(P123) INCIDENT ATRIAL FIBRILLATION PREDICTION AT A SPECIALIZED TERTIARY HEART CENTRE: A DEEP LEARNING APPROACH USING MEDIAN ECG WAVEFORMS

Background: Deep learning (DL) applied to electrocardiograms (ECG) is an emerging modality for predicting incident atrial fibrillation or flutter (AF). This study aims to train a DL model using 12-lead ECG median waveforms to predict 5-year incident AF in a patient population of a tertiary heart centre (THC) and compare it to the traditional approach which utilizes full-length 10-second waveforms.

METHODS AND RESULTS: In a retrospective study, we examined 1.4 million ECG acquired from over 250,000 adults at a THC between 2004 and 2022. ECG were excluded if they were performed within 30 days after cardiac surgery, captured a rhythm other than sinus rhythm, or were acquired in patients with pre-existing AF. Incident AF at 5 years was modelled as a binary outcome and was determined using available outpatient and inpatient clinical databases, and ECG diagnoses. Median waveforms were computed by the MUSE (GE Healthcare) software for each ECG. Included ECG were randomly split by distributing patients into training (70%), validation (10%), and test (20%) sets. Two ResNet-50 models were developed: one using median waveforms (median model) and the other using original 10-second waveforms (original model). Model performance was evaluated on the test sets. Bootstrapping was used to report confidence intervals (CI).
A total of 669,782 ECG (47% of the screened ECG) were included among 145,323 patients. Mean age was 63±15 years and 62% were male. The 5-year incident AF outcome was observed in 12% of ECG and 16% of patients. At the ECG level, The median model reached an area under the receiver operating characteristic curve (AUC) of 0.738 (95% CI: 0.734-0.742), while the original model achieved an AUC of 0.749 (95% CI: 0.744-0.753). The area under the precision-recall curve (PRC), a performance metric better suited for unbalanced datasets, was not statistically different between both models with the original model at 0.310 (95% CI: 0.303-0.318) and the median model at 0.304 (95% CI: 0.297-0.311). Saliency maps applied on the median waveforms highlighted the P-wave area as having the highest influence on the model’s prediction.

Conclusion: A unimodal DL model utilizing median ECG waveforms demonstrated similar performance in predicting 5-year incident AF compared to using original full-length waveforms in a cohort of all-comer patients at a THC. Future studies could further investigate the potential use of median waveforms in ECG-based DL applications to enable model explainability.