Detection and Classification of Obstructive Sleep Apnea Using Audio Spectrogram Analysis

• DOI: 10.3390/electronics13132567
• Publication Date: 2024-07-01T12:17:29Z
• Authors: Salvatore Serrano; Luca Patanè; Omar Serghini; Marco Scarpa
• Keywords: Spectrogram; Sleep; Bioacoustics

Sleep disorders are steadily increasing in the population and can significantly affect daily life. Low-cost noninvasive systems that assist diagnostic process will become increasingly widespread coming years. This work aims to investigate compare performance of machine learning-based classifiers for identification obstructive sleep apnea–hypopnea (OSAH) events, including apnea/non-apnea status classification, index (AHI) prediction, AHI severity classification. The dataset considered contains recordings from 192 patients. It is derived a recently released which contains, amongst others, audio signals recorded with an ambient microphone placed ∼1 m above studied subjects apnea/hypopnea accurate events annotations performed by specialized medical doctors. We employ mel spectrogram images extracted environmental as input machine-learning-based classifier proposed approach involves stacked model utilizes combination pretrained VGG-like classification (VGGish) network bidirectional long short-term memory (bi-LSTM) network. Performance analysis was conducted using 5-fold cross-validation approach, leaving out patients used training validation models testing step. Comparative evaluations presented methods literature demonstrate advantages approach. architecture be useful tool supporting OSAHS diagnoses means low-cost devices such smartphones.