ClinSeek-Bench

ClinSeek-Bench is the evaluation suite introduced in ClinSeekAgent: Automating Multimodal Evidence Seeking for Agentic Clinical Reasoning. It has two parts:

• ClinSeek-Bench-Text: 1,800 text-only EHR questions in 13 task categories.
• ClinSeek-Bench-MM: 989 multimodal EHR + chest X-ray (CXR) questions in 6 task categories.

Every question is evaluated under two paired settings that share the task definition, the prediction time, and the answer label:

• Curated Input: the model answers in one call from the evidence package of the source benchmark, rendered from the patient record available at the prediction time.
• ClinSeekAgent: no evidence is provided. The agent receives the task, the patient identifier, and the prediction time, and must retrieve evidence itself from the patient's raw EHR (and CXR studies, for ClinSeek-Bench-MM) with tools.

This repository releases the metadata needed to rebuild ClinSeek-Bench locally. It does not contain patient records. See Reconstructing the benchmark.

What is released

ClinSeek-Bench is built from credentialed clinical datasets (MIMIC-IV, MIMIC-IV-Note, MIMIC-IV-ED, MIMIC-CXR, MIMIC-CXR-JPG) and from benchmarks derived from them (EHR-Bench, EHRXQA, MedMod). These must be obtained under each user's own credentialed access and data-use agreements. This repository therefore does not redistribute raw MIMIC tables, patient databases, serialized patient timelines, clinical notes, radiology report text, CXR images, or model trajectories.

File	Rows	Contents
ClinSeek-Bench_text.json	1,800	qid, subject_id, hadm_id, prediction_time, latest_event_time, task, task_type, category, instruction, candidates, label, output, task_info, source_file, source_idx; target_hadm_id / target_stay_id for diagnosis-code tasks
ClinSeek-Bench_multimodal.jsonl	989	qid, task, category, question, answer_type, ground_truth, subject_id, hadm_id, stay_id, prediction_time, modalities, CXR study_ids / dicom_ids / image and report path references, scope, snapshot_spec, leakage_policy_version, source references

source_file and source_idx identify each text question's row in the public EHR-Bench release, so the original EHR-Bench input can be joined back if needed. The ClinSeekAgent evaluation does not use it: both settings read the patient record from a locally rebuilt database.

Temporal policy fields

Each question carries a prediction time. At run time, ClinSeekAgent builds a question-specific view of the patient record that contains only what was available in the EHR before that time, and both settings use the same view.

• prediction_time: the cutoff for the question-specific record view.
• target_hadm_id / target_stay_id (80 + 80 text rows): the admission or ED stay whose codes are being predicted. Diagnosis and procedure codes of this encounter are hidden regardless of the cutoff.
• scope, study_ids, dicom_ids, snapshot_spec, leakage_policy_version (multimodal rows): which patient and which linked CXR studies the tools may access, and the rule set used to build the view. The runner rejects multimodal rows that lack these fields.

The full policy is described in Appendix "Temporal Policy" of the paper.

Benchmark composition

ClinSeek-Bench-Text samples 40 questions from each of the 45 EHR-Bench subtasks of EHR-R1 (1,800 questions, 1,563 patients): 720 risk-prediction and 1,080 decision-making questions.

ClinSeek-Bench-MM adapts EHRXQA (497 questions) and MedMod (492 questions), both grounded in MIMIC-IV and MIMIC-CXR.

Benchmark	Category	N	Tasks
Text, risk prediction	Mortality	280	In-hospital death from an ED visit or an inpatient stay; death within 1, 2, 3, 7, or 14 days in the ICU
	Readmission	160	ED return within 72 h; ICU readmission in the same hospitalization; hospital readmission within 30 or 60 days
	Length of stay	160	ICU stay > 7 or 14 days; hospital stay > 3 or 7 days
	Escalation of care	120	ED patient hospitalized, transferred to the ICU within 12 h, or critical outcome
Text, decision making	Diagnosis	160	Next diagnoses of an admission or ED visit (ICD titles or CCS categories)
	Medication	320	Next administered medication, prescriptions, ED medication reconciliation, ED dispensing-cabinet items, ICU input and ingredient items
	Procedure	120	Next procedures (ICD titles or CCS categories) and ICU procedure items
	Laboratory & microbiology	80	Next laboratory and microbiology tests
	Vitals & measurements	160	Next ICU charted, datetime, and output items; next outpatient measurements
	Imaging report	40	Next radiology examination
	Orders	40	Type of the next provider order
	Encounter & flow	120	Next admission type, clinical service, and transfer care unit
	Next event	40	Which EHR table the next event is recorded in
MM	CXR finding presence	215	Whether a finding is present in a CXR study (EHRXQA)
	CXR finding enumeration	182	List the findings of a CXR study (EHRXQA and MedMod)
	CXR change comparison	222	How findings change between two CXR studies (EHRXQA)
	Clinical deterioration	125	ICU death within the next 24 hours (MedMod)
	Mortality	125	Death during the current ICU stay (MedMod)
	Phenotype	120	Acute and chronic conditions of an ICU stay (MedMod)

Results

F1 (%) from the paper. Δ = ClinSeekAgent − Curated Input, paired on questions; bold Δ means the paired-bootstrap 95% CI excludes zero.

Overall

Model	Benchmark	Curated Input	ClinSeekAgent	Δ
Claude Opus 5	ClinSeek-Bench-Text	61.7	63.9	+2.2
GPT-5.6 Sol	ClinSeek-Bench-Text	57.8	59.3	+1.6
Claude Opus 5	ClinSeek-Bench-MM	66.5	68.8	+2.3
GPT-5.6 Sol	ClinSeek-Bench-MM	60.0	62.5	+2.5
Claude Opus 4.6	ClinSeek-Bench-Text	58.3	56.6	−1.7
GLM-5.2	ClinSeek-Bench-Text	51.0	51.7	+0.7
Kimi K2.6	ClinSeek-Bench-Text	50.6	49.9	−0.7
Claude Opus 4.6	ClinSeek-Bench-MM	57.2	62.4	+5.1
Kimi K2.6	ClinSeek-Bench-MM	55.7	57.8	+2.0

By category (Claude Opus 5 and GPT-5.6 Sol)

Benchmark	Category	N	Opus 5 Curated	Opus 5 Agent	Δ	Sol Curated	Sol Agent	Δ
Text	Mortality	280	90.7	90.4	−0.4	87.1	86.1	−1.1
	Readmission	160	63.7	63.7	0.0	57.5	63.1	+5.6
	Length of stay	160	84.4	83.8	−0.6	81.2	82.5	+1.2
	Escalation of care	120	84.2	82.5	−1.7	80.8	82.5	+1.7
	Diagnosis	160	36.4	44.7	+8.3	33.7	36.7	+3.0
	Medication	320	55.7	60.8	+5.1	51.1	53.5	+2.4
	Procedure	120	40.7	43.0	+2.3	40.6	38.2	−2.4
	Laboratory & microbiology	80	63.8	55.6	−8.2	44.5	55.6	+11.2
	Vitals & measurements	160	37.1	48.6	+11.5	38.6	35.5	−3.2
	Imaging report	40	60.0	51.9	−8.1	44.2	45.4	+1.2
	Orders	40	31.8	45.2	+13.3	28.5	34.3	+5.8
	Encounter & flow	120	63.3	63.3	0.0	60.8	64.2	+3.3
	Next event	40	27.5	18.8	−8.7	27.5	22.5	−5.0
MM	CXR finding presence	215	77.3	78.2	+0.9	73.9	76.3	+2.3
	CXR finding enumeration	182	45.1	48.5	+3.4	35.2	34.2	−0.9
	CXR change comparison	222	61.0	59.5	−1.4	52.2	52.5	+0.3
	Clinical deterioration	125	95.2	96.8	+1.6	96.8	96.8	0.0
	Mortality	125	84.8	87.2	+2.4	86.4	88.0	+1.6
	Phenotype	120	40.9	51.5	+10.5	21.4	37.2	+15.8

Δ is computed before rounding, so it can differ by 0.1 from the difference of the rounded scores.

Files

ClinSeek-Bench/
├── README.md
├── ClinSeek-Bench_text.json
└── ClinSeek-Bench_multimodal.jsonl

Reconstructing the benchmark

Reconstruction and evaluation code is in the ClinSeekAgent repository:

• Code: https://github.com/UCSC-VLAA/ClinSeekAgent
• Text data preparation: docs/ClinSeek-Bench_text_data_prepare.md
• ClinSeekAgent text evaluation: docs/ClinSeek-Bench_text_evaluation.md
• Curated Input text evaluation: docs/ClinSeek-Bench_text_curated_input_evaluation.md
• Multimodal data preparation and evaluation: docs/ClinSeek-Bench_multimodal_data_prepare.md, docs/multimodal_benchmark.md

git clone https://huggingface.co/datasets/UCSC-VLAA/ClinSeek-Bench
git clone https://github.com/UCSC-VLAA/ClinSeekAgent.git

Then download the source datasets below and follow the guides.

Required source data

Dataset	Official source	Used for
MIMIC-IV	https://physionet.org/content/mimiciv/	structured EHR tables
MIMIC-IV-Note	https://physionet.org/content/mimic-iv-note/	discharge and radiology notes
MIMIC-IV-ED	https://physionet.org/content/mimic-iv-ed/	emergency department tables
MIMIC-CXR	https://physionet.org/content/mimic-cxr/	CXR metadata and report text
MIMIC-CXR-JPG	https://physionet.org/content/mimic-cxr-jpg/	linked CXR JPG files
EHRXQA	https://physionet.org/content/ehrxqa/	source CXR question-answering rows
MedMod	https://github.com/nyuad-cai/MedMod	source ICU/CXR prediction rows
EHR-Bench (EHR-R1)	https://github.com/MAGIC-AI4Med/EHR-R1	source text-only EHR tasks

Expected counts

• ClinSeek-Bench-Text: 1,800 rows, 45 subtasks, 13 categories, 1,563 patients, 80 rows with target_hadm_id, 80 rows with target_stay_id.
• ClinSeek-Bench-MM: 989 rows (497 EHRXQA, 492 MedMod), 6 categories, scope = patient_scope and leakage_policy_version = multimodal-strict-temporal-scope-v1 on every row.

Changelog

• 2026-10 (arXiv v2). Added category to both files. Added target_hadm_id / target_stay_id to text rows and scope, study_ids, dicom_ids, snapshot_spec, leakage_policy_version to multimodal rows, as required by the temporal policy of the current ClinSeekAgent code. Removed the serialized EHR-Bench input (input, question) from the text file; it can be joined back from EHR-Bench via source_file / source_idx. Results updated to the revised paper. The May 2026 files remain available under the v1 tag.
• 2026-05 (arXiv v1). Initial release.

Citation

@misc{wu2026clinseekagent,
  title         = {ClinSeekAgent: Automating Multimodal Evidence Seeking for Agentic Clinical Reasoning},
  author        = {Wu, Juncheng and Zhang, Letian and Wang, Yuhan and Tu, Haoqin and Chen, Hardy and Wang, Zijun and Xie, Cihang and He, Yufan and Zhao, Can and Xu, Daguang and Tang, Yucheng and Zhou, Yuyin},
  year          = {2026},
  eprint        = {2605.20176},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2605.20176}
}

License and responsible use

The metadata in this repository is released under Apache-2.0. The underlying clinical data remain governed by the PhysioNet credentialed data-use agreements of their source datasets; reconstructing the benchmark requires your own credentialed access.

ClinSeek-Bench is for research on clinical evidence seeking. It is not a medical device and must not be used for clinical diagnosis, treatment, triage, or patient management.