A Tiny Deep Learning Model for Sleep Apnea Detection Based on ECG Signals

Sleep apnea is a prevalent and often underdiagnosed sleep disorder characterized by repeated interruptions in breathing during sleep, which can lead to serious health complications such as hypertension, cardiovascular disease, and daytime fatigue. Traditional methods for diagnosing sleep apnea rely on polysomnography, which is expensive, time-consuming, and requires a clinical setting. To address these limitations, recent approaches have explored automated sleep apnea detection using electrocardiogram (ECG) signals. However, existing approaches rely on ECG feature extraction, such as RR intervals, before apnea event classification. This paper presents a Tiny Deep Learning (TinyDL) approach for detecting sleep apnea on resource-constrained embedded devices by using raw ECG signals. Unlike previous studies that rely on ECG signal preprocessing to extract features such as RR intervals before classification, we propose a lightweight convolutional neural network (CNN) that directly processes raw ECG data, eliminating the need for feature extraction. The model is trained on the PhysioNET Apnea-ECG dataset and optimized for deployment on an STM32 Nucleo-F401RE board. Model quantization is employed to reduce memory footprint and computational complexity while maintaining classification accuracy. Experimental results demonstrate the feasibility of deploying DL-based sleep apnea detection on low-power embedded systems, achieving accuracy comparable with that of TinyML models requiring feature extraction.