Clinical challenges in the adaptation of AI predictive models in radiation oncology for gynaecological cancer: a systematic review by the radiation oncology-AI MITO group.

INTRODUCTION: The Radiation Oncology-AI MITO group conducted a systematic review to map the current landscape of AI-based predictive modelling in patients with gynaecological malignancies treated with radiotherapy (RT). The aim was to evaluate available evidence, methodological quality, and the clinical applicability of existing models. METHODS: Relevant studies were retrieved from PubMed, EMBASE, and Scopus following PRISMA guidelines. Methodological quality was assessed using the Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies (CHARMS) checklist, while risk of bias and applicability were evaluated using the Prediction Model Risk of Bias Assessment Tool (PROBAST). A total of 1402 records were identified, and 1025 unique studies were screened after duplicate removal. RESULTS: Thirty-one studies (1.1%) met eligibility criteria. Of these, 27 (87.0%) focused on cervical cancer, 2 (6.5%) on uterine tumours, and 2 (6.5%) on mixed gynaecological cohorts. Predictive models addressed clinical outcomes (35.5%), treatment-related toxicities (29.0%), or treatment response (29.0%). Studies incorporated 3-114 predictors, using clinical, dosimetric, and radiomic features, mostly from retrospective datasets. LASSO was the most frequent variable-selection method (25.8%). All studies reported internal validation, whereas only two (6.5%) performed external validation. According to CHARMS, 13 studies (41.9%) showed high risk of bias; performance reporting was heterogeneous, commonly limited by small sample sizes and suboptimal predictor selection. CONCLUSIONS: AI-based predictive modelling in gynaecological RT remains constrained by methodological variability, limited external validation, and a lack of clinically deployable tools. Advancing clinical translation requires rigorous model development, transparent reporting, robust validation, and interpretable tools supported by expert consensus.