Machine Learning Prediction of Incomplete Hysteroscopic Myomectomy Using Preoperative Clinical and Imaging Variables.
Authors
Affiliations (3)
Affiliations (3)
- Helen Schneider Hospital for Women, Rabin Medical Center, Petach Tikva, Israel; Faculty of Medical and Health Sciences, Tel Aviv University, Tel Aviv, Israel. Electronic address: [email protected].
- Helen Schneider Hospital for Women, Rabin Medical Center, Petach Tikva, Israel; Faculty of Medical and Health Sciences, Tel Aviv University, Tel Aviv, Israel.
- Beilinson Medical Center Innovation, Artificial Intelligence Center, Rabin Medical Center, Petah Tikva, Israel.
Abstract
To develop and validate a machine-learning (ML) model using preoperative clinical and imaging variables including ultrasound and diagnostic hysteroscopy findings to predict incomplete hysteroscopic myomectomy among women with submucosal leiomyomas. Retrospective cohort study. Tertiary referral center for minimally invasive gynecologic surgery with expertise in operative hysteroscopy (Helen Schneider Hospital for Women, Rabin Medical Center, Petach Tikva, Israel). A total of 345 procedures from 328 women who underwent hysteroscopic myomectomy for submucosal leiomyomas between January 2012 and December 2024 were included. Incomplete resection was defined as any documented residual submucosal myoma at the end of surgery. The overall rate of incomplete myomectomy was 16.2% (56/345). For model development, complete resection was coded as the positive class and individual risk of incomplete resection was obtained as 1-P(complete resection). A CatBoost binary classifier was trained using stratified 5-fold patient-level cross-validation. The model achieved moderate discrimination and high average precision for this imbalanced prediction task (AUROC = 0.72, average precision = 0.93) and outperformed logistic regression trained on identical inputs, with high PPV and sensitivity but only moderate specificity at the prespecified 0.50 probability threshold. FIGO type (2), larger myoma diameter, and multiplicity were the strongest predictors of incomplete resection. SHAP analysis confirmed consistent feature effects across folds, highlighting myoma morphology as the main driver of model predictions. A ML model integrating preoperative clinical and imaging data from ultrasound and diagnostic hysteroscopy predicted incomplete hysteroscopic myomectomy with moderate discrimination and modestly outperform conventional regression. This approach may help guide preoperative counseling and surgical planning by providing clinically useful risk estimates for incomplete hysteroscopic myomectomy.