Integration of Sound and Image Data for Detection of Sleep Apnea

Abstract

To detect sleep apnea syndrome (SAS), we propose a method that integrates sound and image processing results using overnight sound and video data. In sound processing, snoring sounds are initially extracted using an unsupervised method. The zero-cross ratio (ZCR) and power (PWR) are calculated for the snoring sounds, and the total validations (TV) for ZCR and PWR are calculated using the sound data. Then, a support vector machine (SVM) is used to classify SAS and non-SAS subjects using TV values. In image processing, inter-frame absolute difference values are calculated every 1 s from video data, and multiscale entropy (MSE) is calculated using the time series data. Then, an SVM is employed to classify SAS and non-SAS subjects using two MSE values selected based on the Bhattacharyya distance. To integrate sound and image data, we focused on the number of detected snores. When two results do not match and the number of extracted snores is greater than five per hour, the sound processing result is adopted. When the number of snores is less than five per hour, the image processing result is adopted. The proposed integrated method demonstrates better accuracy than that obtained using only sound or image processing results.