Sleep apnea detection from respiratory sounds using time–Frequency features and neural networks 

Abstract

Background Sleep apnea is a prevalent condition that is frequently misdiagnosed. Currently, the diagnosis of sleep apnoea is dependent on polysomnography (PSG), a costly procedure that restricts screening to a limited number of patients. While recent artificial intelligence (AI)-based methods have explored a variety of physiological signals, many remain exclusively focused on obstructive apnoea. Furthermore, these methods are characterised by either reliance on complex architectures or a lack of validation using accessible data suitable for home monitoring. Methods The present study proposes a lightweight deep learning model that detects apnea events using a single acoustic channel. Respiratory audio is transformed into a series of time-frequency features, including Mel-Frequency Cepstral Coefficients (MFCC), Mel-spectrograms, and energy-based descriptors. The model was subjected to experimental validation using real-world data from adult and older-adult populations, with the objective of ensuring that it effectively captures clinically relevant patterns. Results In an independent test set, the proposed model demonstrated an accuracy of 85.84%, an F1-score of 88.96%, and an Area Under the Curve (AUC) of 93.06%. Specifically, the model demonstrated a high sensitivity of 91.21% for the apnea class, prioritising the detection of abnormal respiratory events. Conclusions The findings demonstrate that respiratory sounds contain sufficient information for reliable apnea detection, validating their potential as a non-invasive, low-cost biomarker. This approach provides a scalable alternative for home-based screening and telemonitoring, with the potential to enhance early detection and reduce cardiovascular and metabolic risks.