Screening for obstructive sleep apnea in atrial fibrillation patients using deep learning on ECG

Obstructive Sleep Apnea (OSA) is increasingly recognized as a significant comorbidity in patients with Atrial Fibrillation (AF), with its prevalence varying between 21% and 74%. OSA not only predisposes individuals to AF but also complicates AF therapeutic management, resulting in higher recurrence rates and treatment failure. The detection and management of OSA in AF patients are thus crucial for effective clinical outcomes. Traditional diagnostic methods for OSA, such as polysomnography, are resource-intensive and not always feasible in the typical cardiac care setting. This study explores the use of multimodal deep learning model on 12-lead ECG in sinus rhythm and available clinical covariates such as age, sex and BMI for screening OSA in AF patients.

 

Developing a non-invasive and efficient screening tool for OSA in AF patients using multimodal deep learning on 12-lead ECG and clinical covariates.

Pre-procedural sinus rhythm ECGs from 207 patients undergoing AF ablation were collected, alongside demographic and clinical covariates. These 12-lead ECGs were processed using a novel deep learning architecture, integrating Long Short-Term Memory, Convolutional Neural Networks with a cross-attention module. This multimodal deep learning model was trained using ECG and the following clinical covariates (age, sex, and BMI) to predict OSA. The model's efficacy was assessed via a 5-fold stratified cross-validation, ensuring robustness and reliability.

 

In the studied group, 52 (25%) of the patients were identified with OSA. The multimodal deep learning model showed a 93.33% accuracy rate, achieved a specificity of 94.44%, and attained a sensitivity of 88.89% for OSA detection.

 

Our multimodal deep learning approach shows promise in screening for OSA. This method could streamline the diagnostic process, directing more resource-intensive diagnostic approaches to patients with a higher likelihood of having OSA.