An XGBoost predictive model of ongoing pregnancy in patients following hysteroscopic adhesiolysis

Abstract

Research question: What are the factors influencing the fertility of patients with intrauterine adhesions (IUA) after hysteroscopic adhesiolysis and which assessment system is more efficient in predicting post-operative ongoing pregnancy?

Design: The clinical information of 369 individuals diagnosed with and treated for IUA were obtained from the Multicentre Prospective Clinical Database for the Construction of Predictive Models on Risk of Intrauterine Adhesion (NCT05381376) and randomly divided into the training and validation cohorts. A univariate analysis was performed to identify relevant clinical indicators, followed by a least absolute shrinkage and selection operator (LASSO) regression for regularization and SHapley Additive exPlanation (SHAP) for extreme gradient boosting (XGBoost) predictive model visualization. Finally, receiver operating characteristic (ROC) curves were constructed to assess the model's efficiency.

Results: Univariate analysis and LASSO regression demonstrated that 12 clinical indicators were significantly associated with post-operative ongoing pregnancy in IUA patients. SHAP visualization indicated that post-operative Fallopian tube ostia, blood supply, uterine cavity shape and age had the highest significance. The area under the ROC curve (AUC) of the XGBoost model in the training and validation cohorts was 0.987 (95% CI 0.979-0.996) and 0.985 (95% CI 0.967-1), respectively. These values were significantly higher than those of the American Fertility Society (AFS) classification, the Chinese Society for Gynecological Endoscopy (CSGE) classification and endometrial thickness (all P < 0.001).

Conclusions: The XGBoost model had higher accuracy in predicting post-operative reproductive outcomes in IUA patients. Clinically, the model may be useful for managing and categorizing IUA and determining optimal action to aid in pregnancy.

Keywords: Extreme gradient boosting; Fallopian tube ostia; Hysteroscopic adhesiolysis; Intrauterine adhesions; Machine learning; Ongoing pregnancy.