Prediction of Complicated Ovarian Hyperstimulation Syndrome in Assisted Reproductive Treatment Through Artificial Intelligence

Abstract

Background This study explores the utility of machine learning (ML) models in predicting complicated Ovarian Hyperstimulation Syndrome (OHSS) in patients undergoing infertility treatments, addressing the challenge posed by highly imbalanced datasets. Objective This research fills the existing void by introducing a detailed structure for crafting diverse machine learning models and enhancing data augmentation methods to predict complicated OHSS effectively. Importantly, the research also concentrates on pinpointing critical elements that affect OHSS. Method This retrospective study employed a ML framework to predict complicated OHSS in patients undergoing infertility treatment. The dataset included various patient characteristics, treatment details, ovarian response variables, oocyte quality indicators, embryonic development metrics, sperm quality assessments, and treatment specifics. The target variable was OHSS, categorized as painless, mild, moderate, or severe. The ML framework incorporated Ray Tune for hyperparameter tuning and SMOTE-variants for addressing data imbalance. Multiple ML models were applied, including Decision Trees, Logistic Regression, SVM, XGBoost, LightGBM, Ridge Regression, KNN, and SGD. The models were integrated into a voting classifier, and the optimization process was conducted. The SHAP package was used to interpret model outcomes and feature contributions. Results The best model incorporated IPADE-ID augmentation along with an ensemble of classifiers (SGDClassifier, SVC, RidgeClassifier), reaching a recall of 0.9 for predicting OHSS occurrence and an accuracy of 0.76. SHAP analysis identified key factors: GnRH antagonist use, longer stimulation, female infertility factors, irregular menses, higher weight, hCG triggers, and, notably, higher number of embryos . Conclusion This novel study demonstrates ML's potential for predicting complicated OHSS. The optimized model provides insights into contributory factors, challenging certain conventional assumptions. The findings highlight the importance of considering patient-specific factors and treatment details in OHSS risk assessment.