Streaming model improves real-time personalized diabetes forecasting
SSMCGM-Stream: A Streaming State-Space Model for Personalized Diabetes Management
Machine LearningArtificial Intelligence
Summary
Managing diabetes requires predicting blood sugar levels continuously and accurately. The authors developed a model that updates its understanding of a person’s glucose in real time as new data arrives, making it fast and personalized. This model also explains its predictions and suggests how different actions might affect glucose levels. Their approach performs as well as complex neural networks but is much faster and can run directly on a user’s device.
What this means in practice
- •For diabetes care teams: Provide personalized, real-time glucose forecasts and explanations to support patient management during daily monitoring.
- •For mobile health app developers: Integrate efficient on-device glucose forecasting with interpretability features to enhance diabetes self-management apps.$Commercial implications: Enables creation of advanced diabetes management apps offering personalized, explainable glucose predictions running efficiently on smartphones.
Tested on one dataset.
Authors
Shakson Isaac, Chirag Patel
Abstract
Continuous glucose monitoring supports diabetes management but requires forecasting models that are accurate, interpretable, and efficient enough for real-time deployment. We present SSMCGM-Stream, a streaming state-space model that maintains a personalized hidden state updated with each incoming glucose reading to provide constant-time forecasting, physiologically structured counterfactual reasoning, and interpretable forecasts through hidden-state attribution. Built on this foundation, SSMCGM-Agent delivers risk-aware, natural-language decision support by translating forecasts, explanations, and counterfactuals into grounded recommendations through a guarded language-model interface. Evaluated on the T1DEXI cohort, SSMCGM-Stream achieves forecasting performance comparable to strong neural baselines while substantially reducing computational cost, enabling continuous on-device diabetes decision support.