Zero-shot interpretable biomedical literature appraisal with generative large language models.

Area of Science:

• Artificial Intelligence in Medicine
• Biomedical Informatics
• Clinical Trial Methodology

Background:

• Automating the critical appraisal of randomized controlled trials (RCTs) is crucial for efficient knowledge synthesis.
• Large language models (LLMs) offer potential for automating complex scientific text analysis.

Purpose of the Study:

• To evaluate the performance of two decoder-based Generative Pre-trained Transformer (GPT) models (GPT-4o and GPT-3.5-mini) in automating RCT methodological appraisal.
• To compare GPT models against a fine-tuned encoder-only BioLinkBERT model using various prompting strategies.

Main Methods:

• A stratified random sample of 800 RCT articles was appraised.
• Two prompting schemes were used: classifier (independent assessment) and verifier (validation of BioLinkBERT).
• Assessments considered either title/abstract (TIAB) or full text, with performance measured against human assessments using Matthews correlation coefficient (MCC).

Main Results:

• GPT-4o as a classifier using full text achieved an MCC of 0.429, comparable to BioLinkBERT (MCC 0.466).
• GPT-4o as a verifier using full text showed similar performance (MCC 0.391).
• GPT models provided transparent, criterion-specific justifications, but performance significantly decreased when using only TIAB (MCC ≤0.100).

Conclusions:

• GPT-4o effectively automates RCT critical appraisal when full text is available, offering comparable performance to specialized models.
• GPT models enhance interpretability and transparency through explicit justifications.
• Fine-tuned models may complement GPTs when full texts are unavailable, and prompt optimization is key for clinical adoption.