Key result
ResNet-50 achieves ~49% accuracy for four-class ECG image classification.
Why the study?
In automated deep learning interpretation of 12-lead ECG images, post-hoc explainability is typically presented as illustrative saliency imagery rather than systematically checked against the known anatomical structure of the ECG.
Does a two-phase transfer-learned ResNet-50 pipeline with a five-method explainability suite accurately classify four-class ECG images and localize clinically relevant waveform features?
Does a two-phase transfer-learned ResNet-50 pipeline with a five-method explainability suite accurately classify four-class ECG images and localize clinically relevant waveform features?
This study provides a reproducible explainable deep-learning pipeline for ECG image classification, demonstrating that while ResNet-50 learns discriminative representations (AUC 0.82-0.97) and focuses saliency on the ECG waveform trace via Grad-CAM/Grad-CAM++, resolution downscaling, class-weighting on small samples, and single-lead dataset mismatches severely limit uncalibrated classification accuracy (49.3%) and multi-lead validation.
Answers are grounded in this paper and cite their sources.
Modest F1 performance with overfitting precludes clinical use of this pipeline; leaves open whether larger datasets or refined architectures can advance ECG image classification.
Automated interpretation of 12-lead electrocardiogram (ECG) images with deep learning has repeatedly been reported to approach cardiologist-level performance, yet a persistent gap in the literature is that post-hoc explainability is typically presented as illustrative saliency imagery rather than as a mechanism that is systematically checked against the ECG's known anatomical structure. This paper reports an empirical, reproducible pipeline built specifically to close that gap. A ResNet-50 convolutional network, initialised with ImageNet weights and fine-tuned using a two-phase (frozen-head, then discriminative-learning-rate) transfer-learning schedule, was trained and evaluated on the four-class Mendeley ECG Images dataset (929 patients; Normal, Myocardial Infarction, Abnormal Heartbeat, History of Myocardial Infarction). On a held-out, stratified test set (n = 140) the model attained 49.3% accuracy and a macro-averaged F1 of 0.44 (95% bootstrap confidence interval: 0.367-0.508), with one-vs-rest per-class AUC ranging from 0.83 to 0.97 - evidence that the network's continuous class scores separate the four conditions considerably better than its arg-max decisions realise, a pattern consistent with a classifier that has learned informative representations but suffers a threshold and class-collapse problem compounded by measurable overfitting (training F1 0.81 versus validation F1 0.43). A five-method explainability battery - Gradient-weighted Class Activation Mapping (Grad-CAM), Grad-CAM++, a SHAP gradient explainer, occlusion sensitivity, and a purpose-built clinical lead-alignment analysis - was implemented and applied to correctly classified test examples. Grad-CAM and Grad-CAM++ saliency was observed to concentrate on the drawn ECG trace rather than on background grid lines or report margins, whereas the SHAP gradient explainer produced degenerate, spatially uninformative maps under the memory-constrained consumer-grade hardware used in this study. These results are reported transparently as a partial, in-progress benchmark: two further planned backbones (EfficientNet-B3 and a Vision Transformer, ViT-B/16) had not completed a training run at the time of writing, and the quantitative lead-alignment scoring is presented as a validated methodology rather than a completed numerical result, because the specific Mendeley image variant used is a single-lead (Lead II) rhythm-strip report rather than the full 3x4 twelve-lead grid the alignment method was designed to score. The paper documents the pipeline architecture, experimental protocol, diagnosed causes of the observed accuracy ceiling, and a concrete completion roadmap, positioning the work as a reproducible foundation for clinically grounded explainability research in deep-learning-based cardiac diagnostics.
No takes yet. Share an insight, caveat, or question.
Bhadane et al. (2026) studied Cardiovascular conditions on ECG (Normal, Myocardial Infarction, Abnormal Heartbeat, History of Myocardial Infarction) (n=929). ResNet-50 deep learning pipeline with five-method explainability battery (Grad-CAM, Grad-CAM++, SHAP, occlusion sensitivity, clinical lead-alignment) vs. Cardiologist-verified ground-truth diagnostic categories was evaluated on Four-class ECG image classification accuracy and macro-averaged F1 score on held-out test set (n=140) (Macro F1 0.441 (Accuracy 49.3%), 95% CI 0.367-0.508). ResNet-50 achieved 49.3% accuracy and macro-F1 of 0.441 with per-class AUC 0.82-0.97 on 140 test ECG images.
Synapse has enriched 10 related papers. Consider them for comparative context: