Stage-Structured, Distributional Prediction of IVF Outcomes with Conditional Updating

BackgroundCurrent IVF calculators provide either cumulative success probabilities, such as the CDC IVF Success Estimator [1] and OPIS calculators [2, 3, 4, 5], or stage-specific point estimates such as the Orchid Embryo Banking Calculator [6], but they do not quantify uncertainty and cannot incorporate patient-specific outcomes observed during treatment. ObjectiveTo develop a distribution-based framework that (i) produces full probability distributions at each IVF stage and (ii) allows downstream predictions to update when new stage outcomes are known. MethodsWe constructed a sequential probabilistic model using fresh, autologous IVF cycles from the Human Fertilisation and Embryology Authority (HFEA) registry (2017-2018) [7] for egg retrieval, maturity, and fertilization, and integrated published clinical studies totaling over 435,000 additional observations for blastocyst formation, euploidy, freeze/thaw survival, and live birth after euploid transfer. Models were validated using 70/30 train-test splits with out-of-sample performance metrics. Egg retrieval is modeled with zero-inflated negative binomial (ZINB) regression; downstream stages apply sequential binomial filters. A "known value selection" mechanism conditionally updates predictions when observed counts are entered. ResultsThe model generates full probability distributions at each stage of IVF rather than point estimates. Multi-cycle modeling enables comprehensive family planning assessments, while known value updating on combined distributions maintains cycle-specific biology rather than averaging outcomes. When observed values are entered, downstream distributions update accordingly, helping to guide clinical decisions. Out-of-sample validation demonstrates minimal overfitting with train-test R2 gaps under 0.007. Distribution evaluation confirms well-calibrated prediction intervals (50% coverage: 50.2%, 80% coverage: 79.2%, 95% coverage: 94.9%). The model is available as a web application at https://www.herasight.com/ivf-calculator. ConclusionsA distribution-based, sequential framework with conditional updating addresses key limitations of existing calculators by providing uncertainty-quantified, stage-aware predictions that adapt to patient-specific outcomes observed during care. Study Funding/Competing InterestsFunded by Herasight Inc. Authors are employees or consultants of Herasight.