Time-hybrid OSAformer (THO): A hybrid temporal sequence transformer for accurate detection of obstructive sleep apnea via single-lead ECG signals

Abstract

Background and Objective

Obstructive Sleep Apnea (OSA) is among the most sleep-related breathing disorders, capable of causing severe neurological and cardiovascular complications if left untreated. The conventional diagnosis of OSA relies on polysomnography, which involves multiple electrodes and expert supervision. A promising alternative is single-channel Electrocardiogram (ECG) based diagnosis due to its simplicity and relevance. However, extracting respiratory-related features from ECG is challenging since ECG signals do not directly reflect respiratory patterns. Consequently, the accuracy of most deep learning models that predict OSA using ECG data remains to be improved.

Methods

In this study, we propose the Time-Hybrid OSA transformer (THO), a novel method that leverages single-lead ECG signals for accurate OSA detection. The THO enhances feature extraction using a hybrid architecture combining dilated convolution and Long Short-Term Memory (LSTM), along with a multi-scale feature fusion strategy. Additionally, THO integrates an embedded memory decay mechanism within a multi-head attention model to capture real-time characteristics of time series data. Finally, a voting mechanism is incorporated to enhance decision reliability.

Results

Evaluation of the THO model demonstrates superior performance with prediction accuracy (ACC) and area under the receiver operating characteristic curve (AUC) values of 95.03 % and 96.85 %, respectively, representing improvements of 11 % and 8 % over comparative models. Moreover, the ACC shows a 5 % enhancement relative to state-of-the-art models.

Conclusions

These results prove the THO model's efficacy in predicting OSA, offering a robust alternative to traditional diagnostic approaches.