Lightweight Binarized Neural Network for Real-Time Sleep Apnea Detection on Edge Hardware

This paper proposes a real-time obstructive sleep apnea (OSA) screening device for adults that employs an ECG patch and a pulse oximeter. The system is designed to classify OSA through the inference of a lightweight binarized neural network (L-BNN) model at the edge. The three-layer L-BNN model was trained separately using open-source OSA data obtained from PhysioNET, and hardware validation was conducted on a TinyML microcontroller with a maximum power consumption rate of ∼10 mW. The Larq library was utilized and modified to binarize the convolutional neural network (CNN) model to suit the dataset and its application. The final two L-BNN models accurately detected OSA, achieving an accuracy rate of ∼89% for both the ECG and SpO2 datasets. The memory utilization for both models showed similar results of 16.1 KB of RAM usage and 69KB of flash usage. The classification time for the inference models on the TinyML microcontroller was approximately 205 ms for ECG data and 186 ms for SpO2 data.