A deep learning framework for virtual continuous glucose monitoring and glucose prediction based on life-log data

Abstract

While continuous glucose monitoring (CGM) has revolutionized metabolic health management, widespread adoption remains limited by cost constraints and usage burden, often resulting in interrupted monitoring periods. We propose a deep learning framework for glucose level inference that operates independently of prior glucose measurements, utilizing comprehensive life-log data. The model employs a bidirectional Long Short-Term Memory (LSTM) network with an encoder-decoder architecture, incorporating dual attention mechanisms for temporal and feature importance. The system was trained on data from 171 healthy adults, encompassing detailed records of dietary intake, physical activity metrics, and glucose measurements. The encoder’s hidden state as latent representations were analyzed for distributions of patterns of glucose and life-log sequences. The model showed a 19.49 ± 5.42 (mg/dL) in Root Mean Squared Error, 0.43 ± 0.2 in correlation coefficient, and 12.34 ± 3.11 (%) in Mean Absolute Percentage Eror for current glucose level predictions without any information of glucose at the inference step. The distribution of latent representations from the encoder showed the potential differentiation for glucose patterns. The model’s ability to maintain predictive accuracy during periods of CGM unavailability has the potential to support intermittent monitoring scenarios for users.

Introduction

Continuous glucose monitoring (CGM)^1,2 is one of the prospective technologies to measure interstitial glucose levels in real time through minimally invasive subcutaneous methods. It has been utilized for multiple daily injections (MDI) and automated insulin delivery (AID) for the management of diabetic patients; proved for its efficacy and usefulness for several clinical trials^3–5. Additionally, CGM could be used to detect stage 2 T1D at an early stage and identify individuals at high risk of developing T1D⁶. Nowadays, CGM has started to be utilized for checking glucose trajectories not only for diabetic patients but also for healthy people for improving their lifestyle and behavioral patterns^7–9. Some studies reported that CGM usages were helpful to reduce body weights and correction of improper habits of lifestyles, with entering the domain of preventive healthcare and lifestyle optimization for normoglycemic individuals^10–12. However, the maintenance cost of CGM is relatively expensive and this can induce the user’s financial burden even though real-time measurement within long period of usages may be helpful to capture glucose and behavioral patterns^13–15. With this manner, a Long Short-Term Memory (LSTM)-based model^16,17 for prediction of glucose levels without the previous glucose trajectory for inference is proposed in this study, as a virtual CGM. When real CGM data from the user is available, then the virtual CGM model can be trained to estimate current and future glucose levels based on glucose levels and life-log data including food consumptions and its nutrients, metabolic equivalent (MET) and step counts for exercise, and time. When CGM data is unavailable, then the virtual CGM can infer the glucose levels to compensate for the lack of real CGM devices.

There are some backgrounds fields of the study supporting the direction of this study.

1. Real CGM and Virtual CGM: CGM has become widely used in diabetes management^18,19, providing real-time glucose data and trends, enabling better glycemic control and reduced hypoglycemic events. CGM devices provide valuable insights into how food, exercise, and other lifestyle factors affect glucose levels in real-time²⁰. Contemporary CGM systems^21,22 incorporate advanced sensor technologies, including electrochemical glucose oxidase reactions and fluorescence-based sensing mechanisms, providing real-time glucose measurements with improved accuracy and reduced lag time compared to traditional monitoring methods, so these are expected to be more widespread. These devices have demonstrated utility in identifying postprandial glucose excursions and nocturnal hypoglycemic events, which might otherwise go undetected. However, several significant barriers impede universal CGM adoption. Attachment to the skin through several days may induce various dermatological complications such as irritations, infection, and inflammation of the skin^23,24. The cost of the devices can be burdens^15,25 for many individuals, and the need for frequent sensor replacements contributes to the overall expense. Thus, with several features from life-log, glucose inference at current time stamp mimicking real CGM operations^10,26 has been needed for supporting not only continuous but also intermittent usages of CGM. Meal intakes information, physical activity data, and recorded time can be alternative sources for estimating current glucose levels and future glucose trajectories with pattern analysis. Nowadays various sources from sensors have been utilized for reconstructing glucose levels. In addition, recent advances in sensor technology and data integration methods have expanded the potential data sources for glucose level estimation, including non-invasive techniques such as optical sensing methods²⁷.

2. Glucose prediction models including deep learning-based methods

   Glucose prediction models^28–31 have been proposed through various methods including autoregression, machine learning such as random forests and gradient boosting models, and neural networks-based models. One study³² proposed a seasonal stochastic modeling framework that predicts glucose by incorporating variable-length time-stamped events, such as meals and activity. This highlights the importance of irregular but clinically significant lifestyle factors, which complement continuous physiological monitoring. Deep learning models, especially those designed for time series analysis, have shown significant promise in capturing the temporal dynamics of glucose levels. Recurrent Neural Networks (RNNs), and to be specific, LSTMs are well-suited for handling sequential data and have been successfully applied in various healthcare applications, including glucose prediction; LSTMs can learn long-term dependencies in the data, making them capable of capturing the complex relationships between lifestyle factors and glucose fluctuations over time. While recent advances in natural language processing have introduced transformer-based architectures, which have shown promise in time-series analysis through their self-attention mechanisms, LSTM networks maintain several advantages³³ in glucose prediction applications. These include computational efficiency, particularly in scenarios with limited data availability, and the ability to handle variable-length input sequences effectively. LSTM architecture’s ability to maintain an internal memory state makes it particularly adept at capturing the delayed effects of various factors on glucose levels, such as the extended impact of high-fat meals or prolonged physical activity. One study³¹ developed an incrementally retrained LSTM model that personalizes glucose forecasts using limited CGM data. Their approach adapts predictions over time, refining its accuracy as more glucose readings become available. While this method improves individualization, it still relies on prior glucose values, unlike the proposed virtual CGM model, which operates independently of real-time CGM inputs. Including this perspective would help position the proposed multimodal approach within the broader landscape of glucose prediction models.

3. Utilization of life-log data for personalized glucose management

   The increasing availability of personal life-log data based on sensors and recordings from users through smartphones and wearable devices has opened new avenues for individualized glucose monitoring^34–37. Modern life-logging systems encompass a diverse array of data sources, including continuous activity monitors, dietary logging applications with advanced nutritional analysis capabilities, and sleep quality assessment tools. This rich data ecosystem provides multidimensional contextual information that can significantly enhance the accuracy and reliability of glucose prediction models. Advanced machine learning algorithms can leverage these diverse data streams to identify complex patterns and relationships between lifestyle factors and glycemic responses. However, the integration of life-log data presents several challenges, including data quality variability, temporal alignment issues, and the presence of human error in manual recordings^38–40. The field of sensor fusion and multi-modal data integration has emerged as a crucial area of research, focusing on methods to combine and normalize data from diverse sources while accounting for varying levels of accuracy and reliability. There are a lot of data types captured from diverse sensors; thus this area has the potential to infer a certain bio-signal with compensating for absence or partial measurement from sensor signals.

With this background, the contributions of our study are two-fold:

1. Development of a Practical Virtual CGM Solution: We introduce a streamlined yet effective deep learning model designed for real-world applications with 15-min prediction intervals and intermittent CGM usage patterns. This is based on encoder-decoder structure consisting of LSTM for virtual glucose prediction, utilizing life-log data such as food intake and physical activities as input, without previous glucose trajectory information at the inference step. This approach addresses the limitations of traditional CGM, particularly the cost and burden associated with continuous usage, potentially enabling more accessible and personalized glucose monitoring.

2. Comprehensive Model Enhancement Experiments: We systematically evaluated multiple performance-enhancing techniques to optimize our virtual CGM system. For example, modules for autocorrelations of trajectories and food name embeddings from a language model were implemented and tested; A pretrained model from all trajectories from all subjects as a universal model were updated via fine-tuning from individualized data to capture personalized characteristics and to improve the performance of the model. We also investigated latent representations from the encoder for the potential clustering. Masking features showed influences on prediction performances with considering priority and errors from collecting data. These experiments can be helpful to design the practical model for glucose inference and forecasting regarding CGM.

Materials and methods

We collected glucose data from CGM and life-log data from smartphone-based apps and smart watches. After that, subsequences were extracted from whole trajectories including glucose levels and life-log data, based on the sliding-window technique. An encoder-decoder model was proposed to infer the current and future glucose levels from features of life-log data from sensors and subject reported activities. Predicted and actual glucose levels were compared and used for loss functions to train the model (Fig. 1).

Fig. 1.

An overview of the study. CGM, continuous glucose monitoring, V-CGM, virtual continuous glucose monitoring.

Data acquisition process

Recorded data of CGM, information of food intakes and exercise from 171 healthy subjects using the smartphone app were acquired in the retrospective manner, the period was between Oct. 2023 to Aug. 2024. The subjects were over 18 in their age without diagnosis of diabetes and pre-diabetic diseases. Days for constructing sequences were 6.93 ± 0.28 and 3.99 ± 0.01 days for training and testing sequences, respectively. The number of readings from CGM were 660 ± 31 and 380 ± 27 for training and testing sequences respectively. Detailed demographic information is shown in Table 1. The study received ethical approval from the institutional review boards of Ulsan National Institute of Science and Technology (IRB No. UNISTIRB-24-018-A), and the requirement for informed consent was waived due to the study’s retrospective design and the de-identification of subject data. The study protocol and process adhered to the ethical guidelines of the 1975 Declaration of Helsinki and its subsequent revisions.

Table 1.

Demographic information.

Time stamp	Types	Value (Counts)
Age (years)	18–29	43
	30–39	79
	40–49	38
	50 and above	11
Gender	Female	23
	Male	148
Weight (kg)		60.88 ± 10.78
Height (cm)		163.55 ± 6.22
BMI		22.70 ± 3.48
Glucose levels (mg/dL)		103.02 ± 21.11

BMI: Body Mass Index. Continuous values are represented mean with standard deviation.

Glucose levels from CGM devices (Dexcom G7) were recorded every 15 min and stored in the app service database. Users manually entered information regarding their meal intake, exercise routines, and body weight. Body mass index (BMI) was calculated using the recorded weights and heights of users. Meal intake data included details such as total calories, carbohydrate content, and the proportions of sugar, fat, and protein, along with the names of food and time of consumption using text and numbers. The user was able to input this food information on time or by recalling it from memory. Exercise data comprises metabolic equivalent of task (MET) values from recordings by users and step counts measured by sensors from smart watch. Detailed Exercise information including exercise names and duration was inputted by users via the app.

For construction of the dataset, subject selection was conducted based on two criteria. First, real CGM data had to be recorded for a minimum of 12 days, to construct training as 7 days, gap as 1 day, and testing period as 4 days based on the number of days with temporal validation manner. The gap period was introduced to explicitly separate the training and testing sequences, mitigating potential data leakage when using a sliding window technique. Second, during both the training and testing periods, data coverage needed to exceed 50% of the possible recording points.

Preprocessing

For construction of training sequences, sliding window technique was adopted. Based on the index timestamp (0 min: current time stamp), the trajectory from previous 90 min to 0 min was set as a previous sequence to be utilized in training models or inference, and the trajectory from 0 min to future 90 min was set as a future sequence to be predicted from the model. Considering the purpose of the model as virtual CGM, glucose level of 0 min was the important object to be predicted. In addition, the trajectories containing food intakes were mainly considered and extracted, to focus on the meal absorption effect that is the primary event to make the glucose levels increase as a disturbance from the stable status of glucose dynamics. In addition, all trajectories with and without meal events were also considered for integrity.

In the preprocessing, some features were modified and augmented. First, the time was transformed the value between -1 and 1 with triangular functions, such as sin and cos, to reflect periodicity of our body systems and behaviors based on 24 h. There were 11 features including sin(time), cos(time), step counts, calories, the amounts of fat, sugar, protein, carbohydrates, MET, weight, and BMI. Other features were not exploited considering the privacy and de-identification. Then features were normalized via min–max normalization. Second, difference between values from two adjacent timepoints as first derivative of features were added to input features to reflect changes of features regarding time. This additional feature augmentation can consider the trend of changes of features along time. Recording times of other features were matched to the nearest timestamps of glucose levels for CGM.

The subsequence for training models consisting of two parts: first part is the life-log features without the glucose trajectory from – 90 min to 0 min (current timestamp as index timestamp), and second part is the glucose trajectories from 0 min to + 90 min. The input for the model is the first part of the subsequence, and the output for inference from the model is the second part of the subsequence. The subsequences were extracted based on sliding window technique for every 15 min. The average numbers of subsequences containing at least one meal event within previous trajectories between -90 min to 0 min per subject are 160.79 ± 28.88 and 77.84 ± 23.44 in the training set and testing set, respectively.

Model architecture, training and inference processes

The model consists of an encoder and a decoder (Fig. 2). The encoder takes life-log features without glucose trajectories as input and future glucose trajectories including glucose levels at 0 min are predicted from the decoder. The encoder and decoder utilized LSTMs. The feature and temporal attention modules were combined to consider feature importance for the explanation of the model’s operations. The autocorrelation module was introduced for additional processing for input subsequences. Hidden states can be regarded as latent representations and decoder can restore the future glucose trajectories. With this proposed model, embeddings from RoBERTa (A Robustly Optimized BERT Pretraining Approach)⁴¹ were utilized for modification of hidden states as an additional model. The loss function is the summation of L1 loss, L2 loss, and cos function for loss functions; the glucose levels of time points in predicted and actual future subsequences are inputted into the loss function for training the model. Before constructing individualized models trained by subsequences from just one subject, the universal model was trained in all training subsequences from all subjects; It means that the universal model can capture the characteristics of patterns from all subjects as the pretrained model.

1

2

3

4

Fig. 2.

A Schematic picture of the model architecture (a) Overall model structure (b) internal operation of the encoder (c) temporal attention module (d) embeddings from food’s name. All information from trajectories is preprocessed. Glucose information is not inputted to encoder, just utilized for calculation of loss functions. Feature attention is calculated from the sequence and autocorrelation is introduced to internal calculations. Temporal attention is calculated based on encoder output and decoder’s hidden states from LSTM.

For implementation, Python 3.9.19 and PyTorch 2.2.2 were utilized. Pretrained RoBERTa Model on Korean Language (KLUE RoBERTa base)⁴² with Transformers 4.46.3. from Hugging Face repository was exploited for extracting from food names written by Korean. Recorded trajectories were extracted to sequences consisting of previous and future subsequences with preprocessing. Then the encoder generated hidden states as latent representations that can be utilized to construct future glucose subsequences including glucose level at current time (0 min). In the encoder, feature attention module was implemented, and autocorrelation was additionally calculated and concatenated to construct the encoder’s output. Hidden states were transferred to the decoder’s hidden state and updated; temporal attention values were calculated via generating future subsequences as the decoder’s output. For utilizing the food’s name as the additional feature, the embedding was acquired from the pretrained model of RoBERTa as LLM.

Evaluation metrics

For testing the model’s performance in prediction of current and future glucose levels, RMSE (Root Mean Squared Error), Correlation, MAE (Mean Absolute Error), and MAPE (Mean Absolute Percentage Error) were calculated for glucose levels in all time stamps of the future subsequences. Let N be the number of subsequences and be the number of timestamps of current and future glucose levels, then these metrics can be defined as below.

5

6

7

8

Results

Inference of glucose Levels

The proposed model was evaluated based on metrics, MAE, Correlation, RMSE, and MAPE at current glucose level at 0 min for prediction were 14.93 ± 4.2, 0.43 ± 0.2, 19.49 ± 5.42, and 12.34 ± 3.11, respectively (Table 2, Fig. 3). The proposed model showed the better performance for inferring current glucose levels at current time stamp (0 min) compared to other cases including no-pretraining, the model without autocorrelation module, and the model with additional usages of the food name’s embedding from the language model. For future glucose levels prediction, the model exploiting embeddings from food names showed better performance compared to other models.

Table 2.

Evaluation of glucose level inference.

Time stamp	Types	MAE (mg/dL)	Correlation	RMSE (mg/dL)	MAPE (%)
0 min	No pretrain	15.47 ± 4.49	0.38 ± 0.2	20.20 ± 5.85	12.77 ± 3.22
	Without AC	15.22 ± 4.36	0.43 ± 0.2	19.85 ± 5.62	12.56 ± 3.19
	Proposed	14.93 ± 4.2	0.43 ± 0.2	19.49 ± 5.42	12.34 ± 3.11
	w/food name	15.11 ± 4.39	0.42 ± 0.19	19.74 ± 5.76	12.48 ± 3.26
15 min	No pretrain	16.04 ± 4.61	0.33 ± 0.22	20.75 ± 5.96	13.09 ± 3.23
	Without AC	15.71 ± 4.37	0.37 ± 0.22	20.36 ± 5.61	12.84 ± 3.18
	Proposed	15.50 ± 4.23	0.38 ± 0.22	20.04 ± 5.46	12.69 ± 3.10
	w/food name	15.41 ± 4.42	0.37 ± 0.2	20.00 ± 5.75	12.56 ± 3.2
30 min	No pretrain	16.36 ± 4.68	0.3 ± 0.23	21.07 ± 6.09	13.28 ± 3.27
	Without AC	15.94 ± 4.41	0.35 ± 0.23	20.57 ± 5.66	12.97 ± 3.20
	Proposed	15.76 ± 4.3	0.35 ± 0.23	20.32 ± 5.53	12.86 ± 3.17
	w/food name	15.55 ± 4.55	0.35 ± 0.21	20.08 ± 5.90	12.59 ± 3.26
45 min	No pretrain	16.2 ± 4.67	0.30 ± 0.23	20.98 ± 6.18	13.21 ± 3.3
	Without AC	15.87 ± 4.5	0.36 ± 0.23	20.51 ± 5.77	12.96 ± 3.3
	Proposed	15.68 ± 4.39	0.35 ± 0.23	20.27 ± 5.59	12.84 ± 3.26
	w/food name	15.38 ± 4.54	0.36 ± 0.22	19.93 ± 5.91	12.53 ± 3.29
60 min	No pretrain	15.65 ± 4.56	0.28 ± 0.24	20.37 ± 6.11	12.9 ± 3.26
	Without AC	15.49 ± 4.41	0.34 ± 0.24	20.08 ± 5.74	12.79 ± 3.26
	Proposed	15.33 ± 4.26	0.34 ± 0.23	19.86 ± 5.52	12.67 ± 3.20
	w/food name	15.03 ± 4.45	0.34 ± 0.23	19.46 ± 5.75	12.42 ± 3.31
75 min	No pretrain	15.04 ± 4.52	0.26 ± 0.24	19.6 ± 6.05	12.53 ± 3.31
	Without AC	14.96 ± 4.3	0.30 ± 0.23	19.45 ± 5.70	12.49 ± 3.23
	Proposed	14.78 ± 4.16	0.31 ± 0.24	19.2 ± 5.43	12.35 ± 3.18
	w/food name	14.55 ± 4.38	0.31 ± 0.23	18.83 ± 5.58	12.18 ± 3.35
90 min	No pretrain	14.46 ± 4.39	0.24 ± 0.23	18.83 ± 5.93	12.18 ± 3.34
	Without AC	14.35 ± 4.23	0.28 ± 0.22	18.72 ± 5.71	12.11 ± 3.29
	Proposed	14.2 ± 4.06	0.28 ± 0.23	18.49 ± 5.37	11.98 ± 3.20
	w/food name	14.08 ± 4.3	0.27 ± 0.22	18.23 ± 5.49	11.91 ± 3.37

No pretrain means that each model was trained based on individual trajectories without the universal model trained from all trajectories of all subjects. Without AC means the model that has no autocorrelation module. w/food name model means that additional features for the model are embeddings of food names.

Fig. 3.

Comparison between predicted and actual glucose levels: point plots. Current (0 min) and future (15 min to 90 min) glucose levels are compared based on actual and predicted values.

In addition, models for different conditions such as including all trajectories with and without meal events, utilizing previous 24 h trajectories rather than 90 min trajectories, and utilizing previous glucose level information were tested (Supplementary Table 2).

Explanation from attention mechanisms

The model integrates two modules for temporal and feature attentions. Without glucose information before the index timestamp, the features in previous subsequence were utilized adaptively to infer the current and future glucose levels (Fig. 4).

Fig. 4.

Comparison between predicted and actual glucose levels. The left part of the figure shows the overall trajectories of previous (− 90 ~ 0 min, green), actual (0 ~  + 90 min, blue), and predicted (0 ~  + 90 min, red) glucose trajectories. The overall attention values show that the carbohydrate feature was the most important as well as other nutrients represented significant contributions for inferring glucose trajectories. The right part of the figure shows examples at certain timestamps to observe detailed functions of the model. The food intakes (gray), exercise checked by MET (yellow), and glucose trajectories (green: actual/previous, blue: actual/future, red: predicted/future) in the subsequence are represented.

Distributions of Latent Representations

For hidden states as latent representations from the encoder, PCA (Principal Component Analysis) were conducted for observing patterns of subsequences including glucose levels and life-log data for all subjects. Latent representations (hidden states) with interpretable features such as the averaged value of glucose levels, MET, step counts, calories and carbohydrates, and triangular functions of the timestamps are not evenly distributed in spatial manner, reflecting that the subsequence can be differentiated via data-driven approaches (Fig. 5).

Fig. 5.

Distributions of hidden states with features on PCA space. From principal axes based on PCA, characteristics of subsequences were matched to visualize latent representations from the output of the encoder.

Discussions

Effects of pretraining models

The implementation of transfer learning through the universal model that is pretrained from all training trajectories in all subjects demonstrated some significant improvements in glucose prediction accuracy across multiple evaluation metrics, especially in correlation performances (Supplementary Table 3). Personalized models initialized with pretrained weights of the universal model and subsequently fine-tuned based on sequences of each subject showed the better performances trends. This enhancement can be attributed to the pretraining process to improve model’s ability to understand glucose dynamics from all responses, as other fields such as computer vision usually exploit pretrained models including ResNet and Vision-transformer to achieve specific tasks⁴³.

Effect of additional modules

The integration of an autocorrelation module^44,45 into the architecture yielded substantial improvements in predictive accuracy, particularly in capturing short-term glucose fluctuations. The autocorrelation mechanism, implemented through a dedicated neural network layer, effectively modeled the temporal dependencies inherent in glucose time series data. The incorporation of semantic food information through pretrained word embeddings provided complementary benefits, particularly in predicting postprandial glucose responses. The model leveraging both autocorrelation and food name embeddings demonstrated better performance in capturing meal-induced glycemic excursions. Moreover, the integration of food embeddings demonstrated a utility in handling novel or uncommon food items, where direct nutritional information might be incomplete or imprecise. The semantic relationships encoded in the embedding space enabled the model to make reasonable predictions even for previously unseen food combinations, suggesting improved generalization capabilities. We additionally tested the adoption of the transformer architecture for initial process⁴⁶ in the model’s operations, but the model’s performance was not better than the LSTM-based model (Supplementary Table 1).

Influences of features to infer glucose levels

Before the ablation study of features to the model, correlations from features of the subsequence were investigated. As expected, features from glucose levels and information from food intakes were gathered, respectively. Features of time represented as sine and cosine of triangular functions also showed correlation patterns with food intakes and exercise patterns. Relevant features were gathered in correlations with hierarchical aspects (Fig. 6). Based on these results some features of groups were masked to the input for measuring influences of features (Table 3). Ablations of time, food-related features, and exercise reflecting features (MET, steps) decreased correlations and increased MAE, MAPE, and RMSE. When the time features were removed, MAE, Correlation, RMSE, and MAPE at current glucose level at 0 min for prediction were 15.3 ± 4.20, 0.38 ± 0.2, 20.06 ± 5.57, and 12.55 ± 3.05, respectively. When the meal intake-related features including calorie and carbohydrate, MAE, Correlation, RMSE, and MAPE at current glucose level at 0 min for prediction were 15.85 ± 4.83, 0.22 ± 0.21, 20.57 ± 6.34, and 12.98 ± 3.62, respectively. When the exercise-related features including steps and MET were removed, MAE, Correlation, RMSE, and MAPE at current glucose level at 0 min for prediction were 15.98 ± 4.76, 0.3 ± 0.22, 20.53 ± 5.81, and 13.56 ± 4.38, respectively.

Fig. 6.

Correlation between features in subsequences. fut., Future, BMI, Body Mass Index, sin, Sine, cos, Cosine.

Table 3.

Ablation of features.

Time stamp	Ablated feature	MAE (mg/dL)	Correlation	RMSE (mg/dL)	MAPE (%)
0 min	Time	15.30 ± 4.20	0.38 ± 0.2	20.06 ± 5.57	12.55 ± 3.05
	Meal intake	15.85 ± 4.83	0.22 ± 0.21	20.57 ± 6.34	12.98 ± 3.62
	Exercise	15.98 ± 4.76	0.30 ± 0.22	20.53 ± 5.81	13.56 ± 4.38
15 min	Time	15.88 ± 4.33	0.32 ± 0.2	20.61 ± 5.64	12.93 ± 3.09
	Meal intake	15.96 ± 4.85	0.22 ± 0.21	20.67 ± 6.3	12.92 ± 3.52
	Exercise	16.35 ± 4.77	0.25 ± 0.22	20.95 ± 5.88	13.70 ± 4.25
30 min	Time	16.10 ± 4.44	0.30 ± 0.21	20.8 ± 5.80	13.07 ± 3.15
	Meal intake	16.05 ± 4.77	0.22 ± 0.22	20.71 ± 6.25	12.94 ± 3.48
	Exercise	16.48 ± 5.0,	0.24 ± 0.22	21.09 ± 6.15	13.69 ± 4.39
45 min	Time	15.98 ± 4.53	0.31 ± 0.21	20.66 ± 5.93	13.02 ± 3.17
	Meal intake	15.97 ± 4.68	0.21 ± 0.23	20.67 ± 6.23	12.94 ± 3.49
	Exercise	16.18 ± 4.96	0.25 ± 0.22	20.79 ± 6.28	13.41 ± 4.25
60 min	Time	15.52 ± 4.46	0.29 ± 0.22	20.1 ± 5.88	12.77 ± 3.11
	Meal intake	15.35 ± 4.51	0.20 ± 0.24	20.0 ± 6.14	12.62 ± 3.45
	Exercise	15.45 ± 4.73	0.24 ± 0.22	20.04 ± 6.19	12.84 ± 3.89
75 min	Time	14.88 ± 4.36	0.26 ± 0.23	19.31 ± 5.73	12.37 ± 3.09
	Meal intake	14.65 ± 4.31	0.19 ± 0.24	19.13 ± 5.89	12.23 ± 3.39
	Exercise	14.77 ± 4.52	0.22 ± 0.21	19.26 ± 6.04	12.36 ± 3.65
90 min	Time	14.21 ± 4.26	0.24 ± 0.23	18.46 ± 5.61	11.92 ± 3.09
	Meal intake	14.02 ± 4.05	0.18 ± 0.22	18.27 ± 5.56	11.85 ± 3.31
	Exercise	14.23 ± 4.40	0.2 ± 0.21	18.52 ± 5.93	12.02 ± 3.60

Ablated feature means that the features were masked to the input for the model to remove the influence of those features.

The model’s accuracy changed over different time horizons. This may be attributed to the increasing uncertainty associated with forecasting glucose levels further into the future. For example, the further ahead a prediction is made from index time as 0 min, the higher the likelihood that the individual will engage in activities such as eating, exercising, or other behaviors. Consequently, in general cases, prediction errors are expected to increase with longer forecasting horizons. In contrast, when model training and evaluation are restricted to subsequences that include at least one meal event, the probability of an additional meal intake within the subsequent 90 min may be reduced, given that a meal was already consumed in the preceding 90-min window. Additionally, the physiological impact of the meal may diminish over time. Therefore, for certain metrics, this may lead to an apparent decrease in error over time, reflecting the reduced variability in input conditions.

Additionally, we also investigated the effect of just one-type feature as ablation studies (Supplementary Table 2). We found that food and exercise information are more important^47,48 for inferring 0 min glucose levels. It indicates that even time can reflect periodicity of users’ behaviors, but loggings are more important and technologies to record user’s activity information and reduce errors for loggings should be needed. For food logging, users just input the information of food events manually and it may be not quite special compared to other approaches. It can induce some errors from missing or inaccurate inputs as noises for the prediction model but it is basic and simple. However, features related with food information in our study have significant effects on glucose inference and prediction (Supplementary Table 2), it can indicate that the food logging in this model, even though it is simple, has a positive effect for model’s performance compared to the case of non-logging of food information.

In the correlation analysis between input features, despite the relatively low correlation among the three categories, masking features from one category did not significantly decrease glucose prediction performance. However, in evaluation metrics, correlation between predicted and accurate glucose levels, the absence of food information caused a relatively significant decrease. Considering that food is a major factor in glucose trajectory changes, this can be aligned with this understanding. In other words, it is aligned with the fact that meal absorption function is one of the important subsystems in mathematical modellings for glucose dynamics⁴⁷.

In addition, feature attention scores were checked as feature importance, the carbohydrate feature has the highest value for contribution to model’s output (Supplementary Fig. 1). Based on deep-learning architecture, attention scores can capture complex and non-linear relationships between variables such as the Shapley value analysis and mixed linear models^49–51. Diverse meal compositions can influence the output of our glucose prediction models, with the relative importance of different nutrients estimated through attention scores.

For exercise, it has been known that the exercise’s effect can affect glucose dynamics up to 24 hours⁵², the model utilizing previous 24-h subsequence was tested and showed better performance compared to the proposed model (Supplementary Table 1). If the log data are recorded consistently over a 24-h period, this would serve as an effective approach for predicting glucose levels. However, our method is designed to operate under limited short-term recordings, taking into account potential constraints where logs may be collected intermittently.

For short-term predictions up to approximately 30 min, models utilizing only previous glucose values demonstrated clearly superior performance (Supplementary Table 1) compared to the proposed model utilizing no glucose levels in previous subsequences. However, beyond this time frame, the performance of models using only previous glucose values and those using all features except glucose became comparable. This suggests that while previous glucose values carry strong predictive information, certain aspects of glucose dynamics can also be inferred from other features.

Implications in the usage of CGM for healthy people

We established virtual CGM models for healthy people in this study; There are debates and studies^53–58 for advantages and disadvantages for using CGM to check glucose levels. The growing interest in CGM among individuals without diabetes reflects needs for personalized health and proactive wellness. CGM can provide real-time insights into glycemic responses to user’s activities such as food, exercise, sleep, and stress, allowing users to gain a deeper understanding of how daily behaviors impact their metabolic health. Some studies have shown that CGM use in people with prediabetic conditions or overweight/obesity can improve health-related behaviors such as dietary choices and physical activity, while also enhancing self-awareness and motivation for weight management. Furthermore, novel CGM-derived metrics like "Latest Spike Time" (LST) have shown strong correlations with weight loss outcomes, suggesting that CGM may support personalized behavioral interventions even outside the context of diabetes care. The emergence of over-the-counter (OTC) CGM devices further increases accessibility for a broader population to engage in preventive health practices.

However, the expansion of CGM usage to healthy populations also raises several concerns. First, there is a lack of standardized interpretation criteria for CGM metrics in non-diabetic individuals, which can lead to misinterpretation of normal physiological glucose fluctuations as pathological. This may cause unnecessary anxiety, dietary restrictions, or over-medicalization for healthy individuals. Moreover, inconsistent clinical recommendations among experts when evaluating CGM reports in people without diabetes may reflect limited consensus and insufficient normative data. Cost and accessibility remain additional barriers, particularly for those without insurance coverage or access to wellness programs, because it would cost $60 per week and $240 per month. One advantage of our approaches to utilize virtual CGM is to save those costs when the CGM is intermittently utilized, because virtual CGM can be used to complement the non-existence of real CGM. Without adequate education and guidance, widespread use of CGM in the general population may lead to more confusion than clarity. Therefore, while CGM can be one of the promising applications beyond diabetes, its use in healthy individuals should be investigated with caution and supported by further research and clear clinical frameworks.

Limitation and future direction

Our investigation into virtual CGM demonstrates comparable results for glucose level reconstruction using intermittent CGM data acquisition. Considering the data input/output and storage and battery consumption according to the measurement cycle of smartphone apps, we aimed to pursue practical implementation by setting 15 min as the unit time for CGM. While the model shows potential as a compensatory mechanism during periods of CGM unavailability, there are several important limitations and opportunities for future research.

First, in the perspective of population heterogeneity and generalizability, the model’s training on normoglycemic adults presents inherent limitations in its generalizability. The absence of validation in populations with altered glucose homeostasis (e.g., Type 1 and Type 2 diabetes mellitus, gestational diabetes, impaired glucose tolerance) raises important questions about the model’s applicability across the spectrum of metabolic phenotypes. Glucose regulation mechanisms in these populations exhibit distinct characteristics; Physiological variations^59–61 may necessitate substantial architectural modifications or population-specific model recalibration.

Second, the reliability of the model is inherently constrained by several data-related factors: temporal resolution and consistency of life-log entries, systematic and random errors in self-reported dietary data, variable user engagement and compliance patterns, potential recall bias in activity logging, circadian and seasonal variations in behavioral patterns. These limitations could introduce systematic biases in the model’s predictions, particularly during periods of irregular data collection or misreported information.

Third, even though the latent representations were represented, there were absences of distinct cluster boundaries in the latent space and overlapping phenotype distributions. The structure and loss function for differentiating subjects based on metabolic and behavioral characteristics via representations should be further investigated.

There are several future directions from this study. First, Prospective studies can be expected to show the validity of significant implications of virtual CGM usages. The development of hybrid models incorporating mechanistic glucose-insulin dynamics and announced meal detection models⁶² can overcome the limited numbers of data points and errors from reports by users. Integration of continuous physiological monitoring from diverse sensors can improve the accuracy of glucose level inference if the sensors can be accessible to users. For example, skin temperature and heart rate²⁶ can be surrogates for integration in features. Sleep patterns⁶³ and stress levels may be estimated from those signals, and these can also be features for the model. Second, pre-diabetic and diabetes patients should be considered for subjects in testing virtual CGM models. If the data format and types remain consistent with those used in our study, the model is expected to be effectively trained on other populations, including individuals with diabetes. Furthermore, expanding the dataset to encompass a broader range of glucose impairments may enhance the model’s generalization capability, enabling it to more accurately capture glucose dynamics across diverse conditions. Fine-tuning models for diabetes or prediabetes can also be considered. Third, considering the performance difference between the model trained solely on subsequences containing meal events and the model trained on all subsequences in our study, a stratified approach that accounts for the uneven distribution of meal timing and activity levels may be beneficial. Fourth, for real-time usages, real-time correction functions such as exemplified by the Retrospective Correction feature in Loop⁶⁴ will be helpful to play a significant role in improving model performance.

Conclusion

This study presents a novel deep learning framework for virtual CGM that demonstrates the feasibility of accurate glucose level inference without continuous reliance on physical CGM devices; CGM is just utilized in the manner of calibration of the model in the training process. The proposed LSTM-based architecture, incorporating dual-attention mechanisms and an autocorrelation module, achieved decent predictive performance compared to glucose prediction models using previous trajectories of glucose levels at inference. The integration of transfer learning through universal model pretraining, coupled with the synergistic implementation of food embedding features, significantly enhanced the model’s capacity to capture complex glycemic responses across varying temporal horizons. The latent representation space suggests potential applications in characteristics of glucose dynamics and lifestyle interventions.

The implementation of this virtual CGM framework can represent a significant advancement in accessible glucose monitoring technology, potentially addressing the economic and practical limitations associated with conventional CGM devices. The model’s ability to maintain predictive accuracy during periods of CGM unavailability can support intermittent monitoring scenarios for users.

Author contributions

Min Hyuk Lim: conceptualization, investigation, methodology, data acquirement, analysis, visualization, writing. Hyocheol Chae: conceptualization, data acquirement, investigation, writing, Jeongwon Yoon, Insik Shin: conceptualization, data acquirement, writing.

Data availability

The data supporting the results of this study are not publicly available due to privacy, commercialization, and ethical restrictions. However, data may be made available upon request from the corresponding author.

Declarations

Ethics statements

The study received ethical approval from the institutional review boards of Ulsan National Institute of Science and Technology (IRB No. UNISTIRB-24-018-A), and the requirement for informed consent was waived due to the study’s retrospective design and the de-identification of subject data. The study protocol and process adhered to the ethical guidelines of the 1975 Declaration of Helsinki and its subsequent revisions.

Competing interests

The authors declare no competing interests.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-01367-7.