RABBITS: Robust Assessment of Biomedical Benchmarks Involving drug Term Substitutions

RABBITS is a dataset transformation project that focuses on evaluating the robustness of language models in the medical domain through the substitution of brand names for generic drugs and vice versa. This evaluation is crucial as medical prescription errors, often related to incorrect drug naming, are a leading cause of morbidity and mortality.

Motivation

Medical knowledge is context-dependent and requires consistent reasoning across various natural language expressions of semantically equivalent phrases. This is particularly crucial for drug names, where patients often use brand names like Advil or Tylenol instead of their generic equivalents. To study this, we create a new robustness dataset, RABBITS, to evaluate performance differences on medical benchmarks after swapping brand and generic drug names using physician expert annotations.

By modifying common medical benchmarks such as MedMCQA and MedQA, and replacing drug names, we aim to:

• Test models' robustness in clinical knowledge comprehension.
• Detect signs of dataset contamination or model bias.
• Highlight the importance of precise drug nomenclature in medical AI applications.

Our findings reveal a consistent performance drop ranging from 1-10% when drug names are swapped, suggesting potential issues in the model's ability to generalize and highlighting the need for robustness in medical AI applications.

Setup

Environment

The codebase is designed to run with Python 3.9. Here are the steps to set up the environment:

conda create -n rabbits python=3.9
conda activate rabbits
pip install -r requirements.txt

Datasets

The following datasets are transformed and analyzed in the RABBITS project:

1. MedQA (medqa)
2. MedMCQA (medmcqa)

Each dataset is processed to create three subsets: a subset that is filtered to questions that contain a drug keyword then two other variants, one with brand names replaced by generic names and another with generic names replaced by brand names. These transformations test the model's ability to maintain accuracy despite changes in clinically equivalent terminology.

Drug Names

data/generic_to_brand.csv - contains the list of generic to brand drug names searched for in the dataset src/get_drug_list.py - creates a list of brand and generic drug pairs using the RxNorm dataset (should be in data/RxNorm).

Keyword Extraction

src/drug_mapping - contains the code to create brand-generic drug mappings src/run_filtering.py - runs the keyword extraction and filters the datasets to only entries that contain either a brand or generic drug name src/run_replacement.py - runs the keyword replacement on the filtered datasets src/counts_per_col.py - counts the number of times each word appears in each column of the dataset e.g. question vs answers

Annotations

src/get_rag_names.py - Creates api calls to cohere to check brand names for each generic drug which was used during the quality assurance process. data/annotated_medqa_new.csv and data/annotated_medmcqa_new.csv - The annotated questions for correct replacement of drug names that is used to filter the dataset for creating the final dataset. src/generate_final_questions.py - filters the drug names to only include those questions that were accepted during the quality assurance process

Each row in the annotation files represents a question and includes the following key information:

• Original question and answer choices
• Generic-to-brand (g2b) transformed question and answer choices
• Found keywords in the original question
• Decision to keep or drop the question
• Any additional comments

Example Annotation

Here's an example of how the annotations are structured:

| id | question_orig | opa_orig | opb_orig | opc_orig | opd_orig | cop_orig | choice_type_orig | exp_orig | subject_name_orig | topic_name_orig | found_keywords_orig | local_id_orig | question_g2b | opa_g2b | opb_g2b | opc_g2b | opd_g2b | cop_g2b | choice_type_g2b | exp_g2b | subject_name_g2b | topic_name_g2b | found_keywords_g2b | local_id_g2b | keep/drop | comments |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 006acfff-dc8f-4bb5-97b2-e26144c56483 | PGE1 analogue is ? | Carboprost | Alprostadil | Epoprostenol | Dinoprostone | -1 | single | | Pharmacology | | ['carboprost' 'dinoprostone' 'alprostadil' 'epoprostenol'] | 4101 | PGE1 analogue is ? | hemabate | caverject | flolan | cervidil | -1 | single | | Pharmacology | | ['carboprost' 'dinoprostone' 'alprostadil' 'epoprostenol'] | 4101 | keep | |

In this example:

• The original question asks about PGE1 analogues, with generic drug names as options.
• The g2b transformation replaces these with corresponding brand names (e.g., Carboprost → hemabate).
• The found keywords are listed, showing which drug names were identified.
• The 'keep/drop' column indicates that this question should be kept in the final dataset.

Results

src/create_benchmarks.py - creates a benchmark for Hugging Face and pushes it to the hub src/create_plots.py - contains the code to create the figures in the paper

Harness was used to evaluate the benchmark -- our specific fork of harness used can be found here

To run the benchmark you can run the equivalent code to below in the terminal after installing this fork as a package.

lm_eval --model hf \
    --model_args pretrained=EleutherAI/gpt-j-6B \
    --tasks b4b \
    --device cuda:0 \
    --batch_size 8

This contains the following tasks:

• medmcqa_orig_filtered
• medqa_4options_orig_filtered
• medmcqa_g2b
• medqa_4options_g2b

The overall results used to produce the main figure can be found in rabbit_results.csv.

Citing

@inproceedings{
  anonymous2024language,
  title={Language Models are Surprisingly Fragile to Drug Names in Biomedical Benchmarks},
  author={Anonymous},
  booktitle={The 2024 Conference on Empirical Methods in Natural Language Processing},
  year={2024},
  url={https://openreview.net/forum?id=T5taTxsNb3}
}