Interpretable non-invasive glucose monitoring: an attention-based deep learning framework for visualizing hemodynamic correlates in PPG signals

초록

Background Diabetes mellitus necessitates frequent blood glucose monitoring, yet invasive finger-prick methods remain a barrier to adherence. Although Photoplethysmography (PPG) offers a promising noninvasive alternative, current deep learning approaches largely operate as "black boxes," obscuring the physiological features driving their predictions and limiting clinical trust.Methods To address this interpretability gap, we propose a novel Attention - Guided Convolutional - Recurrent Neural Network (AG-CRNN). Unlike standard deep learning models that treat the entire signal uniformly, our architecture integrates a Temporal Attention Mechanism that dynamically assigns importance weights to specific segments of the PPG waveform. This allows the model to learn and emphasize hemodynamic fluctuations associated with glucose variability, while suppressing motion artifacts. We validated the framework using a dataset of synchronized PPG and glucose measurements from 23 subjects.Results The proposed attention-based model achieved a Mean Absolute Error (MAE) of 11.59 mg/dL, significantly outperforming standard LSTM and GRU baselines. Crucially, our Explainable AI (XAI) analysis revealed that the model preferentially focuses on the systolic decay and diastolic decay regions. This visualization aligns with the established physiological evidence linking hyperglycemia to vascular stiffness and wave reflection timing.Conclusion This study presents a high-accuracy glucose estimation framework that provides visual explanations of its decision-making process. By demonstrating that the model leverages genuine morphological correlates rather than spurious noise, we bridged the gap between AI performance and clinical plausibility.

키워드