Document resource
Objectives Systematic reviews (SR) play a crucial role in synthesizing scientific literature for evidence-based decision-making, but traditional methods remain time-consuming and prone to inconsistencies. To enhance efficiency, we evaluated the feasibility of using GPT-based automated search techniques for retrieving occupational cancer-related studies from the PUBMED database.Material and Methods The assessment was conducted across seven neoplastic sites (i.e. nasopharynx, lymphomas, bladder, larynx, ovary, breast and multiple myeloma) each involving a review of 100 articles using an AI-generated search prompt. The evaluation focused on title and abstract screening, leveraging automated classification to identify relevant studies on occupational exposures and cancer risks. Only case-control, cohort, cross-sectional studies, and meta-analyses indicating an association between occupational sector and neoplasm were included in the selection.Results The data extracted by GPT were subsequently reviewed by a human gold standard expert in occupational epidemiology. Compared to expert classification, GPT missed 5 out of 164 relevant studies (false negative rate: 3.0%) and flagged 69 irrelevant ones as relevant out of 536 true negatives (false positive rate: 11.4%), demonstrating the potential of AI-assisted literature screening in streamlining systematic reviews.Conclusions The findings suggest that GPT-based automated search techniques can effectively support the initial phases of systematic reviews by efficiently identifying relevant occupational cancer studies with a relatively low rate of false negatives. These results support the feasibility of using GPT-based tools for first-pass screening in occupational cancer SRs. While further refinement is needed to reduce false positives, this approach offers a promising balance between speed and sensitivity, with potential for broader integration into systematic review workflows