Raras-AI/rarebench-br-trajectory

RareBench-BR Trajectory v2（RBT-v2）是首个针对罕见病患者轨迹的基准数据集，设计为自相关免疫，基于44,051个真实的巴西统一医疗系统（SUS，DATASUS）罕见病治疗轨迹构建。数据集包含五个任务：T1（在过渡点预测下一个程序代码）、T2（预测下一个事件是否为变化）、T3（预测患者首次出现的程序）、T4（预测患者是否在随访期内中断治疗）和T5（预测下一次治疗变化的时间）。数据来源于2017年至2021年巴西7个州的高复杂度门诊和孤儿药授权记录，覆盖11种罕见病（如戈谢病、肌营养不良症等）和33种不同的SIGTAP程序代码。数据经过去标识化处理（如年龄分桶、仅保留州级信息），符合巴西伦理法规。数据集提供平衡/分层分割、地理外部测试以及基于计数的强基线模型，旨在促进罕见病进展建模和世界模型能力评估。RareBench-BR Trajectory v2 (RBT-v2) is 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. The dataset includes five tasks: T1 (predicting the next procedure code at transition points), T2 (predicting whether the next event will be a change), T3 (predicting the first occurrence of a procedure), T4 (predicting treatment discontinuation within follow-up), and T5 (predicting time to the next treatment change). Data is sourced from DATASUS APAC-SIA (high-complexity outpatient, orphan-drug authorizations) across 7 Brazilian states from 2017 to 2021, covering 11 rare diseases (e.g., Gaucher, DMD, CF) and 33 distinct SIGTAP procedure codes. It is de-identified (e.g., ages bucketed, UF only) and complies with Brazilian ethics regulations. The benchmark features balanced/stratified splits, a geographic-external test, and strong count-based baselines, aiming to advance rare-disease progression modeling and world-model capabilities.