Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SSM-CGM: Interpretable State-Space Forecasting Model of Continuous Glucose Monitoring for Personalized Diabetes Management

Shakson Isaac, Yentl Collin, Chirag Patel

TL;DR

This work tackles the need for interpretable near-term CGM forecasting to support personalized diabetes management by introducing SSM-CGM, a Mamba-based neural state-space model that fuses CGM with wearable signals from the AI-READI cohort. The model achieves improved short-term accuracy over a Temporal Fusion Transformer baseline and provides interpretability through variable selection networks and hidden temporal attention, while enabling counterfactual forecasts by intervening on planned future covariates like heart rate and respiration via a forward-pass implementation of the sequential g-formula under standard causal assumptions. Empirically, SSM-CGM yields MAE improvements on horizons up to 1 hour and reveals distinct HR- and RR-linked pathways influencing glucose, with daytime HR perturbations generally improving glucose tolerance and nighttime RR perturbations associating with adverse metabolic markers. The approach advances actionable, physiologically grounded guidance for personalized diabetes management, though it carries limitations such as imperfect meal annotations and substantial compute requirements, motivating future work on streaming deployment, multivariate counterfactuals, and integration with reinforcement-learning policies.

Abstract

Continuous glucose monitoring (CGM) generates dense data streams critical for diabetes management, but most used forecasting models lack interpretability for clinical use. We present SSM-CGM, a Mamba-based neural state-space forecasting model that integrates CGM and wearable activity signals from the AI-READI cohort. SSM-CGM improves short-term accuracy over a Temporal Fusion Transformer baseline, adds interpretability through variable selection and temporal attribution, and enables counterfactual forecasts simulating how planned changes in physiological signals (e.g., heart rate, respiration) affect near-term glucose. Together, these features make SSM-CGM an interpretable, physiologically grounded framework for personalized diabetes management.

SSM-CGM: Interpretable State-Space Forecasting Model of Continuous Glucose Monitoring for Personalized Diabetes Management

TL;DR

This work tackles the need for interpretable near-term CGM forecasting to support personalized diabetes management by introducing SSM-CGM, a Mamba-based neural state-space model that fuses CGM with wearable signals from the AI-READI cohort. The model achieves improved short-term accuracy over a Temporal Fusion Transformer baseline and provides interpretability through variable selection networks and hidden temporal attention, while enabling counterfactual forecasts by intervening on planned future covariates like heart rate and respiration via a forward-pass implementation of the sequential g-formula under standard causal assumptions. Empirically, SSM-CGM yields MAE improvements on horizons up to 1 hour and reveals distinct HR- and RR-linked pathways influencing glucose, with daytime HR perturbations generally improving glucose tolerance and nighttime RR perturbations associating with adverse metabolic markers. The approach advances actionable, physiologically grounded guidance for personalized diabetes management, though it carries limitations such as imperfect meal annotations and substantial compute requirements, motivating future work on streaming deployment, multivariate counterfactuals, and integration with reinforcement-learning policies.

Abstract

Continuous glucose monitoring (CGM) generates dense data streams critical for diabetes management, but most used forecasting models lack interpretability for clinical use. We present SSM-CGM, a Mamba-based neural state-space forecasting model that integrates CGM and wearable activity signals from the AI-READI cohort. SSM-CGM improves short-term accuracy over a Temporal Fusion Transformer baseline, adds interpretability through variable selection and temporal attribution, and enables counterfactual forecasts simulating how planned changes in physiological signals (e.g., heart rate, respiration) affect near-term glucose. Together, these features make SSM-CGM an interpretable, physiologically grounded framework for personalized diabetes management.

Paper Structure

This paper contains 54 sections, 28 equations, 14 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Dataset.
  4. Meal Detection.
  5. Forecasting Objective.
  6. SSM-CGM architecture.
  7. Interpretability.
  8. Counterfactual Forecasting.
  9. Results
  10. Benchmarking.
  11. Interpretability.
  12. Counterfactuals.
  13. Conclusion
  14. Technical Appendices and Supplementary Material
  15. Datasets
  16. ...and 39 more sections

Figures (14)

  • Figure 1: SSM-CGM Architecture.
  • Figure 2: Interpretability of SSM-CGM. (A) VSN average features importance (B) Past influence on predictions (average, individual) (C) Head average self-attention map of Light Mamba (last 60-step window, one individual)
  • Figure 3: Counterfactual effects and associations. (A) Estimated change in glucose (mg/dL) from a +2 SD perturbation to heart rate (HR) or respiration rate (RR) over a 1-h horizon, stratified by diabetes status (mean $\pm$ 95% CI). (B) Pearson correlations between the counterfactual effects ($\Delta$G-HR, $\Delta$G-RR) and clinical laboratory measures. Stars indicate FDR-adjusted significance (p): * $<\!0.05$, ** $<\!0.01$, *** $<\!0.001$.
  • Figure 4: Meal detection model architecture
  • Figure 5: Predicted meal flag probability on test set window
  • ...and 9 more figures