Leveraging Large Language Models for Translational Research Classification

Zheng, Zhejun

Permalink

https://hdl.handle.net/2142/133001 Copy

Description

Title

Leveraging Large Language Models for Translational Research Classification

Author(s)

Zheng, Zhejun

Issue Date

2026-03-12

Keyword(s)

• Translational medicine
• Large language models
• Text classification
• Prompt engineering

Abstract

Introduction. Classifying biomedical publications along the translational research spectrum(T0–T4) is essential for research evaluation, yet remains challenging due to inconsistent stage definitions and the labor-intensive nature of manual annotation. Although Surkis(2016) developed a 34-item checklist and trained machine learning classifiers, these statistical models are susceptible to model drift and cannot explicitly encode expert-defined classification rules. Large Language Models(LLMs) present a compelling alternative by enabling rule-based classification through prompt engineering. Method. We transformed the 34-item checklist into structured prompt templates corresponding to five translational categories. Seven LLMs(gpt-oss_20b, glm-4.5, deepseek-reasoner, and qwen3 variants) were evaluated on 296 expert-annotated publications using zero-shot, one-shot, and three-shot prompting strategies. Performance was assessed across three binary classification tasks(T0, T1/T2, T3/T4) using precision, recall, F1-score, and AUC. Results. DeepSeek-Reasoner achieved the highest F1-scores for T0(0.888) and T1/T2(0.886), while GLM-4.5 performed best on T3/T4(0.729). The top-performing models exceeded the original baselines, attaining AUCs of 0.987(T0) and 0.946(T1/T2) compared to the previously reported 0.94 and 0.84. However, 40–60% of publications received either multiple labels or no label due to the independent prompting strategy for each category. Conclusion. LLM-based classification effectively operationalizes expert-defined rules and outperforms traditional machine learning approaches for early translational stages.

Publisher

iSchools

Series/Report Name or Number

iConference 2026 Proceedings

Type of Resource

Other

Genre of Resource

Conference Poster

Language

eng

Permalink

https://hdl.handle.net/2142/133001

Copyright and License Information

Copyright 2026 is held by Zhejun Zheng. Copyright permissions, when appropriate, must be obtained directly from the authors.

Owning Collections