Desenvolvimento e validação de um fluxo automatizado de revisão sistemática sobre semaglutida e desfechos perioperatórios em comparação com um padrão de referência humano

Abstract

Systematic reviews are essential tools for synthesizing health evidence. However, their conduct still requires substantial investments of time and manual effort, particularly during bibliographic deduplication, title and abstract screening, and full-text assessment. In this context, the present study aimed to develop and validate an automated workflow to support a systematic review of perioperative semaglutide use. Decisions made by trained reviewers, following PRISMA guidelines, were used as the reference standard. This was a methodological, observational, and analytical study structured into three sequential modules: bibliographic deduplication, title and abstract screening, and full-text assessment. System performance was evaluated through direct comparison with the selections made by human reviewers using sensitivity, specificity, precision, negative predictive value, F1 score, accuracy, Cohen's kappa coefficient, Jaccard index, and manual workload reduction. In addition, the studies ultimately included in the systematic review were tracked throughout all stages of the automated process to determine whether their identification was preserved. The system demonstrated strong performance in deduplication, particularly at the cluster level, and preserved all 12 studies ultimately included throughout the automated workflow. During title and abstract screening, agreement with the broader human full-text retrieval set was moderate, with a sensitivity of 0.703, although no study ultimately included was missed. The exploratory full-text module achieved a sensitivity of 1.000 in the safety analysis and reduced by 68.9% the number of PDFs requiring prioritized human review. In conclusion, SafeScreen has the potential to reduce the workload of systematic reviews while preserving the studies ultimately included and maintaining decision traceability and auditability. The findings support its use as a reviewer-support tool under human supervision, although external validation is required to assess its generalizability. [resumo fornecido pelo autor]