Automated Classification of Sleep Apnea Types Using CNNs and SVMs on Physiological Data

In this study a hybrid model involving CNNs for feature extraction and SVMs for classification was proposed. A significant improvement over traditional standalone approaches is achieved using the proposed model. Very common and associated with very serious health risks, including cardiovascular diseases, diabetes, and cognitive impairments, sleep apnea is a disorder that occurs when you stop breathing during sleep. Classifying sleep apnea types— Obstructive Sleep Apnea (OSA), Central Sleep Apnea (CSA), and Mixed Sleep Apnea (MSA)–accurately and automatically is important for early diagnosis and treatment. The CNN-SVM model achieves 94.8% accuracy over using standalone CNN (91.2%), SVM (85.3%), LSTM (88.7%) and Random Forest (RF) (86.5%) on the PhysioNet Sleep-EDF dataset. For instance, hybrid model precision, recall and F1-score were 93.5%, 94.2%, 93.8% respectively, compared to CNN (89.5%, 90.1%, 89.8%) and SVM (84.7%, 84.8%, 84.7%). In addition, we show the hybrid model has a high AUC of 0.99 for OSA detection and 0.91 for CSA/MSA classification. Confusion matrix analysis was also carried out to the model’s performance further validating low misclassifications (especially between CSA and MSA). We verified the stability and generalizability of the parameters with training and validation curves, and the model converged within 20 epochs. This study shows how SVM’s robust decision boundaries can be integrated with the feature extraction power of deep learning resulting in a scalable and accurate solution for sleep apnea classification. For real time clinical applications, the hybrid CNN-SVM model has great potential in providing personalized, low cost and efficient diagnosis of sleep apnea disorders.