A novel artificial intelligence model predicting assisted reproductive outcomes in polycystic ovary syndrome

Polycystic ovary syndrome (PCOS) shows endocrine–metabolic heterogeneity that clouds in vitro fertilization/intracytoplasmic sperm injection (IVF/ICSI) outcomes. We evaluated a stage‑wise prediction model that integrates baseline and intermediate treatment data and assessed generalizability. We conducted a retrospective cohort study (2019–2024) at a university hospital in southwestern China, including women with PCOS undergoing first IVF/ICSI with embryo transfer under a GnRH‑antagonist protocol (1719 women; 2539 cycles). Predictors covered age, BMI, ovarian reserve, androgen/metabolic indices, and intermediate results (ovarian response, trigger‑day hormones, embryo quality). Learners (tree‑based ensembles and neural networks) were trained in a locked sequential framework that propagated observed and predicted information across stages. Performance used AUC with resampled CIs, plus sensitivity, specificity, Brier score, calibration, and decision‑curve analysis. Feature importance was summarized by Shapley values. External validation applied the locked model without refitting to an independent external cohort of 204 patients. Fresh‑cycle clinical pregnancy was 39.5% (166/420); cumulative clinical pregnancy was 55.9% (961/1,719). The sequential model achieved AUC 0.834 for cumulative clinical pregnancy, outperforming single‑stage comparators, with good calibration and net clinical benefit. Androgen burden, metabolic indices, and trigger‑day follicular/steroid profiles were influential; embryo quality and early miscarriage risk were key sequential drivers. External validation maintained discrimination; sensitivity analyses were consistent. A stage‑wise model fusing baseline with intermediate treatment information yields accurate, clinically valuable, and externally generalizable predictions across the IVF/ICSI pathway for women with PCOS. Trial registration Ethics Committee approval No. 2020YFS0127.