Snoring-based obstructive sleep apnea screening and AHI estimation with an adapted pretrained audio model

Objective. Obstructive sleep apnea (OSA) is a common sleep disorder that, if left untreated, can lead to serious health problems. Snoring is a typical symptom of OSA and is commonly used to develop non-contact automatic detection methods for apnea states during sleep. However, scarce snoring data and subject heterogeneity hinder model generalization to unseen subjects. Approach. This study proposed a snoring-based framework for OSA detection and apnea-hypopnea index (AHI) estimation. First, the OSA detection model is based on pretrained audio model Wav2vec 2.0. Considering the high computational complexity of the pretrained model, we adapt Wav2vec 2.0 through two complementary strategies. A layer dropping approach was proposed to prune redundant higher Transformer encoder layers to reduce computational burden while retaining low-level and mid-level acoustic representations relevant to snoring pathology. A cross-layer attention mechanism was introduced to integrate complementary features from different Transformer layers, balancing fine-grained acoustic detail with contextual information. Second, by simulating clinical AHI calculation processes, we introduced statistics including both absolute counts and ratio-based measures to build a regression model for AHI estimation. Main results. Evaluations under subject-dependent and subject-independent settings showed that the proposed framework achieved high abnormal-snoring detection performance with a trade-off between accuracy and computational cost. For AHI estimation, it achieved a mean absolute error of about 11 events h (Formula presented) (Formula presented) and a Pearson correlation coefficient of 0.81. Significance. These results suggest that the task-adapted pretrained model can enhance generalization across unseen subjects and provide preliminary evidence of the feasibility of snoring-based OSA initial screening and AHI estimation.