Machine-learning estimation of live birth after AIH-based intrauterine insemination and its association with the number of inseminations per cycle

Intrauterine insemination (IUI) is widely used as a first-line treatment for infertility, yet predicting which couples will achieve live birth remains challenging. We aimed to develop an interpretable machine-learning model for live birth after IUI and to evaluate the impact of performing one versus two artificial inseminations by husband (AIH) within a cycle. We retrospectively analyzed 1,796 IUI cycles from 2015 to 2021 for model development and 314 cycles from 2022 for temporal validation at a reproductive center. Five routinely available predictors (female age, infertility duration, polycystic ovary syndrome, treatment protocol, and number of inseminations per cycle) were entered into five algorithms (logistic regression, random forest classifier [RFC], eXtreme Gradient Boosting, CatBoost, and multilayer perceptron). Performance was assessed using AUROC, F1 score, accuracy, and Brier score. Feature importance was interpreted using SHapley Additive exPlanations (SHAP). Propensity score matching (PSM) compared live-birth rates between one and two inseminations. The RFC showed the best discrimination and calibration, with an AUROC of 0.694 and the lowest Brier score in internal testing, and maintained the highest F1 score and accuracy in temporal validation. SHAP highlighted the number of inseminations per cycle, female age, infertility duration, treatment protocol and polycystic ovary syndrome as key predictors. After PSM, two inseminations per cycle were associated with a higher live-birth rate than a single insemination (10.19% vs. 6.70%, P = 0.020). An RFC-based model using five routine variables predicts live birth after IUI and supports considering two inseminations per cycle in selected patients.