Detecting Spin in Clinical Trials with Large Language Models

AI summary

Overview

Research area: Natural language processing applied to biomedical and clinical-trial text, specifically the detection of "spin" (distorted reporting of study results) using open-source large language models.

Technical level: Intermediate — the paper builds on standard semantic-similarity concepts, LLM prompting, and classification metrics, but assumes familiarity with ROC-based threshold selection, F1 scores, and prompt engineering.

Scope: The paper tests whether four open-source LLMs, used without any parameter fine-tuning, can classify pairs of primary and reported clinical trial outcomes as similar or different — the core subtask needed to detect outcome switching.

What This Paper Is About

Clinical trials sometimes report their results in a way that makes treatments look better than the data justify; this is called spin, and manual audits report it in 47–71% of clinical trial abstracts that fail to confirm their hypothesis. One specific form of spin, outcome switching, involves a mismatch between the outcomes a trial said it would measure and the outcomes it actually reports. The authors build and evaluate an automatic system that decides whether a primary outcome and a reported outcome mean the same thing, using open-source large language models instead of dedicated fine-tuned models.

Key Contributions

1. A prompt-engineering study of four open-source LLMs (OLMo-7B-Instruct, Mistral-7B-Instruct-v0.2, BioMistral-7B-DARE, Llama-3-8B-Instruct) as zero-shot binary classifiers for biomedical outcome-pair similarity, compared against five families of baseline text-similarity models.
2. A probability-based classification method that reads the softmax probabilities of the first generated token (the tokens for "Yes" and "No") to produce a continuous similarity score, following Algorithm 1.
3. Automatic threshold selection using the Youden index, applied on the 100-instance tuning set, instead of manually chosen thresholds.
4. Generation and manual qualitative assessment of natural language explanations for already-classified outcome pairs, plus public release of all source code, predictions, and explanations in a GitHub repository.

Main Findings

• Baseline performance: The sentence transformer (all-MiniLM-L12) achieved the best baseline result with an F1 of 0.67 and accuracy of 0.82 on the 2,496-instance test set.
• Best single LLM with the original prompt: BioMistral reached the best F1 of 0.73 with accuracy of 0.88 among LLMs using the initial prompt.
• Best single LLM with the corrected prompt: OLMo' achieved the best F1 of 0.75 and accuracy of 0.89 after the prompt was revised to reject partially reported outcomes.
• Best overall result: The LLM' voting ensemble (OLMo, Mistral, BioMistral with the corrected prompt, majority vote) reached an F1 of 0.78 and accuracy of 0.90, an improvement of 7 percentage points over baseline voting.
• Comparison to fine-tuned BERT: Koroleva et al. (2019) reported F1 values between 0.87 and 0.90 with fine-tuned BERT models on a non-deduplicated dataset; the authors' results trail that by about 10 percentage points on F1.
• Over-strict instruction following: Llama followed the corrected instructions too strictly — its recall dropped to 0.33 and its F1 to 0.49, making it an unusable classifier.
• Ensembles help: Majority voting improved results over any single model in a group, indicating that the models made distinct mistakes and complemented each other.
• Prompt sensitivity differs by domain adaptation: BioMistral performed better with minimal prompts while Mistral benefited from more context.
• Explanation quality: A manual qualitative analysis found that more than 50% of the generated explanations are valuable, while approximately every tenth contains hallucinations.

Methodology in Plain English

The authors started from an existing dataset of primary and reported outcome pairs drawn from 3,938 RCT articles on PubMed Central. One human annotator, after consulting domain experts, had labeled 3,043 pairs as similar (label 1) or different (label 0). The authors deduplicated these, leaving 2,796 pairs, of which 21.8% were similar and 78.2% different. From that pool they randomly sampled 300 pairs — 100 for tuning and 200 for validation — and kept the remaining 2,496 pairs as the final test set. The tuning, validation, and test sets contained 31.0%, 21.5%, and 21.4% similar pairs respectively.

They then compared several families of baseline methods: string measures (Levenshtein distance, Ratcliff-Obershelp sequence similarity), lexical measures (lemma and stem similarity), vector-based measures (spaCy and Word2Vec embeddings), ontology-based measures (path distance, Leacock-Chodorow, Wu-Palmer), and deep-learning encoders (SciBERT, BioBERT, and the all-MiniLM-L12 sentence transformer).

For the LLM side, they ran four open-source, instruction-tuned models locally on a single NVIDIA Quadro RTX 5000 GPU with 16 GB of VRAM, which capped model size at 8B parameters. Rather than parsing generated text for "Yes" or "No," they used the probabilities of the first output token to build a similarity score. They tried six prompt templates of increasing complexity (sentence template, outcome template, role template, article definition template, Wikipedia definition template, random example template), compared them by AUC on the 200-instance validation set, and selected the role-based prompt — which defines the model as a clinical trial reviewer — because its AUC curve resembled an inverted U and its spread across models was smallest.

Thresholds were set automatically by maximizing the Youden index on the tuning set, then applied to produce binary predictions. After manually inspecting OLMo's misclassifications, the authors identified three recurring error types: insufficient domain terminology and acronym knowledge (e.g., HbA1c), poor comprehension of longer sentences, and acceptance of partially reported outcomes. They iteratively added clarifications to the prompt to address the third error type, re-checking that F1 and accuracy did not degrade. Finally, they built three majority-vote ensembles: Baseline voting (stem similarity, sequence similarity, sentence encoder), LLM voting (OLMo, BioMistral, Llama with the initial role prompt), and LLM' voting (OLMo, Mistral, BioMistral with the corrected prompt). Explanations for classified examples were obtained by continuing the conversation with the message "Why? Explain your answer."

Why This Matters

Automatic spin detection affects how clinical evidence is written, reviewed, and consumed. The paper shows that general-purpose open-source LLMs, running locally on modest hardware and requiring only a few hundred labeled instances, can outperform classical text-similarity baselines on this task without any parameter fine-tuning — making the approach accessible to groups without large annotation budgets or commercial API access.

Real-world applications:

• Peer review and editorial screening: Journals could flag manuscripts where reported outcomes do not match registered primary outcomes before publication.
• Systematic reviewers and meta-analysts: Tools could automatically check whether a trial's abstract reports the outcomes it set out to measure.
• Trial registries and regulators: Automated comparison of registry entries against published results could surface outcome switching at scale.
• Clinical guideline developers: Faster triage of which study results are trustworthy when synthesizing evidence across many trials.

Industry relevance: Because the approach keeps data private through local inference and avoids API costs and latency, it is relevant to hospitals, publishers, contract research organizations, and any regulated environment where clinical text cannot leave the premises and model transparency matters — the authors specifically chose OLMo for its fully documented and replicable architecture, training data, and training procedure.

Future Directions

1. Evaluating larger, more capable LLMs, since the 16 GB VRAM constraint limited this study to models of up to 8B parameters.
2. Extending the system into a complete spin detection pipeline covering other types of spin and related tasks such as outcome extraction and discrepancy judgment.
3. Using models with longer context windows to analyze whole clinical trial reports at once, potentially replacing intermediate steps with reasoning.
4. Repeating the experiments with a machine-translated or separately collected dataset to ensure cross-lingual reliability.

Target Audience

This paper is most useful to NLP researchers working on biomedical text and clinical evidence synthesis; methodologists and meta-researchers studying spin and outcome switching; developers building clinical-trial screening or peer-review tools; and practitioners who need a low-cost, privacy-preserving, non-fine-tuned pipeline for comparing primary and reported outcomes.

Authors’ abstract

Spin in clinical trials 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 that fail to reach statistical significance. The comparison of primary and reported outcomes is crucial for detecting several types of spin, including outcome switching. We used 300 pairs of outcomes labeled with semantic similarity to develop a system for automatic detection of outcome switching. We evaluated baseline text similarity models and open-source LLMs using generated similarity scores and the Youden index to determine the classification threshold. The proposed approach involves prompt engineering, classification based on token probabilities, and majority voting for the final decision. The results on the test set of 2,496 examples with an F1 score of 0.78 and an accuracy of 0.90 outperform baseline text similarity models but trail behind fine-tuned versions of BERT. We used LLMs to generate natural language explanations for the classified instances and manually assessed their quality.