Comparing subset selection methods in multi-lead electrocardiogram data

Heart health complications are often diagnosed through the presence of anomalous heartbeat morphologies in the electrocardiogram (ECG). The automated summarization of ECG data can aid clinicians in promoting a more comprehensive, time-sensitive assessment. Given that lead availability may vary in different health-monitoring settings, however, we investigate subset selection in single-lead ECG, 12-lead ECG, vectorcardiogram (VCG), and VCG magnitude representations using data from the St. Petersburg INCART 12-lead Arrhythmia Database. Subsets for each data representation are found using seven different algorithms that can be used to form CUR matrix decompositions, three of which use oversampling from reduced representations of the data. The QR-based discrete empirical interpolation method (Q-DEIM) with 12-lead data yields the highest class detection results among the non-oversampling algorithms. The extended DEIM (E-DEIM) oversampling algorithm performs the best overall with a lower-rank representation of the VCG magnitude data, offering potential computational savings along with its improved class detection. The results of this work provide insight into the summarization of different ECG representations with the goal that such summaries can in turn be used in subsequent models or presented directly to clinicians for improved patient outcomes.