ECG AI models learn visual shortcuts, not heart signals
CNN classifiers achieve high accuracy by exploiting non-clinical cues like arrows and contrast.
Deep Dive
Researchers created six feature sets from ECG images to test whether CNNs learn clinically meaningful waveform morphology or shortcut cues from non-physiological visual elements. They calculated shortcut retention scores and used attribution methods to identify potential Clever Hans behavior.
Key Points
- Six feature sets tested: raw, cropped waveform, masked metadata, red-arrow artifacts, contrast-enhanced, and Gaussian-blurred images.
- Shortcut retention scores and attribution maps showed models relied on non-clinical cues like red arrows and contrast, not ECG morphology.
- Performance persisted even when waveform data was removed, confirming Clever Hans behavior that undermines clinical reliability.
Why It Matters
High-accuracy ECG AI may be learning visual shortcuts, risking misdiagnosis in real clinical settings.