Feature Contributions to ECG-based Heart-Failure Detection: Deep Learning vs. Statistical Analysis

Assessing feature contributions to a specific diagnosis is commonly done by statistical analysis. In the context of heart failure (HF) diagnosis from the electrocardiogram (ECG), this work compares feature contributions assessed by deep learning with those obtained by statistical analysis. Data consists of ECG pairs (baseline and follow-up) from patients with a history of myocardial infarction. When the follow-up ECG was made, controls patients had remained stable, while cases patients had developed HF. The 42 features that characterized each ECG served as inputs of a deep-learning neural network (NN) created by our Repeated Structuring & Learning Procedure. Subject-specific feature ranking was obtained from the local-interpretable model-agnostic explanatory algorithm and processed to obtain feature relevances (FR). Additionally, 42 areas under the curve (AUC) by univariate statistical analysis were obtained. FR and AUC were compared by Pearson's correlation coefficient (p). After training, the NN had a 99% classification performance. FR ranged from 0.32 to 4.47; AUC ranged from 23% to 82%. Correlation analysis yielded no significant association between AUC and FR (ρ=0.18, P-value =0.25). Deep-learning and statistical-analysis feature contributions to HF detection were discordant. Further studies will investigate which of the two approaches better reflects clinical interpretation.