Causal AI Predicts IVF Success and Reveals Which Changes Actually Help

For couples undergoing in vitro fertilization, one of the most agonizing questions is also the most basic: what are the real chances that this treatment will end with a baby? A new study from reproductive medicine specialists in Yantai, China, offers a fresh answer by combining machine learning with causal inference, a branch of statistics designed to separate genuine cause-and-effect relationships from mere correlations. The research, published in the Journal of Ovarian Research, presents a prediction tool for the cumulative live birth rate, the probability that a complete course of IVF, including all associated frozen-thawed embryo transfers, will ultimately deliver a live infant. Unlike most existing models, the new approach does not stop at producing a number; it also identifies which modifiable factors could shift that number upward, and it defines the threshold ranges within which clinical decisions become meaningful.

The team, led by Aiyu Zhang and Dongmei Wang of the Reproductive Medicine Center at Yantaishan Hospital, together with colleagues from Qilu Medical University and Shandong Medical and Pharmaceutical University, assembled a retrospective cohort of 1,058 couples who underwent 1,419 complete IVF cycles. The data were drawn from patients treated in Yantai, in Shandong Province, and were divided chronologically: couples treated between 2019 and 2022 formed the development cohort on which the models were trained, while those treated in 2023 served as a temporal validation cohort. This design matters because a model that performs well only on the data it was trained on is of limited clinical value. Temporal validation, testing the model on patients treated in a later period, provides a stricter test of whether the model’s accuracy will hold up as patient populations, laboratory practices, and clinical protocols evolve over time.

From the patient records, the researchers extracted variables spanning three domains: baseline patient characteristics, clinical parameters collected during treatment, and detailed embryo data. They then built prediction models using eight different machine learning algorithms, a standard strategy in predictive modeling that allows investigators to compare how different learning approaches capture the structure of the data. The winner of this bake-off was a gradient boosting machine, an ensemble technique that builds a strong predictor by sequentially adding simple decision trees, each new tree trained to correct the errors of the combined model so far. Gradient boosting has become a workhorse of modern medical prediction because it handles nonlinear relationships and interactions between variables gracefully, often outperforming simpler methods such as logistic regression when the underlying biology is complex and multidimensional.

The final model, which the authors call CLBR-GBM, relied on just eight predictors, a deliberately parsimonious design that eases implementation in busy clinics. Its performance in the temporal validation cohort was strong across several complementary metrics. The area under the receiver operating characteristic curve, or AUROC, a measure of how well the model distinguishes couples who will achieve a live birth from those who will not, reached 0.851. The F1 score, which balances precision and recall, was 0.801. The Brier score, which penalizes both poor discrimination and poorly calibrated probabilities, was 0.155, with lower values indicating better performance. An additional metric, the ABscore, registered 0.848. Together these figures indicate that the model not only ranked patients correctly but also produced probability estimates that were reasonably well calibrated, a property that is essential when the output is meant to inform real decisions rather than merely satisfy statistical curiosity.

What sets this study apart from the crowded field of IVF prediction models is its second act: a systematic effort to convert prediction into actionable insight. The authors deployed a battery of causal inference techniques, including serial mediation analysis, diverse counterfactual explanations known by the acronym DiCE, double machine learning, and the T-learner approach. Each of these methods addresses the same fundamental problem from a different angle. Standard machine learning models learn associations, but an association between, say, body mass index and live birth does not by itself prove that changing one will change the other. Confounding factors, variables that influence both the predictor and the outcome, can create spurious relationships. Counterfactual explanation methods ask a subtly different question: for a specific patient, what is the smallest change in the input variables that would flip the model’s prediction across a decision threshold? Double machine learning, meanwhile, uses flexible machine learning models to adjust for confounders while estimating treatment effects, and the T-learner estimates effects separately within groups that received different levels of an exposure.

The causal analysis converged on a clear message. The primary pattern that shifted a couple’s predicted probability across the optimal decision threshold toward a live birth was a synergistic combination of two modifiable factors: a decrease in body mass index and an increase in the number of embryos available for transfer. The average treatment effect for increased embryo number was estimated at 0.068, meaning that, on average across the cohort, having more embryos raised the predicted probability of cumulative live birth by about seven percentage points. The average treatment effect for decreased BMI was -0.010, a smaller average effect but one with striking heterogeneity. When the researchers examined subgroups, they found that BMI intervention showed its most robust causal efficacy among young high-responder patients with obesity, where the conditional average treatment effect reached -0.105, roughly a ten-and-a-half percentage point change in predicted probability. In other words, the benefit of weight reduction is not uniform across the population; it is concentrated in a specific, identifiable subgroup, which is precisely the kind of insight that individualized medicine requires.

The third innovation concerns the decision threshold itself. Most prediction studies report a single cutoff, often 0.5, above which patients are classified as positive. But the clinically optimal cutoff depends on the relative costs of false positives and false negatives, which vary with the decision at hand. The researchers used interactive decision curve analysis, or iDCA, to explore how the net benefit of acting on the model’s predictions changes across different thresholds. Net benefit is a metric that quantifies the clinical value of a prediction model by weighing the true positives it identifies against the harm of unnecessary interventions triggered by false positives. For this cohort, the optimal threshold was 0.429, yielding a net benefit of 0.331. Crucially, the model maintained robust net benefit, ranging from 0.312 to 0.379, across a broad threshold band from 0.3 to 0.5. This band structure is arguably more useful to clinicians than any single number, because it defines a zone within which the model remains clinically valuable, giving practitioners flexibility to adjust the cutoff according to individual patient preferences and risk tolerance, the very element the authors identified as missing from prior models.

The translational gap the study targets is a familiar one in medical artificial intelligence. Many published prediction models achieve impressive statistical performance in a journal but never influence a clinical conversation. Part of the problem is that a raw probability, however accurate, does not tell a patient or a physician what to do. By pairing the prediction with counterfactual explanations, the model can tell a specific couple what would need to change for their predicted probability to cross the decision threshold, and by validating the threshold band with decision curve analysis, the study anchors those explanations in a framework of clinical utility. This combination of prediction, explanation, and threshold optimization represents a template that other areas of reproductive medicine, and predictive medicine more broadly, could follow.

Several caveats deserve emphasis. The study is retrospective and non-interventional, which means the causal estimates, however carefully derived, describe effects inferred from observational data rather than from randomized interventions. The cohort comes from a single region of China, and although temporal validation within that setting was robust, transportability to populations with different demographics, treatment protocols, and laboratory standards remains to be demonstrated. The BMI finding, in particular, should be interpreted as evidence that weight is a promising interventional target in a specific subgroup, not as a prescription for all patients. The authors themselves frame the work as identifying causally augmented, temporally validated decision thresholds with balanced discrimination and robust transportability for cumulative live birth strategies, a measured claim consistent with the evidence presented.

Nevertheless, the practical output is immediately accessible. The team deployed the model as a publicly available web platform, hosted at fertility-yts.shinyapps.io/CLBR_GBM, where both patients and clinicians can enter individualized inputs and receive a predicted cumulative live birth probability. For a couple weighing whether to proceed with another transfer, take a pause, or consider weight reduction before a new cycle, a tool that combines a validated probability estimate with an honest account of which changes are likely to matter, and by how much, fills a genuine gap. The study was approved by the Clinical Trial Ethics Committee of Yantaishan Hospital and conducted in accordance with the Declaration of Helsinki, using anonymized data collected after routine informed consent. As machine learning continues to move into fertility clinics, this work suggests that the models most likely to earn clinical trust will be those that can answer not only what will happen, but what could be done about it.

Subject of Research: Causality-enhanced machine learning prediction of cumulative live birth rate in in vitro fertilization

Article Title: Causality enhanced machine learning for cumulative live birth rate prediction with threshold band optimization in in vitro fertilization

Article References: Zhang, A., Huang, X., Han, Z., Yang, Q., Sun, X., & Wang, D. (2026). Causality enhanced machine learning for cumulative live birth rate prediction with threshold band optimization in in vitro fertilization. Journal of Ovarian Research. https://doi.org/10.1186/s13048-026-02251-2

Image Credits: AI Generated

DOI: 10.1186/s13048-026-02251-2

Keywords: in vitro fertilization, cumulative live birth rate, machine learning, gradient boosting, causal inference, counterfactual explanation, double machine learning, decision curve analysis, body mass index, personalized prediction, reproductive medicine, temporal validation