Nature's AI Model for ICD Patient Selection Faces Criticism

A recent Nature study explored using an AI model to determine which patients should receive an Implantable Cardioverter Defibrillator (ICD). In response, @VPrasadMDMPH published a thread raising five major clinical and methodological concerns. His core argument is that such models must do more than just "predict risk"—they must prove they identify truly shockable arrhythmias and ultimately deliver actual survival benefits.

Key Criticisms

@VPrasadMDMPH points out that the model must identify "shockable" fatal arrhythmias. However, the study uses sudden death labels from death certificates, which likely include Ventricular Tachycardia (VT) scenarios that are not salvageable by a defibrillator, such as VT caused by myocardial infarction. If the underlying disease is irreversible, predicting the risk does not mean the clinical outcome will be improved.

Methodological Concerns

Regarding algorithm training, the author highlights two main issues. First, the model must avoid counting patients who later develop ICD indications as "successfully predicted," otherwise it merely labels eventual deaths earlier rather than reducing sudden death risk. Second, if the model is trained on raw ECGs that include patients with existing ICDs, it might learn device artifacts (like pacing spikes) instead of true pathological risk features, thereby overestimating its clinical value.

Conclusion

Based on these points, @VPrasadMDMPH concludes that the study currently demonstrates observational correlation rather than proving that AI-assisted ICD decision-making can improve actual survival outcomes.

2026-07-13 ~ 2026-07-13 · 6 related posts