Machine learning-based glucose prediction with use of continuous glucose and physical activity monitoring data: The Maastricht Study

Use of Machine Learning to Predict Individual Postprandial Glycemic Responses to Food Among Individuals With Type 2 Diabetes in India: Protocol for a Prospective Cohort Study

BACKGROUND

Type 2 diabetes (T2D) is a leading cause of premature morbidity and mortality globally and affects more than 100 million people in the world's most populous country, India. Nutrition is a critical and evidence-based component of effective blood glucose control and most dietary advice emphasizes carbohydrate and calorie reduction. Emerging global evidence demonstrates marked interindividual differences in postprandial glucose response (PPGR) although no such data exists in India and previous studies have primarily evaluated PPGR variation in individuals without diabetes.

OBJECTIVE

This prospective cohort study seeks to characterize the PPGR variability among individuals with diabetes living in India and to identify factors associated with these differences.

METHODS

Adults with T2D and a hemoglobin A1c of ≥7 are being enrolled from 14 sites around India. Participants wear a continuous glucose monitor, eat a series of standardized meals, and record all free-living foods, activities, and medication use for a 14-day period. The study's primary outcome is PPGR, calculated as the incremental area under the curve 2 hours after each logged meal.

RESULTS

Data collection commenced in May 2022, and the results will be ready for publication by September 2025. Results from our study will generate data to facilitate the creation of machine learning models to predict individual PPGR responses and to facilitate the prescription of personalized diets for individuals with T2D.

CONCLUSIONS

This study will provide the first large scale examination variability in blood glucose responses to food in India and will be among the first to estimate PPGR variability for individuals with T2D in any jurisdiction.

TRIAL REGISTRATION

Clinical Trials Registry-India CTRI/2022/02/040619; https://tinyurl.com/mrywf6bf.

INTERNATIONAL REGISTERED REPORT IDENTIFIER (IRRID)

DERR1-10.2196/59308.

Composite activity type and stride-specific energy expenditure estimation model for thigh-worn accelerometry

BACKGROUND

Accurately measuring energy expenditure during physical activity outside of the laboratory is challenging, especially on a large scale. Thigh-worn accelerometers have gained popularity due to the possibility to accurately detect physical activity types. The use of machine learning techniques for activity classification and energy expenditure prediction may improve accuracy over current methods. Here, we developed a novel composite energy expenditure estimation model by combining an activity classification model with a stride specific energy expenditure model for walking, running, and cycling.

METHODS

We first trained a supervised deep learning activity classification model using pooled data from available adult accelerometer datasets. The composite energy expenditure model was then developed and validated using additional data based on a sample of 69 healthy adult participants (49% female; age = 25.2 ± 5.8 years) who completed a standardised activity protocol with indirect calorimetry as the reference measure.

RESULTS

The activity classification model showed an overall accuracy of 99.7% across all five activity types during validation. The composite model for estimating energy expenditure achieved a mean absolute percentage error of 10.9%. For running, walking, and cycling, the composite model achieved a mean absolute percentage error of 6.6%, 7.9% and 16.1%, respectively.

CONCLUSIONS

The integration of thigh-worn accelerometers with machine learning models provides a highly accurate method for classifying physical activity types and estimating energy expenditure. Our novel composite model approach improves the accuracy of energy expenditure measurements and supports better monitoring and assessment methods in non-laboratory settings.

Learning control-ready forecasters for Blood Glucose Management

Type 1 diabetes (T1D) presents a significant health challenge, requiring patients to actively manage their blood glucose (BG) levels through regular bolus insulin administration. Automated control solutions based on machine learning (ML) models could reduce the need for manual patient intervention. However, the accuracy of current models falls short of what is needed. This is due in part to the fact that these models are often trained on data collected using a basal bolus (BB) strategy, which results in substantial entanglement between bolus insulin and carbohydrate intake. Under standard training approaches, this entanglement can lead to inaccurate forecasts in a control setting, ultimately resulting in poor BG management. To address this, we propose a novel algorithm for training BG forecasters that disentangles the effects of insulin and carbohydrates. By exploiting correction bolus values and leveraging the monotonic effect of insulin on BG, our method accurately captures the independent effects of insulin and carbohydrates on BG. Using an FDA-approved simulator, we evaluated our approach on 10 individuals across 30 days of data. Our approach achieved on average higher time in range compared to standard approaches (81.1% [95% confidence interval (CI) 80.3,81.9] vs 53.6% [95%CI 52.7,54.6], p<0.001), indicating that our approach is able to reliably maintain healthy BG levels in simulated individuals, while baseline approaches are not. Utilizing proxy metrics, our approach also demonstrates potential for improved control on three real world datasets, paving the way for advancements in ML-based BG management.

Artificial intelligence in preventive cardiology

Artificial intelligence (AI) is a field of study that strives to replicate aspects of human intelligence into machines. Preventive cardiology, a subspeciality of cardiovascular (CV) medicine, aims to target and mitigate known risk factors for CV disease (CVD). AI's integration into preventive cardiology may introduce novel treatment interventions and AI-centered clinician assistive tools to reduce the risk of CVD. AI's role in nutrition, weight loss, physical activity, sleep hygiene, blood pressure, dyslipidemia, smoking, alcohol, recreational drugs, and mental health has been investigated. AI has immense potential to be used for the screening, detection, and monitoring of the mentioned risk factors. However, the current literature must be supplemented with future clinical trials to evaluate the capabilities of AI interventions for preventive cardiology. This review discusses present examples, potentials, and limitations of AI's role for the primary and secondary prevention of CVD.

Machine Learning and Deep Learning Models for Nocturnal High- and Low-Glucose Prediction in Adults with Type 1 Diabetes

Glucose management at night is a major challenge for people with type 1 diabetes (T1D), especially for those managed with multiple daily injections (MDIs). In this study, we developed machine learning (ML) and deep learning (DL) models to predict nocturnal glucose within the target range (3.9-10 mmol/L), above the target range, and below the target range in subjects with T1D managed with MDIs. The models were trained and tested on continuous glucose monitoring data obtained from 380 subjects with T1D. Two DL algorithms-multi-layer perceptron (MLP) and a convolutional neural network (CNN)-as well as two classic ML algorithms, random forest (RF) and gradient boosting trees (GBTs), were applied. The resulting models based on the DL and ML algorithms demonstrated high and similar accuracy in predicting target glucose (F1 metric: 96-98%) and above-target glucose (F1: 93-97%) within a 30 min prediction horizon. Model performance was poorer when predicting low glucose (F1: 80-86%). MLP provided the highest accuracy in low-glucose prediction. The results indicate that both DL (MLP, CNN) and ML (RF, GBTs) algorithms operating CGM data can be used for the simultaneous prediction of nocturnal glucose values within the target, above-target, and below-target ranges in people with T1D managed with MDIs.

Explainable Machine Learning for Real-Time Hypoglycemia and Hyperglycemia Prediction and Personalized Control Recommendations

BACKGROUND

The occurrences of acute complications arising from hypoglycemia and hyperglycemia peak as young adults with type 1 diabetes (T1D) take control of their own care. Continuous glucose monitoring (CGM) devices provide real-time glucose readings enabling users to manage their control proactively. Machine learning algorithms can use CGM data to make ahead-of-time risk predictions and provide insight into an individual's longer term control.

METHODS

We introduce explainable machine learning to make predictions of hypoglycemia (<70 mg/dL) and hyperglycemia (>270 mg/dL) up to 60 minutes ahead of time. We train our models using CGM data from 153 people living with T1D in the CITY (CGM Intervention in Teens and Young Adults With Type 1 Diabetes)survey totaling more than 28 000 days of usage, which we summarize into (short-term, medium-term, and long-term) glucose control features along with demographic information. We use machine learning explanations (SHAP [SHapley Additive exPlanations]) to identify which features have been most important in predicting risk per user.

RESULTS

Machine learning models (XGBoost) show excellent performance at predicting hypoglycemia (area under the receiver operating curve [AUROC]: 0.998, average precision: 0.953) and hyperglycemia (AUROC: 0.989, average precision: 0.931) in comparison with a baseline heuristic and logistic regression model.

CONCLUSIONS

Maximizing model performance for glucose risk prediction and management is crucial to reduce the burden of alarm fatigue on CGM users. Machine learning enables more precise and timely predictions in comparison with baseline models. SHAP helps identify what about a CGM user's glucose control has led to predictions of risk which can be used to reduce their long-term risk of complications.

In-human testing of a non-invasive continuous low-energy microwave glucose sensor with advanced machine learning capabilities

Continuous glucose monitoring schemes that avoid finger pricking are of utmost importance to enhance the comfort and lifestyle of diabetic patients. To this aim, we propose a microwave planar sensing platform as a potent sensing technology that extends its applications to biomedical analytes. In this paper, a compact planar resonator-based sensor is introduced for noncontact sensing of glucose. Furthermore, in vivo and in-vitro tests using a microfluidic channel system and in clinical trial settings demonstrate its reliable operation. The proposed sensor offers real-time response and a high linear correlation (R2 ∼ 0.913) between the measured sensor response and the blood glucose level (GL). The sensor is also enhanced with machine learning to predict the variation of body glucose levels for non-diabetic and diabetic patients. This addition is instrumental in triggering preemptive measures in cases of unusual glucose level trends. In addition, it allows for the detection of common artifacts of the sensor as anomalies so that they can be removed from the measured data. The proposed system is designed to noninvasively monitor interstitial glucose levels in humans, introducing the opportunity to create a customized wearable apparatus with the ability to learn.

Comparative Analysis of Predictive Interstitial Glucose Level Classification Models

BACKGROUND

New methods of continuous glucose monitoring (CGM) provide real-time alerts for hypoglycemia, hyperglycemia, and rapid fluctuations of glucose levels, thereby improving glycemic control, which is especially crucial during meals and physical activity. However, complex CGM systems pose challenges for individuals with diabetes and healthcare professionals, particularly when interpreting rapid glucose level changes, dealing with sensor delays (approximately a 10 min difference between interstitial and plasma glucose readings), and addressing potential malfunctions. The development of advanced predictive glucose level classification models becomes imperative for optimizing insulin dosing and managing daily activities.

METHODS

The aim of this study was to investigate the efficacy of three different predictive models for the glucose level classification: (1) an autoregressive integrated moving average model (ARIMA), (2) logistic regression, and (3) long short-term memory networks (LSTM). The performance of these models was evaluated in predicting hypoglycemia (<70 mg/dL), euglycemia (70-180 mg/dL), and hyperglycemia (>180 mg/dL) classes 15 min and 1 h ahead. More specifically, the confusion matrices were obtained and metrics such as precision, recall, and accuracy were computed for each model at each predictive horizon.

RESULTS

As expected, ARIMA underperformed the other models in predicting hyper- and hypoglycemia classes for both the 15 min and 1 h horizons. For the 15 min forecast horizon, the performance of logistic regression was the highest of all the models for all glycemia classes, with recall rates of 96% for hyper, 91% for norm, and 98% for hypoglycemia. For the 1 h forecast horizon, the LSTM model turned out to be the best for hyper- and hypoglycemia classes, achieving recall values of 85% and 87% respectively.

CONCLUSIONS

Our findings suggest that different models may have varying strengths and weaknesses in predicting glucose level classes, and the choice of model should be carefully considered based on the specific requirements and context of the clinical application. The logistic regression model proved to be more accurate for the next 15 min, particularly in predicting hypoglycemia. However, the LSTM model outperformed logistic regression in predicting glucose level class for the next hour. Future research could explore hybrid models or ensemble approaches that combine the strengths of multiple models to further enhance the accuracy and reliability of glucose predictions.

Personalized Prediction of Change in Fasting Blood Glucose Following Basal Insulin Adjustment in People With Type 2 Diabetes: A Proof-of-Concept Study

AIMS

For people with type 2 diabetes treated with basal insulin, suboptimal glycemic control due to clinical inertia is a common issue. Determining the optimal basal insulin dose can be difficult, as it varies between individuals. Thus, insulin titration can be slow and cautious which may lead to treatment fatigue and non-adherence. A model that predicts changes in fasting blood glucose (FBG) after adjusting basal insulin dose may lead to more optimal titration, reducing some of these challenges.

OBJECTIVE

To predict the change in FBG following adjustment of basal insulin in people with type 2 diabetes using a machine learning framework.

METHODS

A multiple linear regression model was developed based on 786 adults with type 2 diabetes. Data were divided into training (80%) and testing (20%) sets using a ranking approach. Forward feature selection and fivefold cross-validation were used to select features.

RESULTS

Participants had a mean age of approximately 59 years, a mean duration of diabetes of 12 years, and a mean HbA1c at screening of 65 mmol/mol (8.1%). Chosen features were FBG at week 2, basal insulin dose adjustment from week 2 to 7, trial site, hemoglobin level, and alkaline phosphatase level. The model achieved a relative absolute error of 0.67, a Pearson correlation coefficient of 0.74, and a coefficient of determination of 0.55.

CONCLUSIONS

A model using FBG, insulin doses, and blood samples can predict a five-week change in FBG after adjusting the basal insulin dose in people with type 2 diabetes. Implementation of such a model can potentially help optimize titration and improve glycemic control.

A Stacked Long Short-Term Memory Approach for Predictive Blood Glucose Monitoring in Women with Gestational Diabetes Mellitus

Gestational diabetes mellitus (GDM) is a subtype of diabetes that develops during pregnancy. Managing blood glucose (BG) within the healthy physiological range can reduce clinical complications for women with gestational diabetes. The objectives of this study are to (1) develop benchmark glucose prediction models with long short-term memory (LSTM) recurrent neural network models using time-series data collected from the GDm-Health platform, (2) compare the prediction accuracy with published results, and (3) suggest an optimized clinical review schedule with the potential to reduce the overall number of blood tests for mothers with stable and within-range glucose measurements. A total of 190,396 BG readings from 1110 patients were used for model development, validation and testing under three different prediction schemes: 7 days of BG readings to predict the next 7 or 14 days and 14 days to predict 14 days. Our results show that the optimized BG schedule based on a 7-day observational window to predict the BG of the next 14 days achieved the accuracies of the root mean square error (RMSE) = 0.958 ± 0.007, 0.876 ± 0.003, 0.898 ± 0.003, 0.622 ± 0.003, 0.814 ± 0.009 and 0.845 ± 0.005 for the after-breakfast, after-lunch, after-dinner, before-breakfast, before-lunch and before-dinner predictions, respectively. This is the first machine learning study that suggested an optimized blood glucose monitoring frequency, which is 7 days to monitor the next 14 days based on the accuracy of blood glucose prediction. Moreover, the accuracy of our proposed model based on the fingerstick blood glucose test is on par with the prediction accuracies compared with the benchmark performance of one-hour prediction models using continuous glucose monitoring (CGM) readings. In conclusion, the stacked LSTM model is a promising approach for capturing the patterns in time-series data, resulting in accurate predictions of BG levels. Using a deep learning model with routine fingerstick glucose collection is a promising, predictable and low-cost solution for BG monitoring for women with gestational diabetes.

Quantifying postprandial glucose responses using a hybrid modeling approach: Combining mechanistic and data-driven models in The Maastricht Study

Computational models of human glucose homeostasis can provide insight into the physiological processes underlying the observed inter-individual variability in glucose regulation. Modelling approaches ranging from "bottom-up" mechanistic models to "top-down" data-driven techniques have been applied to untangle the complex interactions underlying progressive disturbances in glucose homeostasis. While both approaches offer distinct benefits, a combined approach taking the best of both worlds has yet to be explored. Here, we propose a sequential combination of a mechanistic and a data-driven modeling approach to quantify individuals' glucose and insulin responses to an oral glucose tolerance test, using cross sectional data from 2968 individuals from a large observational prospective population-based cohort, the Maastricht Study. The best predictive performance, measured by R2 and mean squared error of prediction, was achieved with personalized mechanistic models alone. The addition of a data-driven model did not improve predictive performance. The personalized mechanistic models consistently outperformed the data-driven and the combined model approaches, demonstrating the strength and suitability of bottom-up mechanistic models in describing the dynamic glucose and insulin response to oral glucose tolerance tests.

Long-Term Glucose Forecasting for Open-Source Automated Insulin Delivery Systems: A Machine Learning Study with Real-World Variability Analysis

Glucose forecasting serves as a backbone for several healthcare applications, including real-time insulin dosing in people with diabetes and physical activity optimization. This paper presents a study on the use of machine learning (ML) and deep learning (DL) methods for predicting glucose variability (GV) in individuals with open-source automated insulin delivery systems (AID). A three-stage experimental framework is employed in this work to systematically implement and evaluate ML/DL methods on a large-scale diabetes dataset collected from individuals with open-source AID. The first stage involves data collection, the second stage involves data preparation and exploratory analysis, and the third stage involves developing, fine-tuning, and evaluating ML/DL models. The performance and resource costs of the models are evaluated alongside relative and proportional errors for 17 GV metrics. Evaluation of fine-tuned ML/DL models shows considerable accuracy in glucose forecasting and variability analysis up to 48 h in advance. The average MAE ranges from 2.50 mg/dL for long short-term memory models (LSTM) to 4.94 mg/dL for autoregressive integrated moving average (ARIMA) models, and the RMSE ranges from 3.7 mg/dL for LSTM to 7.67 mg/dL for ARIMA. Model execution time is proportional to the amount of data used for training, with long short-term memory models having the lowest execution time but the highest memory consumption compared to other models. This work successfully incorporates the use of appropriate programming frameworks, concurrency-enhancing tools, and resource and storage cost estimators to encourage the sustainable use of ML/DL in real-world AID systems.

Chinese diabetes datasets for data-driven machine learning

Data of the diabetes mellitus patients is essential in the study of diabetes management, especially when employing the data-driven machine learning methods into the management. To promote and facilitate the research in diabetes management, we have developed the ShanghaiT1DM and ShanghaiT2DM Datasets and made them publicly available for research purposes. This paper describes the datasets, which was acquired on Type 1 (n = 12) and Type 2 (n = 100) diabetic patients in Shanghai, China. The acquisition has been made in real-life conditions. The datasets contain the clinical characteristics, laboratory measurements and medications of the patients. Moreover, the continuous glucose monitoring readings with 3 to 14 days as a period together with the daily dietary information are also provided. The datasets can contribute to the development of data-driven algorithms/models and diabetes monitoring/managing technologies.

Impact of sleep and physical activity habits on real‐life glycaemic variability in patients with type 2 diabetes

The aim of this study was to better characterise whether sleep habits, eating schedule and physical activity in real-life are associated with glycaemic control in patients with type 2 diabetes. A total of 28 patients (aged 60 years [58; 66], 54% female) with type 2 diabetes treated with basal-bolus insulin therapy administered by insulin pumps were analysed. Glycaemic data measured by Flash Glucose Monitor System, physical activity and sleep data measured by accelerometer, and meal schedules were simultaneously collated with insulin pump administration data, for 7 days in real-life. Their respective impact on the time spent in target, in hypoglycaemia, in hyperglycaemia and on glycaemic variability was evaluated. Multiple regressions showed that the total daily dose of meal boluses of insulin was inversely associated with the coefficient of variation (CV; coefficient β = -0.073; 95% confidence interval: -0.130, -0.015; p = 0.016), as well as sleep duration. The higher the sleep duration, the lower the glycaemic variability (coefficient β = -0.012; 95% confidence interval: -0.023, -0.002; p = 0.027). The mean 7 days physical activity of the subjects was very low and was not associated with glycaemic control on the 7 days mean values. However, days with at least 1 hr spent in physical activity higher than 1.5 METs were associated with less glycaemic variability that same day. This real-life observation highlights the importance of sufficient sleep duration and regular physical activity to lessen the glycaemic variability of patients with type 2 diabetes.

A prospective study of the relationships between movement and glycemic control during day and night in pregnancy

Both disturbed sleep and lack of exercise can disrupt metabolism in pregnancy. Accelerometery was used to objectively assess movement during waking (physical activity) and movement during sleeping (sleep disturbance) periods and evaluated relationships with continuous blood glucose variation during pregnancy. Data was analysed prospectively. 15-women without pre-existing diabetes mellitus wore continuous glucose monitors and triaxial accelerometers from February through June 2018 in Sweden. The relationships between physical activity and sleep disturbance with blood glucose rate of change were assessed. An interaction term was fitted to determine difference in the relationship between movement and glucose variation, conditional on waking/sleeping. Total movement was inversely related to glucose rate of change (p < 0.001, 95% CI (− 0.037, − 0.026)). Stratified analyses showed total physical activity was inversely related to glucose rate of change (p < 0.001, 95% CI (− 0.040, − 0.028)), whereas sleep disturbance was not related to glucose rate of change (p = 0.07, 95% CI (< − 0.001, 0.013)). The interaction term was positively related to glucose rate of change (p < 0.001, 95% CI (0.029, 0.047)). This study provides temporal evidence of a relationship between total movement and glycemic control in pregnancy, which is conditional on time of day. Movement is beneficially related with glycemic control while awake, but not during sleep.