A New Approach to Predict Potential Sleep Disorder based on Fractal Analysis from Non-overlapping Single Lead ECG Using Support Vector Machine

Abstract

Sleep disorders are challenging to diagnose. The complexity of records obtained from electrocardiogram (ECG) recordings requires manual inspection by experienced medical practitioners. Meanwhile, ECG records are still widely used to diagnose heart problems during sleep. To resolve the issue, the fractal analysis is a promising means to help identify the characteristics of non-overlapping apnea and non-apnea events based on signal scaling behaviour and QRS wave morphologies. Therefore, we propose a new approach to develop automatic sleep disorder classification to minimalize visual inspection and manual scoring. We employed the monofractal and the multifractal analyses to generate new features such as alpha1, residue1, alpha2, residue2, Dqmin, Dqmax, hqmin, hqmid, hqmax, and hqmaxhqmin. To improve the proposed method's performance, we used the ten new features that have been extracted, which are eventually being used as inputs space to a support vector machine (SVM). Through examining the feature set, we designed an optimum SVM model classifier to explore the usability of patterns to predict potential sleep disorder corresponding to apnea and non-apnea events. Hence, our approach through SVM with radial basis function (RBF) kernel is achieved to have accuracy, sensitivity, specificity of 92.16%, 88.24%, 94.12% respectively.