🤖 machine learning

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

This paper proposes an intervention-aware clinical world model that encodes baseline cardiac imaging into a structured latent state and evolves it through asynchronous post-procedural events to accurately forecast atrial fibrillation recurrence and scar extent within a 90-day recovery window without requiring follow-up MRI data at inference.

Original authors: Yunsung Chung, Yingshuo Liu, Abboud F. Hassan, Han Feng, Mary M. Maleckar, Nassir Marrouche, Jihun Hamm

Published 2026-08-14

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Yunsung Chung, Yingshuo Liu, Abboud F. Hassan, Han Feng, Mary M. Maleckar, Nassir Marrouche, Jihun Hamm

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict the weather. A simple forecast might look at the sky right now and guess if it will rain tomorrow. But what if the weather is actually a chaotic, shifting story? What if a sudden gust of wind, a change in temperature, or a cloud forming at a weird time completely changes the outcome? In the world of medicine, doctors often try to predict a patient's future health by looking at a single snapshot—like an MRI scan taken before a surgery—and guessing what will happen months later. But the human body isn't a static photo; it's a movie. After a procedure, patients take new medicines, have follow-up tests, or even need extra treatments at irregular times. These events are like plot twists in a story that can change the ending. The big question is: how do we build a computer model that doesn't just look at the first page of the book, but keeps reading the story as it unfolds, updating its prediction every time a new event happens?

This is exactly what a team of researchers at Tulane University and Simula Research Laboratory set out to solve. They created something they call an "intervention-aware clinical world model." Think of it as a digital twin of a patient's heart that doesn't just sit still. Instead, it starts with a 3D map of the heart's anatomy (taken from an MRI before surgery) and then "lives" through the recovery period. As the patient takes medication, has an electrical shock to reset their heart rhythm, or gets a repeat procedure, the model updates its internal state. It's like a video game character that changes its stats and appearance based on the items they pick up and the battles they fight. The researchers tested this on patients who had surgery for atrial fibrillation (a type of irregular heartbeat). They found that by watching these "plot twists" in the first 90 days after surgery, their model could predict who would have their heart rhythm go wrong again within a year much better than older methods. Specifically, the model achieved a score of 0.756 (called AUROC) and 0.777 (called AUPRC) for predicting recurrence, and it could even guess how much scar tissue would be left on the heart with an error of just 2.971 percentage points, all without needing to see a follow-up MRI scan at the time of the prediction.

The Problem: The "One-Shot" Guess

For a long time, medical AI has been a bit like a fortune teller who looks at your hand once and tells you your whole future. If you get a heart surgery, doctors often take a picture of your heart before the operation and use that to guess if you'll be fine a year later. But this ignores the messy reality of recovery. Patients don't recover on a perfect schedule. They might take a new pill on day 10, have a scare on day 25, or need a second procedure on day 60. These events happen at irregular times, and they change the risk. Old models usually treat the whole recovery as a single step, ignoring the drama of the middle chapters. They miss the fact that the story changes as new evidence arrives.

The Solution: A Living Digital Twin

The researchers proposed a new way to think about this. Instead of a static guess, they built a "world model." Imagine a video game where you have a character (the patient).

The Starting State: The game begins with a detailed 3D map of the character's heart, created from a pre-surgery MRI. This is the "latent state"—a compressed, digital version of the heart's anatomy.
The Plot Twists: As time passes, the game receives "events." These are real-world things like a medication change, an electrical cardioversion (a shock to reset the heart), or a repeat procedure. The model also looks at ECG readings (heart rhythm traces) taken right before these events.
The Update: Every time an event happens, the model updates the character's internal state. It's not just adding a note to a file; it's actually shifting the 3D digital heart to reflect what that event might do. If a patient takes a new drug, the model simulates how that might change the heart's risk profile.
The Forecast: The model can pause the story at any point (like day 30, day 60, or day 90) and ask, "Based on everything that has happened so far, what is the chance of the heart rhythm failing again in the next year?"

How It Works (The Magic Behind the Curtain)

The model uses a special kind of AI called a Transformer (the same type of technology that powers many modern chatbots) to read the timeline of events. It treats the patient's history like a sentence, where each word is a medical event. It also uses a "latent matching" trick. During training, the model is shown the actual follow-up MRI of the patient. It tries to predict what that future heart would look like based only on the pre-surgery scan and the events that happened in between. If it gets the "future heart" right, it knows it understands the story well. This helps it learn to predict the outcome (recurrence) and the amount of scar tissue without needing to see the future scan when it's actually being used.

What They Found

The team tested this on a group of 91 patients who had complete records (including the tricky 3D maps of where the surgery was done).

Better Predictions: Their model was significantly better at predicting who would have a recurrence of atrial fibrillation than older methods. While simple models that just looked at the pre-surgery scan or the events alone struggled (scoring around 0.51 to 0.65), this new model scored 0.756 for accuracy (AUROC) and 0.777 for precision (AUPRC).
Scar Prediction: They also asked the model to guess how much scar tissue would be on the heart after healing. Even without seeing the follow-up MRI, it guessed the scar extent with an error of only 2.971 percentage points. This is almost as good as if it had been given the follow-up scan to look at (which scored 3.189).
The "What If" Power: One of the coolest features is that the model can look back and edit the story. You can tell the model, "What if this patient hadn't taken that medication?" or "What if the repeat procedure happened two weeks later?" The model then recalculates the risk. This doesn't prove that changing the treatment will work (it's not a time machine), but it shows the model understands how those events are linked to the outcome.

The Limits and the Future

The researchers are careful to say this isn't a magic cure-all yet. The group of patients they tested on was relatively small (91 people), and all the data came from one specific study (DECAAF-II). They also noted that the 3D maps of the surgery were hard to make and sometimes noisy. When they tried to run the model on a larger group of 258 patients who didn't have those detailed 3D maps, the model still worked well (scoring 0.713), but it couldn't prove how much the 3D maps helped in that specific group.

The authors suggest that while this model is a big step forward in understanding how to predict heart recovery by watching the story unfold, we need more data and bigger tests before we can use it to make real-life decisions about changing treatments. For now, it's a powerful tool that shows us that to predict the future of a patient's heart, we have to pay attention to every twist and turn of their recovery, not just the beginning.

Technical Summary: Intervention-Aware Clinical World Model for Post–Op Outcome Forecasting in Cardiology

Problem Statement
Current clinical prediction models often treat post-intervention outcomes as a static, one-step mapping from baseline measurements to a future endpoint. However, patient recovery, particularly following cardiac procedures, is a dynamic process characterized by irregular trajectories. Clinical observations, medication adjustments, repeat interventions, and physiological measurements occur asynchronously and can significantly alter risk assessments over time. Specifically, in atrial fibrillation (AF) ablation, the 90-day "blanking period" involves irregular events (e.g., medication changes, cardioversion) that are rarely integrated with baseline imaging to forecast long-term recurrence. Existing models fail to update predictions dynamically as longitudinal evidence accrues during this critical window.

Methodology
The authors propose an intervention-aware clinical world model that represents each patient via a structured 3D spatial latent state, evolving this state over time based on time-ordered post-intervention events. The framework consists of three primary components:

Latent Anatomical State Initialization:
- Pre-ablation Late Gadolinium Enhancement (LGE) MRI ( $x_0$ ) is encoded into a 3D spatial latent state $z_0$ using a frozen Variational Autoencoder (VAE) pretrained on a larger cohort.
- A 3D Convolutional Neural Network (CNN) encodes the ablation heatmap ( $A$ ) into a spatial representation.
Event-Conditioned Latent Dynamics:
- The model processes a sequence of tokens including static covariates, the ablation anchor, and time-ordered clinical events (medication, cardioversion, repeat procedures) occurring during the 90-day blanking period.
- Each event token $e_i$ encodes normalized time, event flags, medication categories, and peri-event ECG embeddings (derived from ECGFounder models recorded within 7 days prior to the event).
- A Transformer encoder contextualizes these tokens.
- The latent state is updated iteratively. For each valid token $i$ , the state $\hat{z}_i$ is updated from $\hat{z}_{i-1}$ using a 3D residual CNN ( $f_\phi$ ) conditioned on the event context and a multilayer perceptron ( $g_\phi$ ) that models context-dependent drift over elapsed time ( $\Delta t_i$ ). This allows the model to simulate the structural trajectory of the atrium without requiring follow-up imaging at inference.
Forecasting and Training Objective:
- Targets: The model forecasts two outcomes at a selected query horizon $t_T$ : (1) AF recurrence probability ( $y$ ) within 451 days, and (2) scar extent ( $b$ ) as a percentage of scar-positive voxels.
- Training: The model is trained using a composite loss function:
  - Latent Matching Loss: Minimizes the distance between the predicted terminal latent state $\hat{z}_T$ and the ground-truth follow-up latent state $z_{post}$ (derived from follow-up MRI, used only for training).
  - Classification Loss: Binary cross-entropy for recurrence prediction.
  - Regression Loss: Huber loss for scar extent prediction.
- Inference: The model requires only pre-ablation MRI, the ablation heatmap, static covariates, and available pre-event ECGs. It can query risk at any horizon by adjusting the terminal horizon token.

Key Contributions

Intervention-Aware Latent World Model: A novel architecture that evolves a 3D anatomical state conditioned on ablation geometry, irregular post-procedural events, and ECG embeddings, moving beyond static baseline predictions.
Horizon-Token Formulation: A mechanism enabling "anytime" recurrence forecasting, allowing clinicians to query risk estimates based on partial blanking-period histories up to a specific time point.
Comprehensive Evaluation: Internal validation on the DECAAF-II complete-record cohort ( $N=91$ ) and a larger auxiliary cohort ( $N=258$ ) lacking usable ablation geometry, demonstrating the model's robustness across different data availabilities.

Experimental Results
The model was evaluated using 5-fold cross-validation on the DECAAF-II cohort:

Recurrence Prediction: The proposed model achieved an AUROC of 0.756 and AUPRC of 0.777. This represents a significant improvement over matched-input sequence baselines (e.g., LSTM: AUROC 0.653) and static-only models.
Scar Extent Prediction: The model achieved a Mean Absolute Error (MAE) of 2.971 percentage points for scar extent. Notably, this was achieved without using follow-up MRI intensities during inference, performing competitively against an oracle-style baseline that uses follow-up MRI (MAE 3.189).
Ablation Studies: Removing event data or the latent matching loss resulted in the largest performance drops, confirming that state evolution and structural supervision are critical. Removing the ablation map or ECG embeddings had smaller but still measurable impacts.
Auxiliary Cohort: In a cohort without ablation geometry ( $N=258$ ), the model variant achieved an AUROC of 0.713, outperforming sequence baselines by a substantial margin, suggesting the event-conditioned rollout signal is effective even without spatial geometry.
Risk Trajectories: The model demonstrated the ability to differentiate risk profiles as evidence accumulates (e.g., separating early vs. late recurrence risks from day 30 to day 90).

Significance and Claims
The paper claims that this framework addresses a critical gap in clinical modeling by treating recovery as a dynamic, irregular event sequence rather than a static snapshot. By coupling 3D imaging with procedural context and physiological data, the model provides:

Dynamic Risk Updating: The ability to refine recurrence risk estimates as new clinical data (medication changes, cardioversion) becomes available during the blanking period.
Structural Forecasting: The capacity to predict post-procedural scar extent without requiring follow-up MRI scans at the time of inference, potentially reducing clinical burden.
Retrospective Analysis: The ability to perform "what-if" scenarios by editing blanking-period records to observe changes in predicted risk, offering a tool for understanding the sensitivity of risk to specific interventions.

The authors maintain a modest stance, noting that the complete-record cohort is small ( $N=91$ ) and results are internal to DECAAF-II. They emphasize that while input edits show model sensitivity, they do not establish causal treatment effects. The work is presented as a step toward causal modeling of post-ablation interventions, pending larger cohorts and external validation.