A Bayesian Fusion Model for Heart Rate Annotations

Lack of gold standard of truth is a critical problem in medical field. Consensus extraction from noisy labels to estimate ground truth is a common approach. However, significant inter-observer variability and various biases limit the accuracy. For this challenge, we proposed an unsupervised Bayesian framework to aggregate multiple heart rate (HR) annotations from the electrocardiogram (ECG) signal and infer the underlying ground truth and precision of annotators when the ground truth is not available. We further introduced the bias and precision adjustment factor of annotator to model the annotator's performance in each instance. Test on the 2014 PhysioNet/ Computing in Cardiology challenge database showed that the proposed model had an improvement not only over the existing aggregation model (EM-R method), mean and median strategies (the highest of 10.36%,20.26% and 12.43%) but also to the best single annotator (13.75%) used in this paper.