Generative AI-powered multimodal physiological sensors for personalized edge health monitoring

Abstract

This study focuses on personalized edge inference for multi-source physiological time-series signals instead of imaging data. The main contribution is a lightweight model for personalized, calibrated prediction that integrates a generative prior and flexible latent structures for denoising and imputation for short segments of ECG/PPG followed by prediction. Supporting components include federated learning, privacy-preservation mechanisms, and TinyML. Experiments were performed using the public MIMIC-III Waveform Database. The analyzed dataset comprised 10,282 patients and 22,317 patient-disjoint, record-level windows for binary (abnormal/risk) event classification. Positive labels were assigned to the physiological abnormality window intervals, and negative labels were assigned to the stable, non-event interval classes. Separating the data for training, validation, and testing was performed at the patient level using a disjoint 70%/15%/15% split of the patient datasets (7,197/1,542/1,543 patients), with approximately 30% positive-class examples, and ensured that no windows occurred across multiple partitions. Within this primary experimental setting, GenWear achieved an AUROC of 0.956 ± 0.006, AUPRC of 0.735 ± 0.026, F1 of 0.842 ± 0.016, and an expected calibration error (ECE) of 0.017; the best edge baseline achieved an AUROC of 0.951 ± 0.008. An INT8 implementation of GenWear on a GAP9 chip achieved 45 Hz and had a latency of 22/34 ms for p50/p95 with an inference energy of 1.45 mJ. Several studies assessed the robustness of the model against missing data, noise, and domain shift. Given MIMIC-III contains ICU data, the data should only be considered a proof of concept for physiological sensing with wearable devices. A validation study on a real-world ambulatory dataset is still required.