Forecasting Heart Procedure Outcomes Using World Models

Summary & Methodology Analysis

The architecture centers on a clinical world model that maps patient physiological states into a latent (hidden) space. The process starts by pretraining a Variational Autoencoder, a generative model that compresses input data into a structured latent representation, to map MRIs into a 3D spatial state. The system initializes this state using pre-ablation LGE-MRI and incorporates static covariates and ablation heatmaps. To handle longitudinal data, the model uses a transformer-like approach to treat irregular clinical events and pre-event ECG embeddings as context tokens, allowing the model to evolve the latent anatomical state over time using a 3D residual CNN for state transitions and an MLP for time-dependent drift. A terminal horizon token enables anytime forecasting of recurrence and scar extent, providing flexibility in clinical monitoring. During training, the model uses latent forecasting to align with follow-up MRI states, binary cross-entropy for recurrence, and Huber loss, a robust regression loss function that is less sensitive to outliers, for scar extent. The study is constrained by the small 91-patient size of the complete-record cohort and the absence of ablation geometry in the auxiliary cohort, which limits full-model validation. Additionally, manual mapping of electroanatomic data to MRI can introduce localization errors, and the input edits reflect sensitivity to changes rather than direct causal effects. The paper does not specify inference latency, hardware resource requirements, or training costs.