Deep Learning and Machine Learning Fusion for Sleep Apnea Diagnosis and Classification

Sleep apnea is a widespread sleep disorder with life threatening conditions like cardiovascular diseases as well as cognitive impairments. Given that accurate classification of its types (obstructive, central, and mixed) is essential to treatment, a thorough physical examination is highly essential during the initial visit. In this study, we propose to design a hybrid model consisting of Convolutional Neural Networks (CNNs) for feature extraction and Support Vector Machines (SVMs) for classification using physiological data like Electrocardiograms (ECG), pulse oximetry (SpO2), etc. The overall accuracy of the model is 93.0% and precision, recall, and F1-score were 92.4%, 93.0%, and 92.7%, respectively. Its strong discriminatory ability is also validated by the ROC-AUC score of 0.960. Obstructive apnea (95.3%) has the highest accuracy, central apnea (92.8%) has the second highest and mixed apnea (90.7<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup>) has the third highest accuracy. A comparative analysis demonstrates that the hybrid approach further decreases misclassification rates and improves generalization to apnea types compared to standalone classifiers. We perform feature importance analysis which reveals SpO2 variations and heart rate variability (HRV) as the two most important predictors that characterize desaturations and cardiovascular patterns indicative of apnea events. The model shows strong performance, but is limited in generalizability to unseen populations and involves a computational overhead. We propose a CNN-SVM hybrid model which we find to be reliable and cost effective to perform automated sleep apnea classification. This suggests that by utilizing physiological data, this approach can aid in the development of real time diagnostic tools, lower the dependence on polysomnography that is both time and resource intensive, and enhance personalized treatment strategies.