Análisis de variables clínicas de mujeres con hiperrespuesta al estímulo ovárico controlado mediante Machine Learning

Abstract

Hyperresponsiveness to controlled ovarian stimulation presents a significant clinical challenge in in vitro fertilization (IVF) due to the coexistence of clinical risks and the difficulty in individually estimating implantation probability. This project aims to develop and evaluate a predictive model based on explainable machine learning to estimate the clinical implantation probability in patients with ovarian hyperresponsiveness, integrating clinical, endocrine, and embryological variables. The hypothesis is that combining nonlinear models, clinical interpretability, and probabilistic calibration improves the usefulness of predictions for medical decision-making. A retrospective observational study was conducted using anonymized clinical and embryological data provided by Clínica NeoVita, a private assisted reproduction center located in São Paulo, Brazil. The analysis was performed on a final cohort of 1,479 patients with complete and validated clinical and embryological records. Of these, 884 corresponded to patients with ovarian hyperresponsiveness and 559 to the control group. In the hyperresponsive group, 335 patients achieved pregnancy and 549 did not, while in the control group, 234 cases resulted in pregnancy and 325 did not. The CatBoost model achieved solid discriminative performance with a ROC-AUC of approximately 0.84, with adequate probabilistic reliability and clinical interpretability. It is concluded that the developed model constitutes a potentially useful tool to support personalized clinical decisions in assisted reproduction. Keywords: ovarian hyperresponsiveness; clinical implantation; in vitro fertilization; personalized reproductive medicine; explainable machine learning.