Shortcut Learning and Clever Hans Effect in CNN-Based ECG Image Classification

A study published on arXiv (2607.25117) investigates shortcut learning and the Clever Hans effect in convolutional neural networks (CNNs) used for ECG image classification. Researchers created six controlled image feature sets (FS1–FS6) to test whether models rely on non-physiological visual cues rather than actual ECG waveform morphology. FS1 used raw full ECG images; FS2 cropped to waveform-only; FS3 masked waveform to show metadata; FS4 introduced red-arrow artifacts for myocardial infarction class; FS5 applied contrast enhancement for abnormal heartbeat class; FS6 used Gaussian blur for normal class. The findings highlight that deep learning models may achieve high accuracy by exploiting spurious correlations, undermining clinical trust and interpretability. The study underscores the need for robust validation to ensure models learn physiologically relevant features.