Machine Learning-Based Prediction of IVF Outcomes: The Central Role of Female Preprocedural Factors

Abstract

Objectives: We aimed to develop and validate a per-cycle prediction model for in vitro fertilization (IVF) success using only preprocedural clinical variables available at the first consultation. Methods: We retrospectively analysed 1243 IVF/ICSI cycles (University of Szeged, 21 January 2022–12 December 2023). An Extreme Gradient Boosting (XGBoost version 1.7.7.1) classifier was trained on 14 baseline predictors (e.g., female age, AMH, BMI, FSH, LH, sperm concentration/motility, and infertility duration). A parsimonious 9-variable model was derived by feature importance. Model performance was assessed on the untouched test set and, as a final step, on an independent same-centre external validation cohort (n = 92) without re-fitting or recalibration. Results: The 9-variable model achieved an AUC of 0.876 on the internal test set, with an accuracy of 81.70% (95% CI 76.30–86.30%), sensitivity of 75.60%, specificity of 84.40%, PPV of 68.60%, and NPV of 88.50%. In external validation, the model maintained strong performance with an accuracy of 78.30%, confirming consistent discrimination on an independent same-centre cohort. Female age was the dominant high-impact feature, while AMH and BMI acted as “workhorse” predictors, and male factors added incremental value. Conclusions: IVF outcome can be predicted at the first visit using routinely collected preprocedural data. The model showed consistent discrimination internally and in external validation, supporting its potential utility for early, individualized counselling and treatment planning.