Abstract: FR-PO0037
Improving Literature Screening Efficiency with Large Language Model-Assisted Abstract Triage
Session Information
- Artificial Intelligence: Clinical Prediction, Decision Support, and Digital Kidney Care
October 23, 2026 | Location: Exhibit Hall A, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Artificial Intelligence, Digital Health, and Data Science
- 300 Artificial Intelligence, Digital Health, and Data Science
Authors
- Bergling, Karin, Renal Research Institute, New York, New York, United States
- Yueh, Sheng-Han, Renal Research Institute, New York, New York, United States
- Clark, Tatyana, Fresenius Medical Care Holdings Inc, Waltham, Massachusetts, United States
- Sawin, Dixie-Ann, Fresenius Medical Care Holdings Inc, Waltham, Massachusetts, United States
- Usvyat, Len A., Renal Research Institute, New York, New York, United States
- Himmele, Rainer, Fresenius Medical Care Holdings Inc, Waltham, Massachusetts, United States
- Zhang, Hanjie, Renal Research Institute, New York, New York, United States
Background
Efficient literature review is essential, yet keyword-based PubMed searches often lack specificity and require extensive manual screening. Meaningful parameters such as study design, treatment details, and outcomes are difficult to capture with high specificity using standard operators. Large language models (LLMs) may enable more targeted abstract filtering through semantic interpretation.
Methods
We developed an AI-assisted literature search tool combining PubMed querying with GPT-5 (OpenAI) abstract filtering (Fig. 1). Users enter a PubMed query, (Fig. 1A) or a free-text research question, which GPT-5 converts into a PubMed-compatible query. Abstracts and citation metadata are retrieved via the Entrez Python API. Users add natural-language screening criteria for study design, population, interventions, or outcomes (Fig. 1B). GPT-5 classifies abstracts as relevant, maybe relevant, or not relevant, with brief justifications (Fig. 1C). Users can review, select, and download abstracts with citation (Fig. 1D).
Results
In an illustrative case (Fig. 1), the PubMed query Hemodialysis AND Hemodiafiltration AND (Quality of Life OR HRQoL) AND Randomized Clinical Trial retrieved 49 abstracts. The LLM screened for randomized clinical trials, excluding meta-analyses and protocols, that investigated quality-of-life outcomes with at least 200 participants per study arm. Two abstracts were classified as relevant, three as maybe relevant, and 44 as not relevant. Manual review confirmed both relevant abstracts and all 44 exclusions. Of the three maybe relevant abstracts, one was deemed suitable for full-text review, and two were likely not relevant, supporting “maybe relevant” as a useful uncertainty category. Overall, five abstracts were prioritized for review, reducing the initial screening set by 89.8%.
Conclusion
An LLM-assisted tool can reduce manual screening by prioritizing potentially relevant abstracts. Combining transparent PubMed querying, semantic filtering, and user review may support more efficient evidence synthesis.
Acknowledgment
Large language models were used as part of the abstract triage workflow described in this study and to assist with abstract preparation. The authors take full responsibility for the integrity, accuracy, and final content of the abstract.
LLM-assisted literature search workflow: (A) query input, (B) LLM screening criteria, (C) LLM classification with justification, (D) selection and download with citations.
Funding
- Commercial Support – Fresenius Medical Care