Sleep Apnea Detection using Multimodal Physiological Signals

Sleep apnea, characterized by repeated interruptions in breathing during sleep, is a highly prevalent disorder affecting individuals between the ages of 30 to 70. This study proposes an automated approach to sleep apnea detection using physiological signals—electrocardiography (ECG), electroencephalography (EEG), and peripheral oxygen saturation (SpO$_2$). We apply both classical machine learning methods, including Random Forest classifier, and deep learning technique Network to identify apnea events. Discriminative features are extracted from ECG, SpO$_2$ and EECG signals. Moreover, EEG and ECG data are converted into spectrograms to capture stage-specific frequency patterns. For deep learning classification, recurrence plots and spectrograms are used as input to a ResNet-18 convolutional neural network. The models achieve 83$\%$ accuracy in binary classification of apnea versus non-apnea events. This work highlights the potential of combining traditional machine learning with deep neural networks to develop an accessible, non-invasive diagnostic tool for sleep apnea using data from wearable sensors.