Multimodal Transformer Glucose Prediction: From Supervised Learning to Non-Invasive Estimation

This article introduces the multimodal Transformer glucose prediction project open-sourced by the Temple University team, covering two subsystems: supervised (glucose_transformer) and non-invasive (noninvasive_glucose). Core innovations include cross-modal attention mechanism, uncertainty quantification, model calibration, etc. It aims to address pain points of traditional continuous glucose monitoring (CGM) devices such as invasiveness and high cost, providing a complete technical path from supervised learning to fully non-invasive estimation.

Glucose monitoring is a core component of diabetes management. Traditional CGM devices require implantable sensors, which have problems such as high cost, poor comfort, and low compliance. Developing glucose estimation methods based on non-invasive physiological signals (e.g., heart rate, ECG, EMG, etc.) is an important direction in the wearable health monitoring field. This project open-sources a complete technical solution from supervised learning to fully non-invasive estimation, including two subsystems: the supervised prediction system glucose_transformer and the non-invasive estimation system noninvasive_glucose, which share the core architecture but serve different scenarios.

Supervised Prediction System (glucose_transformer): Both training and inference use glucose data as one of the input features, with the goal of predicting glucose values 30/60 minutes in the future. It adopts a phased learning path (gradually introducing multimodality from a single heart rate signal).

Non-Invasive Estimation System (noninvasive_glucose): Glucose data is used as the supervision signal during training, but during inference, it does not rely on glucose input at all, estimating current glucose only through physiological signals to simulate real deployment scenarios.

The two systems reflect the technical evolution idea: the supervised system serves as a learning path, introducing concepts like self-attention; the non-invasive system adds constraints, introducing uncertainty quantification and model calibration.

Cross-Modal Attention Fusion Mechanism: Uses heart rate as the query, other modalities (ECG, EMG, cerebral blood flow) as keys and values. EEG uses summary tokens to reduce complexity.

Multi-Scale EEG Processing: Explores three strategies: frequency band division, block processing, and hierarchical processing, balancing information retention and computational efficiency.

Uncertainty Quantification: The non-invasive system outputs glucose estimation values and log variance to quantify prediction uncertainty.

Model Calibration: Implements techniques like temperature scaling to ensure that predicted probabilities match actual observation frequencies, improving the reliability of uncertainty estimation.

The supervised system is trained and evaluated using the OhioT1DM dataset (multi-day monitoring records of type 1 diabetes patients); the non-invasive system uses the PhysioCGM dataset or a synthetic data fallback scheme.

Evaluation metrics: Root Mean Square Error (RMSE), proportion of predictions in Clarke error grid zones A+B. The current submission version of the non-invasive system has an RMSE of 21.81 mg/dL, with 100% of predictions falling in zones A+B (Note: This is a smoke test result, not a complete convergence report).

Phased Learning Path: Five-stage progressive design

• PartA: Only heart rate signal, learning self-attention and positional encoding
• PartB: Introduce ECG and EMG, learning cross-modal fusion
• PartC: Add EEG and cerebral blood flow, mastering multimodal processing
• PartD: Add user conditioning, learning personalized modeling
• Non-invasive system: Remove glucose input constraints, integrate technologies, and introduce uncertainty quantification

Deployment Considerations: Optimized for 6GB VRAM, supporting consumer-grade GPUs; clear code structure, explicit dependencies, simple environment configuration; provides a complete documentation system (theory, architecture, training, etc.).

This project demonstrates the innovative application of multimodal deep learning in the healthcare field. Through strategies such as cross-modal attention, uncertainty quantification, and phased learning, it achieves competitive glucose prediction performance under resource constraints. For researchers and engineers interested in wearable health monitoring, multimodal machine learning, and medical AI deployment, it is an open-source project of great reference value.