Obstructive Sleep Apnea Detection Using Speech Signals with High Frequency Components

Abstract

In this study, the Obstructive Sleep Apnea (OSA) detection using speech signals during awake is considered. Traditional speech based OSA detection methods adopt traditional features (Formants, MFCC, etc.) on normal speech frequency range (<6kHz). However, it ignores the signal components outside this range that usually appear in pathological voices. In this paper, higher order traditional speech features (with more high frequency components) are adopted for detection. To better characterize OSA patients’ speech, a high frequency feature set is proposed. It consists of the traditional speech features with optimized parameters and a new proposed feature: High frequency energy. Principal Component Analysis (PCA) based Sequence Forward Feature Selection (PCASFFS) are adopted as feature selection. In the simulation using 66 OSA patients’ speech signals, it achieves an accuracy of 84.85% for multi-class (4 levels) detection with the proposed high frequency feature set using quadratic discriminant analysis classifier (QDA).

Obstructive sleep apnea detection during wakefulness: a comprehensive methodological review

“…More recently, a study focused on studying the use of higher frequency range (> 6 kHz) components of the speech signals and their effect on the detection of OSA during wakefulness [ 85 ]. The authors extracted traditional higher-order speech features but added higher-frequency components of the speech signals during awake for a better characterization of OSA patients’ speech.…”

“…The best feature selection classification pairs were found to be GA with Bayesian classification, GA with SVM, and LDP with MLP. Then by combining these by a majority vote yielding 82.9% accuracy, 81.49% sensitivity, and 84.69% specificity in detecting severe OSA for validation [ 52 ] 129 females (AHI 0 to 108.4) Formant frequencies showed a weak correlation (up to − 0.26) with AHI [ 85 ] 66 Subjects (31 subjects for AHI < 5; 13 subjects for 5 ≤ AHI < 15;10 subjects for 15 ≤ AHI < 30; and 12 subjects for AHI ≥ 30) The new optimized feature for the whole frequency range achieves an accuracy of 84.85% for multi-class OSA detection using the QDA classifier [ 86 ] 40 subjects (20 healthy and 20 OSA with AHI > 9) The nonlinear characteristics of vocal tract changes in subjects with OSA can be used as discriminant features, especially for consonants. Vowels feature only were 83.5% using KNN; consonants only were 96% using SVM; and 85.5% accuracy using a blind test set with six consonant features …”

A modified second-order synchroextracting transform

“…(8) where represent the first order time and frequency derivatives of NSTFT respectively. Then its modulation operator [14][15][16] can be expressed as (9)…”