ApneaWhisper: Transformer-Based Audio Segmentation for Fine-Grained Non-Invasive Sleep Apnea Detection

Sleep apnea is a prevalent sleep disorder with serious health implications. This study introduces ApneaWhisper, a Transformer-based audio segmentation model designed for noninvasive detection of sleep apnea subtypes using PSG-Audio data.

We utilized a PSG-Audio dataset from 284 patients. ApneaWhisper leverages a pretrained Whisper encoder to extract 10 ms-resolution frame-level features from 20-second audio clips. A lightweight Transformer decoder with token-based segmentation and a classification head aggregates these features for both frame-level and clip-level predictions. The model was fine-tuned using class-balanced cross-entropy loss to address data imbalance across apnea subtypes.

ApneaWhisper achieved strong performance for sleep apnea detection, with a clip-level F1-score of 0.82 and a frame-level F1-score of 0.70, outperforming conventional baselines including MFCC+DNN, VGGish+bi-LSTM, and VAD-based models. It also showed promising ability in distinguishing between OSA, MSA, CSA, and hypopnea, though with varying success.

The model's fine-grained temporal resolution enables precise apnea event localization, duration estimation, and subtype classification. While ApneaWhisper performs robustly for OSA, challenges remain in distinguishing central (CSA) and mixed (MSA) sleep apnea, due to subtle or ambiguous acoustic patterns. The frame-level segmentation also facilitates accurate apnea-hypopnea index (AHI) estimation, which could reduce dependence on full PSG studies in certain clinical and home-monitoring scenarios. Future improvements may involve multimodal integration (eg, oxygen saturation) and noise-robust training techniques.