Development and internal validation of a post-retrieval machine learning models for OHSS risk stratification in...

This study aimed to explore the feasibility of developing and internal validating machine learning (ML) models for post-retrieval ovarian hyperstimulation syndrome (OHSS) risk stratification in patients undergoing assisted reproductive technology (ART).

In this exploratory, single-center, retrospective study, three ML models were established using clinical and laboratory data from infertile patients undergoing controlled ovarian hyperstimulation (COH) at Jinxin Xinan Women & Children Hospital between 2020 and 2023. The models were evaluated comprehensively using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, F1 score, calibration curves, and decision curve analysis. Model interpretability was achieved using the SHapley Additive exPlanations (SHAP) method to identify key predictors and clarify the basis of model decisions.

A total of 500 infertility patients were enrolled, including 40 diagnosed with OHSS. Participants were randomly divided into a training set (n = 350) and a test set (n = 150). Among the three ML models developed, the RandomForest model demonstrated the best discriminatory performance in the test set, achieving an accuracy of 0.79, sensitivity of 0.75, specificity of 0.80, F1 score of 0.37, and an AUC of 0.81 (95% CI: 0.68-0.95). The Least Absolute Shrinkage and Selection Operator (LASSO) and SHAP analysis identified number of oocytes retrieved, age, P levels on the trigger day, and ratio of baseline LH/FSH as the most influential predictors in the model.

This proof-of-concept study demonstrates that the ML-based approach shows exploratory promise for OHSS risk stratification in infertility patients undergoing ART. By integrating demographic and clinical variables, the model - particularly the Random Forest algorithm - may serve as hypothesis-generating framework for informing post-retrieval risk stratification, ultimately mitigating the risk of OHSS. Combining ML with SHAP provided explicit, individualized risk interpretation that may help clinicians intuitively understand the impact of key predictive features. However, this interpretability does not imply validated clinical utility. These findings are exploratory and hypothesis-generating. They require rigorous external validation in larger, multi-center prospective studies before clinical implementation.