Design: Retrospective study, with analysis employing the XGBoost machine learning algorithm that is capable of using complex data and provides a high level of interpretability using shapley values.

Setting: Tertiary referral center for pelvic pain and endometriosis.

Patients or Participants: All patients who underwent surgery related to endometriosis from 2011 to 2022 by a single surgeon.

Interventions: Retrospective data collection and development of an advanced AI model for analysis. A binary classification model was developed utilizing machine learning to analyze both categorical and continuous data provided by a dataset encompassing relevant patient history, examination findings, imaging, operative findings, and pathology results.

Measurements and Main Results: Two separate XGBoost models were developed using 5-fold cross-validation during training with 80% of the data. Testing was done with the 20% that was held out during training. The first model was a classifier tailored for categorical data and the second model was trained on continuous data.

Performance metrics were calculated for both models. Each model reported accuracy, precision, recall, F1-score, and ROC-AUC. The results indicated strong prediction primarily for the model trained on categorical data, reaching an accuracy of 83.44% and an ROC-AUC of 72.25%. Features related to pain characteristics were suggested to have a stronger and more consistent influence on the model’s prediction as indicated by positive SHAP values.

Conclusion: A useful prediction model for endometriosis is developed from a large dataset using advanced analysis with AI techniques.