CASE-Bench/CASE-Agent-Traj-And-Bench-Partial

CASE: Clinical Agents for Seeking Evidence — Agent Trajectories & Benchmarks (De-identified)

Partial, de-identified release — companion data for the paper From Given to Gathered Evidence: Agentic Learning for Longitudinal Medical Reasoning: recorded agent trajectories, RL training prompts, benchmark banks and example preview images for longitudinal medical imaging. Full imaging data is not included (UK Biobank controlled); participant identifiers have been replaced by uniform anonymous IDs (case0001–case4739); the ID mapping exists but is not publicly disclosed, in accordance with UK Biobank requirements. Rendered example/preview images are provided in images_examples/ (figure headers cropped).

The CASE project trains three cooperating clinical-agent personas on longitudinal medical imaging through an agentic harness (multi-turn tool calling; the model plans, calls tools, reads observations and answers):

• Anatomist — organ segmentation / volumetry across body stations
• Radiologist — cross-visit image comparison (multi-modality, multi-sequence)
• Consultant — report-grounded synthesis and claim verification

Contents

Path	What
trajectories/{Anatomist_agent,Radiologist_agent,Consultant_agent}/	recorded agent trajectories (agent_caseNNNN.jsonl) and per-patient annotations (dense_caseNNNN.json); Consultant additionally ships skills_caseNNNN.jsonl
rl/verl_train_tool.parquet	reinforcement-learning prompts (verl parquet, one row per patient-question; 30,604 rows)
benchmark/	evaluation question banks: benchmark.parquet = current bank (14,795 rows / 482 patients); test.parquet and unified.parquet = earlier banks; each bank ships its participant ID list
images_examples/	rendered preview slices for the benchmark set: per participant × six body stations × baseline/follow-up visits (4,938 images; figure headers cropped)

Formats

• Trajectories — JSONL, one question instance per line with the full message list. Image content appears only as relative path references (images/<caseNNNN>/img_*.png); the image files themselves are not included.

• Parquet (rl/, benchmark/) — one row per (patient, question); columns:

  • data_source — task origin: vista3d_opsd (Anatomist), medpanel_v2_opsd (Radiologist), medboard_opsd (Consultant)
  • prompt — full chat template: system message (persona, tool definitions, output-format rules) + user message (question, options, tool-use instructions)
  • reward_model — answer key, e.g. {"ground_truth": "D"}
  • extra_info — JSON string with patient_id (anonymous case ID), question_idx, question, options, ground_truth, tools_kwargs (patient-scoped tool arguments) and teacher_insight (teacher reasoning notes used for the OPSD distillation signal)
  • agent_name — tool_agent (multi-turn tool-calling loop)

  Row counts — rl/verl_train_tool.parquet: 30,604 (vista3d 9,788 / medpanel 10,408 / medboard 10,408); benchmark.parquet: 14,795.

  No images are embedded in any parquet: image observations are served to the tools at run time and are not part of this release.

Licensing / data access

Built on UK Biobank-derived data under an access-controlled research agreement.

• Participant IDs — all participant references use uniform anonymous IDs (case0001–case4739). An ID mapping to the original identifiers exists but is not disclosed on this platform, in accordance with UK Biobank requirements.
• Images (images_examples/) — rendered preview images are provided for academic display only; redistribution and commercial use are prohibited.
• Complete dataset — the full dataset is not redistributed here; access must be requested through the UK Biobank Access Management System (AMS) to obtain access permission for the returned data.