ECG Mortality/SCD Prediction with Generative Explainability

Ziad Obermeyer, A. Schubert, J. Ross, Sendhil Mullainathan

Summary

Obermeyer et al. trained a large deep neural network on ECGs to predict probability of death / sudden cardiac death (SCD). A separate generative model (trained on ECG distributions, likely diffusion/VAE) then acts as an interpreter of the black-box predictive model by synthesizing counterfactual ECGs — making the AI's decision visible as modified waveforms rather than heatmaps.