Attention-Based Transfer Learning for Efficient Obstructive Sleep Apnea (OSA) Classification on Snore Sound

Polysomnography (PSG) is currently the gold-standard technique for classifying sleep apnea disorders. Yet, it is costly and requires an expert to score the severity, making it impractical for self-screening and home use. Snore sound classification with Deep Learning (DL) is a promising approach and has gained increasing interest due to its relationship with abnormal breathing conditions in both time and frequency domains. This study proposes an attention-based transfer learning model for non-invasive detection of obstructive sleep apnea (OSA) using audio signals. Mel-spectrograms and MFCC features were input into the MobileNetV3-Large to extract deep features. A modified SENet was implemented to provide suitable channel attention for the extracted features. The attention-based features are then classified into normal and abnormal snore events. The study evaluates model performance using 10-fold cross-validation on sound data of the 70 adult OSA patients of an open-source PSG-Audio dataset [1]. Results show that utilizing both Mel-Spectrogram and MFCCs as input features significantly enhances classification performance compared to single-feature models. In addition, the MobileNetV3-Large with modified SENet significantly outperforms the combination of other pre-trained and attention mechanisms. Specifically, the proposed model provides an accuracy of 92.576±0.910%, a sensitivity of 92.906±1.928%, a specificity of 92.269±2.740%, a precision of 92.173±3.024%, and an F1-score of 92.486±1.326% on the binary classification. Its performance also shows statistically significant improvement when benchmarked with other existing OSA classification models. Our proposed model demonstrates suitable potential for portable device-based sleep apnea monitoring applications.