Machine learning model predicts early pregnancy loss after IVF using five routine clinical markers

Researchers have developed a machine learning model that estimates the risk of early pregnancy loss after in vitro fertilization or intracytoplasmic sperm injection using five indicators that fertility clinics already collect routinely. The retrospective cohort study, published in BioData Mining, analyzed data from couples undergoing their first IVF/ICSI cycle and produced an online risk calculator intended to support counseling and risk stratification.

The study included 1,201 couples whose first IVF/ICSI cycle resulted in an ultrasound-confirmed intrauterine clinical pregnancy. Outcomes across the study population comprised 979 live births and 222 cases of early pregnancy loss, providing a substantial basis for model development.

Feature selection combined univariable analysis, Boruta, and LASSO regression to identify the most relevant routine clinical features.

Five core predictors emerged from the selection process: female age, body mass index, basal follicle-stimulating hormone, platelet count, and thyroid-stimulating hormone. All five are standard measurements gathered during fertility workups, which means the model can be applied without additional testing or cost. Among the evaluated algorithms, XGBoost, a gradient-boosted decision tree method, showed the highest area under the receiver operating characteristic curve together with a relatively low Brier score, indicating both strong discrimination and well-calibrated probability estimates.

The SHAP analysis ranked basal follicle-stimulating hormone as the largest contributor to the model’s output, followed by body mass index, thyroid-stimulating hormone, female age, and platelet count. This ordering offers clinicians a transparent view of which factors drive an estimated risk in a given patient.

Because the model relies exclusively on routinely available indicators, the authors suggest it could be integrated into clinical workflows with minimal friction, and a web-based calculator was developed to facilitate its implementation.

The researchers are careful to frame the tool’s status. Their conclusion states that the XGBoost model based on the five indicators may assist in estimating early pregnancy loss risk after IVF/ICSI, that it is intended to support risk stratification and counseling, and that it requires external validation before broader use.

As a retrospective, single-cohort study, the work establishes an internally validated predictive tool rather than a clinically deployed one. The reported performance figures, the identified predictor panel, and the SHAP-based interpretation together describe a model that estimates risk from data clinics already hold, with its real-world value still to be confirmed in independent populations.

Subject of Research: Machine learning prediction of early pregnancy loss after IVF/ICSI using routine clinical features

Article Title: A machine learning–based prediction model for early pregnancy loss after IVF/ICSI: a retrospective cohort study

Article References: Li, H., Zhang, R., & Li, Y. (2026). A machine learning–based prediction model for early pregnancy loss after IVF/ICSI: a retrospective cohort study. BioData Mining. https://doi.org/10.1186/s13040-026-00599-1

Image Credits: AI Generated

DOI: 10.1186/s13040-026-00599-1

Keywords: machine learning, early pregnancy loss, IVF, ICSI, XGBoost, prediction model, SHAP, risk stratification, retrospective cohort, assisted reproduction, BioData Mining, machine