Enhancing Sleep Apnea Detection with Lightweight Models: Interpretable Convolutional Neural Network Deployed on Raspberry Pi

Sleep apnea, or peripheral obstructive sleep apnea (OSA), is a significant health issue characterized by repeated interruptions in breathing during sleep, leading to various cardiovascular complications. Traditional polysomnography (PSG) is costly and disruptive to patients normal sleep patterns. To address these challenges, our study focused on the use of wearable technology and a lightweight deep learning model for real-time OSA detection. The SE-MSCNN convolutional deep learning model, known for its high accuracy and efficiency in analyzing ECG signals, is deployed. In our work, we used Tensor Flow Lite for model optimization, facilitating deployment on resource-constrained devices with the use of the Raspberry Pi 4B, and did a comparative analysis of model inference time on an AMD Core 2 processor with 256 GB of storage. We achieved a segment accuracy of 93%, a reduction in processing time post-optimization, 100% accuracy on testing data and incorporated GradCAM++ for model explainability to ensure transparency in detection.