What does this research mean for the field?

A multimodal machine learning model integrating continuous glucose monitoring with wearable, dietary, and medication data provides more accurate 1- to 3-hour glucose forecasts in adults with Type 2 Diabetes compared to CGM-only models. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

This research aims to develop an advanced model for short-term glucose prediction by integrating various data sources.

June 7, 2026

1956-P: Multimodal Machine Learning for Real-Time Glucose Prediction in Adults with Diabetes

Key Points

This research aims to develop an advanced model for short-term glucose prediction by integrating various data sources.
Prospective data collection from 73 adults with Type 2 Diabetes, utilizing CGM, wearable metrics, dietary logs, and medication data.
Training a NeuralProphet model with hyperparameter tuning using a 60/20/20 train-validation-test split for optimal performance.
Forecasting glucose levels 1-3 hours ahead, validating against traditional CGM-only predictions.
The multimodal model achieved a 1-hour RMSE of 14.47 mg/dL, with a range of errors from 8.2 to 22.38 mg/dL.
Predictive accuracy was maintained with 2-hour and 3-hour RMSEs of 15.03 mg/dL and 15.44 mg/dL, respectively.
Feature importance analysis identified key variables like physical activity and dietary intake that improved prediction accuracy.

Abstract

Introduction and Objective: Accurate short-term glucose prediction is key to preventing hypoglycemia and post-meal spikes, but CGM-only models have notable gaps. This study developed a multimodal ML model within the Lillia Care digital diabetes management that integrates wearable, dietary, and medication data to forecast glucose 60-180 minutes ahead in adults with Type 2 Diabetes. Methods: Data were prospectively collected from 73 adults (64 M) with T2DM (mean age 50 years; BMI 28.9; HbA1c 7.5%; diabetes duration 12.3 years), each contributing complete CGM cycles. Alongside CGM, body composition parameters were recorded, and over a 4 week period data were paired with smartwatch metrics (heart rate, steps, sleep), food logs, and medication adherence. A NeuralProphet deep time-series model was trained with extensive hyperparameter tuning using a 60/20/20 train-validation-test split to ensure stable and accurate 1-hour glucose forecasts. Results: The multimodal model showed better performance than CGM-only baselines, delivering strong and consistent accuracy across all prediction horizons. For 1-hour forecasts, the model achieved an RMSE of 14.47+3.18 mg/dL, with individual errors ranging from 8.2 to 22.38 mg/dL. Predictive reliability remained stable as the horizon extended, with 2-hour and 3-hour RMSEs of 15.03+3.40 mg/dL (range 7.1-22.55) and 15.44+3.35 mg/dL (range 8.4-23.04), respectively. These well-bounded error distributions indicate robust performance across diverse glycemic profiles. Feature-importance analysis further showed that variables such as time since last steps, cumulative calories burned, daily stress levels, and oxygen-desaturation events significantly enhanced forecasting accuracy, underscoring the value of incorporating behavioral and physiological context into glucose prediction models. Conclusion: Integrating CGM with wearable, dietary, and medication signals enables reliable, clinically meaningful short-term glucose forecasting in people with T2DM, maintaining accuracy even as prediction windows extend to 3 hours. Disclosure G. Talbot-Lachance: None. P. Samant: None. A.K. Joshi: None. A. Essat: None. M. Mundra: None. P. Nerkar: None. S. Chakrabarty: None. R.A. Saad: None. Funding QRDI

Perguntar à IA

Bookmark

Perguntar à IA

Bookmark

1956-P: Multimodal Machine Learning for Real-Time Glucose Prediction in Adults with Diabetes

Key Points

Abstract

Cite This Study