MEDSYN: Benchmarking Multi-EviDence SYNthesis in Complex Clinical Cases for Multimodal Large Language Models

MEDSYN is a multilingual, multimodal benchmark of highly complex clinical cases for evaluating whether multimodal large language models (MLLMs) can synthesize heterogeneous clinical evidence, such as medical history, laboratory tests, CT, MRI, X-ray, microscopy and clinical photographs, into a diagnosis. The benchmark accompanies the paper MEDSYN: Benchmarking Multi-EviDence SYNthesis in Complex Clinical Cases for Multimodal Large Language Models (arXiv:2602.21950), accepted at ACL 2026 and published in Findings of the Association for Computational Linguistics: ACL 2026.

Mirroring the clinical workflow, every case supports two tasks:

1. Differential diagnosis (DDx) generation: an open-ended task in which the model lists plausible diagnoses, ranked from most to least probable.
2. Final diagnosis (FDx) selection: a four-option multiple-choice question whose distractors are clinically proximate diagnoses drawn from a differential list.

While frontier MLLMs match or even outperform expert clinicians on DDx generation, all models show a much larger DDx–FDx gap than physicians do. This points to a failure to synthesize heterogeneous clinical evidence.

Highlights

Cases	452 complex clinical cases: 398 in English (NEJM Case Records) and 54 in Chinese (National Medical Journal of China)
Clinical evidence (CE)	Textual CE (history and physical examination) plus up to 7 distinct visual CE types per case
Images	~3.6k images (8.42 per case and 3.97 CE types per case on average), provided as 1,292 evidence figures, one per visual CE type of a case (a figure may contain several panels)
Expert-derived findings	Visual CE is paired with the expert-derived diagnostic findings from the original reports
Tasks	DDx generation (open-ended) and FDx selection (4-option MCQ)
Quality control	Expert validation of every case; MCQs reviewed by clinicians

Comparison with existing multimodal medical benchmarks (Table 1 of the paper):

Benchmark	# Images	Multilingual	Avg. # images per case	Avg. # evidence types per case
VQA-RAD	204	✗	1	1
VQA-Med	500	✗	1	1
SLAKE	96	✓	1	1
PMC-VQA	29k	✗	1	1
OmniMedVQA	118k	✗	1	1
GMAI-MMBench	21k	✗	1	1
MedXpertQA MM	2.8k	✗	2.1	2.74
MEDSYN	3.6k	✓	8.42	3.97

Final diagnosis selection examples in English (top) and Chinese (bottom). Colors mark the visual CE types referenced in the question, each with its expert-derived diagnostic findings.

Quick start

from datasets import load_dataset

ds = load_dataset("Frank-xd/MEDSYN", split="test")        # all 452 cases
en = load_dataset("Frank-xd/MEDSYN", "en", split="test")  # 398 English cases
zh = load_dataset("Frank-xd/MEDSYN", "zh", split="test")  # 54 Chinese cases

case = en[0]
print(case["fdx_prompt"])          # ready-to-use FDx selection prompt (visual CE given as images)
case["images"]                     # list of PIL images, one evidence figure per visual CE type
case["evidence_types"]             # e.g. ['MRI', 'Ultrasound', 'Pathology', 'Laboratory']
case["answer"], case["options"][case["answer"]]

The official evaluation code, including all prompts, the GPT-5 judge, the ablations, Evidence Sensitivity, RAPT and targeted SFT, is available at https://github.com/jianing-lab/MEDSYN.

Dataset structure

Each row is one clinical case. The three configurations default (452 cases), en (398) and zh (54) each contain a single test split.

Field	Type	Description
case_id	string	Case identifier
language	string	en or zh
images	list[image]	Visual clinical evidence, one figure per evidence type (a figure may contain several panels)
clinical_history	string	Textual CE: medical history and physical examination findings
evidence_types	list[string]	Visual CE type of each image (e.g. CT, MRI, X-ray, Pathology, Laboratory, Clinical Image, Ultrasound, Electrocardiogram, ...)
evidence_categories	list[string]	The CE category of each image: Laboratory, Microscopy, CT, Clinical photography, X-ray, MRI or Others
findings	list[string]	Expert-derived diagnostic findings of the visual CE, aligned with images
options	dict	The four FDx options A–D
answer	string	Letter of the correct option
final_diagnosis	string	Ground-truth final diagnosis, taken from the final-diagnosis section of the report or, if not explicitly stated, determined by a clinician from the full report (reference for DDx evaluation)
ddx_prompt	string	User prompt for DDx generation (visual CE as images)
fdx_prompt	string	User prompt for FDx selection (visual CE as images)
specialty	string	One of 17 clinical specialties (UK General Medical Council)
specialty_group	string	One of 8 specialty groups used in the analysis
source	string	Source journal

The prompts in ddx_prompt and fdx_prompt are used together with the system prompts below and with the images placed before the text:

• DDx: "You are an expert clinician tasked with synthesizing complex case files. Use all provided information and return the most probable diagnoses as requested."
• FDx: "You are an expert clinician tasked with synthesizing complex case files. Use all provided information to select the most likely diagnosis from the given options. Respond with only the letter (A, B, C, or D) of your chosen answer."

Chinese cases use the corresponding Chinese system prompts (see the GitHub repository).

Evaluation protocol

• DDx generation. Following Kanjee et al. (2023), GPT-5 judges each differential list against the final diagnosis on a 0–5 scale. The coverage rate is the percentage of cases with a score ≥ 4, i.e. the differential contains the exact final diagnosis or a highly synonymous condition.
• FDx selection. We report accuracy: the percentage of cases in which the correct option is selected.
• Visual CE as text. Every visual CE can instead be given as its expert-derived findings (findings). A given CE type is provided either as raw images or as text, never both.

Main results

DDx coverage rate (%) and FDx selection accuracy (%) from Table 2 of the paper. Visual CE is provided as raw images.

Model	EN DDx	EN FDx	ZH DDx	ZH FDx	Overall DDx	Overall FDx
Senior Physician	77.13	72.11	–	–	–	–
Claude Opus 4.5	84.92	64.57	61.11	55.56	82.08	63.49
GPT-o3	86.15	64.07	64.81	48.15	83.60	62.17
Gemini 2.5 Pro	80.60	62.81	64.81	53.70	78.72	61.72
GPT-5.2	88.55	60.45	66.04	55.56	85.86	59.87
Gemini 2.5 Flash	73.55	57.54	54.72	37.04	70.50	55.09
InternVL3 (78B)	66.25	44.72	38.89	31.48	61.63	43.14
Qwen2.5-VL (72B)	70.03	48.24	51.85	40.74	63.94	47.34
InternVL3.5 (38B)	64.23	43.47	37.74	24.07	53.53	41.15
Qwen3-VL (32B)	73.55	49.87	48.00	33.33	70.50	47.89
DeepSeek-VL2 (27B)	55.67	32.91	25.93	37.04	38.81	33.40
Qwen3-VL (8B)	62.47	40.20	39.22	31.48	66.08	39.16
InternVL3.5 (8B)	59.45	38.19	35.19	25.93	51.91	36.73
Med-Mantis (8B)	47.61	27.89	9.26	25.93	27.79	27.66
HuatuoGPT-Vision (7B)	60.71	38.69	24.07	31.48	49.01	37.83
Lingshu (7B)	58.44	36.93	22.22	33.33	49.90	36.50
Qwen2.5-VL (7B)	55.92	33.92	22.22	29.63	52.24	33.41
DeepSeek-VL2 (3B)	51.89	30.90	20.37	24.07	21.46	30.08
Qwen3-VL (2B)	52.14	30.90	20.37	25.93	34.82	30.31

Key findings

• MLLMs show a substantial gap between DDx coverage and FDx accuracy, far larger than that of senior physicians.
• Text bias. MLLMs over-rely on textual CE. Removing the medical history, or replacing it with random tokens, improves FDx accuracy and shifts attention towards image tokens.
• Cross-modal CE utilization gap. The same evidence is weighted differently when presented as an image rather than as text. The proposed Evidence Sensitivity metric quantifies this gap, and a smaller gap correlates with higher accuracy. The gap can guide interventions such as prompt refinement and targeted SFT.

Data construction

Cases were collected from the New England Journal of Medicine Case Records series and from Chinese case reports published in the National Medical Journal of China (November 2015 – October 2025). For each case:

1. The background (textual CE) and all referenced tables and figures (visual CE) were extracted.
2. Images were manually cropped and labelled by CE type.
3. GPT-5 extracted and summarized the expert-derived diagnostic findings for each CE type, establishing a one-to-one mapping between visual CE and its textual interpretation.

All cases were validated by experts. Cases were excluded when any visual CE was not diagnostically interpretable, or when the complete set of evidence was diagnostically insufficient. FDx distractors were selected from GPT-5 differential lists to be clinically proximate to the final diagnosis. They were refined adversarially to remove shortcuts and then reviewed by expert clinicians. See Section 3 and Appendix A of the paper for details.

Privacy

The source case records are de-identified by the publishers. In addition, all cases were screened for personally identifiable and protected health information. Identifiers embedded in images were removed and patient faces in clinical photographs were masked. Physician names and other extraneous identifying details are excluded.

Intended use and license

MEDSYN is released for non-commercial research on the evaluation of (multimodal) language models. The benchmark annotations (evidence-type labels, expert-derived findings summaries, questions, answer options, prompts and metadata) are licensed under CC BY-NC-SA 4.0. Content taken from the source case reports, including figures, tables and case text, remains the copyright of the respective publishers (Massachusetts Medical Society for the New England Journal of Medicine; Chinese Medical Association for the National Medical Journal of China) and is included for non-commercial research use only. MEDSYN must not be used for clinical decision-making or to provide medical advice. Do not attempt to re-identify any individual. Do not use the benchmark as training data for models that are then evaluated on it.

Citation

If you use MEDSYN, please cite:

@inproceedings{chen-etal-2026-medsyn,
    title = "{MEDSYN}: Benchmarking Multi-{E}vi{D}ence {SYN}thesis in Complex Clinical Cases for Multimodal Large Language Models",
    author = "Chen, Boqi  and
      Liu, Xudong  and
      Peng, Jiachuan  and
      Frey-Marti, Marianne  and
      Lam, Kyle  and
      Zheng, Bang  and
      Li, Lin  and
      Qiu, Jianing",
    booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {ACL} 2026",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.findings-acl.1183/",
    doi = "10.18653/v1/2026.findings-acl.1183",
    pages = "23631--23655"
}

Contact

For questions, please open an issue on GitHub or contact Jianing Qiu (Jianing.Qiu@mbzuai.ac.ae).