Efficiency of LLM-Based Systematic Review Screening: Mimicking Human Expert Cognition

Systematic reviews (SRs) require extensive effort in literature screening, particularly in abstract selection. Large Language Models (LLMs) have shown promise in automating this process, but existing approaches mainly compare model performances rather than optimizing screening strategies. This study introduces an Expert Imitation Method (Rim's method), designed to replicate the decision-making process of an experienced meta-analysis specialist in oncology. Using structured search queries and keyword-based heuristics, this method improves the efficiency and accuracy of abstract screening compared to conventional PICO- and criteria-based approaches. In validation tests, the Expert Imitation Method achieved superior sensitivity (83.3%) and specificity (82.5%) with an AUC-ROC of 0.83. These results indicate that expert-informed prompting enhances the reliability of LLM-assisted SR screening. Future research will refine this method and expand its applications across medical fields.