Ensembles of Large Language Models for Identifying EQ-5D Studies in PubMed Based on Their Abstracts

Researchers developed an ensemble machine learning approach using Google's Gemini and Gemma large language models to automatically identify EQ-5D health quality-of-life studies in PubMed abstracts. The combined model achieved 0.74 F1-score and accuracy, demonstrating that ensemble methods outperform individual LLMs for biomedical document classification tasks.

This research addresses a significant operational bottleneck in systematic literature reviews, where manual screening of thousands of published abstracts consumes substantial resources while remaining prone to human error and inconsistency. The study demonstrates that large language models can effectively automate the identification of specific clinical outcome measures—in this case EQ-5D quality-of-life assessments—directly from published abstracts without requiring full-text access. The ensemble approach combining multiple models achieved superior performance to individual models, suggesting that aggregating different LLM architectures creates more robust decision-making systems.

The broader context reflects accelerating adoption of AI for academic research acceleration. Manual systematic review processes have become increasingly unsustainable as publication volumes grow exponentially across biomedical and other domains. This work exemplifies how LLMs can reduce researcher workload while improving consistency and reducing reviewer bias in study screening phases.

For research institutions and pharmaceutical companies conducting systematic reviews, this approach offers direct operational efficiency gains. Automated screening using ensemble LLMs could reduce screening timelines from months to weeks and lower associated costs. The soft stacking meta-classifier layer adds interpretability, critical for clinical applications where transparency in decision-making is essential for regulatory compliance.

Future developments likely involve extending these methods across different study outcome measures and clinical domains, potentially creating comprehensive AI-driven systematic review pipelines. Validation across diverse datasets and integration with full-text screening phases will determine real-world deployment success.

• →Ensemble LLMs achieved 0.74 F1-score for automated EQ-5D detection, outperforming individual models
• →Combining multiple model architectures improves balance between precision and recall in biomedical classification
• →Soft stacking meta-classifiers enhance reliability and interpretability for clinical applications
• →Automated screening can significantly reduce manual effort in systematic literature reviews
• →LLM-based document classification represents scalable solution for research acceleration