RareBench [KDD2024 ADS Track]

RareBench is a pioneering benchmark designed to systematically evaluate the capabilities of LLMs on 4 critical dimensions within the realm of rare diseases. Meanwhile, we have compiled the largest open-source dataset on rare disease patients, establishing a benchmark for future studies in this domain. To facilitate differential diagnosis of rare diseases, we develop a dynamic few-shot prompt methodology, leveraging a comprehensive rare disease knowledge graph synthesized from multiple knowledge bases, significantly enhancing LLMs’ diagnos- tic performance. Moreover, we present an exhaustive comparative study of GPT-4’s diagnostic capabilities against those of specialist physicians. Our experimental findings underscore the promising potential of integrating LLMs into the clinical diagnostic process for rare diseases.

Github Repo for RareBench: https://github.com/chenxz1111/RareBench Arxiv Paper for RareBench: https://arxiv.org/pdf/2402.06341.pdf

How to use it?

Loading Data

from datasets import load_dataset

datasets = ["RAMEDIS", "MME", "HMS", "LIRICAL", "PUMCH_ADM"]

for dataset in datasets:
    data = load_dataset('chenxz/RareBench', dataset, split='test')
    print(data)

Data Format

{
    "Phenotype": "The list of phenotypes presented in HPO codes",
    "RareDisease": "The list of rare diseases code including OMIM, Orphanet and CCRD format",
    "Department": "(Optional) Only provided in PUMCH_ADM"
}

Evaluation

This repository provides data and mapping files for RareBench. Please refer to our github for further automated evaluation.

Source Data

Data Collection and statistics

This study categorizes datasets into two main groups: publicly available datasets and the Peking Union Medical College Hospital (PUMCH) datasets.

Dataset	RAMEDIS	MME	HMS	LIRICAL	PUMCH_ADM
Countries/Regions	Europe	Canada	Germany	Multi-Country	China
#Cases	624	40	88	370	75
#Disease	74	17	39	252	16
#Department	N/A	N/A	N/A	N/A	5
#Cases per disease
--- Minimum	1	1	1	1	3
--- Median	2	1	1	1	5
--- Maximum	82	11	11	19	8
#HPO terms per case
--- Minimum	3	3	5	3	3
--- Median	9	10.5	17.5	11	16
--- Maximum	46	26	54	95	47

Note: The total number of cases in PUMCH is 1,650. We have currently only made public the 75 cases used in the Human versus LLMs experiment.

Data Processing

We apply reasonable filtering criteria to identify and remove cases of low quality that may be caused by recording errors or missing information, such as those with uncertain or imprecise diagnoses and those lacking sufficient relevant information, i.e., fewer than three phenotypes.

Personal and Sensitive Information

Doctors from PUMCH monitored all cases before uploading text information, ensuring the absence of any potential personal information leaks.

Mapping Files

Files in mapping directory, including:

phenotype_mapping.json: HPO phenotype code mapping to term name

disease_mapping.json: OMIM/Orphanet/CCRD code mapping to disease name

ic_dict.json: HPO phenotype terms' Information Content(IC) values obtained from HPO hierarchical structure

phe2embedding.json: HPO phenotype terms' 256 dimension embedding vectors learned by IC-based random walk

Citation

@article{chen2024rarebench,
  title={RareBench: Can LLMs Serve as Rare Diseases Specialists?},
  author={Chen, Xuanzhong and Mao, Xiaohao and Guo, Qihan and Wang, Lun and Zhang, Shuyang and Chen, Ting},
  journal={arXiv preprint arXiv:2402.06341},
  year={2024}
}