LLMs and Spin in Medical Literature

Code and Data for the paper "Caught in the Web of Words: Do LLMs Fall for Medical Spin?"

Link to Paper: https://arxiv.org/abs/2502.07963

How susceptible are LLMs to spin in medical articles?

:hammer_and_wrench: SETUP

Create conda environment from the environment.yml: conda env create -f environment.yml

Activate the conda environment: conda activate MedLitSpin

:bookmark_tabs: DATA

This project uses the cancer-related medical abstract dataset compiled by Boutron et al. (2014). This dataset comprises 60 real and synthetic abstracts illustrating cases of results reported with and without spin.

More specifically, the dataset includes 30 "base" abstracts, i.e., real articles describing RCTs in the field of cancer which (1) report statistically nonsignificant differences for all primary outcomes measured for the intervention being studied, and, (2) contain spin in the results and conclusion sections of the abstract (intimating more positive findings). The "neutral" abstracts were created by manually editing each of the original 30 abstracts to remove spin.

The data is available in the data folder as csv file.

:speech_balloon: MODELS

We use 22 LLMs, including both open and closed (proprietary) models, for our analyses. We also include both general and specialized (biomedical) LLMs, and models spanning a range of parameter counts.

The generalist closed models can be run on CPU as they are accessed via APIs. The generalist open and biomedical open models require GPUs to run.

:test_tube: EXPERIMENTS

Spin Detection

code/run_spin_detection_evaluation.py

Example script for running this task on GPT-4o:

python3 code/run_spin_detection_evaluation.py \
    --model gpt4o \
    --output_path code/eval_outputs/gpt4o

You can change the arguments to run different models and specify output paths.

Arguments of run_spin_detection_evaluation.py:

• --model: model to evaluate ("gpt35", "gpt4o", "gpt4o-mini", "gemini_1.5_flash", "gemini_1.5_flash-8B", "claude_3.5-sonnet", "claude_3.5-haiku", "olmo2_instruct-7B", "olmo2_instruct-13B", "mistral_instruct7B", "llama2_chat-7B", "llama2_chat-13B", "llama2_chat-70B", "llama3_instruct-8B", "llama3_instruct-70B", "med42-8B", "med42-70B", "openbiollm-8B", "openbiollm-70B", "biomistral7B", "biomedgpt7B", "alpacare-7B")
• --input_path: file path of the dataset to run on. default is data/spin_and_no_spin_abstracts.csv but can pass in data/unspun_abstracts/gpt4o_generated_no_spin_abstracts_formatted.csv to run with the extended, semi-synthetic dataset.
• --output_path: path of directory where json and csv files of the outputs from model and eval metrics should be saved.
• --debug: adding this flag will only run 3 randomly sampled PubMed articles from dataset. This is for debugging purposes.

Script to replicate the paper experiment with 22 LLMs:

scripts/run_detection_evaluation.sh

Interpretation of trial results for spin and unspun abstracts

code/run_spin_interpretation_evaluation.py

Example script for running this task on GPT 3.5:

python3 code/run_spin_interpretation_evaluation.py \
    --model gpt35 \
    --output_path code/eval_outputs/gpt35

Arguments of run_spin_interpretation_evaluation.py:

• --model: model to evaluate ("gpt35", "gpt4o", "gpt4o-mini", "gemini_1.5_flash", "gemini_1.5_flash-8B", "claude_3.5-sonnet", "claude_3.5-haiku", "olmo2_instruct-7B", "olmo2_instruct-13B", "mistral_instruct7B", "llama2_chat-7B", "llama2_chat-13B", "llama2_chat-70B", "llama3_instruct-8B", "llama3_instruct-70B", "med42-8B", "med42-70B", "openbiollm-8B", "openbiollm-70B", "biomistral7B", "biomedgpt7B", "alpacare-7B")
• --input_path: file path of the dataset to run on. default is data/spin_and_no_spin_abstracts.csv but can pass in data/unspun_abstracts/gpt4o_generated_no_spin_abstracts_formatted.csv to run with the extended, semi-synthetic dataset.
• --output_path: path of directory where json and csv files of the outputs from model and eval metrics should be saved.
• --debug: adding this flag will only run 3 randomly sampled PubMed articles from dataset. This is for debugging purposes.

Script to replicate the paper experiment with 22 LLMs:

scripts/run_interpretation_evaluation.sh

Simplifying spun and unspun abstracts

Generation

Generating the simplified (plain language) abstracts can be done by running code/generate_plain_language_summaries.py

Example script for generating simplified abstracts with Llama2 Chat 7B:

python3 code/generate_plain_language_summaries.py \
    --model llama2_chat-7B \
    --output_path code/pls_outputs/llama2_chat-7B \
    --max_new_tokens 300 \
    --prompt_template_name default

Arguments of generate_plain_language_summaries.py:

• --model: model to evaluate ("gpt35", "gpt4o", "gpt4o-mini", "gemini_1.5_flash", "gemini_1.5_flash-8B", "claude_3.5-sonnet", "claude_3.5-haiku", "olmo2_instruct-7B", "olmo2_instruct-13B", "mistral_instruct7B", "llama2_chat-7B", "llama2_chat-13B", "llama2_chat-70B", "llama3_instruct-8B", "llama3_instruct-70B", "med42-8B", "med42-70B", "openbiollm-8B", "openbiollm-70B", "biomistral7B", "biomedgpt7B", "alpacare-7B")
• --input_path: file path of the dataset to run on. default is data/spin_and_no_spin_abstracts.csv but can pass in data/unspun_abstracts/gpt4o_generated_no_spin_abstracts_formatted.csv to run with the extended, semi-synthetic dataset.
• --output_path: path of directory where json and csv files of the outputs from model and eval metrics should be saved.
• --max_new_tokens: maximum number of tokens to generate for the simplified abstract.
• --prompt_template_name: name of the template to use for the prompt. defaults to "default".
• --debug: adding this flag will only run 3 randomly sampled PubMed articles from dataset. This is for debugging purposes.

Script to replicate the paper experiment with 22 LLMs:

scripts/run_generate_pls.sh

Evaluation

Running the interpretation task with simplified (plain language) abstracts can be done by running code/run_spin_interpretation_from_pls_evaluation.py

Example script with Claude 3.5 Sonnet as evaluator on Llama2 Chat 7B outputs:

python3 code/run_spin_interpretation_from_pls_evaluation.py \
      --model claude_3.5-sonnet \
      --input_path code/pls_outputs/llama2_chat-7B/llama2_chat-7B_outputs.csv \
      --output_path code/pls_outputs/_interpretation_eval_results/$evaluator/llama2_chat-7B

Arguments of run_spin_interpretation_from_pls_evaluation.py:

• --model: model to evaluate ("gpt35", "gpt4o", "gpt4o-mini", "gemini_1.5_flash", "gemini_1.5_flash-8B", "claude_3.5-sonnet", "claude_3.5-haiku", "olmo2_instruct-7B", "olmo2_instruct-13B", "mistral_instruct7B", "llama2_chat-7B", "llama2_chat-13B", "llama2_chat-70B", "llama3_instruct-8B", "llama3_instruct-70B", "med42-8B", "med42-70B", "openbiollm-8B", "openbiollm-70B", "biomistral7B", "biomedgpt7B", "alpacare-7B")
• --input_path: directory of where the input data (csv) is stored.
• --output_path: path of directory where json and csv files of the outputs from model and eval metrics should be saved.
• --debug: adding this flag will only run 3 randomly sampled PubMed articles from dataset. This is for debugging purposes.

Script to replicate the paper experiment with Claude 3.5 Sonnet & GPT4o Mini as evalutors evaluating simplified abstracts from all 22 LLMs:

scripts/run_pls_evaluation.sh

Reducing the effect of spin

Adding reference or model output spin labels

The interpretation task with additional information about the given abstract can be done by running code/run_spin_interpretation_evaluation_with_abstract_labels.py

Example script for interpreting results with spin labels from model outputs:

python3 code/run_spin_interpretation_evaluation_with_abstract_labels.py \
    --model llama3_instruct-70B \
    --label_mode model_output_label \
    --output_path code/eval_outputs/llama3_instruct-70B

Arguments of run_spin_interpretation_evaluation_with_abstract_labels.py:

• --model: model to evaluate ("gpt35", "gpt4o", "gpt4o-mini", "gemini_1.5_flash", "gemini_1.5_flash-8B", "claude_3.5-sonnet", "claude_3.5-haiku", "olmo2_instruct-7B", "olmo2_instruct-13B", "mistral_instruct7B", "llama2_chat-7B", "llama2_chat-13B", "llama2_chat-70B", "llama3_instruct-8B", "llama3_instruct-70B", "med42-8B", "med42-70B", "openbiollm-8B", "openbiollm-70B", "biomistral7B", "biomedgpt7B", "alpacare-7B")
• --label_mode: which abstract labels will be used for prompting. options are "gold_label" and "model_output_label".
• --input_path: file path of the dataset to run on. default is data/spin_and_no_spin_abstracts.csv but can pass in data/unspun_abstracts/gpt4o_generated_no_spin_abstracts_formatted.csv to run with the extended, semi-synthetic dataset.
• --output_path: path of directory where json and csv files of the outputs from model and eval metrics should be saved.
• --debug: adding this flag will only run 3 randomly sampled PubMed articles from dataset. This is for debugging purposes.

Joint prompting both detection and interpretation

The interpretation task with spin detection can be done by running code/run_spin_combined_detection_interpretion_evaluation.py

Example script for interpreting results after detecting spin in the same prompt:

python3 code/run_spin_combined_detection_interpretion_evaluation.py \
    --model med42-70B \
    --output_path code/eval_outputs/med42-70B

Arguments of run_spin_combined_detection_interpretion_evaluation.py:

• --model: model to evaluate ("gpt35", "gpt4o", "gpt4o-mini", "gemini_1.5_flash", "gemini_1.5_flash-8B", "claude_3.5-sonnet", "claude_3.5-haiku", "olmo2_instruct-7B", "olmo2_instruct-13B", "mistral_instruct7B", "llama2_chat-7B", "llama2_chat-13B", "llama2_chat-70B", "llama3_instruct-8B", "llama3_instruct-70B", "med42-8B", "med42-70B", "openbiollm-8B", "openbiollm-70B", "biomistral7B", "biomedgpt7B", "alpacare-7B")
• --input_path: file path of the dataset to run on. default is data/spin_and_no_spin_abstracts.csv but can pass in data/unspun_abstracts/gpt4o_generated_no_spin_abstracts_formatted.csv to run with the extended, semi-synthetic dataset.
• --output_path: path of directory where json and csv files of the outputs from model and eval metrics should be saved.
• --debug: adding this flag will only run 3 randomly sampled PubMed articles from dataset. This is for debugging purposes.

Script to replicate the paper experiment running all mitigation strategies:

scripts/run_evaluation_mitigation_strategies.sh

:computer: EXTENSION: EXPERIMENTS ON SEMI-SYNTHETIC DATA

Due to the original evaluation dataset being small, we created a semi-synthetic dataset to test this effect of spin with a larger sample outside of oncology.

This semi-synthetic dataset was generated to expand our evaluation dataset from just 30 to 150 additional pairs of abstracts. The new dataset contains abstracts from 3 different medical fields: Orthopaedic, Emergency Medicine, and Obesity. These datasets contains abstracts of Randomized Controlled Trials (RCTs) that illustrates cases of results reported with spin. The unspun version is generated with GPT-4o.

Further details about the dataset can be found here.

:star: CITATION

@inproceedings{yun2025caught,
  title={Caught in the Web of Words: Do LLMs Fall for Spin in Medical Literature?},
  author={Yun, Hye Sun and Zhang, Karen YC and Kouzy, Ramez and Marshall, Iain J and Li, Junyi Jessy and Wallace, Byron C},
  booktitle={Conference on Health, Inference, and Learning (CHIL) 2025},
  year={2025},
  organization={PMLR}
}