File size: 3,263 Bytes
fef61f9
 
 
 
 
 
 
 
 
 
 
 
 
 
d4dbb9a
ba1fe29
fef61f9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f73251d
 
 
fef61f9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
---
task_categories:
- question-answering
language:
- en
tags:
- medical
- question answering
- large language model
- retrieval-augmented generation
size_categories:
- 100K<n<1M
---
# The StatPearls Corpus in MedRAG

This HF dataset contains the information of the StatPearls corpus used in [MedRAG](https://arxiv.org/abs/2402.13178). It can be used for medical Retrieval-Augmented Generation (RAG).

According to the [privacy policy](https://www.statpearls.com/home/privacypolicy/) of StatPearls, **we are not allowed to distribute the StatPearls content**. Please download the raw data from [NCBI Bookshelf](https://www.ncbi.nlm.nih.gov/books/NBK430685/) and use [our code](https://github.com/Teddy-XiongGZ/MedRAG/blob/main/src/data/statpearls.py) for the chunking.

## Dataset Details

### Dataset Descriptions

[StatPearls](https://www.statpearls.com/) is a point-of-the-care clinical decision support tool similar to [UpToDate](https://www.uptodate.com/). We use the 9,330 publicly available StatPearl articles through [NCBI Bookshelf](https://www.ncbi.nlm.nih.gov/books/NBK430685/) to construct the StatPearls corpus. 
We chunked StatPearls according to the hierarchical structure, treating each paragraph in an article as a snippet and splicing all the relevant hierarchical headings as the corresponding title.
Our chunked corpus contains 301,202 snippets with an average of 119 tokens.

### Dataset Structure
Each row is a snippet of StatPearls, which includes the following features:

- id: a unique identifier of the snippet
- title: the title and subtitles of the StatPearl article from which the snippet is collected
- content: the content of the snippet
- contents: a concatenation of 'title' and 'content', which will be used by the [BM25](https://github.com/castorini/pyserini) retriever

## Uses

<!-- Address questions around how the dataset is intended to be used. -->

### Direct Use

<!-- This section describes suitable use cases for the dataset. -->

```shell
git clone https://github.com/Teddy-XiongGZ/MedRAG.git
cd MedRAG

wget https://ftp.ncbi.nlm.nih.gov/pub/litarch/3d/12/statpearls_NBK430685.tar.gz -P ./corpus/statpearls
tar -xzvf ./corpus/statpearls/statpearls_NBK430685.tar.gz -C ./corpus/statpearls
python src/data/statpearls.py
```

### Use in MedRAG

```python
>> from src.medrag import MedRAG

>> question = "A lesion causing compression of the facial nerve at the stylomastoid foramen will cause ipsilateral"
>> options = {
    "A": "paralysis of the facial muscles.",
    "B": "paralysis of the facial muscles and loss of taste.",
    "C": "paralysis of the facial muscles, loss of taste and lacrimation.",
    "D": "paralysis of the facial muscles, loss of taste, lacrimation and decreased salivation."
}

>> medrag = MedRAG(llm_name="OpenAI/gpt-3.5-turbo-16k", rag=True, retriever_name="MedCPT", corpus_name="StatPearls")
>> answer, snippets, scores = medrag.answer(question=question, options=options, k=32) # scores are given by the retrieval system
```

## Citation
```shell
@article{xiong2024benchmarking,
    title={Benchmarking Retrieval-Augmented Generation for Medicine}, 
    author={Guangzhi Xiong and Qiao Jin and Zhiyong Lu and Aidong Zhang},
    journal={arXiv preprint arXiv:2402.13178},
    year={2024}
}
```