Atrial fibrillation predictor with reject option using belief functions theory

Abstract

This paper proposes a novel approach for predicting atrial fibrillation. The novelty of this approach remains in the reject option. Instead of classifying all instances, it rejects ambiguous observations by leaving them with no decision. In other words, three labels are considered: sick, healthy and unknown. Using physiological signals of patients, the proposed approach extracts several features from the signals in real-time. Then, it uses the features as sources of information in the belief functions framework. Due to the information sources, masses are assigned to the labels, and a risk level of being sick is computed. Prediction is assumed ambiguous if the risk level is within the rejection region and the state is unknown; otherwise, a decision is made. The approach is validated using the MIMIC III database. Due to the reject option, the prediction accuracy increases from 59% to around 78%.