Individualized Prediction of Recurrence Following Uterine-Preserving Pelvic Organ Prolapse Repair Using Internally Validated Machine Learning Models

Background/Objectives: Approximately 11–20% of women affected by pelvic organ prolapse (POP) will undergo surgical intervention during their lifetime and nearly one-third of these women will require an additional procedure due to recurrence or failure of the initial repair. This study aimed to identify risk factors for failure and to develop models to predict subjective, anatomical, and composite failure after primary uterine-preserving POP surgery. Methods: We performed a retrospective cohort study of women undergoing primary uterine-preserving POP repair at a tertiary academic medical center between 2010 and 2024. Failure outcomes were defined as: subjective failure (patient-reported prolapse symptoms), anatomical failure (prolapse beyond the hymen during exam), and composite failure (subjective and/or anatomical failure and/or reoperation for prolapse). Risk factors were evaluated using univariable and multivariable logistic regression with odds ratios and 95% confidence intervals. Prediction models were trained in Python and internally validated using bootstrap optimism correction (500 resamples). Discrimination, overall accuracy, and calibration were evaluated. Results: Among 277 women, subjective failure occurred in 36 (13.0%), anatomical failure in 24 (8.7%), and composite failure in 46 (16.6%). In multivariable risk-factor models, larger genital hiatus was associated with higher odds of failure across endpoints and posterior compartment descent was independently associated with anatomical and composite failure. After bootstrap optimism correction, discrimination remained high for the best-performing models. Conclusions: In this single-center cohort, independent risk factors for failure after primary prolapse surgery were identified and prediction models showed strong internal discrimination. However, calibration limitations necessitate recalibration and external validation before clinical implementation.