Datasets:

Raras-AI
/

rarebench-br-trajectory

RareBench-BR Trajectory v2 — first autocorrelation-immune rare-disease trajectory benchmark. 5 tasks, 44k patients, baselines (repeat-last oracle fails at 12.4%, bigram bar 64.4%).

82353fe verified 2 days ago

preview code

raw

history blame contribute delete

3.22 kB

	# Datasheet — RareBench-BR Trajectory v2 (RBT-v2)

	Following Gebru et al., Datasheets for Datasets (2021).

	## Motivation
	- Purpose. Provide the first public, autocorrelation-immune benchmark for rare-disease patient-trajectory prediction. Existing EHR trajectory tasks are dominated by event repetition; no rare-disease trajectory benchmark existed.
	- Created by. Raras AI. Funded by the team (bootstrapped). No external sponsor.

	## Composition
	- Instances. Decision points within patient trajectories (prefix → target). 44,051 patients; per-task case counts: T1 ≈ 40k, T2 ≈ 80k (balanced), T3 ≈ 10k, T4 ≈ 7.7k (balanced), T5 ≈ 775k.
	- Each instance = `{case_id, split, orpha, sex, uf, prefix_procs, target/target_months/censored}`.
	- Diseases (11): Gaucher (ORPHA:355), MPS I (579), MPS II (580), SMA (70/83330), DMD (98896), CF (586), Wilson (905), Friedreich (95), Marfan (558), NF1 (636), Rett (778), and the Niemann-Pick/Gaucher cluster (646).
	- Procedures: 33 distinct SIGTAP high-complexity / orphan-drug codes.
	- Labels. Derived deterministically from real recorded events (next procedure, change vs continue, first-occurrence, discontinuation gap, time-to-change). Ground truth is observed fact, not annotation.
	- Missing data. Some events lack age / cost; handled by defaults. Trajectories truncated at study window (2017–2021).
	- Splits. Patient-level (no patient appears in two splits). 70/15/15 on the 5 largest UFs + geographic-external test on remaining states.

	## Collection
	- Source. DATASUS SIASUS APAC-SIA public files, CNS-hash linkage (Tier-1). Pulled for 7 states (BA, PE, CE, MG, RJ, SP, RS), 2017–2021.
	- Sampling. All CNS-linked patients matching the rare-disease CID/ORPHA filter with ≥5 treatment events.

	## Preprocessing
	- Events sorted by (year, month); procedure codes normalized; transitions labeled by procedure change + inter-event gap (>6 mo = discontinuation/gap-resumption); first-occurrence flag per patient.
	- Balancing applied to T2 and T4 (down-sample majority class within each split) so the majority baseline is 50%.

	## Uses
	- Intended. Benchmarking rare-disease trajectory / world models; reporting must include the repeat-last and bigram baselines.
	- Not intended. Clinical decision-making; individual-level inference; commercial use (CC-BY-NC).

	## Distribution & maintenance
	- License. CC-BY-NC 4.0. Hosted on HuggingFace Datasets.
	- PHI. None — only de-identified, k-anonymity ≥ 5 aggregate event sequences; CNS hashed; UF-level geography.
	- Maintainer. Raras AI (dimas@raras.ai). Versioned; v2 supersedes the saturated L6 next-event track.

	## Known limitations
	- ICU-free, outpatient-orphan-drug-centric (APAC). Small procedure vocabulary (33 codes) — bigram Top-5 is high; T1 Top-1 is the discriminative metric.
	- Geographic-external test is small (n≈22 patients) — a probe, not a powered external benchmark.
	- Single-payer Brazilian context; cross-jurisdiction transfer is untested (a feature for equity research, a limit for generalization claims).
	- Structured events only (no notes, labs, genomics) — multimodal onset prediction requires a richer substrate.

	# Datasheet — RareBench-BR Trajectory v2 (RBT-v2)

	Following Gebru et al., Datasheets for Datasets (2021).

	## Motivation
	- Purpose. Provide the first public, autocorrelation-immune benchmark for rare-disease patient-trajectory prediction. Existing EHR trajectory tasks are dominated by event repetition; no rare-disease trajectory benchmark existed.
	- Created by. Raras AI. Funded by the team (bootstrapped). No external sponsor.

	## Composition
	- Instances. Decision points within patient trajectories (prefix → target). 44,051 patients; per-task case counts: T1 ≈ 40k, T2 ≈ 80k (balanced), T3 ≈ 10k, T4 ≈ 7.7k (balanced), T5 ≈ 775k.
	- Each instance = `{case_id, split, orpha, sex, uf, prefix_procs, target/target_months/censored}`.
	- Diseases (11): Gaucher (ORPHA:355), MPS I (579), MPS II (580), SMA (70/83330), DMD (98896), CF (586), Wilson (905), Friedreich (95), Marfan (558), NF1 (636), Rett (778), and the Niemann-Pick/Gaucher cluster (646).
	- Procedures: 33 distinct SIGTAP high-complexity / orphan-drug codes.
	- Labels. Derived deterministically from real recorded events (next procedure, change vs continue, first-occurrence, discontinuation gap, time-to-change). Ground truth is observed fact, not annotation.
	- Missing data. Some events lack age / cost; handled by defaults. Trajectories truncated at study window (2017–2021).
	- Splits. Patient-level (no patient appears in two splits). 70/15/15 on the 5 largest UFs + geographic-external test on remaining states.

	## Collection
	- Source. DATASUS SIASUS APAC-SIA public files, CNS-hash linkage (Tier-1). Pulled for 7 states (BA, PE, CE, MG, RJ, SP, RS), 2017–2021.
	- Sampling. All CNS-linked patients matching the rare-disease CID/ORPHA filter with ≥5 treatment events.

	## Preprocessing
	- Events sorted by (year, month); procedure codes normalized; transitions labeled by procedure change + inter-event gap (>6 mo = discontinuation/gap-resumption); first-occurrence flag per patient.
	- Balancing applied to T2 and T4 (down-sample majority class within each split) so the majority baseline is 50%.

	## Uses
	- Intended. Benchmarking rare-disease trajectory / world models; reporting must include the repeat-last and bigram baselines.
	- Not intended. Clinical decision-making; individual-level inference; commercial use (CC-BY-NC).

	## Distribution & maintenance
	- License. CC-BY-NC 4.0. Hosted on HuggingFace Datasets.
	- PHI. None — only de-identified, k-anonymity ≥ 5 aggregate event sequences; CNS hashed; UF-level geography.
	- Maintainer. Raras AI (dimas@raras.ai). Versioned; v2 supersedes the saturated L6 next-event track.

	## Known limitations
	- ICU-free, outpatient-orphan-drug-centric (APAC). Small procedure vocabulary (33 codes) — bigram Top-5 is high; T1 Top-1 is the discriminative metric.
	- Geographic-external test is small (n≈22 patients) — a probe, not a powered external benchmark.
	- Single-payer Brazilian context; cross-jurisdiction transfer is untested (a feature for equity research, a limit for generalization claims).
	- Structured events only (no notes, labs, genomics) — multimodal onset prediction requires a richer substrate.