Intervention-Aware Clinical World Model Forecasts AF Ablation Recurrence at AUROC 0.756 Without Follow-Up MRI

arXiv 2608.13518·low signalXBlueskyLinkedInCopy link

Most clinical prediction models map baseline measurements to a future endpoint in one step, but post-procedure recovery is an irregular trajectory of asynchronously recorded observations, medication changes, and repeat interventions. This model encodes baseline imaging into a 3D spatial latent state and evolves it through time-ordered post-intervention events using procedural context, static covariates, elapsed time, and peri-event physiological embeddings, with follow-up imaging used only as training-time supervision. On DECAAF-II atrial fibrillation ablation data it reached AUROC 0.756 and AUPRC 0.777 for recurrence and a scar-extent MAE of 2.971 percentage points — notably, without requiring follow-up MRI intensities at inference, and supporting risk queries at different horizons.