Evidential Deep Learning Model for Atrial Fibrillation Detection from Holter Recordings

Deep learning (DL) models have shown promising per- formances for detecting atrial fibrillation (AF) and atrial flutter (AFL) from electrocardiograms (ECGs) but often suffer from overconfidence and poor probability calibra- tion. Evidential DL (EDL) addresses this by using evi- dence to parameterize a Dirichlet distribution for uncer- tainty estimation. The main objective of this study was to develop an EDL model for AF and AFL detection from 2-lead Holter ECG recordings, aiming to estimate uncer- tainty without incurring additional computational costs compared to traditional softmax-based DL models. In this study, we developed an evidential residual-based DL model, treating predicted probabilities as subjective opin- ions. The model was trained and tested on a comprehen- sive dataset of 661 Holter recordings. Our experiments showed that the EDL model achieved recalls of 0.953, 0.838, and 0.934 for detecting AF, AFL, and Non-AF, re- spectively. The corresponding AUC scores were about 0.980 for AF and Non-AF, and 0.972 for AFL. In terms of confidence estimation, the EDL model exhibited superior performance with an expected calibration error of 0.09, compared to 0.16 for the softmax-based model. These re- sults indicate that EDL models offer enhanced calibration and effectiveness in detecting AF compared to standard softmax models.