Towards increased usability of noisy ECG signals in HRV-based classifiers

Proliferation of wearable devices and the “quantified self” behavior creates massive amounts of data acquired from physiological sensors. Unfortunately this vast amount of data is to date largely untapped for clinical purposes because of low data reliability caused by issues such as contact noise and motion artifacts. One important physiological signal is ECG and one of its important features is the heart rate, obtained from the QRS complex of the ECG signal or beat detection algorithms. This feature is important in characterizing Heart Rate Variability (HRV) and detecting various pathologies such as arrhythmia, chronic heart failure, or sleep apnea. Many such algorithms simply discard data segments which are deemed “unreliable” due to errors in QRS detection. This paper analyzes the impact of changing the noise level and noise duration on the percentage of data segments that are discarded. The paper proposes an approach to improve the usability of ECG data corrupted by noise by analyzing the impact of noise on features of interest and adapting relevant system parameters accordingly. The proposed approach can be used with any classifier operating on short-term HRV features.