Efficiency of different heartbeat detection methods by using alternative noise reduction algorithms

Marcus Vollmer and Jader Giraldo Guzmán
Institute of Bioinformatics, University Medicine Greifswald


Abstract

Background: ECG noise reduction is an essential step in the ECG preprocessing pipeline. Baseline wander, movement, and muscle artifacts can affect the quality of automated processing, particularly beat annotation and beat classification. This can be much more difficult in arrhythmic cases. The variety of available methods ranges from filter-based techniques (e.g. FIR) to signal decomposition (e.g. wavelets) to neural networks (Cycle-GAN). Methods: We used 707 12-lead ECGs from the SFB/TR19 study on inflammatory cardiomyopathy and 9,989 resting 12-lead ECGs from the Study of Health in Pomerania (SHiP) to assess the accuracy of heartbeat detection methods in relation to the preprocessing methods used. Standard signal libraries (pyECG, neurokit2) were used among other methods to generate 15 differently preprocessed ECGs. More than 20 detectors from standard libraries and state-of-the-art methods from the PhysioNet/CinC Challenge 2020/21 were evaluated. Positive Predictive Values (PPV) and False Negative Rates (FNR) were calculated for each combination of signal lead, preprocessing method and beat detector. Annotations were corrected for fixed delays and scored with a tolerance of 50 ms compared to ground truth annotation. Results: In SFB-ECGs from diseased patients we found better preprocessing methods for each detector that both increase PPV and decrease FNR at the same time. Among the top alternative preprocessing methods were methods from the neurokit2 library: hamilton2002, engzeemod2012, and pantompkins1985. For example, using pantompkins1985 to preprocess before applying Reddy's algorithm improved the PPV from 0.95 to 0.99 and reduced the FNR from 0.05 to 0.03.