State-Dependent Soft-Switching Regression for Hypoglycemia-Aware Multi-Horizon Blood Glucose Prediction in Type-1 Diabetes

Technical Summary: State-Dependent Soft-Switching Regression for Hypoglycemia-Aware Multi-Horizon Blood Glucose Prediction

Problem Statement
Accurate blood glucose (BG) prediction is critical for Type 1 Diabetes (T1D) management, yet standard global models optimized for overall error often fail to predict hypoglycemia (BG < 3.9 mmol/L). Because hypoglycemic episodes are rare in free-living data (approximately 2.7% of training epochs in the studied cohort) but carry the highest clinical risk, models minimizing average error tend to "average away" the low-glucose regime. Furthermore, standard activity tracking often fails to capture the time-dependent interactions between behavior, physiology, and glucose outcomes. There is a need for transparent, state-aware models that explicitly model the low-glucose regime rather than treating all glycemic states as a single homogeneous distribution.

Methodology
The study proposes a per-patient, state-dependent soft-switching Ridge regression framework evaluated on the longitudinal multimodal T1D-UOM dataset (15 participants). The methodology involves three primary components:

1. Multimodal Data Fusion and Preprocessing:

   • A pipeline synchronizes Continuous Glucose Monitoring (CGM), physical activity (from Garmin wearables), and nutritional intake into unified 15-minute epochs.
   • Two "merge builds" were tested to assess temporal alignment sensitivity: a tight window (±451 s) preserving sharp state transitions and a wider window (±1800 s) increasing temporal coverage but smoothing transitions.
   • Feature engineering includes cyclic time encoding, metabolic carbohydrate decay, and lagged variants of activity and glucose signals.

2. State-Dependent Soft-Switching Architecture:

   • Instead of a single global model, the framework trains three dedicated Ridge regression sub-models, one for each glycemic state: Hypoglycemia, Normoglycemia, and Hyperglycemia.
   • Soft Switching: Rather than hard classification, the model uses Gaussian soft-membership functions to blend the outputs of the three sub-models based on the current glucose value. This allows for smooth transitions between states.
   • State-Specific Features: Features are assigned to sub-models based on physiological relevance (e.g., glucose rate-of-change and activity features for the hypoglycemia model; carbohydrate intake and absolute glucose levels for the hyperglycemia model).
   • Asymmetric Regularization: The model employs per-state regularization parameters and square-root inverse-frequency sample weighting to address severe class imbalance, prioritizing accuracy in the rare hypoglycemic band.
   • Debiased Evaluation: The protocol ensures no data leakage by performing feature selection and hyperparameter tuning strictly on the training island (using two-fold temporal cross-validation) before evaluating on a strictly held-out test island.

3. Evaluation Protocol:

   • Performance is assessed across five prediction horizons (15, 30, 45, 60, and 120 minutes).
   • Metrics include Root Mean Squared Error (RMSE) and Mean Absolute Relative Difference (MARD), reported both for the full cohort and specifically for the hypoglycemic band.

Key Contributions

• State-Aware Modeling: Introduction of a switching regression architecture that explicitly models distinct physiological dynamics for hypoglycemia, normoglycemia, and hyperglycemia, rather than averaging them.
• Transparent Decision Support: The use of linear Ridge sub-models combined with Gaussian membership functions ensures the model remains interpretable and auditable, unlike many deep learning alternatives.
• Rigorous Debiased Protocol: Implementation of a strict train/test split where feature selection and tuning occur only on the training data, preventing the data leakage common in previous formulations.
• Multimodal Integration: A robust early-fusion pipeline aligning heterogeneous sensor streams (CGM, activity, nutrition) for supervised learning.

Results

• Hypoglycemia-Specific Improvement: On the primary (451 s) merge build, the switching model reduced hypoglycemic-band RMSE by 13.8% at 15 minutes and up to 18.6% at 30–45 minutes compared to a global Ridge baseline. Compared to a persistence baseline, reductions ranged from 11.5% to 38.5%.
• Overall Performance: The switching model matched or exceeded the global Ridge baseline in overall weighted RMSE for 56 of 75 patient–horizon cases under a 2% relative tolerance.
• Trade-off Analysis: The improvement in the hypoglycemic band was achieved with a marginal concession in the normoglycemic band (0.5–1.4% higher RMSE), which the authors characterize as a deliberate reallocation of model capacity toward the high-risk regime.
• Merge Window Sensitivity: The advantage of the switching model was significantly attenuated in the wider (1800 s) merge build. This suggests that for state-aware modeling, the temporal fidelity of multimodal alignment is more critical than the sheer volume of training data, as wider windows blur the sharp state transitions the model relies on.
• Data Dependency: The model's benefits were most pronounced in participants with sufficient hypoglycemic training data. For participants with no hypoglycemic training samples or highly stable glucose profiles, the model degraded gracefully toward the global baseline or persistence.

Significance and Claims
The paper claims that the primary contribution is improved predictive reliability specifically in the rare, high-risk hypoglycemic regime, which is the most clinically consequential state for T1D management. The authors argue that while global models minimize average error, they inherently under-serve the low-glucose tail. By reallocating modeling capacity to this minority regime through a transparent, state-switching architecture, the proposed method offers disproportionate value for clinical decision support and closed-loop correction.

The study emphasizes that the model's success is contingent on the presence of distinct glycemic regimes and sharp transitions in the data. It does not claim universal superiority over all models in all contexts but highlights that for conditions where the cost of error is concentrated in the low-glucose tail, state-aware switching provides a necessary improvement that global models cannot achieve without sacrificing interpretability or introducing data leakage. The authors conclude that this approach advances transparent clinical decision support for T1D care.