Systematic reviews are critical for evidence-based neurology but are time-consuming and resource-intensive, with screening phases often creating significant bottlenecks. While large language models(LLMs) show potential to streamline this process, their real-world impact on reviewer performance in a supportive, human-in-the-loop role is understudied in neurology.

Four neurology trainees were grouped into two pairs based on previous screening experience. Pair A(A1, A2) consisted of less experienced trainees(1–2 SR), while Pair B(B1, B2) consisted of more experienced trainees(≥3 SRs). Within each pair, one reviewer was assigned to a traditional screening method(A2, B2), while the other was assigned to a generative AI-assisted method(A1, B1). The AI-assisted screening utilized PICOS(Population, Intervention/Exposure, Comparison, Outcome, Study design) summaries derived from titles and abstracts using an open-source LLM (Mistral-Nemo-Instruct-2407). All reviewers independently screened the same set of 1,003 articles against predefined criteria.  Screening times were recorded, and performance metrics were calculated. Post-screening surveys assessed usability, confidence, and perceived cognitive workload.