Optimal vowels measurements for Obstructive Sleep Apnea Detection Using Speech Signals

Abstract

In Obstructive Sleep Apnea (OSA) detection using speech signal during awake, traditional speech-based methods adopt speech features such as Formants and MFCC. As the OSA voice is pathological, the parameters for normal speech processing/recognition is not optimal for the detection. In this paper, we investigate the effects of Linear Predictive coder (LPC) order to the OSA detection. We further propose to adopt dual LPC for feature extractions. In the simulation using 66 OSA patients' voice signals, we achieve the best accuracy of 95.45% and 86.36% with the proposed parameters using quadratic discriminant analysis classifier for multi-class (4 levels) OSA severity classification using resubstitution and leave-one-out method respectively. As compared to the typical parameters setting, the improvement of resubstitution and leave-one-out are 6.06% and 9.09% respectively.