Desing and implementation of a computer aided detection system for diagnosis of sleep apnea

In the thesis study, a computer aided diagnostic system capable of diagnosing sleep apnea is proposed and validated on the PhysioNet Apnea-ECG database. In the preprocessing phase of the proposed system, electrocardiogram signals in the database is subjected to z-score normalization, bandpass filtered and divided into one-minute segments. Then, in the feature extraction stage, one-minute segments are described with seven different feature groups in both the spatial plane and the frequency plane. In addition, feature selection is performed by applying principal component analysis, common vector approach and distinguished common vector approach methods to each feature group. In the classification stage, 2- and 3-class diagnostic studies are performed using the logistic linear classifier, linear differential classifier, fisher linear differential analysis, bayes classifier, k-neighbor classifier. As a result, In the 2-class diagnostic study, it is determined that maximum 72.29% accuracy is achieved by using heart rate variability features, and 67.00% accuracy is achieved after size reduction of heart rate variability features by principal component analysis method. In the 3-class diagnostic study, in which the patient is classified as borderline/healthy, it is determined that the maximum accuracy is 68.67% using heart rate variability features, and 68.67% accuracy is achieved after size reduction with principal component analysis of hybrid features.