DRW-CNN: A Dilated Residual Wavelet CNN for ECG-based Severe Heart Failure Screening

Abstract

Electrocardiogram (ECG)-based heart-failure screening requires clinically valid labels, record-independent evaluation, and uncertainty analysis at the level of independent records. This study evaluates a Dilated Residual Wavelet Convolutional Neural Network (DRW-CNN) for discrimination of clinically confirmed severe Congestive Heart Failure (CHF) from normal sinus rhythm. The primary cohort comprises all 15 available BIDMC CHF records and 18 MIT-BIH Normal Sinus Rhythm Database records. Records are partitioned before segmentation, yielding 528 training, 96 validation, and 168 held-out test segments from mutually exclusive records. The network applies level-4 Daubechies-4 wavelet denoising, z-score normalization, residual blocks with dilation factors 1, 2, and 4, global average pooling, and sigmoid classification. Class-weighted binary cross-entropy addresses imbalance without deleting independent records. At the prespecified threshold of 0.5, held-out predictions achieve 95.238% accuracy, 95.833% sensitivity, 94.792% specificity, 93.243% precision, 94.521% F1-score, MCC 0.903, ROC-AUC 0.988, and average precision 0.989. Record-cluster bootstrap intervals, calibration, threshold sensitivity, record-level aggregation, leave-one-record-out stability, and auxiliary non-HF domain-shift stress analyses are also reported. The study provides a leakage-resistant severe-CHF screening evaluation that separates disease-specific evidence from auxiliary abnormality testing and quantifies uncertainty at record level.