INTRODUCTION: Endometriosis excision surgery carries procedure-specific risks that are not well captured by existing surgical risk calculators. Existing tools fail to account for the complexity of this population, leaving surgeons without an evidence-based method to inform patients of individualized risks, optimize surgical planning, and allocate resources. OBJECTIVE: To develop and internally validate artificial intelligence (AI) and machine learning–based models to predict major, minor, colorectal-specific, and genitourinary-specific perioperative complications in patients undergoing endometriosis excision surgery. METHODS: We performed a retrospective cohort study of patients undergoing endometriosis excision surgery at a tertiary center (2017–2024). Predictors, including demographics, comorbidities, surgical history, imaging findings, and perioperative assessments, and outcomes, including major complications (e.g., surgical site infections, sepsis, thromboembolic events, reoperation, and nerve injury), minor complications (e.g., urinary tract infection, superficial surgical site infection, readmission, transfusion, and pneumonia), and colorectal- and genitourinary-specific adverse events, were extracted from the EHR using Structured Query Language (SQL) and natural language processing (NLP), with manual chart adjudication by two blinded reviewers. Predictive models were developed using the H2O AutoML platform. The dataset was randomly divided into training (80%) and testing (20%) cohorts, and model training incorporated 5-fold cross-validation to optimize performance and minimize overfitting. Model performance was assessed with c-statistics, sensitivity, specificity, and predictive values across demographic subgroups. Discrimination was summarized with the area under the receiver operating characteristic (ROC) curve (AUC). RESULTS: Among 1,717 cases, major complications occurred in 12.0%, minor complications in 8.8%, colorectal-specific events in 4.4%, and genitourinary-specific events in 4.7% (Table 1). The models demonstrated strong discrimination (AUCs: major 0.89, minor 0.90, colorectal 0.91, genitourinary 0.93), with uniformly high negative predictive values (≥0.94) (Table 2). Predictor-importance analyses showed that BMI, Area Deprivation Index (ADI), and age were consistently among the most influential variables, with previous abdominal surgeries and prior endometriosis surgeries also contributing substantially (Figure 1). A user interface was developed to facilitate clinical application (Figure 2). CONCLUSIONS: To our knowledge, this is the first AI-based perioperative risk calculator tailored to endometriosis excision surgery. The model demonstrated strong predictive performance, with high discrimination across outcomes. This continuously evolving tool may help deliver individualized, complication-specific risk estimates using routinely available clinical and imaging data to enhance surgical planning and counseling. Future work will focus on multi-institutional external validation to enhance generalizability, application of continuous learning algorithms for model refinement, and prospective evaluation with calibration and decision curve analyses to assess clinical utility, with the ultimate goal of integration into a user interface for point-of-care use.Figure 1Table 1Table 2
Bar-El et al. (Fri,) studied this question.