MedConceptsQA is a dedicated open source benchmark for medical concepts question answering. The benchmark comprises of questions of various medical concepts across different vocabularies: diagnoses, procedures, and drugs. The questions are categorized into three levels of difficulty: easy, medium, and hard. We conducted evaluations of the benchmark using various Large Language Models. Our benchmark serves as a valuable resource for evaluating the abilities of Large Language Models to interpret medical codes and distinguish between medical concepts. Our findings showed most of the state-of-the-art CLLMs, despite being pre-trained on medical data, achieved accuracy levels close to random guessing on this benchmark. However, general-purpose models (Llama3-70B and GPT-4) outperformed CLLMs. Notably, GPT-4 exhibited the best performance, although its accuracy remained insufficient for certain datasets in our benchmark.
Highlights:
- MedConceptsQA is an open-source benchmark for medical concepts question answering.
- Covers three difficulty levels and includes diagnoses, procedures, and medications.
- Shows challenges Clinical LLMs face in interpreting and distinguishing medical codes.
- Most of Clinical LLMs performed similarly to random guessing on the benchmark.
- General LLMs outperformed Clinical LLMs, despite their focus is not the medical domain.
Task Type | Language | Val | Test | File Format | Size |
---|---|---|---|---|---|
QA | English | 60 | 820,000 | .arrow | 300MB |
Vocabulary | Level | Number of Questions |
---|---|---|
ATC | Easy | 6,440 |
ATC | Medium | 6,440 |
ATC | Hard | 5,938 |
ICD10-CM | Easy | 94,580 |
ICD10-CM | Medium | 81,757 |
ICD10-CM | Hard | 88,013 |
ICD10-PROC | Easy | 190,987 |
ICD10-PROC | Medium | 190,987 |
ICD10-PROC | Hard | 88,582 |
ICD9-CM | Easy | 17,736 |
ICD9-CM | Medium | 17,736 |
ICD9-CM | Hard | 16,858 |
ICD9-PROC | Easy | 4,670 |
ICD9-PROC | Medium | 4,670 |
ICD9-PROC | Hard | 4,438 |
The MedConceptsQA dataset is organized into several structured files, each containing medical question-answering instances designed for training and evaluation of models in the medical domain. Here’s a brief overview of the dataset structure:
MedConceptsQA
│
├── atc_easy
│ ├── dev-00000-of-00001.parquet
│ └── test-00000-of-00001.parquet
│
├── atc_hard
│ ├── dev-00000-of-00001.parquet
│ └── test-00000-of-00001.parquet
│
├── icd10cm_easy
│ ├── dev-00000-of-00001.parquet
│ └── test-00000-of-00001.parquet
│
├── icd10cm_hard
│ ├── dev-00000-of-00001.parquet
│ └── test-00000-of-00001.parquet
│
├── icd10proc_easy
│ ├── dev-00000-of-00001.parquet
│ └── test-00000-of-00001.parquet
│
└── icd10proc_hard
├── dev-00000-of-00001.parquet
└── test-00000-of-00001.parquet
Ofir Ben Shoham (Ben-Gurion University, Israel)
Nadav Rappoport (Ben-Gurion University, Israel)
Official Website: https://nadavlab.github.io/MedConceptsQA-website/
Download Link: https://huggingface.co/datasets/ofir408/MedConceptsQA
Article Address: https://www.sciencedirect.com/science/article/pii/S0010482524011740
Publication Date: 09/2024. Published in the journal: Computers and Biology in Medicine.
@article{SHOHAM2024109089,
title = {MedConceptsQA: Open source medical concepts QA benchmark},
journal = {Computers in Biology and Medicine},
volume = {182},
pages = {109089},
year = {2024},
issn = {0010-4825},
doi = {https://doi.org/10.1016/j.compbiomed.2024.109089},
url = {https://www.sciencedirect.com/science/article/pii/S0010482524011740},
author = {Ofir Ben Shoham and Nadav Rappoport}
Original introduction article is here.