNN)	PPG, systolic/diastolic interstitial pressure, demographics (age, sex, BMI, height, weight)	Estimated glucose levels (Non-temporal) and severity classification (normal, warning, dangerous)	Real-time, updated every 10 seconds	None	Yes	Combined	Yes	BET: R = 0.90, MAE = 22.90 mg/dL, RMSE = 30.74 mg/dL; KNN: accuracy = 98.12%, F1-score = 0.98	Feature importance ranking
⁶⁴	Seo, 2024	Generative Adversarial Network-Based Data Augmentation for Improving Hypoglycemia Prediction	To improve hypoglycemia prediction performance using GAN-based synthetic data augmentation in training data	BiLSTM, CNN-BiLSTM, and GAN for data augmentation	CGM data, augmented with GAN-simulated sequences	Hypoglycemia event	30 minutes	Medtronic iPro2	None		Yes	BiLSTM+GAN AUC = 0.928, CNN-BiLSTM+GAN AUC = 0.956	No
³⁰	Ramazi, 2021	Predicting progression patterns of type 2 diabetes using multi-sensor measurements	To predict future HbA1c, HDL, LDL, and triglyceride levels in T2D patients using wearable and static data	CNN-GRU deep learning model; compared to RF, XGBoost, Wide & Deep	CGM (Dexcom G4), tri-axial ActiGraphy, demographic and lab data	HbA1c, HDL, LDL, triglycerides	1 year	Dexcom G4 Platinum Pro CGM	ActiGraph wGT3X-BT (accelerometer)	Combined	Yes	HbA1c RMSE = 1.37; Accuracy=0.90; AUROC = 0.88; better than all baselines (RF, XGBoost, Wide & Deep)	Yes
⁶⁵	Yin, 2017	A Health Decision Support System for Disease Diagnosis Based on Wearable Medical Sensors and Machine Learning Ensembles	To develop a hierarchical, closed-loop health decision support system (HDSS) combining wearable sensors and CDSSs to enable disease diagnosis and tracking	Naive Bayes, Bayes Network, SVM, k-NN, Decision Trees, Random Forest, MLP; ensemble methods: AdaBoost, DECORATE, bagging, stacking, voting	Physiological signals from wearable sensors: ECG, HR, BP, SpO2, BT, BG, GSR, etc. (Tier-1); Clinical/lab tests and physician insights (Tier-2/3)	Binary/multi-class disease classification	Non-temporal	No	Location (GPS), motion (accelerometers), temperature, GSR, lab tests	Combined	Yes	Accuracies: arrhythmia (86%), T2D (78%), bladder disorder (99.6%), nephritis (94%), hypothyroid (95%). Multi-threshold models <67% accuracy. Improvements up to 28.6% over thresholding. Ensemble + feature filtering enhanced performance.	Partially
⁶⁶	Kim 2024	Impartial feature selection using multi-agent reinforcement learning for adverse glycemic event prediction	Develop a deep learning model for predicting adverse glycemic events in hospitalized T2D patients using CGM, EMR, and MARL-based feature selection	Seq2seq neural network with Bahdanau attention; Multi-agent reinforcement learning (MARL) for feature selection; Time2Vec encoding for insulin and meal time irregularity	CGM, insulin admin status, meal intake time, 10 EMR featues (age, BMI, HbA1c, etc.)	Predict glycemic events (normo-, hypo- or hyper-glycemia) within 30-minute window	30 minutes	Dexcom G5/G6	Insulin logs; meal logs; time; EMR	Combined	Yes	Metric: Normo, Hypo, Hyper Prec.: 92.8, 50.5, 86.7 Recall:85.5, 75.7, 93.1 F1: 89.0, 60.6, 89.8	MARL assigns rewards based on each agent’s contribution, enabling interpretability of EMR feature impact
⁶⁷	Kashif 2024	GLSTM: On Using LSTM for Glucose Level Prediction	To predict glucose levels in individuals with prediabetes using a personalized Long Short-Term Memory (LSTM) model trained on multimodal wearable sensor data	Long Short-Term Memory (LSTM)	CGM, heart rate, dietary intake (sugar and carb)	Predicted glucose levels, 3-hour forecast (36 lookback steps)	3 hours	Dexcom G6	Heart rate; dietary intake	Combined	Yes	With heart rate and dietary info: RMSE: 1.83 MSE: 3.44 MAE: 1.24 R²: 0.99	Not addressed
⁶⁸	Kannenberg, 2024	Personalized Lifestyle Therapy for Type 2 Diabetes Through a Predictive Algorithm-Driven Digital Therapeutic	To evaluate the clinical impact of a CGM-AI-powered digital therapeutic on glycemic management and weight in non-insulin-treated T2D patients	Random forest regressor; Deep learning neural network (multi-layer, deep architectures)	CGM data, meal logs, individual characteristics	Predicted postprandial glucose levels	Meal-by-meal; 3–6 months	Validated CGM (unspecified)	Nutrition diary, physical activity data	Combined	Yes	HbA1c 0.67% overall; 1.08% in poorly controlled group; weight 6.84 kg (6 mo)	Partially
⁶⁹	Hotta 2024	Optimizing postprandial glucose prediction through integration of diet and exercise	To improve postprandial glucose prediction in GDM patients by integrating diet and exercise data using Bayesian transfer learning informed by randomized controlled trial data.	Bayesian Transfer Learning Framework	CGM, energy expenditure (exercise), carbohydrate intake	Postprandial glucose level trajectory (within 90 minutes after eating)	90 minutes	Medtronic Guardian Connect System	Energy expenditure; carbohydrate intake	Combined	Yes	For postprandial segments (with exercise), best-performing model achieved: RMSE: 0.85 MAE: 0.65 AUC error: 53.57 Max glucose error: 0.77	The Bayesian framework supports transparent, interpretable modeling, allowing direct examination of person-specific parameters (e.g., coefficients for diet, exercise, and their interaction); These interpretable parameters facilitate personalized behavioral recommendations.
⁷⁰	Metwally, 2024	Predicting Type 2 Diabetes Metabolic Phenotypes Using Continuous Glucose Monitoring and a Machine Learning Framework	To identify distinct metabolic subphenotypes in individuals with normoglycemia or prediabetes using OGTT-derived CGM data and machine learning to inform precision prevention of T2D.	Random Forest, SVM (linear/RBF), L1-regularized Logistic Regression; Feature engineering and PCA-based reduced representations for classification tasks	OGTT time-series glucose data from CGM and plasma	Classification of metabolic subphenotypes (muscle IR, Î²-cell dysfunction, incretin defect, hepatic IR)	Non-temporal	Dexcom G6 Pro	OGTT (plasma), C-peptide, insulin suppression, IIGI	Combined	Yes	Muscle IR auROC=0.95 (OGTT), 0.88 (CGM); Î²-cell defect auROC=0.89 (OGTT), 0.84 (CGM); incretin defect auROC=0.88-0.90	Yes; PCA, feature correlations, curve interpretability
⁷¹	Goncharov 2024	Insertable Glucose Sensor Using a Compact and Cost-Effective Phosphorescence Lifetime Imager and Multi-Parametric Data-Driven Calibration	Demonstrate feasibility of a low-cost, compact phosphorescence lifetime imager integrated into an insertable glucose sensor system	U-Net-like CNN architecture	Optical lifetime signals from a phosphorescence-based glucose sensor	Continuous (real-time, image-based) glucose concentration (mM) quantified from phosphorescence lifetime	Non-temporal	Novel insertable glucose sensor (phosphorescence-based)	Phosphorescence lifetime imaging system (optical readout)	Combined	Yes	MAE: 0.18 mM; RMSE: 0.22 mM; R²: 0.999 across 0–30 mM range	Not addressed
⁷²	Dénes-Fazakas 2024	Physical Activity Detection for Diabetes Mellitus Patients Using Recurrent Neural Networks	Detect physical activity in T1DM patients using CGM, heart rate, and step data with RNNs	BiLSTM and BiGRU	CGM, HR, Step Count	Physical activity (yes/no) classification. Lookback window of 2 hours (24 ×5-min steps)	Up to 2 hours (24 — 5-minute intervals)	Dexcom G4	Heart rate, step count	Combined	Yes	BiLSTM F1: 0.988; Acc: 0.987; Prec.: 0.986; Recall: 0.988; AUC: 0.990 (glucose + HR/steps)	Not addressed
⁷³	Dave, 2024	Detection of Hypoglycemia and Hyperglycemia Using Noninvasive Wearable Sensors: Electrocardiograms and Accelerometry	To detect hypoglycemia and hyperglycemia using noninvasive ECG and accelerometer data as alternatives to CGMs	Random Forest (RF), Quantile Regression Forest (QRF)	ECG features (HRV), accelerometer features	Hypoglycemia or hyperglycemia detection	15-minute CGM interval	FreeStyle Libre Flash CGM	ECG (Medtronic Zephyr), accelerometer	Combined	Yes	QRF: 76% sens/spec for hypo; 79% sens/spec for hyperglycemia (fusion model); improved over ECG-only models	Partially
⁷⁴	Chowdhury, 2024	Multi-modal Approach to Estimate interstitial Glucose Using Multi-Stream and Cross-Modality Attention	To improve non-invasive interstitial glucose estimation by integrating CGM, ECG, and PPG using multi-stream transformer networks	Multi-stream Transformer with Cross-Modality Attention	CGM, ECG, PPG	interstitial glucose levels	30, 60 minutes	Dexcom G6	ECG (AD8232), PPG (MAX30101)	Combined	Yes	RMSE improved from 14.28 to 12.81 with modality fusion	Yes
⁷⁵	Chikwetu, 2024	Carbohydrate Content Classification Using Postprandial Heart Rate Responses from Non-Invasive Wearables	Classify carbohydrate meal content (low, medium, high) using heart rate response from wearables	XGBoost (best), Random Forest, Logistic Regression	Postprandial heart rate (from Fitbit)	Carbohydrate content classification (low, medium, high). Postprandial: up to 120 min post-meal HR data used	60 seconds (with 25% overlap sliding window)	None	Heart rate (Fitbit); meal carbohydrate content	Separately	N/A	XGBoost Accuracy: 90.6% ; Precision: 0.91; Recall: 0.91; F1: 0.91	SHAP used to evaluate feature importance of HR time segments
⁷⁶	Bonet, 2024	Smart Algorithms for Efficient Insulin Therapy Initiation in Individuals With Type 2 Diabetes: An in Silico Study	To develop and evaluate new AI-driven algorithms for faster and safer basal insulin titration in insulin-naive T2D individuals	Logistic regression (for classification), rule-based CGM and insulin titration algorithms	OGTT-derived glucose & C-peptide, CGM data, FPG, insulin dose history	Optimal insulin dose (OID), time in range (TIR)	Up to 52 weeks (12 months)	Simulated CGM	OGTT test	Combined	Yes	SMART-CGM-BASED reached OID faster; TIR †‘ to 64%; TAR ; minimal TBR1/TBR2 increase; dose changes	Yes
⁷⁷	Beolet 2024	End-to-end offline reinforcement learning for glycemia control	Develop personalized offline RL models for basal insulin control using real-world CGM data from a commercial closed-loop system	Twin Delayed Deep Deterministic Policy Gradient + Behavior Cloning; Batch-Constrained Q-learning; Conservative Q-Learning	CGM, insulin history, IOB, COB, TDD, time of day, body weight	basal insulin rate. Effective RL horizon approx. 8 hours, 5-min steps	8 hours (5-min timestep, Î³=0.99)		Insulin history, meal data, physical activity (limited use), body weight, time of day	Combined	Yes	TD3-BC (best): TIR + 4.49%, TBR − 0.86%, Mean BG − 8 mg/dL, CV 36.9%; TIR ↑ in 84% of participants	Not addressed
⁷⁸	Allam, 2024	Using nonlinear auto-regressive with exogenous input neural network in interstitial glucose prediction	To improve long-horizon interstitial glucose prediction using a modified RNN architecture.	Nonlinear autoregressive exogenous input neural network	CGM data (prior glucose readings)	Predicted glucose level at 15, 30, 45 minutes tested; best at 15 min	15, 30, 45 minutes tested; best at 15 min	Guardian RT CGM system	None			RMSE (min-max): 0.74-1.34 mmol/L; FIT: 75.2%-84.7%; NPE: 10.6%-13.3%	No
⁷⁹	Zanelli, 2023	Type 2 Diabetes Detection With Light CNN From Single Raw PPG Wave	To detect type 2 diabetes using a lightweight CNN applied to raw PPG pulses, with and without transfer learning.	Lightweight CNN; Transfer Learning (with Shape and Hypertension PPG models)	Single PPG pulse (1 second); optionally with age, sex, or PPG handcrafted features	Type 2 diabetes classification (yes/no)	Non-temporal	None	Age, biological sex, PPG handcrafted features	Combined	Yes	Best AUC: 75.5 (majority voting, no transfer learning, with age/sex); Sensitivity: 75%, Specificity: 76%	No
⁸⁰	Yang, 2023	Glucose trend prediction model based on improved wavelet transform and gated recurrent unit	To improve CGM-based glucose trend prediction using an enhanced wavelet denoising algorithm combined with GRU.	IWT-GRU (Improved Wavelet Transform + Gated Recurrent Unit), compared with RNN, LSTM, GRU, SVR	CGM data (5-minute intervals over 7 days, denoised using wavelet transform)	Predicted interstitial glucose values	15-60 minutes; optimal at 45 minutes	Silicon-based kinetic CGM (not brand-named)	None			IWT-GRU: RMSE = 0.5537 mmol/L, MAPE = 2.2147%, R² = 0.989, runtime=37.2 s; CEGA Zone A: 98.69%	No
⁸¹	Tao, 2023	A deep learning nomogram of continuous glucose monitoring data for the risk prediction of diabetic retinopathy in type 2 diabetes	To predict the risk of diabetic retinopathy in T2D patients using a deep learning nomogram based on CGM data.	Deep feedforward neural network + LASSO + nomogram	CGM metrics (TIR, SDBG, GRADE), Duration, HbA1c, other clinical features	Probability of diabetic retinopathy	Non-temporal	Medtronic iPro2	Clinical biomarkers: HbA1c, lipid profile, etc.	Combined	Yes	AUC = 0.86 (training), 0.85 (testing); F1 = 0.78/0.76; Sensitivity=0.77/0.75; Specificity=0.81/0.80	Yes
⁸²	Site, 2023	Machine-learning-based diabetes prediction using multi-sensor data	To compare diabetes prediction performance using single vs. multi-sensor combinations from wearable data (glucose, ECG, accelerometer, breathing).	XGBoost, Gradient Boosting, Random Forest, SVM, MLP, Logistic Regression, Adaboost, Decision Tree	Glucose, ECG, accelerometer, breathing sensors (feature engineered: time-domain, interval-based)	Diabetes classification (yes/no)	Non-temporal	Intermittent glucose monitor (not CGM; device name NR)	ECG, accelerometer, breathing sensor	Both tested	Yes	XGBoost with glucose+ECG + ACC: Accuracy = 98.2%, FP = 1.07%, FN = 2.77%	No
⁸³	Nazha, 2023	Portable Infrared-Based Glucometer Reinforced with Fuzzy Logic	To develop a non-invasive portable glucometer that estimates glucose using infrared sensors and fuzzy logic, incorporating both finger and tear measurements.	Fuzzy Logic (FL) with Clarke Error Grid modeling	Infrared sensor output voltages from finger and tear glucose measurements	Estimated interstitial glucose level (mg/dL)	Non-temporal	None (IR-based custom non-invasive glucometer)	Tear fluid glucose detection using IR sensor	Combined	Yes	97.5% Zone A, 2.2% Zone B, 0.3% Zone C on Clarke Error Grid; error <3%	Yes
⁸⁴	Lin, 2023	Prediction of interstitial Glucose Concentration Based on OptiScanner and XGBoost in ICU	To improve ICU interstitial glucose prediction using mid-IR spectral data collected from OptiScanner and XGBoost, accounting for hetastarch treatment effects.	XGBoost, LASSO, Partial Least Squares Regression (PLS), SVM, RF, KNN, Logistic Regression	Mid-infrared spectral data from OptiScanner (25 wavelength bands between 7-10 µm)	Plasma glucose concentration (continuous value)	N/A (not time-series prediction; point estimation)	None (mid-IR OptiScanner used instead)	None	Single modality	Yes	XGBoost showed lowest RMSE among models; SHAP used for interpretability; Wavelength 3 most predictive	Yes
⁸⁵	Lee, 2023	An Integrated Digital Health Care Platform for Diabetes Management With AI-Based Dietary Management	To evaluate whether an integrated AI-based dietary management platform improves glycemic management and weight in T2D patients	Deep learning (CNN for image-based food recognition)	Meal photographs, glucometer, scale, sphygmomanometer, pedometer data (via Bluetooth)	HbA1c, weight	48 weeks	Dexcom G5	FoodLens (photo-based diet), BP monitor, weight scale, smartband pedometer	Combined	Yes	HbA1c reduced by ˆ’0.28% (B) to ˆ’0.49% (C) vs ˆ’0.06% (A); greater weight loss in groups B and C vs A	Partially
⁸⁶	Lee, 2023	Glucose Transformer: Forecasting Glucose Level and Events of Hyperglycemia and Hypoglycemia	To develop a deep learning model using the Transformer encoder to predict glucose levels and classify events of hypoglycemia/hyperglycemia in hospitalized T2D patients.	Transformer encoder, ECGGAN (BiLSTM + Conv1D GAN)	CGM data (5-min intervals, 3–7 days)	Predicted glucose level; classification into hypoglycemia, normoglycemia, or hyperglycemia	15, 30, and 45 minutes	Dexcom G5 (T2D); Dexcom G4 (T1D)	ECGGAN for augmentation	Combined	Yes	MAPE: best 12.78%; classification F1 improved with ECGGAN and T1D transfer learning	No
⁸⁷	Kistkins, 2023	Comparative Analysis of Predictive Interstitial Glucose Level Classification Models	Compare the efficacy of ARIMA, logistic regression, and LSTM models in classifying glucose states (hypo-, eu-, hyperglycemia) 15 min and 1 h in advance.	ARIMA, Logistic Regression, LSTM	CGM data; engineered features (e.g., rate of change, volatility, range oscillator); insulin; carbohydrate intake	Glucose classification: hypoglycemia, normoglycemia, hyperglycemia	15 minutes, 1 hour	NR (real-world CGM used; brand not specified)	Insulin dosing, carbohydrate intake	Combined	Yes	15-min: Logistic regression best (Accuracy 0.93); 1-h: LSTM best (Accuracy 0.73)	No
⁸⁸	Arbi, 2023	interstitial glucose estimation based on ECG signal	To develop a method to estimate interstitial glucose concentration (BGC) using ECG signal features instead of invasive CGMs.	CNN for segmentation; Regression models including Exponential Gaussian Process Regression (GPR), Linear, Nonlinear, Ensemble (Bagged Trees), Tree (Medium Tree)	ECG parameters: P wave, QRS, T wave intervals, QTc, heart rate (HR), T wave amplitude	Estimated interstitial glucose concentration (BGC)	Non-temporal	iPro2 Professional CGM	ECG (Zephyr BioHarness 3)	Combined	Not applicable	Exponential GPR: RMSE 0.32-0.67; R²: 70%-98% across patients	No
²³	Avram, 2020	A digital biomarker of diabetes from smartphone-based vascular signals	To detect prevalent diabetes using smartphone-based photoplethysmography and deep learning	Deep Neural Network (39-layer CNN)	Smartphone PPG recordings	Diabetes (prevalent)	Non-temporal	None	None		Yes	AUC = 0.766 (user-level), Sens=75%, Spec=65%, NPV = 97.4%, PPV = 13.3% (test set)	Yes
⁸⁹	Andellini, 2023	Artificial intelligence for non-invasive glycaemic-events detection via ECG in a pediatric population	To validate an AI-based algorithm for detecting glycaemic events (hypo/hyperglycemia) using ECG signals in T1D children	Deep learning models and machine learning models (SVM, ANN, Decision Trees)	ECG signals, respiration rate, 3D accelerometer data, CGM glucose levels	Glycaemic events (hypo, hyper, normal)	5-15 min ECG segments	Freestyle Libre, Dexcom G6, Medtronic Guardian	ECG, respiration, activity (via accelerometer)	Combined	Yes	Pending (validation phase; pilot study showed high accuracy)	Partially
⁹⁰	Alvarado, 2023	Combining wavelet transform with convolutional neural networks for hypoglycemia events prediction from CGM data	To detect and predict hypoglycemia events up to 24 hours in advance using deep learning and wavelet transforms	CNN (AlexNet, DenseNet-121) with wavelet preprocessing	24-hour CGM glucose time series	Hypoglycemia event (yes/no)	24 hours	Abbott FreeStyle Libre	None			Accuracy = 89.4%, Sens = 88.8%, Spec = 90.1%, Precision = 90.3%	No
⁹¹	Ahmed, 2023	Performance of artificial intelligence models in estimating interstitial glucose level among diabetic patients using non-invasive wearable device data	To evaluate AI model performance in estimating interstitial glucose levels using non-invasive wearable device data	RF, SVR, MLP, ANFIS (DL + traditional ML)	Heart rate, SPO2, diastolic BP, systolic BP, body temp, sweating/shivering (from smart band), CGM glucose readings	interstitial glucose level	Non-temporal	Freestyle LibrePro (CGM)	Riversong Wave O2 Smart Band (heart rate, SPO2, BP, temp, sweating, shivering)	Combined	Yes	RF (best RMSE); ANFIS (best MAE); RMSE: 0.099-0.197, MAE: 0.097-0.112	No
⁹²	Zale, 2022	Machine Learning Models for Inpatient Glucose Prediction	Review clinical evidence on machine learning models for predicting glucose trends in hospitalized patients and assess their predictive performance and clinical applicability.	Logistic regression, random forest, XGBoost, stochastic gradient boosting, recurrent neural network, boosted trees	EHR data (labs, vitals, meds), CGM data, insulin dose, GFR, demographic data, medication history	Hypoglycemia ( < 70 mg/dL), hyperglycemia ( > 180 mg/dL), controlled glucose range, and quantitative glucose predictions	Next glucose reading (median ~2–4 h), 24-hour risk, admission-level risk, ICU stay, 30-minute ahead prediction	Not specified by brand; CGM used in some studies for short-term prediction	EHR variables (labs, vitals), insulin dosing, glucocorticoids, nutrition, medications	Combined	Yes	AUC ranged from 0.71-0.96 for various models; best performing models (e.g., XGBoost, stochastic gradient boosting) had AUC ~ 0.90 or higher; CGM-based RMSE ~ 21.5 mg/dL; classification model by Zale 2022 had sensitivities 0.64-0.78 and specificities 0.80–0.90 across validation hospitals	No
⁹³	Xiong, 2022	Identifying daily activities of patient work for type 2 diabetes and co-morbidities: a deep learning and wearable camera approach	To classify self-management activities of people with type 2 diabetes and comorbidities using wearable camera data and deep learning.	Ensemble model combining Linear Discriminant Analysis (LDA), ResNet, ResNeXt, WideResNet, 3D ResNet, and YOLOv3	Video images captured by wearable camera (Edesix VB-300)	Classification of 12 daily self-management activity categories (e.g., Sleeping, Exercise, Food-related, Socializing, etc.)	Non-temporal	None	None			Top-1 accuracy: 82.7%, Top-5 accuracy: 87.1%	No
⁹⁴	Nemat, 2022	interstitial Glucose Level Prediction: Advanced Deep-Ensemble Learning Approach	To improve interstitial glucose prediction in Type 1 diabetes using deep-ensemble learning models and novel meta-learning techniques	Linear regression, Vanilla LSTM, Bidirectional LSTM; ensemble: stacking, multivariate LSTM, ConvLSTM	CGM data	interstitial glucose level	30 and 60 minutes	Medtronic Enlite	None			RMSE = 19.63 (best ensemble), MCC = 0.756, SE = 0.204	No
⁹⁵	Malerbi, 2022	Diabetic Retinopathy Screening Using Artificial Intelligence and Handheld Smartphone-Based Retinal Camera	To evaluate the diagnostic accuracy of a DL algorithm and a handheld retinal camera in detecting diabetic retinopathy in a real-world, underserved setting.	Convolutional Neural Network (CNN) - modified Xception architecture	Retinal fundus images (699—699 RGB) from smartphone-based camera (Phelcom Eyer)	Binary classification: more than mild diabetic retinopathy (mtmDR) vs. no/mild DR	Non-temporal	None	Handheld retinal camera (Phelcom Eyer)	Single modality		Sensitivity: 97.8%, Specificity: 61.4%, AUC: 0.89, PPV: 48.5%, NPV: 98.7%	Yes
⁹⁶	Lobo, 2022	A Data-Driven Approach to Classifying Daily Continuous Glucose Monitoring (CGM) Time Series	To identify a finite set of representative daily CGM profiles (motifs) for use in classification, modeling, and automated systems.	Unsupervised clustering based on motif extraction using RMSE in risk space	CGM	Daily CGM profile classification (motif label)	One day (24-hour daily profile)	Dexcom (various, not individually named)	None	Single modality		99.0% of test data classified correctly	Yes
⁹⁷	Lim, 2022	Multi-Task Disentangled Autoencoder for Time-Series Data in Glucose Dynamics	To develop a multi-task disentangled VAE (MD-VAE) for glucose forecasting, event detection, and temporal clustering	Disentangled Variational Autoencoder with dual decoders and attention modules	CGM glucose levels, basal insulin rates, time features	Reconstructed CGM sequences, event labels	30-60 minutes	Dexcom G6 (real), simulated CGM (T1DMS)	Basal insulin rates, time of day (encoded)	Combined	Yes	RMSE: 16.56-28.84 mg/dL; AUROC and F1-scores reported for event detection; cluster analysis using PCA of latent representations	Yes
²⁹	Kim, 2022	Identification of Healthy and Unhealthy Lifestyles by a Wearable Activity Tracker in Type 2 Diabetes	To classify healthy vs. unhealthy lifestyle patterns in T2D using data from wearable activity trackers and unsupervised clustering	Expectation-Maximization (EM) clustering + SHAP	Fitbit data: steps, heart rate, sleep, circadian variation	Lifestyle group (A = healthy, B = unhealthy)		None	Steps, HR, sleep variability metrics (from Fitbit)	Combined	Yes	SHAP values identified sleep regularity as key factor; clear separation of clusters with HbA1c differences	Yes
⁹⁸	Aloraynan, 2022	A Single Wavelength Mid-Infrared Photoacoustic Spectroscopy for Noninvasive Glucose Detection Using Machine Learning	To develop a noninvasive glucose detection system using single-wavelength MIR photoacoustic spectroscopy with ML classification	Ensemble classification model with random subspace sampling	Unprocessed acoustic spectrum (10-30 kHz range, 134 features) from artificial skin phantoms	Discrete glucose concentration classes (75-300 mg/dL)	Non-temporal	None	Photoacoustic signals using single-wavelength QCL laser (1080 cm)	Single modality	Not applicable	Accuracy = 90.4%; F1 = 94.5%; Clarkes EGA (100% in zones A + B after majority voting)	No
²⁴	Yin, 2021	DiabDeep: Pervasive Diabetes Diagnosis Based on Wearable Medical Sensors and Efficient Neural Networks	Develop an efficient, wearable-sensor-based diabetes diagnosis system for edge and server environments	DiabNN: stacked sparsely connected layers (server), H-LSTM recurrent layers (edge), grow-and-prune paradigm	WMS data (GSR, BVP, IBI, temperature, motion, environment), smartphone sensors, demographics	Binary (diabetic vs healthy); Multiclass (T1D, T2D, healthy)	Non-temporal	None	Smartphone ambient sensors, accelerometer, gyroscope, questionnaire demographics	Combined	Yes	Accuracy: 96.3% binary, 95.7% multiclass; edge: 95.3% binary, 94.6% multiclass; edge model 130— smaller	No
⁹⁹	Garcia-Tirado, 2021	Advanced hybrid artificial pancreas system improves on unannounced meal response - In silico comparison to currently available system	To evaluate a fully automated, adaptive hybrid artificial pancreas (RocketAP) with a novel bolus priming system for managing unannounced meals in type 1 diabetes using in silico simulations	Adaptive Model Predictive Control with Bolus Priming System	CGM data, insulin data, simulated unannounced meals	Plasma glucose levels	6-hour glucose control post-meal	Not specified	Insulin pump data	Combined	Yes	RocketAP outperformed USS-Virginia for unannounced meals: Postprandial TIR †‘, time >180 mg/dL; RMSE improved with personalization	Partially
²⁵	Deng, 2021	Deep transfer learning and data augmentation improve glucose levels prediction in type 2 diabetes patients	To improve short-term prediction of glucose levels in T2D using transfer learning and data augmentation on imbalanced datasets	CNN, RNN (GRU), SAN; with Transfer Learning (4 strategies); compared to GP, SVM, FNN	30-min CGM history (7 values, every 5-min), some physiological markers (e.g., insulin, hormones), demographics	Future glucose levels; hypoglycemia vs. no hypoglycemia; 3-class classification	5 to 60 minutes	Unspecified CGM (5-min intervals)	Hormone levels (e.g., cortisol, leptin), insulin, HbA1c	Combined	Yes	Accuracy >98% (binary), >89% (3-class); AUROC > 0.9 (binary); MAE: 13.5, RMSE: 19.1 (30-min horizon)	No
¹⁰⁰	Bent, 2021	Engineering digital biomarkers of interstitial glucose from noninvasive smartwatches	To develop digital biomarkers from noninvasive wearable data and food logs to predict personalized interstitial glucose excursions and real-time glucose values.	Decision tree, gradient boosting, random forest (for feature importance), logistic regression	Smartwatch data (PPG, EDA, temperature, accelerometry), food logs, demographics, HbA1c	Interstitial glucose levels; classification: PersLow, PersNorm, PersHigh	5-min intervals	Dexcom G6	Empatica E4 smartwatch, food logs, demographics	Combined	Yes	Classification accuracy = 84.3%; Glucose RMSE ~ 21 mg/dL; MAPE ~ 13-14%	Yes
¹⁰¹	Baig, 2021	Early Detection of Prediabetes and T2D Using Wearable Sensors and IoT-Based Monitoring Applications	To develop and evaluate an AI-powered early detection system for prediabetes and T2D using wearable technology and IoT-based monitoring	Adaptive Neuro-Fuzzy Inference System (ANFIS)	Heart rate, HR variability, breathing rate, ventilation, steps, cadence, demographics, BMI, HbA1c	Prediabetes and T2D diagnosis/classification	Non-temporal	No CGM used	Hexoskin smart vest, manual glucose/BMI/HbA1c records	Combined	Yes	Accuracy 91%, Kappa = 0.75, sensitivity = 94%, specificity = 90%, predictability = 72%	Yes
²⁷	Zhu, 2020	An Insulin Bolus Advisor for Type 1 Diabetes Using Deep Reinforcement Learning	To optimize mealtime insulin dosing in T1D using deep reinforcement learning bolus advisor with CGM data	Deep Reinforcement Learning (DDPG, actor-critic with DNN)	CGM, CARB intake, insulin on board, mealtime	Insulin bolus suggestion	Up to 5 hours post-meal	NR	Meal input (manual)	Combined	Yes	TIR†‘ (adults: 74.1†’80.9%, adolescents: 54.9†’61.6%), TBR, LBGI, HBGI, CV (all p < 0.05 or p < 0.01)	No
¹⁰²	Xie, 2020	Benchmarking Machine Learning Algorithms on interstitial Glucose Prediction for Type I Diabetes	To compare ML algorithms and classical ARX models for BG prediction in T1D using the same dataset	Elastic Net, Gradient Boosting Trees, Huber, Lasso, Random Forest, Ridge, SVR (Linear, RBF), LSTM, TCN	CGM, insulin, carbohydrate intake, heart rate, physical activity, finger-stick BG, temp, GSR, steps	interstitial glucose level	30-minute ahead	Medtronic Enlite (530 G)	Basis Peak fitness band, smartphone app	Combined	Sometimes	ARX & Ridge RMSE = 19.48 (lowest); LSTM TG = 5.51 min (highest); SVR RBF lowest ESODn; TCN more robust to BG oscillations	Partially
²⁶	He, 2020	interstitial glucose concentration prediction based on kernel canonical correlation analysis with particle swarm optimization and error compensation	To improve short-term interstitial glucose prediction using nonlinear modeling and personalized error compensation	Kernel CCA (KCCA), PSO, EC-CCA (error-compensated CCA)	Historical CGM values only	interstitial glucose concentration	5, 10, 15, 20, 25, 30 minutes	Medtronic MiniMed CGMS	None	Single modality		RMSE = 8.01-16.40 mg/dL; R² = 0.95-0.98; EC-CCA R² = 0.9954; Sensitivity = 94.37%, Specificity = 92.25%	Partially
²⁸	Sun, 2019	A Dual Mode Adaptive Basal-Bolus Advisor Based on Reinforcement Learning	To develop a personalized, dual-mode basal-bolus insulin advisor (ABBA) using reinforcement learning, adaptable to CGM or SMBG input	Actor-Critic Reinforcement Learning	SMBG or CGM data + meal timing/CARB estimates	Basal rate (BR) and 3 daily CIRs	Daily	Dexcom CGM (simulated)	SMBG (simulated alternative)	Separate	Yes	Reduced LBGI, stable HBGI, comparable % time in range CGM vs SMBG, improved MAGE and insulin use	Indirectly (via feature tuning, policy learning)
¹⁰⁴	Samadi, 2018	Automatic Detection and Estimation of Unannounced Meals for Multivariable Artificial Pancreas System	To develop a fuzzy-logic-based system to automatically detect unannounced meals and estimate CARB intake in T1D patients	Fuzzy Logic Qualitative Analysis	CGM data, insulin delivery, physiological state	CARB estimation and bolus insulin	2 hours (CARB estimation window)	CGM (5-min intervals)	BodyMedia & Zephyr (sleep, activity signals)	Combined	Yes	Sensitivity=93.5% (meals), 68% (snacks); FP rate=20.8%; MAE ± 28 g for CARB estimation	Yes
¹⁰⁵	Zecchin, 2015	Jump Neural Network for Real-Time Prediction of Glucose Concentration	To predict short-term (30 min) future glucose levels in T1D patients using a jump neural network incorporating CGM and meal information	Jump Neural Network (feed-forward)	CGM history, glucose rate of change, estimated glucose appearance from CARB intake (RaG), smoothed RaG trend	Future glucose concentration (mg/dL)	30 minutes	Dexcom SEVEN PLUS	CARB from meal pictures and RaG model	Combined	Yes	RMSE = 16.6 ± 3.1 mg/dL; TG = 18.5 ± 3.4 min; ESODnorm=9.6 ± 1.6; 94% hyper events predicted with 23 min anticipation	Partially
¹⁰⁶	Zecchin, 2014	Jump neural network for online short-time prediction of interstitial glucose…	To evaluate a novel jump neural network architecture for short-term glucose forecasting after meals	Jump Neural Network	CGM readings, meal intake info (CARB amount and time)	interstitial glucose levels	30, 60, 90 minutes	Abbott FreeStyle Navigator	Carbohydrate intake info	Combined	Yes	RMSE at 30-min horizon = 20.27 mg/dL (lowest)	No

Open in a new tab

ACC Accuracy, AI Artificial Intelligence, AUC Area Under the Curve, AUROC Area Under the Receiver Operating Characteristic Curve, ANN Artificial Neural Network, ANFIS Adaptive Neuro-Fuzzy Inference System, ARIMA AutoRegressive Integrated Moving Average, BG Blood Glucose, BGC Blood Glucose Concentration, BiLSTM Bidirectional Long Short-Term Memory, BMI Body Mass Index, BP Blood Pressure, BVP Blood Volume Pulse, CARB Carbohydrate, CCA Canonical Correlation Analysis, CEGA Clarke Error Grid Analysis, CGM Continuous Glucose Monitoring, COB Carbohydrates on Board, CNN Convolutional Neural Network, CV Coefficient of Variation, DNN Deep Neural Network, DSS Decision Support System, EC-CCA Error-Compensated Canonical Correlation Analysis, ECG Electrocardiogram, ECGGAN ECG + GAN Combined Model, EDA Electrodermal Activity, EHR Electronic Health Record, EMR Electronic Medical Record, F1 F1-score (Harmonic Mean of Precision and Recall), Fit. Fitness Tracker, GAN Generative Adversarial Network, GRU Gated Recurrent Unit, GSR Galvanic Skin Response, HbA1c Hemoglobin A1c, HR Heart Rate, IBI Inter-Beat Interval, IOB Insulin on Board, IR Insulin Resistance, KNN K-Nearest Neighbors, LASSO Least Absolute Shrinkage and Selection Operator, LSTM Long Short-Term Memory, MAE Mean Absolute Error, MAPE Mean Absolute Percentage Error, MARD Mean Absolute Relative Difference, MARL Multi-Agent Reinforcement Learning, MCC Matthews Correlation Coefficient, MLP Multilayer Perceptron, MIR Mid-Infrared, MSE Mean Squared Error, NR Not Reported, NPV Negative Predictive Value, OGTT Oral Glucose Tolerance Test, PCA Principal Component Analysis, PLS Partial Least Squares, PPG Photoplethysmography, Prec. Precision, PSO Particle Swarm Optimization, QCL Quantum Cascade Laser, QRF Quantile Regression Forest, RaG Rate of Glucose Appearance, RF Random Forest, R² Coefficient of Determination, RBF Radial Basis Function, RMSE Root Mean Square Error, ROC Receiver Operating Characteristic, RR Respiratory Rate, Sens. Sensitivity, SHAP SHapley Additive exPlanations, SMBG Self-Monitoring of Blood Glucose, Spec. Specificity, SpO₂ Peripheral Capillary Oxygen Saturation, SVR Support Vector Regression, SVM Support Vector Machine, T1D Type 1 Diabetes, T2D Type 2 Diabetes Mellitus, TAR Time Above Range, TBR Time Below Range, TDD Total Daily Dose (insulin), TG Time Gain (prediction lead time), TIR Time in Range, VAE Variational Autoencoder.

Studies were conducted across diverse geographical regions, with 45% (27 of 60) conducted in North America, primarily in the United States; 30% (18 of 60) in Asian countries, notably China and South Korea; 20% (12 of 60) in Europe; and 5% (3 of 60) in other regions, such as Australia and the Middle East. Notably, regions like Africa and South America were underrepresented. Sample sizes varied from five to over 1000 participants, with a median size of 150 participants. Forty percent (24 of 60) of studies included fewer than 100 participants, potentially limiting generalizability. Most studies focused on adults with T2D, with an average participant age of 55 years. Gender distribution was relatively balanced, with 48% female participants. However, only 7% (4 of 60) of studies reported racial and ethnic demographics, with low representation of minority populations.

Wearable devices used in the studies included CGMs in 70% (42 of 60) of studies to provide real-time glucose monitoring. Other devices, such as fitness trackers and smartwatches, were utilized in 20% (12 of 60) of studies to capture physical activity, heart rate, and other metrics. Less common wearables, including photoplethysmography (PPG) sensors and electrodermal activity monitors, accounted for 10% (6 of 60) of studies. Data collected typically included minute-by-minute glucose readings, heart rate variability, physical activity, and sleep patterns, often integrated into AI models for predictive analysis.

AI architectures varied significantly across the studies, reflecting a shift toward advanced modeling techniques. Deep learning models, particularly recurrent neural networks (RNNs) and long short-term memory (LSTM) networks, were employed in 45% (27 of 60) of studies, owing to their ability to process time-series data from wearables. Traditional machine learning models, such as random forests and support vector machines (SVMs), were used in 30% (18 of 60) of studies for their interpretability. Emerging architectures like temporal fusion transformers and hybrid models constituted the remaining 25% (15 of 60), highlighting a trend toward sophisticated AI solutions.

Model performance and predictive accuracy

Most studies reported performance metrics, including root mean square error, mean absolute error, and area under the receiver operating characteristic curve, with prediction accuracies varying based on model complexity and data input quality. Sixty percent (36 of 60) of the studies achieved root mean squared error (RMSE) values within the clinically acceptable range for glucose prediction, typically under 15 mg/dL. Predictive accuracy was generally higher in studies utilizing large, high-frequency datasets from CGMs, with some advanced models achieving accuracy rates above 85% in predicting glucose levels within a 1–2-hour window.

Forty percent (24 of 60) of studies incorporated interpretability measures such as Shapley Additive Explanations (SHAP), and feature importance analysis to improve clinician trust and model transparency. Despite this, 60% (36 of 60) of studies used complex “black-box” models, posing barriers to clinical adoption due to limited transparency.

Results of individual studies

The 60 manuscripts reviewed explored a range of AI models with applications to data from wearable devices in T2D, encompassing early detection, diagnosis, real-time glucose monitoring, lifestyle interventions, and personalized insulin management. Several studies, such as those by Avram et al.²³. and Yin et al.²⁴, developed digital biomarkers using smartphone-based PPG and other wearable sensors to enable non-invasive diabetes detection. Deep neural networks and machine learning frameworks were leveraged to improve diagnostic accuracy, but limitations in data diversity and model generalizability were noted. Similarly, CGM devices, integrated with AI models such as recurrent neural networks and transfer learning approaches demonstrated by Deng et al.²⁵ and He et al.²⁶, showcased the capability to predict interstitial glucose trends; however, challenges related to data imbalance and patient-specific adaptation persisted. Other approaches, including ensemble models and temporal convolution networks, improved prediction accuracy, though the need for broader clinical validation remained a significant concern.

Beyond glucose monitoring, several studies explored real-time intervention strategies, including the use of reinforcement learning and fuzzy logic models for optimizing insulin dosing. For example, studies by Zhu et al.²⁷ and Sun et al.²⁸ demonstrated promising results in managing glycemic variability through AI-guided insulin adjustments. Integrating lifestyle data (e.g., sleep, exercise) from wearables, studies such as those by Kim et al.²⁹ and Ramazi et al.³⁰ provided insights into individualized glycemic management, underscoring the value of multimodal data in tailoring diabetes management.

Key limitations noted in the studies included interpretability of complex AI models and sources of bias. Studies focused on clustering algorithms and decision support systems highlighted the need for clearer model explanations to facilitate clinical adoption. Furthermore, gaps in real-world validation and the lack of broader demographic representation were consistently cited, underscoring the need for future research to address these limitations and improve the equity and clinical applicability of AI-driven solutions for T2D management.

Various biases observed in the manuscripts reviewed included selection bias, racial and ethnic bias, and other sources of bias. Selection bias was the most prevalent, stemming from homogeneous study populations often skewed toward certain racial or ethnic groups. This limited generalizability and resulted in the underrepresentation of minority populations, which impacted AI model performance across subpopulations and raised equity concerns. Additional biases, such as data source integrity, technological limitations (e.g., sensor accuracy), and outcome reporting bias, affected intervention consistency and result accuracy. While some studies employed robust methodologies, many lacked external validation and did not adequately address demographic imbalances, underscoring key areas for improvement.

Discussion

This systematic review highlights the increasing role of AI and wearable devices in managing T2D. The majority of studies utilized CGMs to provide granular, real-time glucose data, enabling AI models to predict glycemic variability and detect early hypo- or hyperglycemic events. Other wearables, such as fitness trackers and smartwatches, were used to monitor physiological parameters, including physical activity and heart rate, broadening the data scope available for AI-driven insights. AI models predominantly used deep learning architectures, including recurrent neural networks and long short-term memory networks, which excelled at capturing temporal data patterns from wearables. Traditional machine learning models, such as random forests, remained prevalent due to their interpretability, while emerging architectures demonstrated the potential for integrating high-dimensional data but often lacked transparency. Efforts to enhance model interpretability included tools like SHAP values, which improved transparency but were applied inconsistently across studies. However, several key gaps were identified. Only 7% of studies reported racial and ethnic demographics, with limited representation of racial and ethnic minority populations, particularly in U.S.-based studies. While the geographic distribution of studies was broad, including several from Asian countries, most did not report disaggregated demographic data, making it difficult to assess inclusion of underrepresented populations in any given context. Many studies did not include external validation, and smaller sample sizes limited the generalizability of their findings. None of the studies conducted long-term follow-ups, leaving gaps in understanding the prolonged impact of AI-driven interventions on T2D outcomes. While these advancements underscore the potential of AI and wearable devices to revolutionize diabetes management, they also reveal critical gaps in the existing body of research. Addressing these limitations is essential to ensure the equitable and effective application of AI-driven interventions in diverse clinical populations.

This systematic review fills an important gap in the literature by being the first to comprehensively synthesize studies that combine artificial intelligence with wearable-derived physiological data for the management of T2D. While many prior reviews have focused broadly on digital health or AI in diabetes care, they have not systematically assessed the combined application of wearable technology and AI modeling in T2D populations. In addition to mapping this intersection, our review introduces a framework that explicitly addresses issues of demographic inclusivity, model interpretability, and sensor fusion—areas that are often discussed in isolation but rarely integrated in prior literature reviews. These features differentiate our review and provide specific guidance for future development of equitable, explainable, and data-rich AI systems for diabetes care.

Broader limitations in the studies reviewed have significant implications for the development and application of AI models with wearable devices for T2D management. One prominent limitation is the small sample sizes used in many studies, which can lead to model overfitting and limit the generalizability of findings to broader populations^31,32. Small datasets may fail to capture the full range of physiological variability seen in individuals with T2D, reducing the robustness and reliability of AI predictions in clinical settings. Additionally, many studies demonstrated demographic homogeneity, with a significant proportion of research conducted on specific regional populations, such as predominantly Chinese or white non-Hispanic cohorts. This lack of diversity may result in biased AI models that underperform in underrepresented groups, raising concerns about equity in diabetes care.

Another notable limitation relates to data quality and imbalance. Wearable data, while rich in granularity, often suffers from issues such as missing data, noise, and inconsistencies due to patient adherence or device fatigue, sensor malfunctions, or differing data collection protocols. Such data quality issues can undermine model accuracy and skew predictions, particularly for rare events like hypoglycemia^33,34. The wearable devices themselves present additional challenges, as non-invasive sensors can produce less reliable data under specific conditions, such as varying skin tones, high physical activity, or temperature fluctuations. This impacts the real-time accuracy and reliability of AI models dependent on wearable inputs thereby reducing their overall utility.

Limited external validation of AI models is another critical constraint observed in the reviewed studies. Many models were trained and tested on the same datasets without independent validation, reducing their applicability in real-world settings. The lack of external validation poses a significant challenge for translating AI-driven interventions into routine clinical practice³⁵. Furthermore, the complex “black-box” nature of many AI models, such as deep neural networks, often leads to limited interpretability, hindering clinician trust and adoption^36,37. While efforts such as SHAP values³⁸ and feature importance ranking have been employed to improve model transparency, these methods were applied inconsistently across studies.

Finally, few diabetes-specific studies have directly assessed differential AI model performance across diverse populations, prior research in other health domains such as imaging, genomics, and EHR-based prediction has demonstrated disparities by race, sex, and age^39–44. As such, ensuring diversity in data and transparency in validation remains an important research priority to avoid inequities as AI-driven wearables advance. These inclusive models aim to minimize performance disparities across subgroups and reduce algorithmic bias in healthcare settings. In the short term, inclusivity can be improved by mandating standardized reporting of participant demographics and evaluating model performance across subgroups. In the long term, efforts should focus on the development of large, federated, and representative datasets; incorporation of fairness-aware modeling techniques; and the establishment of regulatory standards for equity auditing. These strategies are critical to ensuring that AI-driven wearable solutions are equitable, generalizable, and beneficial to all individuals with T2D, especially those historically underrepresented in research and clinical care.

Historical clinical practices and demographic imbalances in data can lead to biased model predictions, with potentially harmful consequences for specific patient subgroups^45–49. Additionally, integrating data from multiple wearable sensors presents challenges, as differences in data formats, sampling rates, and reliability complicate fusion and interpretation. Addressing these challenges is critical to ensuring equitable and effective application of AI-driven wearable technologies in clinical practice.

The integration of AI-driven wearable technologies into clinical practice for T2D management offers significant potential for improving patient outcomes but also poses practical challenges. AI-enabled wearables facilitate real-time glucose tracking and predictive intervention, reducing glycemic variability and preventing acute complications such as hypoglycemia or hyperglycemia. By offering timely, personalized insights, these technologies support proactive diabetes management, minimizing the need for emergency care and enhancing long-term outcomes. Wearables also empower individuals with diabetes to take an active role in their care, fostering better adherence and self-management by providing real-time feedback on how daily lifestyle factors impact glucose levels⁵⁰. However, these benefits are tempered by implementation challenges, including barriers to clinical adoption that must be addressed to maximize the potential of AI-driven wearable devices.

A major challenge is the limited interpretability of complex AI models. Clinicians may be reluctant to rely on AI-driven recommendations when they cannot fully understand or explain the model’s decision-making process to individuals with diabetes. Improving transparency through tools like SHAP³⁸ values and feature importance analysis is essential to gaining clinician trust and ensuring effective integration into practice.

In addition to interpretability, logistical challenges, such as infrastructure development and clinician training, must also be overcome to support the adoption of AI-enabled wearables. Moreover, interpreting and responding to AI-generated outputs requires additional clinician time, which may not be reimbursed under current healthcare models. Without appropriate policy changes, this may inadvertently increase clinician workload rather than alleviate it. Implementing AI-driven wearables requires robust infrastructure and targeted clinician training programs. Many healthcare professionals are unfamiliar with AI’s technical aspects, which may hinder effective use and interpretation of wearable data. Comprehensive training focused on AI fundamentals, data interpretation, and clinical implications will be crucial. Cost also remains a significant barrier, as high-quality wearables may be prohibitively expensive for underserved populations. Expanding insurance coverage and improving accessibility are necessary for equitable adoption. Lastly, over-reliance on wearable-generated data poses risks, as AI insights cannot replace holistic clinical assessments or personalized patient care plans. Clinicians must balance technology use with traditional care to ensure comprehensive diabetes management. Addressing these challenges will require targeted research to close existing gaps and optimize AI-driven solutions.

Future research should prioritize several areas to enhance the effectiveness, equity, and reliability of AI-driven wearable solutions. Developing benchmark datasets with diverse patient data from multiple sources is essential to improve model validation and reproducibility⁵¹. Increasing diversity in patient populations included in AI model development will ensure that models better reflect the physiological and lifestyle factors relevant to T2D. Additionally, expanding multimodal AI techniques to integrate data from CGMs, heart rate monitors, and activity trackers can provide richer, context-aware predictions. Such advancements would enable more personalized and equitable care.

Emerging AI models, such as liquid neural networks (LNNs), exemplify the potential for innovation in this field⁵². By dynamically adapting to real-time data, LNNs excel at capturing rapid changes in physiological signals, such as fluctuating glucose levels, making them highly effective for wearable device applications. Their resilience against noisy or incomplete data and low computational overhead enables efficient on-device processing⁴⁶. However, challenges such as complex training requirements and regulatory concerns must be addressed to fully realize their potential. Future efforts should focus on balancing these limitations with their adaptability to ensure practical applications in healthcare.

Cost-effectiveness analyses comparing AI-driven interventions to traditional care approaches are also needed, as they provide evidence on the economic benefits and efficiency of AI technologies, helping to inform healthcare policy and drive broader adoption of these innovative solutions. Finally, developing practical interpretability tools for complex models and investigating how clinician trust in AI can be enhanced will ensure AI-driven wearables are effectively and equitably integrated into clinical practice. Research in these areas will be critical for transforming AI-driven wearables from innovative concepts into trusted, impactful tools in T2D care.

There is also a pressing need to move beyond simplistic metrics such as time-in-range, which often fail to capture the rich variability in physiological patterns. Leveraging the entirety of time-series data generated by wearables can provide deeper insights into individual health, enabling more precise anomaly detection and trend prediction. To achieve this, researchers must develop scalable algorithms capable of integrating contextual data while addressing critical issues such as data privacy and ethical considerations. These efforts will be essential to unlocking the full potential of wearable technologies in diabetes care.

This systematic review has limitations that should be acknowledged. First, the exclusion of gray literature and non-English studies may have resulted in the omission of relevant research, potentially introducing selection bias. Additionally, the review relied on studies with variable methodologies, data quality, and reporting standards, which may have introduced heterogeneity that was challenging to fully address in the synthesis. Lastly, the absence of a meta-analysis, due to study heterogeneity, restricted the quantitative integration of results, which could have strengthened the overall conclusions. These limitations highlight the need for more comprehensive and standardized research in this field.

This systematic review emphasizes the transformative potential of AI and wearable devices in managing T2D. Despite challenges, these technologies have shown significant promise in enabling real-time glucose monitoring, early intervention, and personalized feedback, offering new opportunities to improve glycemic management and patient self-management. However, several critical gaps must be addressed to ensure their equitable and effective implementation in clinical practice. These include addressing limited demographic diversity in study populations, use of benchmark datasets for validating and comparing AI models, and improving model interpretability and transparency. Future research should prioritize the development of inclusive AI models that account for diverse patient populations, the enhancement of complex algorithm explainability, and the exploration of multimodal AI approaches for integrating data from various wearable devices. Addressing these gaps is vital to unlocking the full potential of AI-driven wearables and transforming diabetes care into a more personalized, data-driven, and effective approach for diverse populations.

Methods

Information sources and search strategy

This systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines⁵³. The review protocol was registered with PROSPERO under registration number 1009318 before conducting the review. A comprehensive search strategy was developed to identify relevant studies on the application of AI models in wearable devices for T2D management. The strategy included controlled vocabulary (e.g., MeSH terms) and free-text keywords, combined using Boolean operators. Search concepts included:

Type 2 Diabetes Mellitus: Terms related to type 2 diabetes, including common synonyms such as “Non-Insulin Dependent Diabetes Mellitus” and “Maturity-Onset Diabetes,” were used⁵⁴.
Wearable Devices: Terms related to wearable technology, including “wearable devices,” “smartwatches,” “continuous glucose monitors,” and “fitness trackers” were used^17,55,56.
Artificial Intelligence: A comprehensive list of AI-related terms was used, including “machine learning,” “deep learning,” “neural networks,” and “predictive analytics”^8,9.

Database-specific syntax and filters were employed to maximize retrieval efficiency, with filters specifically used to exclude gray literature. The last comprehensive search was conducted on September 30, 2024, covering PubMed, IEEE Xplore, ACM Digital Library, and Embase. No additional studies were identified through manual reference searches. Specific keyword searches for each database are provided in Supplementary Tables 1–4.

Manuscript eligibility criteria

The eligibility criteria for manuscripts included were defined based on language, study type, population, intervention, and outcomes. Specific inclusion and exclusion criteria were established to ensure relevance and focus on the application of AI models in wearable devices for diabetes management. Although our primary focus was on AI models and wearable devices in T2D, we also considered studies involving other diabetes populations (e.g., T1D, gestational diabetes mellitus, prediabetes) under specific conditions. These studies were included if (1) they had mixed cohorts that included individuals with T2D, (2) they evaluated AI algorithms or wearable-device applications that are translatable or generalizable to T2D care (e.g., glucose prediction models based on CGM data, which are physiologically relevant across diabetes types), or (3) they provided methodological insights or technological innovations with clear implications for advancing T2D management. For example, AI techniques developed in T1D populations for hypoglycemia prediction may be adaptable to T2D, particularly as CGM use expands among people with T2D. Similarly, studies in prediabetes populations were included if they addressed preventive strategies directly applicable to delaying T2D onset. Conversely, studies that focused exclusively on populations or outcomes without relevance to T2D—such as GDM interventions unrelated to longer-term metabolic risk or T1D algorithms designed solely for closed-loop insulin delivery without broader implications—were excluded.

Studies focused solely on mobile health interventions, digital coaching platforms, or app-based behavioral prompts—without application of AI algorithms to wearable-derived physiological data—were excluded. This review emphasizes AI-powered models that utilize data streams from wearable sensors, such as CGMs or smartwatches, to support tasks like prediction, classification, detection, and clinical decision-making relevant to Type 2 Diabetes management. Inclusion criteria were: (1) Peer-reviewed articles published in English from January 2014 to September 2024, (2) Manuscripts involving pre-diabetes, Type 1 and Type 2 diabetes mellitus, and relevant applications in healthy individuals, artificial skin phantoms, or in silico models, (3) Use of wearable devices, including CGMs, smartwatches, fitness trackers, or health-tracking wearables, to capture physiological data relevant to diabetes management (e.g., glucose levels, heart rate, physical activity), (4) Application of AI techniques (e.g., machine learning, deep learning, predictive analytics) on wearable data, (5) Reporting of diabetes-related clinical or self-management outcomes (e.g., glycemic management, glycemic events, medication adherence, or prevention of complications), and (6) Inclusion of experimental (e.g., randomized controlled trials) and observational study designs (e.g., cohort, cross-sectional, case-control).

Exclusion criteria included: (1) Manuscripts not in English or published prior to January 2014, (2) Non-peer-reviewed literature (e.g., conference abstracts, opinion pieces, gray literature), (3) Manuscripts unrelated to diabetes management or lacking relevant outcomes., (4) Manuscripts without wearable devices or using stationary monitors, (5) Manuscripts not applying AI or machine learning to wearable data, and (6) Non-empirical studies (e.g., reviews, theoretical papers, meta-analyses).

Study selection

The study selection process is illustrated in a PRISMA flow diagram in Fig. 2, detailing the number of studies screened, retrieved, and included. A systematic search of PubMed, IEEE Xplore, ACM Digital Library, and Embase yielded a total of 5152 records. Following the removal of 214 duplicates and 527 ineligible records, 1,111 records underwent title and abstract screening. Of these, 1042 records were excluded based on the predefined criteria, leaving 69 reports for retrieval. Six manuscripts could not be retrieved due to inaccessible links or unavailable full texts from the publisher despite repeated effort and 3 were excluded because they did not use AI models. After the full-text review, 60 studies met the inclusion criteria and were included in the final analysis^{8,11,22–30,57–106}.

The selection of studies followed a two-phase process. Initially, titles and abstracts of retrieved articles were screened independently by two reviewers using the predefined inclusion and exclusion criteria listed above. Articles that did not meet the eligibility criteria were excluded at this stage. Any discrepancies between reviewers were resolved through discussion, or, if necessary, by consulting a third reviewer. In the second phase, one reviewer conducted a detailed assessment of the full texts of articles that passed the initial screening. The inclusion and exclusion criteria were rigorously applied to ensure only relevant studies were included. A Fig. 2 provides a PRISMA flow diagram documenting the number of studies screened, assessed for eligibility, and included along with the specific reasons for exclusions.

Data extraction and data synthesis

Data extraction was conducted manually using Microsoft Excel, with a predefined template to ensure consistency. From each manuscript, we extracted 25 structured fields including study metadata (author, year, title, country), study objectives, AI/ML model types, model inputs and outputs, prediction horizon, population size, glycemic and metabolic profiles, demographics (age, sex, race/ethnicity), comorbidities, study setting and duration, CGM devices and other modalities used, data sources, dataset type, modality integration and impact, model performance metrics, and whether interpretability was addressed. One reviewer conducted the data extraction using a predefined template, with best practices including piloting the extraction process, revisiting source material for verification, and documenting all decisions to ensure accuracy and consistency.

Data from the included studies were synthesized using both qualitative and quantitative approaches, with a primary emphasis on narrative synthesis to accommodate the substantial variability in study designs, AI model architectures, and wearable device characteristics. Thematic grouping of studies facilitated the identification of trends and patterns without relying on meta-analysis, which was precluded by the heterogeneity in methodologies and populations. Quantitative results, including key performance metrics such as accuracy and sensitivity, were summarized descriptively to provide context to the narrative findings. Variations in AI models, such as neural networks versus regression-based approaches, and differences in wearable device functionalities were qualitatively explored to ensure a comprehensive understanding of study outcomes. This approach emphasized transparency and rigor in addressing and reporting the methodological and contextual diversity of the included studies.

Supplementary information

PRISMA 2020 Checklist^{(184.7KB, pdf)}

Acknowledgements

Funding: Effort for this study was partially supported by the American Diabetes Association (Grant #7-25-ICTSPC-412, PI: Fraser), National Institute of Diabetes and Digestive Kidney Disease (R01DK118038, R01DK120861, PI: Egede, K01DK131319, PI: Campbell) and the National Institute for Minority Health and Health Disparities (R01MD013826, PI: Egede/Walker, R01MD017574, R01MD018012, PI: Egede/Linde, R01MD018721, PI: Walker).

Author contributions

R.F., R.W., and J.C. conceived the study and designed the systematic review framework. R.F. and J.C. conducted the literature search and data extraction. R.W. and R.F. performed the data synthesis. R.F. provided expertise in AI modeling and wearable technology evaluation. L.E. and O.E. contributed to the interpretation of findings and critical revisions. R.F., drafted the manuscript with input and editing from all authors. All authors reviewed and approved the final manuscript.

Data availability

No datasets were generated or analysed during the current study.

Code availability

No custom code or software was used in the data extraction or analysis for this systematic review.

Competing interests

The authors declare no competing interests.

Footnotes

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary information

The online version contains supplementary material available at 10.1038/s41746-025-02036-9.

References

1.Bahari, N. I. et al. Issues and challenges in the primary prevention of type 2 diabetes mellitus: A systematic review. J. Prev.44, 105–125 (2023). [DOI] [PubMed] [Google Scholar]
2.Roman, S. H. & Harris, M. I. Management of diabetes mellitus from a public health perspective. Endocrinol. Metab. Clin.26, 443–474 (1997). [DOI] [PubMed] [Google Scholar]
3.International Diabetes Federation. IDF Diabetes Atlas. Available from: www.diabetesatlas.org (2021).
4.Institute for Health Metrics and Evaluation. Global diabetes prevalence and forecasts. Available from: www.healthdata.org (2023).
5.Zhou, B. et al. Worldwide trends in diabetes prevalence and treatment from 1990 to 2022: a pooled analysis of 1108 population-representative studies with 141 million participants. Lancet404, 2077–2093 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
6.Centers for Disease Control and Prevention. About type 2 diabetes. U.S. Department of Health and Human Services. Available at: https://www.cdc.gov/diabetes/about/about-type-2-diabetes.html (Accessed Dec 12, 2024).
7.Sarma, A. D. & Devi, M. Artificial intelligence in diabetes management: transformative potential, challenges, and opportunities in healthcare. Hormones24, 307–322 (2025). [DOI] [PubMed] [Google Scholar]
8.Ahmed, A., Aziz, S., Abd-Alrazaq, A., Farooq, F. & Sheikh, J. Overview of Artificial Intelligence-Driven Wearable Devices for Diabetes: Scoping Review Review. J. Med. Internet Res.24, e36010 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
9.Makroum, M. A. A. M., Bouzouane, A. & Ibrahim, H. Machine learning and smart devices for diabetes management: Systematic review. Sens. (Basel, Switz.)22, 1843 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
10.Iftikhar, M. et al. Artificial intelligence-driven transformations in diabetes care: a comprehensive literature review. Ann. Med. Surg.86, 5334–5342 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
11.Anjum, M., Saher, R. & Saeed, M. N. Optimizing type 2 diabetes management: AI-enhanced time series analysis of continuous glucose monitoring data for personalized dietary intervention. PeerJ Comput Sci.10, e1971 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
12.Wong, C. et al. Glycaemic outcomes in people living with diabetes under 65 and over 65 years old using an intermittently scanned continuous glucose monitoring system. Ther. Adv. Endocrinol. Metab.15, 20420188241269133 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
13.Shen, Y., Zou, X. & Hu, G. Digital technology in the management and prevention of diabetes. Front Endocrinol.15, 1511236 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
14.Guerci, B. Insulin-treated diabetes in general practice. La Rev. du Praticien.74, S5–S12 (2024). [PubMed] [Google Scholar]
15.Thomsen, C. H. et al. The intelligent diabetes telemonitoring using decision support to treat patients on insulin therapy (DiaTRUST) trial: study protocol for a randomized controlled trial. Trials25, 744 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
16.Vettoretti, M., Cappon, G., Facchinetti, A. & Sparacino, G. Advanced diabetes management using artificial intelligence and continuous glucose monitoring sensors. Rev. Sens. (Basel, Switz.)20, 3870 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
17.Rodriguez-León, C., Villalonga, C., Munoz-Torres, M., Ruiz, J. R. & Banos, O. Mobile and wearable technology for the monitoring of diabetes-related parameters: Systematic review. JMIR mHealth uHealth.9, e25138 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
18.Alhalafi, A. et al. Utilizing artificial intelligence among patients with diabetes: A systematic review and meta-analysis. Cureus16, e58713 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar] [Retracted]
19.Jin, X., Cai, A., Xu, T. & Zhang, X. Artificial intelligence biosensors for continuous glucose monitoring. Interdiscip. Mater.2, 290–307 (2023). [Google Scholar]
20.Jung, H. H., Lee, H., Yea, J. & Jang, K.-I. Wearable electrochemical sensors for real-time monitoring in diabetes mellitus and associated complications. Soft Sci. 4, 15 (2024).
21.Ribeiro G., Monge J., Postolache O., Pereira J. M. D. A novel AI approach for assessing stress levels in patients with type 2 diabetes mellitus based on the acquisition of physiological parameters acquired during daily life. Sensors (Basel, Switzerland) 24. 10.3390/s24134175 (2024). [DOI] [PMC free article] [PubMed]
22.Bergoeing M. et al. Exploring the potential of an AI-integrated cloud-based mHealth platform for enhanced type 2 diabetes mellitus management. 100-111 (Springer, 2023).
23.Avram, R. et al. A digital biomarker of diabetes from smartphone-based vascular signals. Nat. Med.26, 1576–1582 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
24.Yin, H., Mukadam, B., Dai, X. & Jha, N. K. DiabDeep: Pervasive Diabetes Diagnosis Based on Wearable Medical Sensors and Efficient Neural Networks. IEEE Trans. Emerg. Top. Comput.9, 1139–1150 (2021). [Google Scholar]
25.Deng, Y. et al. Deep transfer learning and data augmentation improve glucose levels prediction in type 2 diabetes patients. NPJ Digit. Med.4, 109 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
26.He, J & Wang Y. Blood glucose concentration prediction based on kernel canonical correlation analysis with particle swarm optimization and error compensation. Comput. Methods Programs Biomed. 2020;196. 10.1016/j.cmpb.2020.105574. [DOI] [PubMed]
27.Zhu, T., Li, K., Kuang, L., Herrero, P. & Georgiou, P. An insulin bolus advisor for type 1 diabetes using deep reinforcement learning. Sensors20, 5058 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
28.Sun, Q. et al. A dual mode adaptive basal-bolus advisor based on reinforcement learning. IEEE J. Biomed. Health Inform.23, 2633–2641 (2019). [DOI] [PubMed] [Google Scholar]
29.Kim, K. J. et al. Identification of healthy and unhealthy lifestyles by a wearable activity tracker in type 2 diabetes: A Machine Learning-based Analysis. Endocrinol. Metab.37, 547–551 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
30.Ramazi, R., Perndorfer, C., Soriano, E. C., Laurenceau, J. P. & Beheshti, R. Predicting progression patterns of type 2 diabetes using multi-sensor measurements. Smart Health21, 100206 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
31.Vabalas, A., Gowen, E., Poliakoff, E. & Casson, A. J. Machine learning algorithm validation with a limited sample size. PloS one14, e0224365 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
32.Larracy R., Phinyomark A., Scheme E. Machine learning model validation for early stage studies with small sample sizes. 2314-2319 (IEEE, 2021). [DOI] [PubMed]
33.Gudivada, V., Apon, A. & Ding, J. Data quality considerations for big data and machine learning: Going beyond data cleaning and transformations. Int. J. Adv. Softw.10, 1–20 (2017). [Google Scholar]
34.Gong, Y., Liu, G., Xue, Y., Li, R. & Meng, L. A survey on dataset quality in machine learning. Inf. Softw. Technol.162, 107268 (2023). [Google Scholar]
35.Slieker, R. C., van der Heijden, A. A. W. A., Siddiqui, M. K., Langendoen-Gort, M., Nijpels, G., Herings, R., Feenstra, T. L., Moons, K. G. M., Bell, S., Elders, P. J., 't Hart, L. M., & Beulens, J. W. J. Performance of prediction models for nephropathy in people with type 2 diabetes: systematic review and external validation study. BMJ 374 (2021) [DOI] [PMC free article] [PubMed]
36.MoDastoni, D. A. Exploring methods to make AI decisions more transparent and understandable for humans. Adv. Eng. Innov.3, 32–36 (2023). [Google Scholar]
37.Beltramin D., Lamas E., Bousquet C. Ethical issues in the utilization of black boxes for artificial intelligence in medicine. Advances in Informatics, Management and Technology in Healthcare. 249-252, (IOS Press, 2022). [DOI] [PubMed]
38.Lundberg S.M. & Lee SI. A unified approach to interpreting model predictions. Advances in neural information processing systems, 30 (2017).
39.Pham, Q., Gamble, A., Hearn, J. & Cafazzo, J. A. The need for ethnoracial equity in artificial intelligence for diabetes management: Review and recommendations. J. Med Internet Res.23, e22320 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
40.Bozkurt, S., Cahan, E. M. & Seneviratne, M. G. Reporting of demographic data and representativeness in machine learning models using electronic health records. J. Am. Med Inf. Assoc.27, 1878–1885 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
41.Cirillo, D., Catuara-Solarz, S., Morey, C. & Guney, E. Sex and gender differences and biases in artificial intelligence for biomedicine and healthcare. NPJ Digit. Med.3, 81 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
42.Murphy, A., Bowen, K., El Naqa, I. & Yoga, B. Bridging health disparities in the data-driven world of artificial intelligence: a narrative review. J. Racial Ethn. Health Disparities.12, 2367–2379 (2025). [DOI] [PubMed] [Google Scholar]
43.Cau, R., Pisu, F., Suri, J. S. & Saba, L. Addressing hidden risks: Systematic review of artificial intelligence biases across racial and ethnic groups in cardiovascular diseases. Eur. J. Radiol.183, 111867 (2025). [DOI] [PubMed] [Google Scholar]
44.Perets O., Stagno E., Ben Yehuda E. Inherent bias in electronic health records: a scoping review of sources of bias. ACM Trans Comput Healthcare; 10.1145/3757924 (2025).
45.Parate, A. P., et al. Review of data bias in healthcare applications. Int. J. Online Biomed. Eng. 20, 124–136 (2024).
46.Babu, A. & Joseph, A. P. Artificial intelligence in mental healthcare: transformative potential vs. the necessity of human interaction. Front. Psychol.15, 1378904 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
47.Huang K., Wang Y & Xu M. Investigating the fairness of deep learning models in breast cancer diagnosis based on race and ethnicity. Proceedings of the AAAI Symposium Series, 4, 303–307 (2024)
48.Gu, T. et al. Mitigating bias in AI mortality predictions for minority populations: a transfer learning approach. BMC Med. Inform. Decis. Mak.25, 30 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
49.Moukheiber M., Moukheiber L., Moukheiber D & Lee H.-C. Unmasking societal biases in respiratory support for ICU patients through social determinants of health. Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence AI for Good. 7421–7429 (2024).
50.Topol, E. J. High-performance medicine: the convergence of human and artificial intelligence. Nat. Med.25, 44–56 (2019). [DOI] [PubMed] [Google Scholar]
51.Nguyen, T. et al. A novel decentralized federated learning approach to train on globally distributed, poor quality, and protected private medical data. Sci. Rep.12, 8888 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
52.Hasani R., Lechner M., Amini A., Rus D & Grosu R. Liquid time-constant networks. Proceedings of the AAAI Conference on Artificial Intelligence.35, 7657–7666 (2021).
53.Page M. J. et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ 372, 10.1136/bmj.n71 (2021). [DOI] [PMC free article] [PubMed]
54.Schellenberg, E. S., Dryden, D. M., Vandermeer, B., Ha, C. & Korownyk, C. Lifestyle interventions for patients with and at risk for type 2 diabetes: a systematic review and meta-analysis. Ann. Intern. Med.159, 543–551 (2013). [DOI] [PubMed] [Google Scholar]
55.Ahmed, A. et al. The effectiveness of wearable devices using artificial intelligence for blood glucose level forecasting or prediction: Systematic review. J. Med. Internet Res.25, e40259 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
56.Piet, A. et al. Non-invasive wearable devices for monitoring vital signs in patients with type 2 diabetes mellitus: a systematic review. Bioengineering10, 1321 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
57.Zhu, T. et al. Population-Specific Glucose Prediction in Diabetes Care With Transformer-Based Deep Learning on the Edge. IEEE Trans. Biomed. Circuits Syst.18, 236–246 (2024). [DOI] [PubMed] [Google Scholar]
58.Zhu, T., Afentakis, I., Li, K., Armiger, R., Hill, N., Oliver, N., & Georgiou, P. Multi-Horizon Glucose Prediction Across Populations With Deep Domain Generalization. J. Biomed. Sci. Health Informatics, 29, 5424–5437 (IEEE, 2025). [DOI] [PubMed]
59.Warren, M., Bergenstal, R. M., Hager, M., Bashan, E. & Hodish, I. A scalable application of artificial intelligence-driven insulin titration program to transform type 2 diabetes management. Diab. Technol. Ther.26, 556–565 (2024). [DOI] [PubMed] [Google Scholar]
60.Wang, X. et al. A novel hypoglycemia alarm framework for type 2 diabetes with high glycemic variability. Int J. Numer Method Biomed. Eng.40, e3799 (2024). [DOI] [PubMed] [Google Scholar]
61.Tucker, A. P., Erdman, A. G., Schreiner, P. J., Ma, S. & Chow, L. S. Neural networks with gated recurrent units reduce glucose forecasting error due to changes in sensor location. J. Diab. Sci. Technol.18, 124–134 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
62.Tao, R. et al. DDLA: a double deep latent autoencoder for diabetic retinopathy diagnose based on continuous glucose sensors. Med. Biol. Eng. Comput.62, 3089–3106 (2024). [DOI] [PubMed] [Google Scholar]
63.Shuzan M. N. I. et al. QU-GM: An IoT based glucose monitoring system from photoplethysmography, blood pressure and demographic data using machine learning. (IEEE, 2024).
64.Seo W., Kim N., Park S. W., Jin S. M., Park S. M. Generative adversarial network-based data augmentation for improving hypoglycemia prediction: A proof-of-concept study. Biomed. Signal Processing Control 92. 10.1016/j.bspc.2024.106077 (2024).
65.Yin, H. & Jha, N. K. A health decision support system for disease diagnosis based on wearable medical sensors and machine learning ensembles. IEEE Trans. Multi-Scale Comput. Syst.3, 228–241 (2017). [Google Scholar]
66.Kim S. H., Kim D. Y., Chun S. W., Kim J., Woo J. Impartial feature selection using multi-agent reinforcement learning for adverse glycemic event prediction. Comput. Biol. Med. 173. 10.1016/j.compbiomed.2024.108257 (2024). [DOI] [PubMed]
67.Kashif, M., Flesca, S. & Veltri, P. GLSTM: On using LSTM for glucose level prediction. Stud. health Technol. Inform.314, 103–107 (2024). [DOI] [PubMed] [Google Scholar]
68.Kannenberg S. et al. Unlocking potential: Personalized lifestyle therapy for type 2 diabetes through a predictive algorithm-driven digital therapeutic. J. Diabetes Sci. Technol. 10.1177/19322968241266821 (2024). [DOI] [PMC free article] [PubMed]
69.Hotta S., Kytö M., Koivusalo S., Heinonen S., Marttinen P. Optimizing postprandial glucose prediction through integration of diet and exercise: Leveraging transfer learning with imbalanced patient data. PLoS ONE;19. 10.1371/journal.pone.0298506 (2024). [DOI] [PMC free article] [PubMed]
70.Metwally A. A. et al. Predicting type 2 diabetes metabolic phenotypes using continuous glucose monitoring and a machine learning framework. medRxiv. (2024).
71.Goncharov, A. et al. Insertable glucose sensor using a compact and cost-effective phosphorescence lifetime imager and machine learning. ACS Nano.18, 23365–23379 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
72.Dénes-Fazakas L. et al. Physical activity detection for diabetes mellitus patients using recurrent neural networks. Sensors (Basel, Switzerland);24. 10.3390/s24082412 (2024). [DOI] [PMC free article] [PubMed]
73.Dave, D. et al. Detection of hypoglycemia and hyperglycemia using noninvasive wearable sensors: Electrocardiograms and accelerometry. J. Diab. Sci. Technol.18, 351–362 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
74.Chowdhury M. H., Chowdhury M. E. H., Alqahtani A. MMG-net: Multi modal approach to estimate blood glucose using multi-stream and cross modality attention. Biomed. Signal Processing Control. 2024;92. 10.1016/j.bspc.2024.105975.
75.Chikwetu L., Younes R. Carbohydrate content classification using postprandial heart rate responses from non-invasive wearables. Sensors (Basel, Switzerland);24. 10.3390/s24165331 (2024). [DOI] [PMC free article] [PubMed]
76.Bonet J., Visentin R., Dalla Man C. Smart algorithms for efficient insulin therapy initiation in individuals with type 2 diabetes: An in silico study. J. Diabetes Sci. Technol.10.1177/19322968241245930 (2024). [DOI] [PMC free article] [PubMed]
77.Beolet T., Adenis A., Huneker E., Louis M. End-to-end offline reinforcement learning for glycemia control. Artif. Intell. Med.;154. 10.1016/j.artmed.2024.102920 (2024). [DOI] [PubMed]
78.Allam, F. Using nonlinear auto-regressive with exogenous input neural network (NNARX) in blood glucose prediction. Bioelectron. Med.10, 11 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
79.Zanelli, S., El Yacoubi, M. A., Hallab, M. & Ammi, M. Type 2 diabetes detection with light CNN from single raw PPG wave. IEEE Access.11, 57652–57665 (2023). [Google Scholar]
80.Yang, T., Yang, Q., Zhou, Y. & Wen, C. Glucose trend prediction model based on improved wavelet transform and gated recurrent unit. Math. Biosci. Eng.20, 17037–17056 (2023). [DOI] [PubMed] [Google Scholar]
81.Tao, R. et al. A deep learning nomogram of continuous glucose monitoring data for the risk prediction of diabetic retinopathy in type 2 diabetes. Phys. Eng. Sci. Med.46, 813–825 (2023). [DOI] [PubMed] [Google Scholar]
82.Site, A., Nurmi, J. & Lohan, E. S. Machine-Learning-Based Diabetes Prediction Using Multisensor Data. IEEE Sens. J.23, 28370–28377 (2023). [Google Scholar]
83.Nazha H. M. et al. Portable infrared-based glucometer reinforced with fuzzy logic. Biosensors (Basel);13. 10.3390/bios13110991 (2023). [DOI] [PMC free article] [PubMed]
84.Lin, C. H. & Liu, C. L. Prediction of blood glucose concentration based on OptiScanner and XGBoost in ICU. IEEE Access.11, 116524–116533 (2023). [Google Scholar]
85.Lee, Y.-B. et al. An integrated digital health care platform for diabetes management with AI-based dietary management: 48-week results from a randomized controlled trial. Diab. Care.46, 959–966 (2023). [DOI] [PubMed] [Google Scholar]
86.Lee, S. M., Kim, D. Y. & Woo, J. Glucose transformer: Forecasting glucose level and events of hyperglycemia and hypoglycemia. IEEE J. Biomed. Health Inform.27, 1600–1611 (2023). [DOI] [PubMed] [Google Scholar]
87.Kistkins S. et al. Comparative analysis of predictive interstitial glucose level classification models. Sensors (Basel);23. 10.3390/s23198269 (2023). [DOI] [PMC free article] [PubMed]
88.Fellah Arbi, K., Soulimane, S., Saffih, F., Bechar, M. A. & Azzoug, O. Blood glucose estimation based on ECG signal. Phys. Eng. Sci. Med.46, 255–264 (2023). [DOI] [PubMed] [Google Scholar]
89.Andellini, M. et al. Artificial intelligence for non-invasive glycaemic-events detection via ECG in a paediatric population: Study protocol. Health Technol.13, 145–154 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
90.Alvarado J., Velasco J. M., Chavez F., Fernández-de-Vega F., Hidalgo J. I. Combining wavelet transform with convolutional neural networks for hypoglycemia events prediction from CGM data. Chemometrics Intell. Laboratory Syst. 243. 10.1016/j.chemolab.2023.105017 (2023).
91.Ahmed, A., Aziz, S., Qidwai, U., Abd-Alrazaq, A. & Sheikh, J. Performance of artificial intelligence models in estimating blood glucose level among diabetic patients using non-invasive wearable device data. Comput. Methods Prog. Biomed. Update3, 100094 (2023). [Google Scholar]
92.Zale, A. & Mathioudakis, N. Machine learning models for inpatient glucose prediction review. Curr. Diab. Rep.22, 353–364 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
93.Xiong, H. et al. Identifying daily activities of patient work for type 2 diabetes and co-morbidities: A deep learning and wearable camera approach. J. Am. Med. Inform. Assoc.29, 1400–1408 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
94.Nemat, H., Khadem, H., Eissa, M. R., Elliott, J. & Benaissa, M. Blood glucose level prediction: Advanced deep-ensemble learning approach. IEEE J. Biomed. Health Inform.26, 2758–2769 (2022). [DOI] [PubMed] [Google Scholar]
95.Malerbi, F. K. et al. Diabetic retinopathy screening using artificial intelligence and handheld smartphone-based retinal camera. J. Diab. Sci. Technol.16, 716–723 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
96.Lobo, B., Farhy, L., Shafiei, M. & Kovatchev, B. A data-driven approach to classifying daily continuous glucose monitoring (CGM) time series. IEEE Trans. Biomed. Eng.69, 654–665 (2022). [DOI] [PubMed] [Google Scholar]
97.Lim, M. H., Cho, Y. M. & Kim, S. Multi-task disentangled autoencoder for time-series data in glucose dynamics. IEEE J. Biomed. Health Inform.26, 4702–4713 (2022). [DOI] [PubMed] [Google Scholar]
98.Aloraynan A., Rassel S., Xu C., Ban D. A single wavelength mid-infrared photoacoustic spectroscopy for noninvasive glucose detection using machine learning. Biosensors;12. 10.3390/bios12030166(2022). [DOI] [PMC free article] [PubMed]
99.Garcia-Tirado, J., Lv, D., Corbett, J. P., Colmegna, P. & Breton, M. D. Advanced hybrid artificial pancreas system improves on unannounced meal response - In silico comparison to currently available system. Comput Methods Prog. Biomed.211, 106401 (2021). [DOI] [PubMed] [Google Scholar]
100.Bent, B. et al. Engineering digital biomarkers of interstitial glucose from noninvasive smartwatches. npj Digit. Med. 4, 89 (2021). [DOI] [PMC free article] [PubMed]
101.Baig, M. M., GholamHosseini, H., Gutierrez, J., Ullah, E. & Lindén, M. Early Detection of prediabetes and T2DM using wearable sensors and internet-of-things-based monitoring applications. Appl. Clin. Inform.12, 1–9 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
102.Xie, J. & Wang, Q. Benchmarking machine learning algorithms on blood glucose prediction for type i diabetes in comparison with classical time-series models. IEEE Trans. Biomed. Eng.67, 3101–3124 (2020). [DOI] [PubMed] [Google Scholar]
103.Li, K., Liu, C., Zhu, T., Herrero, P. & Georgiou, P. GluNet: A deep learning framework for accurate glucose forecasting. IEEE J. Biomed. Health Inform.24, 414–423 (2020). [DOI] [PubMed] [Google Scholar]
104.Samadi, S. et al. Automatic detection and estimation of unannounced meals for multivariable artificial pancreas system. Diab. Technol. Ther.20, 235–246 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
105.Zecchin C., Facchinetti A., Sparacino G., Cobelli C. Jump neural network for real-time prediction of glucose concentration. Methods Mol. Biol.1260, 245-259 (2015). [DOI] [PubMed]
106.Zecchin, C., Facchinetti, A., Sparacino, G. & Cobelli, C. Jump neural network for online short-time prediction of blood glucose from continuous monitoring sensors and meal information. Comput. Methods Prog. Biomed.113, 144–152 (2014). [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

PRISMA 2020 Checklist^{(184.7KB, pdf)}

Data Availability Statement

No datasets were generated or analysed during the current study.

No custom code or software was used in the data extraction or analysis for this systematic review.

[CR1] 1.Bahari, N. I. et al. Issues and challenges in the primary prevention of type 2 diabetes mellitus: A systematic review. J. Prev.44, 105–125 (2023). [DOI] [PubMed] [Google Scholar]

[CR2] 2.Roman, S. H. & Harris, M. I. Management of diabetes mellitus from a public health perspective. Endocrinol. Metab. Clin.26, 443–474 (1997). [DOI] [PubMed] [Google Scholar]

[CR3] 3.International Diabetes Federation. IDF Diabetes Atlas. Available from: www.diabetesatlas.org (2021).

[CR4] 4.Institute for Health Metrics and Evaluation. Global diabetes prevalence and forecasts. Available from: www.healthdata.org (2023).

[CR5] 5.Zhou, B. et al. Worldwide trends in diabetes prevalence and treatment from 1990 to 2022: a pooled analysis of 1108 population-representative studies with 141 million participants. Lancet404, 2077–2093 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR6] 6.Centers for Disease Control and Prevention. About type 2 diabetes. U.S. Department of Health and Human Services. Available at: https://www.cdc.gov/diabetes/about/about-type-2-diabetes.html (Accessed Dec 12, 2024).

[CR7] 7.Sarma, A. D. & Devi, M. Artificial intelligence in diabetes management: transformative potential, challenges, and opportunities in healthcare. Hormones24, 307–322 (2025). [DOI] [PubMed] [Google Scholar]

[CR8] 8.Ahmed, A., Aziz, S., Abd-Alrazaq, A., Farooq, F. & Sheikh, J. Overview of Artificial Intelligence-Driven Wearable Devices for Diabetes: Scoping Review Review. J. Med. Internet Res.24, e36010 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR9] 9.Makroum, M. A. A. M., Bouzouane, A. & Ibrahim, H. Machine learning and smart devices for diabetes management: Systematic review. Sens. (Basel, Switz.)22, 1843 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR10] 10.Iftikhar, M. et al. Artificial intelligence-driven transformations in diabetes care: a comprehensive literature review. Ann. Med. Surg.86, 5334–5342 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR11] 11.Anjum, M., Saher, R. & Saeed, M. N. Optimizing type 2 diabetes management: AI-enhanced time series analysis of continuous glucose monitoring data for personalized dietary intervention. PeerJ Comput Sci.10, e1971 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR12] 12.Wong, C. et al. Glycaemic outcomes in people living with diabetes under 65 and over 65 years old using an intermittently scanned continuous glucose monitoring system. Ther. Adv. Endocrinol. Metab.15, 20420188241269133 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR13] 13.Shen, Y., Zou, X. & Hu, G. Digital technology in the management and prevention of diabetes. Front Endocrinol.15, 1511236 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR14] 14.Guerci, B. Insulin-treated diabetes in general practice. La Rev. du Praticien.74, S5–S12 (2024). [PubMed] [Google Scholar]

[CR15] 15.Thomsen, C. H. et al. The intelligent diabetes telemonitoring using decision support to treat patients on insulin therapy (DiaTRUST) trial: study protocol for a randomized controlled trial. Trials25, 744 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR16] 16.Vettoretti, M., Cappon, G., Facchinetti, A. & Sparacino, G. Advanced diabetes management using artificial intelligence and continuous glucose monitoring sensors. Rev. Sens. (Basel, Switz.)20, 3870 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR17] 17.Rodriguez-León, C., Villalonga, C., Munoz-Torres, M., Ruiz, J. R. & Banos, O. Mobile and wearable technology for the monitoring of diabetes-related parameters: Systematic review. JMIR mHealth uHealth.9, e25138 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR18] 18.Alhalafi, A. et al. Utilizing artificial intelligence among patients with diabetes: A systematic review and meta-analysis. Cureus16, e58713 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar] [Retracted]

[CR19] 19.Jin, X., Cai, A., Xu, T. & Zhang, X. Artificial intelligence biosensors for continuous glucose monitoring. Interdiscip. Mater.2, 290–307 (2023). [Google Scholar]

[CR20] 20.Jung, H. H., Lee, H., Yea, J. & Jang, K.-I. Wearable electrochemical sensors for real-time monitoring in diabetes mellitus and associated complications. Soft Sci. 4, 15 (2024).

[CR21] 21.Ribeiro G., Monge J., Postolache O., Pereira J. M. D. A novel AI approach for assessing stress levels in patients with type 2 diabetes mellitus based on the acquisition of physiological parameters acquired during daily life. Sensors (Basel, Switzerland) 24. 10.3390/s24134175 (2024). [DOI] [PMC free article] [PubMed]

[CR22] 22.Bergoeing M. et al. Exploring the potential of an AI-integrated cloud-based mHealth platform for enhanced type 2 diabetes mellitus management. 100-111 (Springer, 2023).

[CR23] 23.Avram, R. et al. A digital biomarker of diabetes from smartphone-based vascular signals. Nat. Med.26, 1576–1582 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR24] 24.Yin, H., Mukadam, B., Dai, X. & Jha, N. K. DiabDeep: Pervasive Diabetes Diagnosis Based on Wearable Medical Sensors and Efficient Neural Networks. IEEE Trans. Emerg. Top. Comput.9, 1139–1150 (2021). [Google Scholar]

[CR25] 25.Deng, Y. et al. Deep transfer learning and data augmentation improve glucose levels prediction in type 2 diabetes patients. NPJ Digit. Med.4, 109 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR26] 26.He, J & Wang Y. Blood glucose concentration prediction based on kernel canonical correlation analysis with particle swarm optimization and error compensation. Comput. Methods Programs Biomed. 2020;196. 10.1016/j.cmpb.2020.105574. [DOI] [PubMed]

[CR27] 27.Zhu, T., Li, K., Kuang, L., Herrero, P. & Georgiou, P. An insulin bolus advisor for type 1 diabetes using deep reinforcement learning. Sensors20, 5058 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR28] 28.Sun, Q. et al. A dual mode adaptive basal-bolus advisor based on reinforcement learning. IEEE J. Biomed. Health Inform.23, 2633–2641 (2019). [DOI] [PubMed] [Google Scholar]

[CR29] 29.Kim, K. J. et al. Identification of healthy and unhealthy lifestyles by a wearable activity tracker in type 2 diabetes: A Machine Learning-based Analysis. Endocrinol. Metab.37, 547–551 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR30] 30.Ramazi, R., Perndorfer, C., Soriano, E. C., Laurenceau, J. P. & Beheshti, R. Predicting progression patterns of type 2 diabetes using multi-sensor measurements. Smart Health21, 100206 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR31] 31.Vabalas, A., Gowen, E., Poliakoff, E. & Casson, A. J. Machine learning algorithm validation with a limited sample size. PloS one14, e0224365 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR32] 32.Larracy R., Phinyomark A., Scheme E. Machine learning model validation for early stage studies with small sample sizes. 2314-2319 (IEEE, 2021). [DOI] [PubMed]

[CR33] 33.Gudivada, V., Apon, A. & Ding, J. Data quality considerations for big data and machine learning: Going beyond data cleaning and transformations. Int. J. Adv. Softw.10, 1–20 (2017). [Google Scholar]

[CR34] 34.Gong, Y., Liu, G., Xue, Y., Li, R. & Meng, L. A survey on dataset quality in machine learning. Inf. Softw. Technol.162, 107268 (2023). [Google Scholar]

[CR35] 35.Slieker, R. C., van der Heijden, A. A. W. A., Siddiqui, M. K., Langendoen-Gort, M., Nijpels, G., Herings, R., Feenstra, T. L., Moons, K. G. M., Bell, S., Elders, P. J., 't Hart, L. M., & Beulens, J. W. J. Performance of prediction models for nephropathy in people with type 2 diabetes: systematic review and external validation study. BMJ 374 (2021) [DOI] [PMC free article] [PubMed]

[CR36] 36.MoDastoni, D. A. Exploring methods to make AI decisions more transparent and understandable for humans. Adv. Eng. Innov.3, 32–36 (2023). [Google Scholar]

[CR37] 37.Beltramin D., Lamas E., Bousquet C. Ethical issues in the utilization of black boxes for artificial intelligence in medicine. Advances in Informatics, Management and Technology in Healthcare. 249-252, (IOS Press, 2022). [DOI] [PubMed]

[CR38] 38.Lundberg S.M. & Lee SI. A unified approach to interpreting model predictions. Advances in neural information processing systems, 30 (2017).

[CR39] 39.Pham, Q., Gamble, A., Hearn, J. & Cafazzo, J. A. The need for ethnoracial equity in artificial intelligence for diabetes management: Review and recommendations. J. Med Internet Res.23, e22320 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR40] 40.Bozkurt, S., Cahan, E. M. & Seneviratne, M. G. Reporting of demographic data and representativeness in machine learning models using electronic health records. J. Am. Med Inf. Assoc.27, 1878–1885 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR41] 41.Cirillo, D., Catuara-Solarz, S., Morey, C. & Guney, E. Sex and gender differences and biases in artificial intelligence for biomedicine and healthcare. NPJ Digit. Med.3, 81 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR42] 42.Murphy, A., Bowen, K., El Naqa, I. & Yoga, B. Bridging health disparities in the data-driven world of artificial intelligence: a narrative review. J. Racial Ethn. Health Disparities.12, 2367–2379 (2025). [DOI] [PubMed] [Google Scholar]

[CR43] 43.Cau, R., Pisu, F., Suri, J. S. & Saba, L. Addressing hidden risks: Systematic review of artificial intelligence biases across racial and ethnic groups in cardiovascular diseases. Eur. J. Radiol.183, 111867 (2025). [DOI] [PubMed] [Google Scholar]

[CR44] 44.Perets O., Stagno E., Ben Yehuda E. Inherent bias in electronic health records: a scoping review of sources of bias. ACM Trans Comput Healthcare; 10.1145/3757924 (2025).

[CR45] 45.Parate, A. P., et al. Review of data bias in healthcare applications. Int. J. Online Biomed. Eng. 20, 124–136 (2024).

[CR46] 46.Babu, A. & Joseph, A. P. Artificial intelligence in mental healthcare: transformative potential vs. the necessity of human interaction. Front. Psychol.15, 1378904 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR47] 47.Huang K., Wang Y & Xu M. Investigating the fairness of deep learning models in breast cancer diagnosis based on race and ethnicity. Proceedings of the AAAI Symposium Series, 4, 303–307 (2024)

[CR48] 48.Gu, T. et al. Mitigating bias in AI mortality predictions for minority populations: a transfer learning approach. BMC Med. Inform. Decis. Mak.25, 30 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR49] 49.Moukheiber M., Moukheiber L., Moukheiber D & Lee H.-C. Unmasking societal biases in respiratory support for ICU patients through social determinants of health. Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence AI for Good. 7421–7429 (2024).

[CR50] 50.Topol, E. J. High-performance medicine: the convergence of human and artificial intelligence. Nat. Med.25, 44–56 (2019). [DOI] [PubMed] [Google Scholar]

[CR51] 51.Nguyen, T. et al. A novel decentralized federated learning approach to train on globally distributed, poor quality, and protected private medical data. Sci. Rep.12, 8888 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR52] 52.Hasani R., Lechner M., Amini A., Rus D & Grosu R. Liquid time-constant networks. Proceedings of the AAAI Conference on Artificial Intelligence.35, 7657–7666 (2021).

[CR53] 53.Page M. J. et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ 372, 10.1136/bmj.n71 (2021). [DOI] [PMC free article] [PubMed]

[CR54] 54.Schellenberg, E. S., Dryden, D. M., Vandermeer, B., Ha, C. & Korownyk, C. Lifestyle interventions for patients with and at risk for type 2 diabetes: a systematic review and meta-analysis. Ann. Intern. Med.159, 543–551 (2013). [DOI] [PubMed] [Google Scholar]

[CR55] 55.Ahmed, A. et al. The effectiveness of wearable devices using artificial intelligence for blood glucose level forecasting or prediction: Systematic review. J. Med. Internet Res.25, e40259 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR56] 56.Piet, A. et al. Non-invasive wearable devices for monitoring vital signs in patients with type 2 diabetes mellitus: a systematic review. Bioengineering10, 1321 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR57] 57.Zhu, T. et al. Population-Specific Glucose Prediction in Diabetes Care With Transformer-Based Deep Learning on the Edge. IEEE Trans. Biomed. Circuits Syst.18, 236–246 (2024). [DOI] [PubMed] [Google Scholar]

[CR58] 58.Zhu, T., Afentakis, I., Li, K., Armiger, R., Hill, N., Oliver, N., & Georgiou, P. Multi-Horizon Glucose Prediction Across Populations With Deep Domain Generalization. J. Biomed. Sci. Health Informatics, 29, 5424–5437 (IEEE, 2025). [DOI] [PubMed]

[CR59] 59.Warren, M., Bergenstal, R. M., Hager, M., Bashan, E. & Hodish, I. A scalable application of artificial intelligence-driven insulin titration program to transform type 2 diabetes management. Diab. Technol. Ther.26, 556–565 (2024). [DOI] [PubMed] [Google Scholar]

[CR60] 60.Wang, X. et al. A novel hypoglycemia alarm framework for type 2 diabetes with high glycemic variability. Int J. Numer Method Biomed. Eng.40, e3799 (2024). [DOI] [PubMed] [Google Scholar]

[CR61] 61.Tucker, A. P., Erdman, A. G., Schreiner, P. J., Ma, S. & Chow, L. S. Neural networks with gated recurrent units reduce glucose forecasting error due to changes in sensor location. J. Diab. Sci. Technol.18, 124–134 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR62] 62.Tao, R. et al. DDLA: a double deep latent autoencoder for diabetic retinopathy diagnose based on continuous glucose sensors. Med. Biol. Eng. Comput.62, 3089–3106 (2024). [DOI] [PubMed] [Google Scholar]

[CR63] 63.Shuzan M. N. I. et al. QU-GM: An IoT based glucose monitoring system from photoplethysmography, blood pressure and demographic data using machine learning. (IEEE, 2024).

[CR64] 64.Seo W., Kim N., Park S. W., Jin S. M., Park S. M. Generative adversarial network-based data augmentation for improving hypoglycemia prediction: A proof-of-concept study. Biomed. Signal Processing Control 92. 10.1016/j.bspc.2024.106077 (2024).

[CR65] 65.Yin, H. & Jha, N. K. A health decision support system for disease diagnosis based on wearable medical sensors and machine learning ensembles. IEEE Trans. Multi-Scale Comput. Syst.3, 228–241 (2017). [Google Scholar]

[CR66] 66.Kim S. H., Kim D. Y., Chun S. W., Kim J., Woo J. Impartial feature selection using multi-agent reinforcement learning for adverse glycemic event prediction. Comput. Biol. Med. 173. 10.1016/j.compbiomed.2024.108257 (2024). [DOI] [PubMed]

[CR67] 67.Kashif, M., Flesca, S. & Veltri, P. GLSTM: On using LSTM for glucose level prediction. Stud. health Technol. Inform.314, 103–107 (2024). [DOI] [PubMed] [Google Scholar]

[CR68] 68.Kannenberg S. et al. Unlocking potential: Personalized lifestyle therapy for type 2 diabetes through a predictive algorithm-driven digital therapeutic. J. Diabetes Sci. Technol. 10.1177/19322968241266821 (2024). [DOI] [PMC free article] [PubMed]

[CR69] 69.Hotta S., Kytö M., Koivusalo S., Heinonen S., Marttinen P. Optimizing postprandial glucose prediction through integration of diet and exercise: Leveraging transfer learning with imbalanced patient data. PLoS ONE;19. 10.1371/journal.pone.0298506 (2024). [DOI] [PMC free article] [PubMed]

[CR70] 70.Metwally A. A. et al. Predicting type 2 diabetes metabolic phenotypes using continuous glucose monitoring and a machine learning framework. medRxiv. (2024).

[CR71] 71.Goncharov, A. et al. Insertable glucose sensor using a compact and cost-effective phosphorescence lifetime imager and machine learning. ACS Nano.18, 23365–23379 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR72] 72.Dénes-Fazakas L. et al. Physical activity detection for diabetes mellitus patients using recurrent neural networks. Sensors (Basel, Switzerland);24. 10.3390/s24082412 (2024). [DOI] [PMC free article] [PubMed]

[CR73] 73.Dave, D. et al. Detection of hypoglycemia and hyperglycemia using noninvasive wearable sensors: Electrocardiograms and accelerometry. J. Diab. Sci. Technol.18, 351–362 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR74] 74.Chowdhury M. H., Chowdhury M. E. H., Alqahtani A. MMG-net: Multi modal approach to estimate blood glucose using multi-stream and cross modality attention. Biomed. Signal Processing Control. 2024;92. 10.1016/j.bspc.2024.105975.

[CR75] 75.Chikwetu L., Younes R. Carbohydrate content classification using postprandial heart rate responses from non-invasive wearables. Sensors (Basel, Switzerland);24. 10.3390/s24165331 (2024). [DOI] [PMC free article] [PubMed]

[CR76] 76.Bonet J., Visentin R., Dalla Man C. Smart algorithms for efficient insulin therapy initiation in individuals with type 2 diabetes: An in silico study. J. Diabetes Sci. Technol.10.1177/19322968241245930 (2024). [DOI] [PMC free article] [PubMed]

[CR77] 77.Beolet T., Adenis A., Huneker E., Louis M. End-to-end offline reinforcement learning for glycemia control. Artif. Intell. Med.;154. 10.1016/j.artmed.2024.102920 (2024). [DOI] [PubMed]

[CR78] 78.Allam, F. Using nonlinear auto-regressive with exogenous input neural network (NNARX) in blood glucose prediction. Bioelectron. Med.10, 11 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR79] 79.Zanelli, S., El Yacoubi, M. A., Hallab, M. & Ammi, M. Type 2 diabetes detection with light CNN from single raw PPG wave. IEEE Access.11, 57652–57665 (2023). [Google Scholar]

[CR80] 80.Yang, T., Yang, Q., Zhou, Y. & Wen, C. Glucose trend prediction model based on improved wavelet transform and gated recurrent unit. Math. Biosci. Eng.20, 17037–17056 (2023). [DOI] [PubMed] [Google Scholar]

[CR81] 81.Tao, R. et al. A deep learning nomogram of continuous glucose monitoring data for the risk prediction of diabetic retinopathy in type 2 diabetes. Phys. Eng. Sci. Med.46, 813–825 (2023). [DOI] [PubMed] [Google Scholar]

[CR82] 82.Site, A., Nurmi, J. & Lohan, E. S. Machine-Learning-Based Diabetes Prediction Using Multisensor Data. IEEE Sens. J.23, 28370–28377 (2023). [Google Scholar]

[CR83] 83.Nazha H. M. et al. Portable infrared-based glucometer reinforced with fuzzy logic. Biosensors (Basel);13. 10.3390/bios13110991 (2023). [DOI] [PMC free article] [PubMed]

[CR84] 84.Lin, C. H. & Liu, C. L. Prediction of blood glucose concentration based on OptiScanner and XGBoost in ICU. IEEE Access.11, 116524–116533 (2023). [Google Scholar]

[CR85] 85.Lee, Y.-B. et al. An integrated digital health care platform for diabetes management with AI-based dietary management: 48-week results from a randomized controlled trial. Diab. Care.46, 959–966 (2023). [DOI] [PubMed] [Google Scholar]

[CR86] 86.Lee, S. M., Kim, D. Y. & Woo, J. Glucose transformer: Forecasting glucose level and events of hyperglycemia and hypoglycemia. IEEE J. Biomed. Health Inform.27, 1600–1611 (2023). [DOI] [PubMed] [Google Scholar]

[CR87] 87.Kistkins S. et al. Comparative analysis of predictive interstitial glucose level classification models. Sensors (Basel);23. 10.3390/s23198269 (2023). [DOI] [PMC free article] [PubMed]

[CR88] 88.Fellah Arbi, K., Soulimane, S., Saffih, F., Bechar, M. A. & Azzoug, O. Blood glucose estimation based on ECG signal. Phys. Eng. Sci. Med.46, 255–264 (2023). [DOI] [PubMed] [Google Scholar]

[CR89] 89.Andellini, M. et al. Artificial intelligence for non-invasive glycaemic-events detection via ECG in a paediatric population: Study protocol. Health Technol.13, 145–154 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR90] 90.Alvarado J., Velasco J. M., Chavez F., Fernández-de-Vega F., Hidalgo J. I. Combining wavelet transform with convolutional neural networks for hypoglycemia events prediction from CGM data. Chemometrics Intell. Laboratory Syst. 243. 10.1016/j.chemolab.2023.105017 (2023).

[CR91] 91.Ahmed, A., Aziz, S., Qidwai, U., Abd-Alrazaq, A. & Sheikh, J. Performance of artificial intelligence models in estimating blood glucose level among diabetic patients using non-invasive wearable device data. Comput. Methods Prog. Biomed. Update3, 100094 (2023). [Google Scholar]

[CR92] 92.Zale, A. & Mathioudakis, N. Machine learning models for inpatient glucose prediction review. Curr. Diab. Rep.22, 353–364 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR93] 93.Xiong, H. et al. Identifying daily activities of patient work for type 2 diabetes and co-morbidities: A deep learning and wearable camera approach. J. Am. Med. Inform. Assoc.29, 1400–1408 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR94] 94.Nemat, H., Khadem, H., Eissa, M. R., Elliott, J. & Benaissa, M. Blood glucose level prediction: Advanced deep-ensemble learning approach. IEEE J. Biomed. Health Inform.26, 2758–2769 (2022). [DOI] [PubMed] [Google Scholar]

[CR95] 95.Malerbi, F. K. et al. Diabetic retinopathy screening using artificial intelligence and handheld smartphone-based retinal camera. J. Diab. Sci. Technol.16, 716–723 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR96] 96.Lobo, B., Farhy, L., Shafiei, M. & Kovatchev, B. A data-driven approach to classifying daily continuous glucose monitoring (CGM) time series. IEEE Trans. Biomed. Eng.69, 654–665 (2022). [DOI] [PubMed] [Google Scholar]

[CR97] 97.Lim, M. H., Cho, Y. M. & Kim, S. Multi-task disentangled autoencoder for time-series data in glucose dynamics. IEEE J. Biomed. Health Inform.26, 4702–4713 (2022). [DOI] [PubMed] [Google Scholar]

[CR98] 98.Aloraynan A., Rassel S., Xu C., Ban D. A single wavelength mid-infrared photoacoustic spectroscopy for noninvasive glucose detection using machine learning. Biosensors;12. 10.3390/bios12030166(2022). [DOI] [PMC free article] [PubMed]

[CR99] 99.Garcia-Tirado, J., Lv, D., Corbett, J. P., Colmegna, P. & Breton, M. D. Advanced hybrid artificial pancreas system improves on unannounced meal response - In silico comparison to currently available system. Comput Methods Prog. Biomed.211, 106401 (2021). [DOI] [PubMed] [Google Scholar]

[CR100] 100.Bent, B. et al. Engineering digital biomarkers of interstitial glucose from noninvasive smartwatches. npj Digit. Med. 4, 89 (2021). [DOI] [PMC free article] [PubMed]

[CR101] 101.Baig, M. M., GholamHosseini, H., Gutierrez, J., Ullah, E. & Lindén, M. Early Detection of prediabetes and T2DM using wearable sensors and internet-of-things-based monitoring applications. Appl. Clin. Inform.12, 1–9 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR102] 102.Xie, J. & Wang, Q. Benchmarking machine learning algorithms on blood glucose prediction for type i diabetes in comparison with classical time-series models. IEEE Trans. Biomed. Eng.67, 3101–3124 (2020). [DOI] [PubMed] [Google Scholar]

[CR103] 103.Li, K., Liu, C., Zhu, T., Herrero, P. & Georgiou, P. GluNet: A deep learning framework for accurate glucose forecasting. IEEE J. Biomed. Health Inform.24, 414–423 (2020). [DOI] [PubMed] [Google Scholar]

[CR104] 104.Samadi, S. et al. Automatic detection and estimation of unannounced meals for multivariable artificial pancreas system. Diab. Technol. Ther.20, 235–246 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR105] 105.Zecchin C., Facchinetti A., Sparacino G., Cobelli C. Jump neural network for real-time prediction of glucose concentration. Methods Mol. Biol.1260, 245-259 (2015). [DOI] [PubMed]

[CR106] 106.Zecchin, C., Facchinetti, A., Sparacino, G. & Cobelli, C. Jump neural network for online short-time prediction of blood glucose from continuous monitoring sensors and meal information. Comput. Methods Prog. Biomed.113, 144–152 (2014). [DOI] [PubMed] [Google Scholar]

PERMALINK

Integration of artificial intelligence and wearable technology in the management of diabetes and prediabetes

Raphael A Fraser

Rebekah J Walker

Jennifer A Campbell

Obinna Ekwunife

Leonard E Egede

Abstract

Introduction

Results

Study selection

Fig. 1.

Study characteristics

Table 1.

Table 2.

Model performance and predictive accuracy

Results of individual studies

Discussion

Methods

Information sources and search strategy

Manuscript eligibility criteria

Study selection

Fig. 2.

Data extraction and data synthesis

Supplementary information

Acknowledgements

Author contributions

Data availability

Code availability

Competing interests

Footnotes

Supplementary information

References

Associated Data

Supplementary Materials

Data Availability Statement

ACTIONS

PERMALINK

RESOURCES

Similar articles

Cited by other articles

Links to NCBI Databases