Robust detection of heart beats using dynamic thresholds and moving windows

Background: This contribution relates to the PhysioNet/CinC Challenge 2014 on Robust Detection of Heart Beats in Multimodal Data. The aim is to locate heart beats in continuous long-term data. Methods: The beat detection system is build up of several parts. Preprocessing consists of high pass filtering followed by standardization. Extrema of a moving window were used to capture the heart beat impulse. A windowed approach led to dynamic thresholds. Valid parts of the channels were determined and the locations of beats were extracted. The beat locations of various channels were compared during the multichannel fusion procedure and dynamic delay correction. Doubtful locations were checked using RR distances. Results: The algorithm was tested on the training data set for this challenge (one hundred 10-minute recordings) and on several freely available PhysioNet databases which were annotated by physicians. The algorithm had the best score applied to the hidden Phase 1 dataset of the 2014 PhysioNetlCinC challenge. Conclusion: The developed algorithm presents a promising approach to detect heart beats in multivariate records.