Automated Detection of Cardiac Syncope Causes in 12-lead ECGs Using Residual Neural Networks

Background: Syncope, a transient loss of consciousness due to cerebral hypoperfusion, is a common emergency department (ED) presentation, with incidence increasing with age. Detecting cardiac causes is crucial, as cardiac-related syncope is linked to higher one-year mortality. Traditional telemetry monitoring often generates excessive alarms, leading to alarm fatigue and diagnostic delays. This study explores a two-step deep learning approach using a residual neural network (ResNet) with an Efficient Channel Attention (ECA) module to improve automatic detection of syncope-related arrhythmias in 12-lead electrocardiograms (ECGs).

Methods: A cardiologist-guided selection process identified the most clinically relevant arrhythmias associated with syncope. The study utilized the Chapman Ningbo dataset, comprising 45,152 ECG recordings. Although the dataset included 48 distinct diagnoses, not all were relevant to syncope. Consequently, eight diagnoses deemed critical for syncope detection were selected, resulting in a refined subset of 11,432 multi-labeled ECG recordings for model training. A novel two-stage classification framework was developed. The first stage employed a binary classification model to differentiate normal from abnormal ECGs. ECG recordings were classified as normal if they exhibited a sinus rhythm with no additional abnormalities, while all other recordings were labeled as abnormal. The second stage applied a multi-classification model to the subset identified as abnormal, further categorizing ECGs based on the presence of specific cardiac conditions associated with syncope. The model's performance was optimized using Bayesian hyperparameter tuning, class-weighted adjustments to address dataset imbalances, and early stopping techniques to prevent overfitting. Performance evaluation was conducted using recall, precision, F1-score, confusion matrices, and receiver operating characteristic (ROC) curves.

Findings: Through expert consultation, the arrhythmias most relevant to syncope were identified as prolonged QT interval, atrial fibrillation, left bundle branch block, second-degree AV block, low QRS voltages, complete heart block, and Wolff-Parkinson-White (WPW) syndrome. The two-stage classification approach demonstrated superior performance in detecting key arrhythmias associated with syncope, outperforming single-stage models from prior research in several critical areas. The binary classification model achieved high sensitivity, with macro-averaged recall, precision, and F1-scores of 0.99 on the test set. The multi-classification model also produced promising results, particularly in identifying second-degree AV block, left bundle branch block, and WPW syndrome, where recall scores showed notable improvements (>0.05) compared to the single-stage classification approach. These findings underscore the potential of this deep learning framework as a valuable clinical tool for detecting high-risk arrhythmias in patients with syncope.

Discussion: This study demonstrates that a two-stage deep learning framework improves sensitivity in detecting cardiac syncope, particularly high-risk conditions like left bundle branch block and second-degree AV block, highlighting its potential as a clinical decision-support tool. By reducing alarm fatigue and enhancing arrhythmia detection, this method could aid real-time syncope risk stratification. Challenges remain, such as optimizing sensitivity for sinus rhythm and addressing dataset imbalances, but the model’s performance supports its clinical potential. Future research should focus on expanding datasets, refining training, and exploring real-time telemetry applications. With further development, this approach could improve early syncope detection, enhancing patient safety and clinical efficiency.