Noise resistance of several top-scored heart beat detectors

Background: PhysioNet provides the MIT-BIH Noise Stress Test Database, which consists of only two different clean ECG recordings for which realistic noise signals were added to the ECG in different signal-to-noise ratios (SNR). For the evaluation of the noise resistance of multisignal heart beat detectors this dataset is not sufficient. Methods: Our stress test is based on 100 clean records from the training dataset (set-p) of the PhysioNet/CinC Challenge 2014. Each record is composed of a different number of several biosignals (ECG, respiration, blood pressure). Multi-signal heart beat detectors were able to identify the heart beats easily without a predictive loss. We added noise to the 10 minute records using the nst-function of the WFDB Toolbox with SNR ranging from 24 to -6 dB. Since noise was added only between the 5th and 7th and after the 9th minute, clear segments can serve as learning periods. Results: The evaluation of the noise resistance of several top-scored multi-signal heart beat detectors includes the algorithms by Pangerc et al., Johnson et al., Hoog Antink et al., de Cooman et al. and Vollmer. The noise resistance is expressed as sensitivity and positive predictivity in noisy segments of the beat detection at different SNR levels. The most resistant algorithms were published by Hoog Antink and Vollmer. We are able to observe the effect of signal quality based constraints in the algorithms by Pangerc, Johnson and Vollmer, which mitigates a further decrease in the positive predictive value.