ClinicalCode

Animating the Clinical Source Code
of Disease Progression for Medical World Models

Qi Chen1Alan L. Yuille1Jieneng Chen2

1Johns Hopkins University2Stanford University
Four separate parts: CT, parsing with organ and lesion masks, structured ClinicalCode, and simulation of anatomy and lesion evolution.

Overview

Medical world models should describe how disease state evolves under candidate interventions. ClinicalCode uses a typed, executable representation for anatomy, lesions, imaging phenotype, clinical context, time, uncertainty, missingness, and provenance.

Compilers build this state from CT, MRI, and pre-decision observations. An action-conditioned dynamics model rolls the state forward, while validation checks expose inconsistent fields and temporal changes. A calibrated decoder then connects the trajectories to documented treatment actions.

This explicit interface lets researchers inspect individual variables, edit a state, and distinguish errors in perception, state evolution, and treatment decoding. The framework is evaluated across CT treatment planning and longitudinal glioma MRI.

Video demonstration

Explore a liver lesion in 4D

Liver lesion · arterial CT

CT

Parsing

ClinicalCode

{
  "anatomy": {
    "organ": "liver",
    "source": "reference_mask"
  },
  "lesion": {
    "id": "lesion_1",
    "baseline_ml": 72.384,
    "volume_ml": 63.171,    "necrotic_fraction": 0.456,    "viability_proxy": 0.582  },
  "action": "action_1",
  "time_days": 45,
  "uncertainty_proxy": 0.160,  "provenance": {
    "baseline": "CT + masks",
    "slice_index": 32,
    "dynamics": "illustrative"
  }
}

Tumor Growth

45 day
Loading anatomy…Lesion detail
Lesion volume63.2 mL
Necrotic fraction46%
Viability proxy0.58
Day 0306090
Day 45

Results

CT treatment planning, F1 in percent
Method / configurationInternal F1 (%) ↑External F1 (%) ↑

Citation

@misc{chen2026clinicalcode,
  title  = {Animating the Clinical Source Code of Disease
            Progression for Medical World Models},
  author = {Chen, Qi and Yuille, Alan L. and Chen, Jieneng},
  year   = {2026},
  note   = {Research manuscript},
  url    = {https://qic999.github.io/projects/ClinicalCode/}
}