Enhancing convolutional neural network predictions of electrocardiograms with left ventricular dysfunction using a novel sub-waveform representation

Abstract

Background: Electrocardiogram (ECG) deep learning (DL) has promise to improve the outcomes of patients with cardiovascular abnormalities. In ECG DL, researchers often use convolutional neural networks (CNNs) and traditionally use the full duration of raw ECG waveforms that create redundancies in feature learning and result in inaccurate predictions with large uncertainties.

Objective: For enhancing these predictions, we introduced a sub-waveform representation that leverages the rhythmic pattern of ECG waveforms (data-centric approach) rather than changing the CNN architecture (model-centric approach).

Results: We applied the proposed representation to a population with 92,446 patients to identify left ventricular dysfunction. We found that the sub-waveform representation increases the performance metrics compared to the full-waveform representation. We observed a 2% increase for area under the receiver operating characteristic curve and 10% increase for area under the precision-recall curve. We also carefully examined three reliability components of explainability, interpretability, and fairness. We provided an explanation for enhancements obtained by heartbeat alignment mechanism. By developing a new scoring system, we interpreted the clinical relevance of ECG features and showed that sub-waveform representation further pushes the scores towards clinical predictions. Finally, we showed that the new representation significantly reduces prediction uncertainties within subgroups that contributes to individual fairness.

Conclusion: We expect that this added control over the granularity of ECG data will improve the DL modeling for new artificial intelligence technologies in the cardiovascular space.

Keywords: Deep learning, Cardiology, Electrocardiograms, Sub-waveform representation, Machine Learning.

Figure 1

Electrocardiography (ECG) waveform representations. A: Full-waveform representation that directly comes from an ECG device. B–D: Rhythmic discretization (B) and sliding of the lead I full waveform (C) to create the lead I sub-waveform representation (D). E: The same discretization and sliding as lead I is applied to other 7 leads to create sub-waveform representation for all 8 leads.

Figure 2

Systematic experiments for evaluating performance of sub-waveform with respect to full-waveform representation (baseline) for left ventricular dysfunction (LVD) case study for 9244 patients in holdout test set. A: Minimum sliding is used and is equal to 0.74 seconds. B: Maximum sliding is used and is equal to sub-waveform duration. For both experiments in panels A and B, we examine 2 sets of a number of sub-waveforms: 1 sub-waveform and maximum number of sub-waveforms. The optimal sub-waveform is highlighted by dashed circle and has a duration of 1.48 seconds with 10 sub-waveforms. Our sub-waveform representation provides more accurate predictions with smaller uncertainties.

Figure 3

Explanation for enhanced learning by sub-waveform representation for 9244 patients in holdout test set. A: Full-waveform heart rate for each patient (top) and heart rate histogram (bottom) for rhythmic and arrhythmic groups. Heart rate coefficient of variation (CV) is set to 0.01 to define the 2 groups. The error bars come from averaging across different heartbeats of full-waveform representation. B: Eighty-eight lead I waveforms are taken from rhythmic group corresponding to heart rate with maximum count and waveforms are left aligned for each representation. C: The aligned waveforms and their saliency maps are averaged. The absolute values of a saliency map are min-max normalized for each waveform. As predicted by higher averaged maximum importance and lower CV, features are more aligned with less uncertainty in sub-waveform representation.

Figure 4

Interpretation of electrocardiographic (ECG) features for full-waveform and sub-waveform representations. A: ECG features: P wave, PR segment, QRS complex, ST segment, and T wave. The bounding boxes show the extents of each feature. B: Predicted probability for each patient and corresponding confusion matrix. C: Positive saliency maps for top 5 patients with positive outcome that are classified as false-negative by full-waveform representation and turned into true positive by sub-waveform. Top-5 shows the ranking with respect to the difference in probabilities of sub-waveform and full-waveform representations. D: For each patient, P, PR, QRS, ST, and T importance scores are calculated and shown as a bar chart for both representations.

Figure 5

Impact of electrocardiographic waveform representation on 14 subgroups in holdout test set. Top plot shows prevalence of arrhythmia in each subgroup. Area under receiver operating characteristic (AUROC) and area under precision-recall curve (AUPRC) for full-waveform (green) and sub-waveform (orange) representations are shown in middle and bottom plots. The number of patients in each subgroup is shown in the parenthesis. Less prevalence of arrhythmia in a subgroup may induce redundancies owing to higher number of rhythmic full waveforms that cause higher deep learning prediction uncertainties. Sub-waveform representation provides individual fairness by reducing these disparities within a subgroup.