Intervention-Aware Clinical World Model for Post-Op Outcome Forecasting in Cardiology

The management of atrial fibrillation (AF) has seen rapid advances with the integration of medical imaging, electrophysiology, and machine learning techniques. Early models for recurrence prediction primarily relied on static features such as preoperative MRI, electroanatomic mapping, and clinical scores like CHA2DS2-VASc. These approaches, while useful, could not account for the dynamic and irregular nature of post-ablation recovery, where patient status evolves through asynchronous events like medication adjustments, cardioversions, and repeat procedures. Recent developments in deep learning, especially in the form of world models, have demonstrated the ability to simulate complex systems by learning latent dynamics from multimodal data. In the context of cardiology, some efforts have focused on static scar quantification or short-term risk prediction, but few have addressed the challenge of long-term, multi-horizon forecasting that incorporates asynchronous, multimodal data streams. This gap is critical because effective management requires understanding how various interventions and physiological changes influence future outcomes over extended periods. The current work builds on this foundation by proposing a structured, event-driven latent state model that captures the evolving anatomy and physiology of the heart, aiming to provide clinicians with a powerful tool for personalized, dynamic risk assessment.

Despite the availability of rich multimodal data, existing predictive models for AF recurrence are limited by their static nature and inability to handle asynchronous interventions. The core problem is to develop a model that can dynamically update the patient's disease state as new evidence arrives, accurately forecast long-term recurrence risk, and support scenario analysis through retrospective input editing. The challenge lies in integrating heterogeneous data types—imaging, procedural geometry, ECG, and clinical events—each with irregular sampling and variable temporal gaps. Moreover, the model must support multi-horizon predictions, enabling clinicians to assess risk at different future points, which is essential for planning follow-up strategies and interventions. Achieving this requires a novel architecture that can learn structured latent representations, evolve them through complex, asynchronous events, and provide interpretable, reliable predictions. Addressing these issues is crucial for advancing personalized medicine in cardiology, reducing recurrence rates, and optimizing resource allocation.

The primary innovations of this work include: 1) The development of a structured 3D spatial latent state that encodes patient anatomy from baseline MRI, pretrained on large datasets for robust generalization. 2) An event-driven state evolution mechanism that incorporates procedural geometry, static covariates, and peri-event ECG embeddings, allowing the model to update patient status asynchronously and reflect real-world clinical workflows. 3) The use of a Transformer encoder to process irregular, asynchronous event sequences, capturing complex temporal dependencies beyond simple Markov assumptions. 4) A multi-horizon forecasting framework based on a horizon-token mechanism, enabling risk predictions at arbitrary future times, supporting clinical decision-making at multiple points. 5) A latent space matching loss that enforces consistency between predicted and actual follow-up states, improving stability and structural accuracy. These innovations collectively enable a dynamic, flexible, and interpretable model tailored for long-term disease trajectory prediction, setting a new standard in medical world modeling.

• �� The process begins with a pre-trained variational autoencoder (VAE) that encodes baseline MRI into a high-dimensional 3D latent state z0, representing the patient's initial anatomy.
• �� Static covariates such as age, gender, and clinical history are embedded into vectors using a learned embedding function ϕs.
• �� The ablation heatmap, derived from procedural geometry, is encoded via a 3D CNN (ϕa) to produce A, capturing spatial scar distribution.
• �� Each clinical event (e.g., medication change, cardioversion, repeat procedure) is mapped to a token ei, which encodes event time, type flags, and peri-event ECG embeddings obtained through mean pooling within a 7-day window before the event.
• �� Starting from the initial latent state z0, the model updates the state iteratively for each valid event i using a residual CNN (fϕ) that incorporates the conditioning vector ci (from the event token and covariates) and the ablation map A. The update formula (equation 4) combines the previous state, event-conditioned residual, and elapsed time drift.
• �� To handle asynchronous event sequences, a Transformer encoder processes the event tokens, producing contextualized embeddings hi, which are broadcast over the spatial grid for the latent state updates.
• �� During training, the model minimizes a composite loss function: latent forecasting loss (matching predicted terminal state to zpost from follow-up MRI), recurrence classification loss (binary cross-entropy), and scar extent regression loss (Huber loss). The latent forecast ensures the model captures structural evolution, while the classification guides risk estimation.
• �� The model supports multi-horizon risk queries by changing the terminal horizon token, enabling clinicians to assess recurrence probabilities at different future time points.
• �� Additionally, the framework allows for retrospective input editing, where the presence or timing of events can be modified to simulate alternative intervention scenarios and evaluate their impact on predicted outcomes.

• �� The experimental setup involves the DECAAF-II dataset, comprising 91 patients with complete pre- and post-ablation MRI, procedural records, ECG, and recurrence labels. An auxiliary cohort of 258 patients without complete ablation geometry was used to evaluate the model’s robustness without spatial information.
• �� The pre-trained MRI encoder was trained on a larger dataset of 732 patients with incomplete records, ensuring generalization of the spatial latent representation.
• �� The main evaluation employed 5-fold cross-validation, measuring AUROC and AUPRC for recurrence prediction, and MAE for scar extent prediction.
• �� Baseline comparisons included static feature models, sequence models (GRU, LSTM, Transformer), and ablation variants removing key components like event information or latent matching.
• �� Ablation studies systematically assessed the impact of conditioning signals, event encoding, and structural priors, confirming the importance of each component.
• �� The model’s ability to perform multi-horizon risk predictions was tested at D30, D60, D90, and beyond, analyzing the evolution of risk over time and the model’s responsiveness to new evidence.

• �� The proposed model achieved AUROC 0.756 and AUPRC 0.777, outperforming static models (AUROC 0.511) and MRI-only models (AUROC 0.525), demonstrating the advantage of dynamic, event-driven state evolution.
• �� It accurately predicted scar extent with a mean absolute error of 2.971 percentage points, close to the oracle baseline (3.189), despite not using follow-up MRI intensities at inference.
• �� Ablation results showed that removing event information caused a significant drop in AUROC (by 0.133), highlighting the importance of asynchronous event modeling.
• �� The risk trajectories over multiple horizons revealed that early recurrence risk increased rapidly within 30 days post-procedure, while late recurrence risk remained relatively stable, aligning with clinical expectations.
• �� The model’s ability to support retrospective input editing demonstrated its potential for scenario analysis, with risk estimates shifting significantly when event timings or presence were altered.

• �� The model can be integrated into clinical workflows for real-time risk monitoring, enabling personalized follow-up schedules and intervention planning.
• �� It supports scenario analysis, allowing clinicians to simulate the impact of different post-procedure interventions, such as medication adjustments or additional procedures.
• �� The approach can be extended to other chronic diseases with irregular event patterns, such as heart failure or stroke, facilitating broader personalized disease management platforms.

• �� The dataset used for training and validation is relatively small and from a single source, limiting the assessment of model generalizability.
• �� Dependence on high-quality multimodal data, including MRI, ECG, and procedural records, poses challenges in real-world settings where data quality varies.
• �� The interpretability of the latent anatomical state remains limited; further efforts are needed to relate latent features to clinical concepts.
• �� Computational complexity may hinder real-time deployment, requiring optimization for clinical environments.