An Artificial Intelligence-Driven Predictive Model for Microinvasive Ovarian Carcinoma and Serous Tubal Intraepithelial Carcinoma (STIC) Risk in BRCA-Mutated Women undergoing risk-reducing salpingo-oophorectomy

Background Pathogenic variants in the BRCA1 and BRCA2 genes significantly increase the lifetime risk of developing breast and ovarian cancer. Current guidelines recommend biannual transvaginal ultrasound and serum CA-125 surveillance, alongside Risk-Reducing Salpingo-Oophorectomy (RRSO), typically performed around age 35-40 for BRCA1 carriers and age 40-45 for BRCA2 carriers. While RRSO is highly effective, reducing the risk of ovarian cancer by 80-96%, it imposes premature surgical menopause, leading to potential metabolic, cardiovascular, and bone health complications, thus negatively impacting the quality of life (QoL). Different series about pathologic findings at RRSO have shown that, in women with a pathogenic BRCA1/2 variant, 4.5% to 9% were found to have occult gynecologic neoplasia, including invasive carcinomas and intraepithelial lesions, through thorough pathologic examinations1 of their ovaries and fallopian tubes. There is a critical need for non-invasive tools to identify early-stage or precursor lesions, potentially optimizing the timing of surgical intervention. Purpose The primary objective of this study is to develop an Artificial Intelligence (AI)-driven diagnostic algorithm capable of achieving early, non-invasive detection of microinvasive ovarian cancer and its precursor lesion, STIC, in asymptomatic women carrying a germline BRCA mutation. Methods A retrospective cohort of 344 BRCA1/2-positive patients who underwent RRSO between 2007 and 2025 across three Italian centers was analyzed. The dataset integrated clinical parameters (parity, BMI, menopausal status), longitudinal CA-125 biomarkers, and ultrasound findings. Statistical associations were assessed via univariate analysis and multivariable logistic regression. Subsequently, an explainable ML model (Decision Tree) was developed using the Orange data mining suite to stratify risk. Results Occult malignancy was identified in 7.3% of the cohort (STIC: 4.1%; invasive cancer: 3.2%). Univariate analysis identified preoperative CA-125, parity, menopausal status, and estroprogestinic therapy (EPT) as significant predictors. A derived clinical point-score achieved an Area Under the Curve (AUC) of 0.874, with a CA-125 threshold of 20.03 U/mL showing superior sensitivity compared to standard clinical cut-offs. The Decision Tree model prioritized CA-125 as the primary discriminant (threshold 27.30 U/mL) and achieved high sensitivity (0.94), emphasizing the reduction of false negatives in a screening context. Conclusion The integration of clinical scoring and ML algorithms enhances the detection of occult adnexal malignancies. By utilizing a lower CA-125 threshold and incorporating reproductive history, these models provide a robust framework for personalized surgical counseling and risk stratification in high-risk populations. Our study demonstrates that integrating clinical variables, specifically a lowered CA-125 threshold (20.03-27,3 U/mL), parity, menopausal status, and EPT use, significantly enhances the detection of occult neoplasia in BRCA carriers. The synergy between the Clinical Score (NPV 0.97) and Machine Learning algorithms (Sensitivity 0.94) provides a robust framework for personalized risk stratification. While derived from a preliminary cohort, these findings suggest that a multi-parametric diagnostic approach could replace static age-based guidelines, enabling clinicians to identify the optimal surgical timing for RRSO. Scalability to larger, multicenter cohorts is warranted to validate this model as a standard of care in precision preventive oncology.