Raras-AI/rarebench-br-trajectory

RareBench-BR Trajectory v2 (RBT-v2)

The first rare-disease patient-trajectory benchmark designed to be autocorrelation-immune. Built from 44,051 real CNS-linked Brazilian SUS (DATASUS) rare-disease treatment trajectories. Five tasks, balanced/stratified splits, geographic-external test, strong count-based baselines, full datasheet.

Authors: Raras AI · License: CC-BY-NC 4.0 · Contact: dimas@raras.ai Companion architecture: Raras-AI/gemeo-arch

Why this benchmark exists

Patient-trajectory prediction on EHR data is dominated by event autocorrelation: in a rare-disease orphan-drug trajectory, ~82% of events are repeats of the same monthly dispensing code. A model that just copies the patient's last code scores near-perfectly on naive next-event tasks — the documented "repeated event tokens inflate metrics" pitfall (RAVEN, arXiv 2603.24562). There was no public rare-disease trajectory benchmark, and our own first attempt (an L6 next-event track) was 99.5% one class — trivially won by "always repeat".

RBT-v2 is built so that a repeat-last baseline cannot win: the core task scores only transition points (where the trajectory actually changes), binary tasks are balanced 50/50, and we ship the autocorrelation oracle as an explicit baseline so you can see it fail.

The five tasks

Task	Definition	Metric	Why it's hard
T1 — Next-proc @ transition	Given a prefix, predict the next procedure code at a point where the trajectory changes (switch / gap-resumption)	Recall@1/5, MRR	Repeats excluded — the autocorrelation oracle scores only 12.4%
T2 — Will-change	Will the next event be a change vs a continuation? (balanced 50/50)	Balanced acc, AUROC	Majority baseline = 50.0% by construction
T3 — New-onset	Predict the first occurrence of a procedure the patient has never had	Recall@1/5	First-occurrences only (RAVEN-style)
T4 — Discontinuation	Will the patient discontinue therapy (>6-mo gap) within follow-up? (balanced)	Balanced acc, AUROC	Treatment dropout — clinically critical; majority = 50.0%
T5 — Time-to-transition	Months until the next treatment change (right-censored)	C-index, Brier	Genuine time-to-event / world-model capability

Baselines (the bar to beat)

Computed on train, evaluated on test (95% bootstrap CI). Count-based baselines are strong (per arXiv 2511.00782) — beating the bigram is the real bar.

Task	Baseline	test
T1	frequency Top-1	26.9% [25.8, 28.0]
T1	bigram Top-1	64.4% [63.2, 65.6] ← the bar
T1	bigram Top-5	96.0% [95.5, 96.5]
T1	repeat-last (autocorrelation oracle)	12.4% [11.6, 13.2] ← fails by design
T2	majority / always-continue	50.0%
T3	frequency Top-1 / Top-5	17.5% / 57.9%
T4	majority	50.0%

The repeat-last oracle scoring 12.4% (not ~99%) is the proof of autocorrelation-immunity. A real model must beat the bigram, not just copy.

GEMEO leads every novelty & long-context task

The flagship gemeo-sus world model (recurrence-aware) sets the current bar — see LEADERBOARD.md:

Task	GEMEO	Strong baseline	Margin
New-onset prediction (Top-1)	53.7%	38.2% (frequency)	+15.5 pp
T2 — will-change (AUROC)	0.906	0.889 (count-based)	+0.017
T5 — transition-within-12mo (AUROC)	0.827	0.790 (count-based)	+0.037
T4 — treatment discontinuation (AUROC)	0.838	0.696 (count-based)	+0.142

The world model's learned representation pulls clearly ahead on the context-rich tasks that matter most in rare disease. For single-step Markov transitions (T1), the count-based bigram remains near-optimal — the expected task-dependent split (arXiv 2511.00782).

Data provenance & ethics

• Source: DATASUS APAC-SIA (high-complexity outpatient, orphan-drug authorizations), CNS-hash linked. 7 Brazilian states, 2017–2021.
• Cohort: 44,051 patients with ≥5 treatment events; 11 rare diseases (Gaucher, MPS I/II, SMA, DMD, CF, Wilson, Friedreich, Marfan, NF1, Rett, …); 33 distinct SIGTAP procedure codes.
• De-identification: ages bucketed, UF only (no município), CNS hashed, k-anonymity ≥ 5.
• Ethics: Brazilian Res. CNS 466/2012 + 510/2016; LGPD-compliant.
• Splits: patient-level 70/15/15 (train/val/test) on the 5 largest UFs + a geographic-external test on the remaining states (external validity / equity probe; note ext_test is small, n≈22 patients — interpret as a probe, not a powered test).

Files

tasks/
├── T1_next_proc_transition.{train,val,test,ext_test}.jsonl
├── T2_will_change.{...}.jsonl                 (balanced 50/50)
├── T3_new_onset.{...}.jsonl
├── T4_discontinuation.{...}.jsonl             (balanced 50/50)
└── T5_time_to_transition.{...}.jsonl
baselines.json     # all baseline numbers + bootstrap CI
stats.json         # cohort + vocab statistics
DATASHEET.md       # full datasheet-for-datasets

Each case: {case_id, split, orpha, sex, uf, prefix_procs:[...], target:..., ...}.

How to evaluate your model

1. Train on *.train.jsonl (+ *.val.jsonl).
2. Predict on *.test.jsonl (and *.ext_test.jsonl for external validity).
3. Report the metric in the table above with bootstrap 95% CI.
4. You must report the repeat-last and bigram baselines alongside your model — a result that doesn't beat the bigram on T1 is not a positive result.

Citation

@misc{rarebench_br_trajectory_v2_2026,
  title  = {RareBench-BR Trajectory v2: An Autocorrelation-Immune
            Rare-Disease Patient-Trajectory Benchmark from Brazilian SUS},
  author = {Timmers, Dimas and the Raras AI team},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Raras-AI/rarebench-br-trajectory},
  note   = {First public rare-disease trajectory benchmark. CC-BY-NC 4.0.}
}

⚠️ Research only. Not a medical device. Derived from de-identified aggregate SUS data.