GEMEO — A Reference Architecture for Patient World Models

Generative dynamics models of rare-disease patient trajectories. In the lineage of Dreamer, Sora, and Genie — specialised to clinical event streams, grounded in a biomedical knowledge graph, verified by an agentic swarm, and validated on real Brazilian SUS (DATASUS) data.

Research preview. Not a medical device.

Authors: Dimas Timmers (Raras Health), Alexandre Melo Kawassaki (Raras Health; Hospital Israelita Albert Einstein; A.C.Camargo Cancer Center), Joao Bosco Oliveira (Co-Human Genomics). Correspondence: dimas@raras.ai

Cite as: Timmers D, Kawassaki AM, Oliveira JB. GEMEO: the first patient world model for rare disease, grounding generative clinical trajectories in the genome and a biomedical knowledge graph. Zenodo. DOI: 10.5281/zenodo.20092130 (concept; always resolves to the latest version).

---

What this is

GEMEO is not a single model — it is a three-pillar architecture (Propose → Simulate → Verify) for patient world models, plus a family of open instances validated on Brazilian SUS rare-disease data.

Patient history
   │
   ▼  Pillar A — Graph Proposer       KG zero-shot → first-onset candidates
   │     (PrimeKG; Marfan → FBN1, the causal gene)
   ▼  Pillar B — World-Model Scorer   Diffusion-Forcing transformer +
   │     recurrence-aware loss → predicts NOVEL events, not repeats
   ▼  Pillar C — Swarm Verifier       case-adaptive multi-agent panel,
   │     3-valued voting, traceable KG evidence paths
   ▼
Ranked new-onset forecast + intervention plan, with evidence

Targets Level 3 (counterfactual rollout) on the NeurIPS 2025 clinical world-model rubric (arXiv 2511.16333), closing the four gaps that survey names.

Run it on your own data: ADAPTING_TO_A_NEW_DATABASE.md — instantiate gemeo-<your-substrate> on any MEDS v0.4.1 EHR in ~5 min of GPU.

Headline results (baselines on the same candidate space + 95% bootstrap CI)

Task	GEMEO	Strong baseline	Margin
New-onset prediction (Top-1)	53.7%	38.2% (frequency)	+15.5 pp
Will-change (AUROC)	0.906	0.889 (count-based)	+0.017
Transition-within-12mo (AUROC)	0.827	0.790 (count-based)	+0.037
Treatment discontinuation (AUROC)	0.838	0.696 (count-based)	+0.142

GEMEO leads on every novelty and long-context task. The recurrence-aware objective makes the model predict novel events, not repeats — so these are real signal, not autocorrelation. The world model's learned representation pulls clearly ahead exactly where the 2026 EHR literature predicts it should: context-rich tasks like treatment discontinuation (dropout drives bad outcomes in rare disease), where it beats count-based methods by +0.142 AUROC (arXiv 2511.00782).

Repository layout

architecture/    GEMEO Architecture spec v1 + v2, diagram, gemeo_bench.py (conformance CLI)
pillars/         Pillar A (KG proposer) + Pillar C (swarm verifier) + demos
reference_impl/  the Diffusion-Forcing world-model code (AGPL-3.0)
benchmark/       RareBench-BR Trajectory v2 — datasheet, leaderboard, baselines
paper/           the paper (md + pdf) + figures
reproducers/     Modal scripts for every experiment (~$6 total GPU)

Open artifacts

• Architecture + conformance suite: https://huggingface.co/Raras-AI/gemeo-arch
• Flagship model (recurrence-aware): https://huggingface.co/Raras-AI/gemeo-sus
• Application layer (6 inference modes): https://huggingface.co/Raras-AI/gemeo-twin-stack
• Benchmark (first rare-disease trajectory benchmark): https://huggingface.co/datasets/Raras-AI/rarebench-br-trajectory

The six conformance principles (v1) + three pillars (v2)

1. Diffusion-Forcing backbone with per-token σ
2. Gated cross-attention to a real biomedical KG (PrimeKG)
3. MEDS v0.4.1 event substrate
4. Bootstrap-then-learn pattern per inference mode
5. Bidirectional health-system grounding (PCDT / formulary)
6. Audit-driven training
7. (v2) Recurrence-aware onset objective (defeats autocorrelation)
8. (v2) Competing-risks first-onset head
9. (v2) KG zero-shot onset proposer
10. (v2) Agentic verification with traceable evidence

Run python architecture/gemeo_bench.py check <checkpoint> to test conformance.

Licensing & scope

Open the recipe, keep the spice. See LICENSING.md for full terms.

• Source code (architecture, pillars, reference implementation, reproducers): AGPL-3.0 — adopt and build on it freely, with attribution. A reference architecture only becomes a standard if people can use it.
• Model weights (gemeo-sus, …) and the RareBench-BR benchmark: CC-BY-NC 4.0 on Hugging Face — no commercial reuse of our trained artifacts/data without a separate agreement.
• Held back on purpose (not in this repo): the proprietary DATASUS ETL (raw SIH/SIA/APAC/SIM extraction, CNS-hash linkage, trajectory construction), the cohort/preprocessing heuristics that confer an advantage, and the future multimodal (Mayo) substrate. The public meds_export.py consumes an already-built trajectory file — it does not produce one.

Citation

@misc{gemeo_2026,
  title  = {GEMEO:             Validated on Brazilian SUS Data},
  author = {Timmers, Dimas and Kawassaki, Alexandre and the Raras AI team},
  year   = {2026},
  url    = {https://github.com/rarasAI/gemeo}
}

Builds on: Diffusion Forcing (Chen NeurIPS 2024), PrimeKG (Chandak Nature 2023), RAVEN (arXiv 2603.24562), CAMP (arXiv 2604.00085), DeepRare (Nature 2026), the clinical world-model rubric (arXiv 2511.16333).

⚠️ Research only. Not a medical device. No clinical use without physician oversight and applicable regulatory clearance. Derived from de-identified, k-anonymity ≥ 5 aggregate SUS data.