Predicting Subclinical Atrial Fibrillation using Artificial Intelligence and validate using propensity-score matching and Explainable AI

Abstract

Introduction: With increased usage of wearables and implantable cardiac devices subclinical atrial fibrillation (SAF) is increasingly discovered. While SAF is asymptomatic, it is not a benign disease as it is associated with approximately a 3-fold increased stroke risk.

Methods: We developed two approaches using artificial intelligence (AI). One convolutional neural network (CNN) and a non-linear regression model on latent features produces by a variational auto-encoder (VAE + Ensemble Network) on electrocardiogram (ECG) median beat data. We included all patients with one AF or atrial flutter (AFL) ECG and one sinus rhythm ECG in the AF cohort. Precipitating factors, and anti-arrhythmic medication data was obtained. We allocated the ECG data in a 80:10:5:5 ratio between the training-, validation- and testing cohorts. We used integrative gradients and Shapely values to create increase transparency into both models.

Results: A total of 730 676 ECGs were included. 12 covariates were selected for propensity score matching. The random matched test cohort yielded a similar area under the receiver operator curve (AUC) for the VAE + Ensemble Network and the CNN 0.79 and 0.82. The propensity matched test cohort yielded a significant drop in performance. Shapley value analysis showed activation patterns for negative T waves, longer median beats, longer QT interval. Integrative gradient analysis showed primarily activation during the p-wave and late T wave with small deflections during the start- and ending of the median beat and the QRS complex.

Conclusion: AI based Networks can adequately identify AF based on a sinus rhythm ECGs.