Detecting spin in clinical trials with large language models

Thesis: https://repozitorij.uni-lj.si/IzpisGradiva.php?id=169018&lang=eng

Abstract: Spin in research reports includes reporting practices that distort the presentation of results. This is particularly critical in medicine, where spin is present in more than 50% of randomized controlled trials (RCT) that fail to reach the threshold of statistical significance. Comparing declared and reported outcomes is crucial for detecting various types of spin, such as selective reporting. We developed a system for automatic detection of spin in clinical trials. We used 300 pairs of outcomes, labeled with semantic similarity. We evaluated baseline statistical models, masked language models (MLM) and generative large language models (LLM). We generated similarity scores and used Youden index to determine the classification threshold. The proposed approach to comparing outcomes using LLMs involves prompt engineering, generating similarity scores based on token probabilities and majority voting. The results on the test set of 2500 examples, with 90% accuracy and F1 score of 78%, outperform dedicated models for semantic similarity evaluation, but trail behind fine-tuned versions of BERT model. An advantage of our approach is the ability to generate explanations for the classified examples.

2026 update 🆕: We evaluated MedGemma-4b-it and got results similar to BioMistral. Maybe this model would be useful for explaining medical terminology and classified examples.

Language models

OLMo-7B-Instruct

• Description: Open training and weights general purpose model.
• Model size: 6.9B parameters
• Vocabulary size: 50k tokens
• Release date: February 2024
• Link: https://huggingface.co/allenai/OLMo-7B-Instruct-hf

Mistral-7B-Instruct-v0.2

• Description: Open weights general purpose model.
• Model size: 7.2B parameters
• Vocabulary size: 32k tokens
• Release date: December 2023
• Link: https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2

BioMistral-7B-DARE

• Description: Open weights medical domain adapted model.
• Model size: 7.2B parameters
• Vocabulary size: 32k tokens
• Release date: February 2024
• Link: https://huggingface.co/BioMistral/BioMistral-7B-DARE

MedGemma-4b-it

• Description: Trained for performance on medical text and images.
• Model size: 4.3B parameters
• Release date: May 2025
• Link: https://huggingface.co/google/medgemma-4b-it

Meta-Llama-3-8B-Instruct

• Description: State of the art open weights model.
• Model size: 8B parameters
• Vocabulary size: 128k tokens
• Release date: April 2024
• Link: https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct

Gpt-4-Turbo

• Description: Closed weights API model.
• Model size: Not disclosed
• Vocabulary size: Not disclosed
• Release date: April 2024
• Link: https://platform.openai.com/docs/models/turbo#gpt-4-turbo-and-gpt-4

Dataset

• Labeled pairs of primary and reported outcomes of PubMed clinical trial abstracts.
• 100 pairs for few shot examples and threshold tuning.
• 2600 pairs for prompt tuning and model evaluation.
• Biased towards negative examples, where the reported outcome does not match the primary outcome.

Note: Authors of the original paper used 3043 pairs of outcomes. After deduplication by attributes out1, out2 and label the dataset consists of 2796 rows.

Prompt templates

Zero shot

• Sentence template:

<|endoftext|><|user|>
Are the sentences semantically similar? Answer with either Yes or No.

First sentence: {out1}

Second sentence: {out2}

Answer:
<|assistant|>

• Outcome template:

<|endoftext|><|user|>
Does the reported outcome match the defined primary outcome? Answer with either Yes or No.
                                                
Primary outcome: {out1}

Reported outcome: {out2}

Answer:
<|assistant|>

• Role template:

<|endoftext|><|user|>
You are a clinical trial report reviewer. Your task is to detect incorrectly reported outcomes.

---

Does the reported outcome match the defined primary outcome? Answer with either Yes or No.
                                                
Primary outcome: {out1}

Reported outcome: {out2}

Answer:
<|assistant|>

• Article definition template:

<|endoftext|><|user|>
You are a clinical trial report reviewer. Your task is to detect incorrectly reported outcomes.

---

Outcome switching is an unjustified change of the predefined trial outcomes, leading to reporting only the favourable outcomes that support the hypothesis of the researchers. Outcome switching is one of the most common types of spin. It can consist in omitting the primary outcome in the results and conclusions of the abstract, or in the focus on significant secondary outcomes.

---

Does the reported outcome match the defined primary outcome? Answer with either Yes or No.

Primary outcome: {out1}

Reported outcome: {out2}

Answer:
<|assistant|>

• Wikipedia definition template:

<|endoftext|><|user|>
You are a clinical trial report reviewer. Your task is to detect incorrectly reported outcomes.

---

Outcome switching is the practice of changing the primary or secondary outcomes of a clinical trial after its initiation. An outcome is the goal of the clinical trial, such as survival after five years for cancer treatment. Outcome switching can lead to bias and undermine the reliability of the trial, for instance when outcomes are switched after researchers already have access to trial data. That way researchers can cherry pick an outcome which is statistically significant.

---

Does the reported outcome match the defined primary outcome? Answer with either Yes or No.

Primary outcome: {out1}

Reported outcome: {out2}

Answer:
<|assistant|>

• Chain of thought template:

<|endoftext|><|user|>
You are a clinical trial report reviewer. Your task is to detect incorrectly reported outcomes.

---

Does the reported outcome match the defined primary outcome? Lets think step by step.

Primary outcome: {out1}

Reported outcome: {out2}

Steps:
<|assistant|>

One shot

• Random and similar example template:

<|endoftext|><|user|>
You are a clinical trial report reviewer. Your task is to detect incorrectly reported outcomes.

---

Does the reported outcome match the defined primary outcome? Answer with either Yes or No.

Primary outcome: {example_out1}

Reported outcome: {example_out2}

Answer: {example_ans}

---

Does the reported outcome match the defined primary outcome? Answer with either Yes or No.

Primary outcome: {out1}

Reported outcome: {out2}

Answer:
<|assistant|>

Few shot

More provided examples did not improve the auc score.

Separation results

Baseline models

Large language models

Note: Model role is a part of the user prompt, except in the case of Llama that uses a dedicated system prompt.

Key takeaways:

• Domain specific BioMistral is most effective with minimal prompt engineering.
• State of the art Llama 3 excels at understanding complex prompts.
• Olmo and Mistral are relatively stable in their predictions.

Classification results

Olmo

Mistral

BioMistral

Llama

Note: Model bias is corrected by adjusting the threshold based on Youdens J statistic.

Top models

Note: Combination of diverse language models achieves the highest classification accuracy.

Mistake analysis

Key takeaways:

• BioMistral is more effective for examples containing medical terminology.
• Llama is more effective for examples requiring complex reasoning.

Edge analysis

Adjusted Olmo threshold is 0.1 and examples close to this value are considered to be on the edge of classification.

Key takeaways:

• Olmo edge examples are often based on phrase similarity.
• BioMistral edge examples are often based on concept similarity.

Mistake fix

The prompt is adjusted to reject generalized reported outcomes.

• Detail template:

<|endoftext|><|user|>
You are a clinical trial report reviewer. Your task is to detect incorrectly reported outcomes.

---

Primary outcome is defined at the start of a clinical trial and is the observed variable of the study.
This outcome is later reported in the results section of the report.
The reported outcome should match the defined primary outcome.
That means it should include all its components and details.

---

Does the reported outcome match the defined primary outcome? Answer with either Yes or No.

Primary outcome: {out1}

Reported outcome: {out2}

Answer:
<|assistant|>

Key takeaways:

• Llama is the only model that is too sensitive to the prompt change.
• Combination of the rest achieves by far the best f1 score.

Explanation results

Open ended

• Explanation template:

<|endoftext|><|user|>
Why? Explain your answer in one sentence.
            
Explanation:
<|assistant|>

Multiple choice

Probabilistic selection of the provided explanations did not reach the desired results.

References

• Measuring semantic similarity of clinical trial outcomes using deep pre-trained language representations
• SciBERT: A Pretrained Language Model for Scientific Text
• A Comprehensive Overview of Large Language Models
• OLMo: Accelerating the Science of Language Models
• Unleashing the potential of prompt engineering in Large Language Models: a comprehensive review
• Large Language Models are Few-Shot Clinical Information Extractors
• Analyzing Spin in Abstracts of Orthopaedic Randomized Controlled Trials With Statistically Insignificant Primary Endpoints
• BioMistral: A Collection of Open-Source Pretrained Large Language Models for Medical Domains
• MedGemma Technical Report
• COMPare: A prospective cohort study correcting and monitoring 58 misreported trials in real time
• Outcome switching
• Youdens J statistic
• Breaking the Ceiling of the LLM Community by Treating Token Generation as a Classification for Ensembling
• Measuring Massive Multitask Language Understanding
• Can Large Language Models Explain Themselves? A Study of LLM-Generated Self-Explanations