Learning generalizable representations with adversarial domain adaptation for snoring-based sleep apnea detection

Sleep apnea (SA) is a typical sleep disorder that, if untreated, can lead to severe health complications. Snoring is a key symptom of SA and can be used to develop a non-contact automatic SA detection method. However, due to inter-subject variability, the acoustic characteristics of snoring can vary significantly across individuals. This inter-subject variability poses a significant challenge for model generalization and often leads to degraded performance when applied to unseen subjects. To address this challenge, we propose an unsupervised domain adaptation framework based on Wasserstein Generative Adversarial Network (WGAN) and introduce a lightweight Subject-Invariant Feature Refiner (SIFR) module. A two-stage training strategy is adopted. In the first stage, we pretrain a feature extractor, which integrates CNN and Transformer Encoder, using snoring data of existing subjects to extract a generalized snoring feature representation. In the second stage, the pretrained feature extractor is frozen. The SIFR module is introduced and combined with adversarial training to conduct domain adaptation with a small amount of unlabeled data from new subjects. We evaluate the model performance on three main SA-related tasks: SA event detection, SA patient diagnosis, and Apnea-Hypopnea Index (AHI) estimation. On the subject-independent dataset, our model achieves 74.24% accuracy in SA event detection and 83.59% accuracy in SA patient diagnosis. For AHI estimation, our model achieves a mean absolute error (MAE) of 10.430 and a Pearson correlation coefficient (PCC) of 0.86, indicating a strong correlation between the estimated and real AHI values. This study provides an effective solution for SA detection in new subjects in practical applications.