Temporal and Behaviour-Aware Multimodal Modelling for Hour-Ahead Hypoglycaemia Prediction During Ramadan Fasting in...

Area of Science:

• Artificial Intelligence in Medicine
• Endocrinology and Metabolism
• Digital Health

Background:

• Ramadan fasting significantly impacts meal timing, sleep, and activity, increasing hypoglycemia risk in adults with type 1 diabetes (T1D).
• Current continuous glucose monitoring (CGM) systems offer reactive alerts with limited prediction horizons, insufficient for fasting-related disruptions.
• Behavioral and circadian changes during Ramadan necessitate advanced predictive tools for proactive hypoglycemia management in T1D.

Purpose of the Study:

• To evaluate behavior-aware, temporally enriched deep learning models for one-hour-ahead hypoglycemia forecasting in T1D adults during Ramadan.
• To assess the utility of multimodal data, including CGM and wearable signals, for improving predictive accuracy.
• To investigate the impact of explicit temporal and circadian feature engineering on model performance.

Main Methods:

• An observational cohort study in Qatar monitored 33 adults with T1D using CGM and wearables during Ramadan and the post-fasting period.
• Recurrent deep learning models (LSTM, BiLSTM) were trained on aggregated hourly features with a 36-hour lookback window, incorporating temporal and circadian proxies.
• Model performance was evaluated using ROC AUC, precision-recall AUC, recall, and calibration on a naturally imbalanced test set, including cross-phase analysis.

Main Results:

• The best multimodal model achieved an ROC AUC of 0.867 and identified 77% of next-hour hypoglycemic events.
• Temporal feature enrichment and a 36-hour lookback window improved model discrimination and calibration.
• Wearable-derived signals alone provided comparable or superior performance to CGM-only models in some configurations, demonstrating robustness.

Conclusions:

• Behavior-aware, temporally enriched multimodal models can provide calibrated, hour-ahead hypoglycemia risk estimates for T1D adults during Ramadan.
• Explicit modeling of circadian and behavioral dynamics enhances predictive performance under real-world class imbalance.
• Integrating wearable data alongside CGM improves predictive value, supporting proactive hypoglycemia management.