Feature Normalization for Improving the Performance of Sleep Apnea Detection System

Sleep apnea is a common sleep disorder characterized by intermittent cessation of breath during sleep. Diagnosis of this disorder requires, prolonged and expensive sleep study test. The unavailability of such high-end diagnosis setup in rural areas, makes such disorders undiagonised. This paper introduces a low cost system which can detect sleep apnea by a rural health worker independent of sleep state, using electrocardiography (ECG) and respiratory effort signal (RES). The baseline-system uses statistical features derived from heart rate variability (HRV) and respiratory rate variability (RRV) data, with SVM used as a backend classifier. The feature vectors extracted from ECG and RES, carries patient and stage specific variations that does not contain information about apnea condition. Any effort to minimize these variations on the feature vectors can improve the performance. We explore two approaches to minimize these variations in the input features to improve the system performance. In the first approach we used nuisance attribute projection (NAP) in which we consider these variations as nuisance, and removed the components that are adversely effecting the performance of the classifier. Individual systems that are patient and stage independent were developed up on performing NAP algorithm and we got 81.25% sensitivity, 68.75% specificity and an overall accuracy of 75% absolute. Further using covariance normalization (CVN) we obtained an improvement of 16% absolute in the overall accuracy compared to the baseline-system. We further combined the NAP and CVN, and did not find any encouraging results.