Wearable Edge AI System for Real-Time Obstructive Sleep Apnea Detection and Severity Classification

Abstract

Obstructive sleep apnea (OSA) is a sleep disorder characterized by repeated upper airway obstruction, which, if left undetected, contributes to hypertension, cardiovascular disease, stroke, and cognitive decline. This work presents a real-time, single-lead ECG-based system for automated OSA detection and severity classification. The system is evaluated on the OSASUD dataset, comprising 30 stroke-unit patients with 961,357 annotated seconds of single-lead ECG at 80 Hz, under leave-one-out cross-validation. A comprehensive feature set incorporating heart rate variability (HRV) metrics, nonlinear dynamical features, and morphological statistics is processed by a transformer-based deep learning (DL) model operating on 3-minute ECG windows with a 30-second stride, enabling per-second apnea event detection. The hardware subsystem includes biopotential signal acquisition via surface electrodes, followed by analog amplification, bandpass filtering, and high-resolution analog-to-digital conversion (ADC), with real-time signal processing executed on an embedded singleboard computer (SBC). The system achieves 96.7% binary classification accuracy of presence or absence of apnea and 80.0% four-class severity classification accuracy across Normal, Mild, Moderate, and Severe OSA categories via Apnea-Hypopnea Index (AHI) estimation. These results demonstrate that low-cost, single-lead ECG screening can be a clinically viable alternative to traditional overnight sleep studies, with the potential to reduce the cardiovascular and neurological burden of undiagnosed OSA.