Quantifying Uncertainty of a Deep Learning Model for Atrial Fibrillation Detection from ECG Signals

Recently, deep learning (DL) demonstrated capable to identify atrial fibrillation (AF) from electrocardiograms (ECGs) with significant performance. Nevertheless, these models may present an exaggerated self-confidence in their predictions and showing poor calibration in their output probabilities. In addition, such models cannot quantify the uncertainty of the predictions: a fundamental property in the clinical practice. In this study, we compared two DL models with the same architecture, but the second one had the first and last layers trained using a Bayesian approach, i.e., variational inference (VI), allowing the estimate of uncertainty of the predictions. We then compared the models in terms of predictive uncertainty H and Expected Calibration Error (ECE). Our experiments showed that the both models performed very well on the MIT-BIH Atrial Fibrillation dataset (sensitivity and specificity> 96%). The first model proved being i) more confident than the second one (H: 0.006 vs 0.090); and ii) more poorly calibrated (ECE: 0.360 vs 0.028). Despite the computational demand required for using the Bayesian approach in DL, our study demonstrated the importance of quantifying uncertainty of DL-based predictions for AF detection from ECG signals.