Limitations of Pacemaker Spike Detection in Capacitive ECGs via Deep Learning

Abstract Pacemaker spike detection is an important step in monitoring paced patients. Capacitive ECG facilitates unobtrusive monitoring of subjects during daily routines such as driving. Robust algorithms are required to deal with low signal quality and artifacts, e.g. by employing fusion of multiple signal channels. Due to the low signal-to-noise ratio of the measurement, there are limitations to detection accuracy compared to conventional ECG monitors. Especially low voltage stimulations such as bipolar pacemaker spikes are hard to detect. We present a convolutional network approach to improve on recent signal processing algorithms.We show a realistic evaluation of its performance using leave-one-subject-out cross validation (LOOCV), its dependence on the size of the receptive field, and an estimation of an upper performance bound.