Predicting the spontaneous cardioversion of atrial fibrillation using artificial intelligence-enabled electrocardiography.

Aims: Spontaneous cardioversion (SCV) is commonly observed in patients presenting to emergency departments (EDs) with primary atrial fibrillation (AF). Predicting SCV could facilitate timely discharge and avoid costly admissions. We sought to evaluate whether SCV could be predicted using artificial intelligence-enabled electrocardiograms (AI-ECGs) and whether this could produce cost savings.

Methods and results: We recruited patients presenting to EDs with primary AF throughout 2022-23. Patients were excluded if the outcome of their AF episode was unclear, or the ECG was not accessible. Spontaneous cardioversion prediction was attempted using ResNet50, EfficientNet, and DenseNet convolutional neural network (CNN) architectures and subsequently an ensemble learning model. We then performed a cost-minimization analysis to estimate the cost effect of a prediction-guided 'wait-and-see' protocol. There were 1159 presentations to the ED, of which 502 had sufficient data for inclusion. The median age was 74.0 years and 54.0% were women. Spontaneous cardioversion occurred in 227 (45.2%) patients and was more frequent in younger patients (P < 0.001). The ensemble learning model outperformed individual CNNs, achieving an accuracy of 69.7% (SD 5.91) and a receiver operating characteristic area under the curve (ROC AUC) of 0.742 (SD 0.037) with a sensitivity and specificity of 0.736 (SD 0.068) and 0.657 (SD 0.150), respectively. The per patient cost was $4681 if all patients were admitted, which reduced to $3398 with a prediction-guided 'wait-and-see' protocol with a 33.3% reduction in overall hospitalization.

Conclusion: Artificial intelligence-enabled electrocardiogram can predict SCV in patients presenting to EDs with primary AF, and a prediction-guided 'wait-and-see' protocol utilizing AI-ECG could lead to substantial cost savings and reduced hospitalization.