Intelligent Audio-based Signal Processing for Automatic Detection of Obstructive Sleep Apnea

Abstract

In this work, we propose a new acoustic-based method for the screening of obstructive sleep apnea (OSA) which employs breath and respiratory sounds recorded using an smartphone. In our proposed method, a set of acoustic parameters aimed at characterizing the respiratory and snore patterns of the patient are extracted from the sleep sound recordings. These include Snore Rate Variability (SVR), SET (Snore Energy Trench) parameters and Snore-to-Snore Intervals (SSI). Data fusion techniques were investigated, as well as the demographic characteristics of the subjects, which were assessed from the apneahypopnea index (AHI) estimated from all nightly recordings. Subsequently, a multiclass classification of each patient according to their OSA level was performed using several classifier methods, namely TabTransfomer, Support Vector Machines (SVM) and XGBoost. Real recordings made during home sleep apnea tests were used to develop and evaluate the proposed system. The TabTransformer-based classifier obtained the best results in estimating AHI severity, achieving a specificity of 0.65, accuracy rate of 0.65 and an sensitivity of 0.64, with an AUC score of 0.78. This offers the prospect of at-home screening for OSA.