Electronic health record-derived machine learning model for hypoglycemia risk prediction in type 2 diabetes mellitus patients: development and validation

Abstract

Background

Hypoglycemia is a serious complication of diabetes. Early recognition of hypoglycemia can improve clinical prognosis, however, traditional diagnostic tools are often limited. Machine learning offers a promising approach for predicting adverse outcomes in diabetic patients.

Objective

This study aims to develop and validate machine learning-based models to predict the risk of hypoglycemia in type 2 diabetes mellitus (T2DM) patients.

Methods

A cohort study design was employed. Clinical data were collected from the electronic health record system. The dataset was randomly partitioned into training and validation subsets using a 7:3 ratio. Four machine learning algorithms, logistic regression (LR), Extreme Gradient Boosting (XGBoost), random forest (RF), and support vector machine (SVM) were implemented to develop hypoglycemia risk prediction models. Predictive performance was assessed using sensitivity, specificity, accuracy, precision, F1 score, and the area under the receiver operating characteristic curve (AUC).

Results

831 T2DM patients were included, the hypoglycemia incidence was 22.0%. In the training cohort, the AUC for the LR, XGBoost, SVM, and RF models were 0.82, 0.86, 0.84, and 0.80, and corresponding AUCs were 0.76, 0.78, 0.72, and 0.75 in the validation cohort. The XGBoost demonstrated the highest overall predictive performance. Feature importance analysis based on the XGBoost model identified creatinine, triglycerides, albumin, HbA1c, C-peptide, aspartate aminotransferase, hemoglobin, and sulfonylurea use as the most influential predictors of hypoglycemia risk.

Conclusions

The XGBoost model exhibited superior predictive performance for achieving the higher AUC, F1 score, greater accuracy, sensitivity and specificity. This model enables effective identification of T2DM patients who may require intensified monitoring or targeted interventions to prevent hypoglycemic events.

Clinical trial number

Not applicable.

Introduction

Diabetes has become a major threat to global health and represents a critical public health challenge requiring coordinated international action. Current estimates suggest that more than 589 million people worldwide are living with diabetes, with the vast majority of cases attributable to type 2 diabetes mellitus (T2DM) [1]. Clinical evidence indicates that treatment with sulfonylureas or insulin can achieve effective glycemic control when compared with conventional antidiabetic therapies [2–3]. However, these treatment strategies are also associated with an increased risk of acute diabetic complications, most notably hypoglycemia. Studies have reported highly variable hypoglycemia incidence rates among individuals with T2DM, ranging from 0.072 to 16,360 episodes per 1,000 person-years [4]. And approximately 47.5% of hospitalized T2DM patients experienced at least one hypoglycemic event [2]. Hypoglycemia causes more severe and immediate harm than hyperglycemia. Evidence suggests that hypoglycemic episodes can trigger coronary heart disease in elderly patients with T2DM and are associated with increased morbidity and mortality [5–6]. Furthermore, patients with a history of hypoglycemia have a 54% higher risk of developing dementia compared with those without hypoglycemia, with the risk increasing by approximately 30% for each additional episode [7].

Although many patients experience typical symptoms during hypoglycemic episodes like hunger, diaphoresis, palpitations, and dizziness, recurrent hypoglycemia can lead to defective glucose counterregulation and the development of impaired awareness of hypoglycemia [8]. This impaired awareness diminishes patients’ ability to recognize early warning signs of declining blood glucose levels. Using continuous glucose monitoring (CGM) to track real-time glucose fluctuations, Gehlaut et al. reported that approximately 75% of hypoglycemic episodes detected by CGM in patients with T2DM were not perceived by the patients themselves [9]. Collectively, these findings indicate that the true prevalence of hypoglycemia among individuals with T2DM substantially exceeds both clinician estimates and patient self-recognition.

Despite numerous risk factors for hypoglycemia in patients with T2DM have been identified [10, 11], many of these factors are dynamic and context dependent, making accurate risk identification persistently challenging. In recent years, machine learning (ML) techniques have been increasingly applied in diabetes research, offering new avenues for individualized hypoglycemia prediction. For instance, Jin et al. combined deep learning approaches with natural language processing (NLP) to develop a clinically applicable model for the automated detection of hypoglycemic events [12]. Similarly, Shi et al. proposed a novel model to predict the one-year risk of severe hypoglycemia in older adults with diabetes [13]. Nevertheless, the predictive performance of existing models remains inconsistent, and their accuracy and generalizability across diverse clinical settings require further validation. In the present study, we aimed to develop and validate machine learning-based models for predicting hypoglycemia in hospitalized patients with T2DM using routinely available electronic health record (EHR) data. We further anticipated that the proposed model could be integrated into a clinical decision support system to enable precise and timely risk stratification in this patient population.

Materials and methods

Study design

This cohort study was conducted and reported in accordance with the TRIPOD + AI (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis, Extended for Artificial Intelligence) guidelines [14]. Our study was approved by the Ethics Committee of the Second Hospital of Chongqing Medical University (No: 2025.778), in accordance with the Declaration of Helsinki. All procedures complied with ethical standards, and strict measures were implemented to ensure patient anonymity and data confidentiality.

Study population

The target population was derived from a tertiary hospital from Jan 2024 to Jan 2026 in Chongqing, China. Eligible participants met the following inclusion criteria: (1) diagnosis of T2DM and with a disease duration of at least one year [15]; (2) hospitalization lasting ≥ 24 h; (3) age over 18 and under 65 years old. Patients were excluded if they: (1) had secondary diabetes or gestational diabetes; (2) had a documented history of cognitive impairment or psychiatric disorders; (3) were admitted primarily due to hypoglycemia. Sample size estimation was performed using G*Power software (version 3.1.9.7). For a two-tailed test comparing two independent proportions, the following parameters were assumed: type I error rate (α) = 0.05, power (1-β) = 0.80, allocation ratio = 1:1. Based on previous literature, the anticipated incidence of hypoglycemia in the non-exposed group was 18%. To detect an odds ratio of 1.6, the minimum required sample size was calculated to be 650.

Data collected

All patients’ data were extracted from the hospital’s EHR system. Demographic variables included sex, age, history of alcohol consumption, and body mass index (BMI), which was calculated as weight in kilograms divided by height in meters squared (kg/m²). Disease-related clinical information comprised diabetes duration, number of antidiabetic medications, use of sulfonylureas and insulin, presence of hypertension or kidney disease, prior history of hypoglycemia, and nutritional status. Laboratory parameters collected for analysis included hemoglobin, creatinine, aspartate aminotransferase (AST), albumin, total cholesterol (TC), triglycerides (TG), glycated hemoglobin (HbA1c), estimated glomerular filtration rate (GFR), and C-peptide. These variables were selected based on evidence from published literature [13, 16–17], expert consensus, and their availability within the EHR system. Nutritional status was assessed using the Nutritional Risk Screening 2002 (NRS 2002) [18], which evaluates disease severity, nutritional impairment, and age. Total scores range from 0 to 7, with scores of 0–2 indicating normal nutritional status and scores ≥ 3 indicating the presence of nutritional risk. The primary outcome of this study was the hypoglycemic events, defined as any documented blood glucose measurement < 3.9 mmol/L (70 mg/dL) during hospitalization. Hypoglycemia events were identified through the hospital’s hypoglycemia adverse event reporting system.

Data preparation

The dataset was randomly partitioned into training (70%) and internal validation (30%) sets, in order to eliminate the risk of leakage associated with overlapping patient data across sets, we partitioned the dataset strictly by patient ID, ensuring that all samples from the same patient belong exclusively to either the training or the test set. And the training set was pre‑processed using the Synthetic Minority Oversampling Technique (SMOTE) to address class imbalance and equalize category distributions. Outliers defined as values beyond 1.5 times the interquartile range (IQR) from the first or third quartile were identified via the IQR method and handled using statistical approaches or domain knowledge [19]. Variables with > 15% missing values were excluded from the final cohort, and remaining missing values were imputed using multiple imputation [20]. To ensure data integrity and model robustness, numerical attributes were standardized to eliminate scale disparities, and categorical features were dummy-encoded into a numerical format. For continuous measurements, continuity was retained as recommended by the PROBAST statement to avoid discretization bias [21]. These preprocessing steps prevent unequal weighting of attributes during data analysis, thereby accelerating computation and enhancing model performance.

Feature selection

Feature selection is an essential step in transforming raw data into model training data. Its primary task is to filter out redundant features from the constructed features and identify the most relevant features, thereby reducing the dimensionality of the original feature space and preventing the model from overfitting. Two-step approach was employed in our study, namely univariate analysis and random forest classifier (RFC) tree-based model Screening. First, potential predictors of hypoglycemia were preliminarily screened based on the results of univariate analysis. Subsequently, feature importance was assessed using a RFC tree‑based model to remove less important features within each collinear group [22–23]. This approach helps ensure that the strongest predictive features are retained while maintaining model performance.

Model development and performance evaluation

Hypoglycemia risk prediction models were constructed using logistic regression (LR), random forest (RF), Extreme Gradient Boosting (XGBoost), and support vector machine (SVM) algorithms. RF is a representative algorithm in ensemble learning, fundamentally constructing multiple distinct decision trees through bootstrap sampling and integrating them into the final model. Its advantages lie in random feature selection and sampling. The XGBoost algorithm is an ensemble learning method based on gradient boosting trees, serving as an optimized model that fuses linear models with boosting trees. It effectively handles complex variables with interactions and collinearity issues, delivering high accuracy and stability. SVM is a classic binary classification algorithm. This model is characterized by its broad applicability, delivering excellent results for small-sample datasets, non-linear problems, and high-dimensional classification tasks. Five-fold cross-validation was applied during model training to enhance robustness and reduce overfitting. Following model development, predictive performance was assessed using sensitivity, specificity, accuracy, precision, F1 score, and the area under the receiver operating characteristic curve (AUC). The optimal model was selected based on overall predictive performance across these metrics.

Statistical analysis

In the descriptive analysis of interest variables, the normality test was conducted using Kolmogorov-Smirnov test and visual inspection of Q-Q plots and histograms. Continuous variable of normal distribution were summarised by means with standard deviations, while discrete variables were presented by frequency and percentage. Comparisons between groups were performed using two-tailed t tests for continuous variables and chi-square tests for categorical variables. All statistical analyses were conducted using R software (version 4.1.2; R Core Team), and statistical significance was defined as a two-sided P value < 0.05. In addition, Python (version 3.7) was used for machine learning model development. Variables demonstrating statistically significant differences (P < 0.05) in univariate analyses were selected as input features, while the occurrence of hypoglycemia served as the outcome label.

Results

Baseline characteristics of study participants

A total of 831 patients were included, comprising 349 males and 482 females, with ages ranging from 26 to 64 (55.3 ± 9.0) years. Among these patients, 183 (22.0%) experienced hypoglycemia during hospitalization. The baseline characteristics of study population are summarized in Table 1. Comparative analysis between patients with and without hypoglycemia revealed significant differences in BMI, history of alcohol consumption, prior hypoglycemia, use of sulfonylureas or insulin, presence of kidney disease, and levels of hemoglobin, creatinine, albumin, AST, TG, C-peptide, and HbA1c.

Table 1.

Demographics of patients with T2DM (N = 831)

Items	Hypoglycemia Group	Non-hypoglycemic group	P
	Mean ± SD or n (%)	Mean ± SD or n (%)
Counts	183	648
Age (years)	55.9 ± 8.6	55.1 ± 9.1	0.323
BMI, (kg/m2)	21.8 ± 3.0	24.5 ± 3.5	<0.001
Gender			0.085
Male	87(47.5)	262(40.4)
Female	96(52.5)	386(59.6)
Insulin usage			0.008
No	91(49.7)	393(60.6)
Yes	92(50.3)	255(39.4)
Alcohol			0.050
No	104(56.8)	315(48.6)
Yes	79(43.2)	333(51.4)
NRS 2002			0.865
<3 score	131(71.6)	468(72.2)
≥ 3 score	52(28.4)	180(27.8)
Hypoglycemia history			<0.001
No	128(69.9)	534(82.4)
Yes	55(30.1)	114(17.6)
Hypertension			0.089
No	83(45.4)	340(52.5)
Yes	100(54.6)	308(47.5)
DM Course(years)	8.9 ± 5.2	8.1 ± 6.8	0.638
Antidiabetic drugs			0.089
<3 types	83(45.4)	340(52.5)
≥ 3 types	100(54.6)	308(47.5)
Sulfonylurea usage			<0.001
No	66(36.1)	450(69.4)
Yes	117(63.9)	198(30.6)
Kidney Disease			<0.001
No	20(10.9)	166(25.6)
Yes	163(89.1)	482(74.4)
HbA1c (%)	8.6 ± 2.1	9.4 ± 2.2	<0.001
C-peptide (ug/L)	1.5 ± 0.9	2.2 ± 1.0	<0.001
AST (U/L)	16.9 ± 4.4	20.4 ± 6.9	<0.001
TG (mmol/L)	1.8 ± 1.1	2.1 ± 1.5	0.027
TC(mmol/L)	4.4 ± 1.7	4.6 ± 1.3	0.547
Hemoglobin(g/L)	124.1 ± 15.9	130.9 ± 17.2	<0.001
Creatinine(µmol/L)	73.8 ± 27.1	69.5 ± 25.5	0.050
Albumin(g/L)	38.8 ± 5.5	40.4 ± 4.9	<0.001
GFR (mL/min)	85.9 ± 24.7	89.3 ± 23.7	0.086

SD standard deviation, BMI body mass index, NRS2002 Nutritional Risk Screening 2002, DM diabetes mellitus, HbA1c glycated hemoglobin, AST aspartate aminotransferase, TG triglyceride, TC total cholesterol, GFR glomerular filtration rate

Machine learning model development and validation

We performed a series of preliminary experiments using various machine learning algorithms to construct hypoglycemia prediction models. Using hypoglycemia as the dependent variable and the 13 risk factors identified as statistically significant in univariate analyses as independent variables, predictive models were developed for hospitalized patients with T2DM employing LR, XGBoost, RF, and SVM algorithms. Validate each constructed model in the training cohort, the results showed the AUC values for the LR, XGBoost, SVM, and RF models were 0.82(95%CI:0.71–0.94), 0.86(95%CI:0.78–0.95), 0.84(95%CI: 0.73–0.95), and 0.80(95%CI: 0.70–0.90), all greater than 0.80. To further evaluate the robustness of the predictive models, their performance was assessed on the validation dataset. We found XGBoost exhibited the highest predictive performance, achieving an AUC of 0.78, accuracy of 0.83, sensitivity of 0.70, specificity of 0.84, and F1 score of 0.74. The XGBoost model outperformed the other three models in all metrics except precision, where its 0.72 score was slightly lower than the RF model’s 0.74. Overall, it demonstrated the best performance. The predictive metrics for all four machine learning algorithms in the validation set are summarized in Table 2; Fig. 1. The XGBoost model importance ranking results indicate that the feature importance of creatinine is significantly higher than that of other features. The relative importance rankings of the remaining features are shown in Fig. 2.

Table 2.

Predictive performance of each model for developing hypoglycemia in hospitalized T2DM patients in the validation set

Model	AUC	Precision	Accuracy	F1-score	Sensitivity	Specificity
RF	0.75	0.74	0.82	0.71	0.68	0.81
SVM	0.72	0.71	0.80	0.70	0.69	0.80
LR	0.76	0.72	0.80	0.73	0.70	0.83
XGBoost	0.78	0.72	0.83	0.74	0.70	0.84

LR logistic regression, XGBoost Extreme Gradient Boosting, RF random forest, SVM support vector machine

Fig. 1.

Comparison of predictive performance of four machine learning models

Fig. 2.

Ranking of importance of risk factors in XGBoost model

Discussion

In this study, the incidence of hypoglycemia among T2DM patients was 22.0%, which aligns closely with recent reports of 21.7% [24], but is lower than the 47.5% observed in González et al. [2]. One potential explanation for this discrepancy is the difference in patients’ age, the mean age in our cohort was 55 years, compared with 70 years in González’s study. Previous studies have reported hypoglycemia rates in T2DM patients ranging from 21.7% to 41.2% [24–26], with variation likely attributable to differences in study settings, population demographics, and geographic factors. These findings underscore the importance of early identification and prediction of hypoglycemia risk, which allows clinicians to proactively adjust or personalize treatment strategies to prevent adverse events. Integrating machine learning-based decision support systems with routinely collected patient data may enable more accurate and timely detection of patients at elevated risk of hypoglycemia.

Studies have shown that patients treated with insulin or sulfonylureas are at an increased risk of hypoglycemia, consistent with our findings [26–27]. Sulfonylureas cannot fully replicate the physiological dynamics of endogenous insulin, thus it is recommended to use the closed-loop systems that integrate CGM with continuous subcutaneous insulin infusion (CSII). These systems provide real-time glucose readings and trend information, automatically adjusting or pausing insulin delivery when glucose levels approach hypoglycemic thresholds, not only maintain optimal glycemic control but also substantially reduce the incidence of insulin-related hypoglycemia. We also found that T2DM patients with lower C-peptide are more susceptible to hypoglycemia, likely due to increased glucose sensitivity and variability, which predisposes them to larger fluctuations in blood glucose. Previous researches have shown that individuals with C-peptide levels < 200 pmol/L experience a higher rate of hypoglycemia compared with those with higher C-peptide levels, despite similar mean glucose concentrations, the proportion of hypoglycemia was 94% versus 41% in the low versus high C-peptide groups (P < 0.001) [28–29], respectively.

Notably, our study also observed that HbA1c levels were associated with hypoglycemia, whereas studies reported no significant differences in those two [11, 30]. This discrepancy may stem from differences in outcome definitions, prior research often focused exclusively on severe hypoglycemia, whereas we included hypoglycemic events of any severity [11]. These results indicate that glycemic levels alone may not reliably predict hypoglycemia risk, highlighting the value of large datasets to develop predictive models, which can reveal that clinically relevant variables may not always contribute substantially to model performance. The study also confirms serum creatinine as an independent risk factor for hypoglycemia in T2DM patients, consistent with Cai et al. [3]. Creatinine is primarily excreted through glomerular filtration, elevated serum creatinine typically reflects reduced renal excretory and metabolic function, which may predispose to hypoglycemia through mechanisms including slowed drug clearance, accumulation of hypoglycemic agents, and impaired renal gluconeogenesis. Moreover, severe hypoglycemia can trigger stress hormone release, potentially reducing renal blood flow and further compromising fragile renal function.

In the training set, the AUC values were 0.82, 0.86, 0.84, and 0.80 for LR, XGBoost, SVM, and RF, respectively. In the validation set, the corresponding AUCs were 0.76, 0.78, 0.72, and 0.75, and the corresponding F1 scores were 0.73, 0.74, 0.70, and 0.71, the corresponding specificity were 0.83, 0.84, 0.80, and 0.81, respectively. Considering overall AUC, accuracy, precision, specificity, sensitivity and F1 score, the XGBoost model demonstrated the best predictive performance. To facilitate the model into real‑world clinical deployment, we propose a pragmatic integration framework for clinical decision support systems (CDSS). The model interfaces with electronic health records (EHR) via standard Application Programming Interfaces (e.g., Fast Healthcare Interoperability Resources), using routinely collected laboratory and clinical measurements as predictors. Before live deployment, a prospective pilot study in a simulated clinical environment is recommended to assess usability and net benefit, followed by periodic recalibration with new data.

This study has several limitations. First, the single‑center design may introduce selection bias and limit model generalizability, necessitating external validation across diverse hospitals and regions. Future work with larger multi‑center data, including TabNet and deep architectures, would further strengthen the robustness of our findings. Second, lifestyle-related factors like dietary carbohydrate intake and physical activity were not captured in our dataset. Future studies could incorporate these variables along with more detailed medication characteristics (e.g., current insulin use, dosage, type) to enhance the comprehensiveness of risk prediction. Third, hypoglycemic events in this study were identified using the hospital’s adverse event reporting system, constraining sample size and event counts. CGM offers more comprehensive glucose profiling and allows for prolonged monitoring periods, it is recommended to employ CGM to reduce the likelihood of undetected hypoglycemic events. Finally, state-of-the-art model explanation methods failed to account for nonlinear dependencies among features, which inevitably introduced associated bias.

Conclusions

In this study, we leveraged multiple routinely available features from the EHR to develop hypoglycemia risk prediction models using LR, XGBoost, SVM, and RF algorithms. Among these, the XGBoost model demonstrated superior predictive performance. This model may aid healthcare providers in identifying subgroups of T2DM patients at elevated risk of hypoglycemia at an early stage. By focusing on the key predictive factors, it can support the design of evidence-based and targeted intervention strategies to prevent or delay the onset of hypoglycemia, ultimately reducing adverse clinical outcomes.

Author contributions

QR, XZ and WX conceived and designed the study. XQ, HC and LL undertook the literature review and data curation. XQ, LL, and YL interpreted the data. QR, YL, and XQ wrote the first draft of the manuscript, with revision by XZ, WX and HC. All authors have read and approved the final version of the manuscript and had final responsibility for submitting it for publication.

Funding

The study was supported by the National Key Clinical Specialty (Clinical Nursing) Construction Project of China and the Nursing Scientific Research Project of the Second Affiliated Hospital of Chongqing Medical University(HL2024-17), and the funders had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript.

Data availability

The data sets used and analyzed during the study are available from the corresponding author upon reasonable request.

Declarations

Human ethics and consent to participate

The study was conducted according to the guidelines of the Declaration of Helsinki and was approved by the Ethics Committee of the Second Affiliated Hospital of Chongqing Medical University (No: 2025.778). All participants provided written or verbal informed consents.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.