Stage-structured, distributional prediction of IVF outcomes with conditional updating

To develop a stage-structured, distribution-based prediction framework for in vitro fertilization (IVF) that generates full probability distributions at each treatment stage and enables conditional updating of downstream predictions when observed outcomes become known.

We conducted an observational modeling study using de-identified UK Human Fertilisation and Embryology Authority (HFEA) registry data (2017-2018; up to 101,217 model-ready cycles after stage-specific filtering) to model egg retrieval, maturation, and fertilization. Egg retrieval was modeled using a zero-inflated negative binomial specification. Downstream transitions such as blastocyst formation, euploidy, vitrification survival, and live birth after euploid transfer were modeled using stage-specific logistic regressions calibrated to published cohorts and national registry summaries (429,507 additional observations). Predictive performance of HFEA-derived models was evaluated on held-out HFEA test sets using point-prediction accuracy, train-test gaps, prediction-interval coverage, and calibration across predicted outcome strata.

The framework propagates full probability distributions across sequential IVF stages rather than point estimates. Held-out HFEA validation showed minimal train-test degradation ( $R^2$ gaps under 0.007), with modest expected-count accuracy for egg retrieval and stronger expected-count accuracy for maturity and fertilization. Egg-retrieval prediction intervals showed near-nominal coverage (50%: 50.2%; 80%: 79.2%; 95%: 94.9%), and observed mean outcomes were close to predicted means across predicted-yield/rate strata. When observed stage outcomes were entered, downstream distributions updated appropriately, reducing uncertainty and preserving cycle-specific biological parameters in both single- and multi-cycle scenarios.

A sequential, distribution-based IVF prediction model with conditional updating provides uncertainty-quantified, stage-aware predictions that dynamically adapt to patient-specific outcomes, supporting more individualized counseling and treatment planning.