Artificial intelligence-driven study selection in systematic reviews of randomized controlled trials, emulated trials...

Systematic reviews (SRs) are key to evidence-based medicine but are often labor-intensive, especially in the study selection step. This study assessed the use of large language models (LLMs) to automate SR study screening and selection. Five SR projects were included: two published therapeutic SRs (SR1-2), two ongoing emulated-trial SRs (SR3-4), and one economic evaluation SR (SR5). The total number of studies screened for each SR was 3,966, 3,147, 695, 3,096 and 485, respectively, with 20, 24, 46, 32 and 70 eligible studies. Three LLMs-Gemini 2.0 Flash, Llama 3.1, and Qwen 2.5-were evaluated using training sets (five studies), title/abstract datasets, and full-text datasets predicted as relevant. Prompts based on the PICOS framework were iteratively refined using a recall-first strategy. Outputs were compared with human reviewer classifications using recall, number needed to screen (NNS), and percentage reduced workload with 95% confidence intervals. In the title/abstract screening phase, Llama 3.1 and Gemini 2.0 Flash achieved consistently high recall (90.00%-100.00% and 90.48%-100.00%), with workload reduction of 61.92%-97.10% and 65.21%-97.03%, respectively. Qwen 2.5 achieved the highest workload reduction (76.16%-99.19%) but showed the lowest recall (76.67%-88.89%). In the full-text selection phase, Llama 3.1 achieved the highest recall (93.33%-100.00%) with workload reductions of 73.15%-97.41%, but slower processing time (approximately 2.3-3.6 minutes per document). Qwen 2.5 yielded lower recall (66.67%-89.71%), despite the highest workload reduction (80.82%-99.44%) and similarly slow inference times (approximately 3.0-4.2 minutes per document). Gemini 2.0 Flash balanced high recall (83.33%-100.00%) with substantial workload reduction (76.71%-98.91%) and markedly faster inference (approximately 4-8 seconds per document). LLMs-particularly Llama 3.1 and Gemini 2.0 Flash-can substantially reduce SR screening workload while maintaining high recall when guided by a recall-first prompting framework. Remaining challenges include reproducibility in closed-source models and generalizability across diverse SR topics.