An ensemble machine learning model demonstrated greater predictive accuracy than conventional clinical formulas for determining initial gonadotropin doses in ovarian stimulation antagonist protocols, according to a retrospective study of 1,970 women treated at Peking University People’s Hospital. The Light Gradient Boosting Machine model achieved a test set mean absolute error of 53.25 international units and a coefficient of determination of 0.527, substantially outperforming the baseline formula benchmark.
Determining the starting dose of gonadotropin presents critical clinical challenges within in vitro fertilization treatments utilizing gonadotropin-releasing hormone antagonist protocols. Although antagonist protocols have gained widespread adoption due to their shorter overall treatment duration, reduced risk of ovarian hyperstimulation syndrome, and greater convenience for patients compared with long agonist regimens, initial dose selection remains particularly difficult. The shorter suppression phase and accelerated follicular recruitment characteristic of antagonist cycles require precise dosage calibration from the start of ovarian stimulation.
Inappropriate starting dosages in these assisted reproduction cycles carry substantial clinical risks for patients. Administering excessive amounts of gonadotropins can trigger severe complications such as ovarian hyperstimulation syndrome, whereas insufficient dosing frequently results in a poor ovarian response or premature cancellation of the treatment cycle. Despite these significant risks, standard clinical practice has continued to depend on simple linear regression nomograms and traditional formula-based calculations, which exhibit limited predictive performance, particularly in complex clinical presentations that subsequently necessitate mid-treatment dosage adjustments.
To address these dosing limitations, researchers analyzed clinical records collected between October 2016 and December 2021 from 1,970 female patients undergoing antagonist protocols at Peking University People’s Hospital. The research team evaluated multiple computational algorithms, including standard linear regression, random forest, and support vector regression frameworks. Following preliminary assessments of model performance, the investigators selected two advanced computational architectures—the Light Gradient Boosting Machine and a Multi-Layer Perceptron neural network—specifically designed to capture high-dimensional medical features and nonlinear physiological relationships.
Across the test dataset, the Light Gradient Boosting Machine, an ensemble learning method operating on gradient-boosted decision trees, consistently surpassed standard predictive benchmarks. While the traditional baseline formula produced a test set mean absolute error of 61.03 international units and an R² of 0.298, the gradient boosting model reduced the error to 53.25 international units while increasing the explained variance to 0.527. Concurrently, the Multi-Layer Perceptron neural network attained the lowest overall mean relative error at 0.212, improving upon the baseline formula relative error of 0.271.
The gradient boosting model proved particularly advantageous in challenging patient scenarios where clinical teams originally had to modify medication doses during stimulation. In these complex cases requiring dose modifications, the Light Gradient Boosting Machine achieved a mean absolute error of 67.53 international units. This algorithmic performance was substantially lower than the observed clinical adjustment error of 87.07 international units, demonstrating that the machine learning algorithm delivered more precise dosing recommendations than retrospective clinician estimates under difficult therapeutic conditions.
The investigators concluded that integrating machine learning algorithms, notably gradient boosting architectures, into clinical decision support systems could substantially improve initial gonadotropin dose optimization in antagonist regimens. Such computational tools could potentially enhance assisted reproductive outcomes, reduce patient complication rates, and lower overall fertility treatment costs. The authors noted that translating these models into everyday reproductive medicine will require prospective multicenter validation studies, the creation of accessible clinical user interfaces, and the investigation of deeper neural network architectures to further advance predictive precision.
Subject of Research: Machine learning prediction of initial gonadotropin dosage in IVF antagonist protocols
Article Title: Machine learning-based prediction of initial gonadotropin dosage in antagonist protocols
Article References: Wu, D., Yu, Y.-Q., Zhang, Z., Shen, H., Wang, Y.-B., Han, H.-J., Zhang, M., & Gao, F.-M. (2026). Machine learning-based prediction of initial gonadotropin dosage in antagonist protocols. Journal of Ovarian Research. https://doi.org/10.1186/s13048-026-02348-8
Image Credits: AI Generated
DOI: 10.1186/s13048-026-02348-8
Keywords: machine learning, gonadotropin dosing, antagonist protocol, in vitro fertilization, ovarian hyperstimulation syndrome, Light Gradient Boosting Machine, Multi-Layer Perceptron, follicular recruitment, Peking University People's Hospital, retrospective cohort study, mean absolute error, dose modification, decision support systems, poor ovarian response, cycle cancellation, clinical dose optimization
News Source: Teresa Odom. (November 18, 2026). Machine Learning Model Improves Initial Gonadotropin Dose Accuracy in Antagonist Protocols. Scienmag.



