Towards Artificial Intelligence-Based Decision Support for Large-Scale Screening for Atrial Fibrillation

Atrial fibrillation is a prevalent cardiac arrhythmia, significantly increasing the risk of stroke, heart failure, and mortality. Early detection, especially during asymptomatic and paroxysmal stages, is essential for effective intervention. This study explores the application of deep neural networks in simplified ECG screening to enhance population-wide detection of atrial fibrillation. A handheld device, MyDiagnostick, was employed for large-scale ECG data acquisition within a pharmacy-based clinical trial on 7295 subjects aged 65 years and older. Automated diagnosis yielded 6.08% of AF prevalence in the given dataset. The data were then analyzed using a validated deep neural network model for the detection of cardiac arrhythmia in 12-lead ECG data for feature extraction and detection of atrial fibrillation. In addition, we investigate the capabilities of explainable artificial intelligence to provide diagnostic support for cardiologists and assess the feasibility of implementing deep neural networks in wearable devices for continuous monitoring. The study also emphasizes the importance of interpretability in artificial intelligence models for medical applications, leveraging explainable artificial intelligence to highlight ECG segments indicative of atrial fibrillation. Our findings demonstrate the efficacy of deep neural networks in atrial fibrillation detection with an F1-score of 86% vs. 81% of the automated ECG stick analysis and the potential for their integration into wearable technology by successfully reducing the number of weights by 99% without significant loss of accuracy, providing a robust tool for early diagnosis and continuous monitoring of atrial fibrillation.