Distribution‐Aware Edge AI for Real‐Time Abnormality Detection in Wearable Body Area Networks

Wearable body area networks (WBANs) enable continuous physiological monitoring, but reliable on‐device abnormality screening remains constrained by noisy multimodal streams, missing observations, scarce abnormal labels, and limited computing resources. This letter proposes a distribution‐aware lightweight edge AI framework for WBAN‐assisted preliminary abnormality screening rather than clinical diagnosis. The method models normal physiological windows as a high‐density manifold and identifies abnormal windows as out‐of‐distribution or near‐distribution deviations. It integrates mask‐aware feature construction, transformed‐prior variational modeling, latent‐space anomaly enhancement, and INT8 edge inference. Experiments on a real measured WBAN dataset collected from 24 human participants under controlled wearable‐monitoring conditions show 95.9% sensitivity, 95.2% specificity, 94.8% F1‐score, and 0.982 ROC‐AUC while reducing latency and memory cost compared with six baselines reimplemented on the same dataset. Additional clarification is provided on sensor sampling, participant‐level physiological variability, abnormality categories, and comparison protocols.