AI-based ECG analysis outperforms predefined ECG features to predict AF recurrences in patients undergoing AF ablation: Data from the ISOLATION registry

Ablation is an effective rhythm control strategy for patients with atrial fibrillation (AF); however, many patients experience recurrences of AF following the procedure. Accurately identifying patients who are most likely to benefit from ablation could be a crucial step towards personalized AF treatment. Traditionally, predicting ablation outcomes relies on predefined electrocardiogram (ECG) and clinical features. This approach, however, has shown only limited performance so far.

We hypothesized that modern deep learning techniques can analyze subtle differences in normal sinus rhythm ECG (SR-ECG) to predict AF ablation outcomes more effectively than traditional techniques with predefined ECG features. We also hypothesized that incorporating supplementary clinical data (demographics, medical history) could enhance these prediction models.

A 10-second 12-lead SR-ECG was collected from 232 AF patients scheduled for ablation. Predefined features—such as amplitude, duration, and complexity of the P-waves, along with QRS, T-wave, QTc, and PR interval durations—were extracted from beat-averaged ECG. A gradient boosting model was trained using these features to predict AF recurrence after a 12-month follow-up. The same ECG signals were also used to train an AI-based approach utilizing a convolutional neural network (AI-ECG). Both models were evaluated using 10-fold cross-validation across multiple metrics. Clinical data were integrated into the AI model to enhance performance. Finally, using saliency scores, we analyzed on which ECG features the AI-ECG model focused for its decisions.

The AI-ECG approach outperformed the predefined ECG model, achieving an AUROC of 0.71±0.06, sensitivity of 0.72±0.21, specificity of 0.76±0.26, accuracy of 0.75±0.15, and F1-score of 0.52±0.08, compared to an AUROC of 0.61±0.06, sensitivity of 0.74±0.30, specificity of 0.63±0.28, accuracy of 0.66±0.13, and F1-score of 0.52±0.08 for the predefined approach (Table 1). Incorporating clinical data further improved both models, with a greater enhancement observed for the AI-ECG model. The best performance was achieved when combining all three approaches—AI-ECG, predefined ECG, and clinical data—yielding an AUROC of 0.78±0.05, sensitivity of 0.84±0.13, specificity of 0.74±0.15, accuracy of 0.76±0.08, and F1-score of 0.64±0.04. Beat-averaged saliency analysis showed that the AI model significantly focused on the P-waves in ECGs during correct predictions (p<0.05).

Our findings demonstrate that the AI-ECG model outperforms traditional ECG feature-based approaches in predicting AF recurrence after ablation, particularly when combined with clinical data. The model's reliance on P-wave highlights its ability to capture valuable predictive information from P-waves that may be missed by conventional features. These results suggest that AI-augmented ECG analysis could play a crucial role in personalized AF treatment planning.  

Results