Answers are grounded in this paper and cite their sources.
We evaluated Jev, a non-generative decision model, for title and abstract screening using 4527 records from two systematic reviews of bipolar disorder treatments previously used in our 2024 Journal of Medical Internet Research study. Human decisions to retain records for full-text assessment served as the reference. In the primary analysis, Jev answered yes-or-no questions (Noul); at a prespecified 50% cutoff, sensitivity and specificity were 99.6% and 98.9% for light therapy and 93.2% and 93.5% for adjunctive pharmacotherapy. For direct model comparison, Jev used explicit include/exclude options (Choice) and the same prompts as GPT-5 mini and GPT-6 Astra. Jev Choice achieved sensitivity and specificity similar to or higher than those of GPT-5 mini and GPT-6 Astra, while API charges per 1000 records were US$0.043 for Jev, US$0.57 for GPT-5 mini, and US$7.39 for GPT-6 Astra. Jev retained all reports ultimately included in both reviews across all tested conditions and runs. Although these findings do not establish that decision models can replace human reviewers, they support Jev as a promising, low-cost alternative to the tested generative LLM configurations for assisting human title and abstract screening. Independent validation across diverse review topics is needed to assess Jev’s generalizability and utility in human-supervised screening workflows.
No takes yet. Share an insight, caveat, or question.
Matsui et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: