Development of a semi–real-time electrocardiogram monitoring system integrating artificial intelligence and wearable devices for atrial fibrillation screening

Background Atrial fibrillation (AF) is a common arrhythmia associated with substantial morbidity, particularly ischemic stroke. Conventional AF screening using rule-based algorithms is limited by suboptimal accuracy and the need for extensive clinician review when processing large volumes of electrocardiogram (ECG) data. Artificial intelligence (AI) has emerged as a promising approach to improve detection efficiency and scalability. Objectives To develop and internally evaluate a deep learning model for AF screening using real-world 24-hour Holter ECG data, and to evaluate a semi–real-time monitoring workflow integrating AI and wearable ECG devices in high-risk patients. Methods This retrospective study included 1,489 Holter ECG recordings collected at Nguyen Trai Hospital. A Residual Network (ResNet) model was trained on 1,089 recordings and evaluated on an independent dataset of 400 recordings using case-level classification with patient-level data splitting. A semi–real-time screening workflow integrating AI analysis and wearable ECG devices was prospectively implemented in a high-risk cohort. Results From the training dataset, a total of 29,765 minutes of cardiologist-annotated AF episodes were identified across 82 AF-positive recordings. These episode-level annotations were subsequently mapped to fixed 60-second segments for model training using a predefined sampling strategy. On the independent evaluation dataset (n = 400; AF prevalence 6.3%), the model achieved 100.0% sensitivity, 88.0% specificity, and 35.7% positive predictive value for case-level AF detection. In the pilot cohort (n = 167), AF was detected in 24 patients (14.4%), including 20 (12.0%) identified after 24 hours of monitoring. Conclusions The proposed framework demonstrated high sensitivity for AF detection in real-world Holter ECG data. Integration with wearable devices in a semi–real-time clinician-in-the-loop workflow was feasible and may support AF screening in high-risk populations.