Patient-Centric In Vitro Fertilization Prognostic Counseling Using Machine Learning for the Pragmatist

Next, we present model design considerations using models reported in the literature to illustrate key points. As much as possible, we focus on pretreatment models designed to support patient counseling prior to starting the first IVF cycle. These model design considerations could be easily extrapolated to address other clinical scenarios—after failing one or more IVF treatments and when considering egg freezing or donor egg IVF, etc.

The IVF data source may be a single center, group of multiple centers, or national registry database. The choice depends on whether the model usage will be limited to one center, one group, or applied nationally. For optimal results, the training and test data should be representative of the same patient population. Although some researchers advocate testing whether an established model can be adapted and applied elsewhere, if feasible a center-specific model may yield better model performance metrics. 43 45 84 Since the model applicability may be heavily affected by the time period chosen, data should be from a more recent, shorter time period than to use more years of data. For example, intracytoplasmic sperm injection (ICSI) usage and other innovations improving embryology outcomes (e.g., extended embryo culture, blastocyst vitrification, and elective single embryo transfer [eSET]) became widely adopted in the mid-late 1990s and mid-late 2010s, respectively. 89 90 91 92 93 The center-specific implementation dates of those technologies should also be considered. For example, in recent years, freeze-all followed by serial transfers of single cryopreserved-thawed blastocysts until live birth is reached is increasingly applied when possible. Therefore, the selection of dataset parameters including time periods may affect the appropriate clinical usage of the resulting model. The relative usefulness of clinical predictors in each model depends on which other clinical predictors are being used, since clinical variables often have overlapping or redundant contributions to LBP models in IVF. Here, we use age and AMH to demonstrate a few practical points, but these concepts can be extrapolated to other clinical predictors as well. For example, if age and AMH are available for most IVF cycles in a dataset, then the relative importance of age and AMH in the resulting model may reflect their true relative contribution to LBP. However, if AMH is available only in a third of the sample, then the resulting model may still perform very well and may best serve that center's own patients even though the model may rely on age more than AMH. Compounding the above is that the relative weighting and scoring of AMH and age are expected to vary depending on the clinical profiles of patients seen locally at each center. Therefore, comments such as “the coefficient/relative importance of AMH is such and such in predicting IVF LBP” should be qualified by the specific patient population, time period, and other practice contexts. On the topic of clinical outcome to be modeled, model design requirements are distinct from the ongoing controversy in the literature on whether to use live birth or clinical ongoing pregnancy (COP; defined as reaching 12 weeks of gestation with documented fetal cardiac activity on ultrasound) as primary outcomes in clinical trials. 94 95 96 While we do agree with the goal of using live birth outcome as the outcome to be predicted, we caution that when using binary outcomes (e.g., yes or no, positive or negative), restricting the “positive class” to only live birth necessitates labeling COP as negative class. Treating COP as “no live birth” may compromise model performance and clinical applications because in layman's term, that would be “inaccurate” as only an estimated 5% of COPs do not result in live birth. 94 Using an example of 200 first IVF cycles in a dataset, let us say that 100 cycles have documented live birth outcomes and 20 cycles have documented COP outcomes. Let us assume that 5% of those 20 cycles (1 cycle) later are found to have ended as second or third-trimester pregnancy losses. By restricting positive outcomes to live births only, the positive class is 50% of the dataset. By applying positive outcomes to both live births and COP outcomes, the positive class is 60%. Later, when the fates of all COPs are confirmed, the positive class is 59%. A model trained using 60% of the data as a positive class is much closer to the truth (59%) than assuming that only 50% of the data has a positive class. Furthermore, it is not only the positive class rate that would be vastly underestimated but the bigger problem is that the model would be trained based on incorrect relationships between predictors and outcomes. You may ask, “Then why don't we wait until all the COPs have been followed up and we can use a dataset with confirmed final live birth vs. no live birth outcomes confirmed?” The answer to that question is explained in the section “Addressing the Risk of Data Drift ” later in this article. It may be helpful to note that the objective of a clinical prediction model for patient counseling is vastly different from clinical trials aiming to determine the efficacy and safety of a clinical intervention. Therefore, although the design of clinically practicable prediction models relies on the application of expert clinical research and modeling knowledge, it has requirements and best practices different from conventional clinical trial design. Last but not least, we should mention that the quality of the data preprocessing and processing steps are critical to successful validation of any prediction models, though they are typically given the least amount of attention in the literature. Establishing and adhering to rules and logic in the data pipeline and frequent code reviews and updates are important tasks in maintaining top-quality data for modeling.

Model review and validation are crucial steps before applying models in clinical practice to demonstrate performance, thereby building trust in model accuracy. Model review involves monitoring descriptive and analytical statistics for various variables, with experts reviewing any irregularities to detect and resolve issues throughout the modeling process. In this section and in Table 5 , we provide an overview of important model metrics with references for further details. Regardless of the metric used, there should be a control model for comparison. The AUC of the ROC may be the most widely reported model metric. The ROC AUC measures the ability of the model to discriminate or rank predictions showing the trade-off between the true positive rate (TPR) and false-positive rate (FPR). 116 117 118 119 While the ROC AUC measures a model's ability to rank predictions, it has significant drawbacks. For instance, it may not detect clinically meaningful improvements in the model. Moreover, the AUC can be artificially inflated by including specific patient groups, such as those older than 42 years or those with very low live birth rates, giving a false sense of reassurance about the model's performance. This metric is also not suitable for highly imbalanced datasets. To address limitations of ROC AUC, additional metrics that measure different attributes of the model can be considered ( Table 5 ). In particular, we created the metric, the Posterior Log of Odds Ratio Compared with Age Model (PLORA), for measuring predictive power in the specific context of IVF LBP. 2 36 37 38 86 PLORA compares the log-likelihood of the IVF prediction model to an age-based control model: that is, “how much more likely will this new model fit the observed data and outcomes compared to the age control model?” This metric is sensitive to dataset size and model improvements and can be communicated in linear scale ( e PLORA ) for easier understanding by clinicians. Observing a positive PLORA in conjunction with other model metrics provides a comprehensive indication of model performance. In addition to ROC AUC and PLORA, we employ other important metrics such as IVF LBP distribution, reclassification, and dynamic range to further evaluate models. 131 132 133 Reclassification examines whether more patients as a group receive higher live birth probabilities better reflecting actual live birth rates, while the dynamic range evaluates the highest and lowest possible live birth probabilities that the model may predict. These metrics provide insight into the strengths and limitations differentiating candidate models. 36 37 38 45 85 86 87 Precision, Recall, F1 Score, and PR AUC are also effective at detecting improvements in predicting positive live birth outcomes, especially in imbalanced datasets. 116 117 134 Precision, or positive predictive value, indicates the likelihood that predicted positive outcomes are correct, ensuring patients are not misled about their chances. Recall, or sensitivity, measures the model's ability to identify actual positive outcomes, ensuring that patients with high live birth probabilities are accurately identified. The F1 Score (i.e., the harmonic mean of precision and recall) balances both metrics to provide a comprehensive evaluation of the model's performance. The PR AUC plots precision versus recall without requiring a specific threshold and offers detailed views of the model's predictive capabilities across different thresholds. These metrics help demonstrate a model's ability to support clinical care and business operations. By leveraging a combination of validation metrics, we can provide more reliable prognostics, ultimately improving clinical decision-making and operational efficiency.

Effective IVF prognostic counseling requires accurate, personalized prognoses and clear, consistent communication of the prognoses. 25 26 The literature has focused on the reporting of IVF live birth rates, 27 28 patients' psychology, 17 25 26 and concerns over patients' overestimates of their personal IVF treatment success probabilities. 29 30 Furthermore, the communication of prognosis to patients tends to be unsupported and relegated to be a matter of personal style, under the label of provider autonomy as commonly seen in other areas of medicine. 31 However, patients' psychology may vary depending on the transparency or information symmetry between patient and provider and it is not possible to measure whether patients under- or overestimated their IVF prognoses in the absence of a validated model and effective communication. As the efficacy of IVF improves, patients should know their personal prognoses whether poor or excellent, based on their own health profiles. For example, patients with excellent IVF LBP may miss out on an opportunity to have a family if they were to underestimate their prognoses. For patients for whom IVF (with own or donor eggs) and IUI are both possible treatment options, patient counseling is especially important, as there is a wavering consensus on whether IUI or IVF should be offered as first-line treatment for patients with unexplained infertility. 32 33 34 On the personal front, patients may differ in how they perceive personal tradeoffs such as financial cost, time from work, and side effects versus having a family. Finally, a common complaint from patients is their perception of IVF as a gamble based on the uncertainty and lack of transparency about IVF treatment success on a personal level. Scaling IVF access and removing health inequities may require IVF prognostic counseling to be delivered by healthcare providers beyond fertility specialists. Motivated by a need to address the clinical and socioeconomic challenges for patients and society at large, we have developed an ML technology platform to support patient counseling, treatment protocol personalization, transparency-driven value-based IVF pricing design, and advancement of precision medicine through the use of accurate, validated clinical prediction models as summarized in the platform schematics. 35 This platform has generated published research, some of which will be further examined later. 2 36 37 38 39 40

For providers in countries that do not have a national registry-based online calculator, McLernon et al recommended performing a series of statistical testing, recalibration, and adaptations of the LR models produced using US SART or UK HFEA data by McLernon et al (US SART) and Ratna et al (UK HFEA), respectively. 43 83 84 122 123 However, Cai et al challenged this recommendation by showing that several MLCS models developed de novo using their own center's data, outperformed the US SART and UK HFEA models based on cross-validation results. In addition, Cai et al showed that the US SART- and UK HFEA-adapted models gave poorer validation results for patients younger than 35 years compared with patients 35 years old or older. 43 45 83 84 When using a multicenter model, it is important to understand whether there are variations in patients across centers and if so to quantify this variability. Using age–AMH–ovulatory disorder diagnosis as a multivariate measure of clinical profiles, Swanson et al reported that inter-center variation of clinical profiles is quantifiable and correlates to live birth outcomes. 124 Specifically, five distinct clinical profiles were demonstrated in 7,742 patients who received IVF treatment from 9 North American centers located in 33 cities across 11 U.S. states and Ontario, Canada. The proportion of patients having each of five distinct clinical profiles varied significantly across centers and the odds of having an IVF live birth varied across these profiles. 124 Also variations in local IVF treatment may contribute toward intercenter variations. Despite the perception that larger datasets enable more generalizable prediction models, it is important to consider the applicability of models for specific patients in specific centers. In other areas of medicine, there have been varying levels of success in applying ML to clinical registry data, with reports of success in producing clinically applicable models and registry data being inadequate for maximizing the utility of ML. 125 126 127 128 129 130 National IVF registries were designed for monitoring outcomes and safety, not for supporting individualized prognostication. In Table 3 (Group 3), the models reported by Banerjee et al and Nelson et al were early prototypes followed by the training and validation by our group of over 50 MLCS models many of which have also been deployed for clinical utilization or to provide operational insights to individual fertility centers. We previously reported that a validated, center-specific ML model computes personalized IVF live birth probabilities with improved discrimination, dynamic range, and posterior log of odds ratio compared to age control models with a significant percentage of patients having higher live birth probabilities than would have been expected based on age alone. 2 36 37 Even controlling for the data source country (the United States, the United Kingdom), the U.S. and UK national registry based-models (the McLernon models) are different from the MLCS models reported by Banerjee et al (the United States) and Nelson et al (the United Kingdom) in many ways—data source, time period, single versus multiple centers, age limit, and AMH availability in the IVF cycles used, training using LR versus GBM, and validation metrics. 36 37 43 83 84 The key observation is that MLCS modeling using far fewer IVF cycles achieved comparable or better model performance compared with national registry-based models ( Table 3 ). Exploring possible causes for differences in multicenter versus single-center model performance may inform research and model improvement efforts. First, Ratna et al acknowledged that the UK McLernon model suffered from lack of AMH values in the HFEA data. 84 122 Second, in our experience when patients aged 42+ years are included in the training and test sets, the ROC AUC which reflects whole-model performance may not be representative of model performance for younger women as older women have a disproportionate number of true negatives. Consistent with our experience, Cai et al showed when applying the McLernon models to the Chinese dataset, and the McLernon models performed better in the older age group (≥35 years) than in the younger age group (<35 years). 45 When using multicenter data, one must control contributions to the training and test sets by each center, to avoid having a model that is overrepresented by centers with large volumes. Potential solutions to be tested include sampling and controlling for each center's proportional contribution to the final training and test sets or creating a center label to represent center-specific factors that are not captured by variables in the dataset. For example, Choi et al showed a method for each center to contribute predictive elements or trees rather than IVF cycles to make up the multicenter dataset. 38

Now we consider 12 publications to illustrate model design considerations in support of patient counseling prior to starting the first IVF treatment 36 37 38 45 82 83 84 85 86 87 (see Table 3 ). Accordingly, these publications (a subset highlighted in blue, Table 2 ) are selected based on the following: (1) limiting predictors to information available at the time of pretreatment counseling prior to the first IVF cycle; (2) using a dataset with known live birth outcomes; (3) validation of the model using an independent test set or cross-validation. In addition, we include models with known clinical usage even if they do not satisfy all these criteria (a subset highlighted in green, Table 2 ). 43 88 However, we did not include models that have not been reported in peer-reviewed research literature. Abbreviations: AUC, area-under-the-curve of receiver operating characteristic curve; GBM, gradient boosted machine; GLM, generalized linear model; LGBM, light gradient boosted machine; LR, logistic regression; MLCS, machine learning, center-specific model; PLORA, posterior log of odds ratio compared to age control; RF, random forest; SVM, support vector machine. U.S. Patents including U.S. Patent Number 9,458,495B2; foreign counterparts; and patents issued. A separate study by Leijdekkers et al, 2018 (155), performed external validation of the McLernon IVF pretreatment model and subsequently updated/recalibrated the model to correct for slight overestimation. External validation of the recalibrated model was not specified.

Although many ML models have been reported in many areas of medicine, there are additional requirements for successful implementation in clinical care as summarized by patient-centric communications; provider collaboration, usability, and explainability; relevance and model performance; ability to handle complex data; scalability including maintenance of quality, user experience, and economic feasibility as usage scales; and adhering to best practice and compliance throughout the product life cycle from raw data processing to production model deployment. 119 135 136

Models trained and validated on historical data risk not being true for patients treated today or in the future. Such a risk, called data drift, can occur through input data drift, concept drift, and clinical context-related data drift, though these data drift subtypes can affect one another. 120 121 Input data drift includes changes in diagnostic labeling (e.g., documenting polycystic ovarian syndrome as an ovulatory disorder rather than the more specific PCOS diagnosis), demographic changes (e.g., having an increase in the proportion of women younger than 35 years). Data drift related to clinical context of use includes changes in patient management and disease prevalence (e.g., a center had recommended IVF as a first-line treatment to patients with unexplained infertility but now changes to recommend IUI as first-line treatment instead; thus, IVF patients with unexplained infertility will now be patients who already failed at least one to two IUI treatments and the IVF live birth outcomes for this altered patient group may be lower than observed in the historical data). Concept drift, a consequence of input data drift or clinical context-related data drift, refers to a changed relationship between predictor and live birth outcome (e.g., in the previous example, the concept drift is that unexplained infertility as a predictor may now be associated with a lower IVF live birth probability due to different clinical management of those patients). Fortunately, providers tend to be conservative; so, major changes in protocols or treatment recommendations tend not to occur over a short period of time. Having validated plans for monitoring, reducing, and addressing data drift if needed can help prevent provider data drift concerns. First, providers should understand that healthcare practitioners are only expected to give patients “the best available current prognostic information.” In the absence of a validated model, it is more difficult to speak to the quality of the prognostic information. Providers have a simple task of alerting model makers if they make significant treatment changes or if there is a change in the patients treated. In addition, we recommend updating the IVF live birth prediction model using the latest available data every 2 to 3 years or sooner if there are changes to patient demographics or treatment. At the time of model update, we also perform “LMV,” validation of the deployed, “live” model using the more recent, and out-of-time data as a test set. Despite data drift deserving caution, fertility centers are typically extremely cautious and avoid making sudden, significant changes to protocols. We have confirmed LMV for fertility centers that have requested it; further, based on a formal reporting of LMV for a sample of six fertility centers, we advocate testing or providing LMV for a larger sample of centers to determine whether this observation can be generalized. 86

A review of predictors of success after IVF by Shingshetty et al following a comprehensive literature search between 1978 and 2023 identified 1,810 publications meeting initial keyword search requirements from which 43 articles were selected for detailed review. 41 However, pragmatically before considering prognostic models it is important to first specify the model objective, which will in turn define relevant clinical variables, outcomes, and other dataset attributes. Table 1 shows the importance of defining the model objectives (e.g., pretreatment counseling, research) and, in the case of pretreatment counseling, the exact clinical contexts (e.g., prior to the first IVF cycle, after a failed IVF cycle). The allowable clinical variables, required outcomes, and data segments to consider for exclusion can then be easily determined. The clinical variables that are commonly tested for predictive value in pretreatment IVF prognostics include age, BMI, ovarian reserve tests (e.g., AMH, AFC, Day 3 FSH), reproductive history, prior IUI or IVF treatment history, and clinical diagnoses such as PCOS, tubal factor, male factor, endometriosis, diminished ovarian reserve, smoking, and the duration of infertility. 42 For example, if the research objective was to understand clinical factors impacting IVF live birth outcomes, it is reasonable to use a dataset comprising clinical variables obtained from pretreatment, ovarian stimulation, and embryology. However, if the objective is to create a prognostic model to counsel patients at the pretreatment stage to consider using IVF for the first time, the variables should be restricted to information that is known and available at the time of patient counseling. The model outcome should be selected to enable the provider to respond to patients' needs. For example, patients wish to know their probability of having a live birth, not a positive pregnancy test. The concept of “reading the ending first” is key, otherwise you may build an excellent model with no clinical utility. This more pragmatic approach diverges from conventional scientific methods requiring the researcher to pose and test a series of hypotheses. Notes: This table is limited to model using structured data and does not attempt to address AI tools aimed to identify blastocysts for transfer. Also, this table serves to illustrate key principles and does not aim to provide an exhaustive list of possible models for prognostic counseling. For example, some scenarios of third-party reproduction are not shown. Taking a pragmatic approach we reviewed the 43 articles identified in Shingshetty et al and added further 7 articles—5 published prior to 2024, one published in 2024, and one submitted in 2024. 36 37 38 39 40 41 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 We reviewed all 50 articles and assigned them to subgroups based on clinical context (e.g., applicability of model to current IVF practice), presumed model objective (e.g., research or pretreatment counseling for a particular scenario), modeling method (e.g., logistic regression [LR] vs. other ML techniques), and whether an independent test set was used for validation or testing (see Table 2 ). Although outside of the scope of the IVF prognostic model for patient counseling, we appreciate the 19 articles contributing to the understanding of clinical factors influencing IVF outcomes and/or modeling methods without utility for patient counseling, referred to as “research objective” in Table 2 as they required information only available following treatment. 44 54 60 63 67 68 69 70 71 72 73 75 76 77 78 79 80 81

Despite reports of IVF live birth prediction models and online calculators for providers and patients, there are relatively few reports on such real-world use. 43 83 84 140 141 142 One study compared a group of Australian patients' perceived IVF LBPs against the U.S. SART calculator and another study compared a group of French patients' perceived IVF LBPs against the French registry data. 29 30 Limited reporting of IVF prognostic tool usage may reflect a variety of challenges such as difficulty in assembling the necessary expertise, limited AI real-world studies, or difficulty in fitting such studies into conventional medical research journals. We have reported clinical implementation of IVF live birth prediction models and our team's multisite implementation experience using the framework recommended by Goldstein et al. 2 86 137 143 A real-world study of the retrospective multicenter experience of 24,238 new IVF patients suggested that usage of a patient-centric, MLCS-based prognostic report was associated with increased IVF conversion among new fertility patients. 143 This study suggested to investigate factors influencing treatment decision-making and real-world optimization of patient-centric workflows incorporating MLCS prognostic reports. 143 Medical research journals might consider a category dedicated to AI applications encouraging AI publications pertaining to medicine. In addition, it may be helpful to support reviewers with articles on the topic of AI implementation, usage, and guidelines. Indeed, the following guidelines facilitate informed and productive review processes. TRIPOD-AI, STARD-AI, PRISMA-AI, CONSORT-AI, and SPIRIT-AI are extensions of widely used TRIPOD, STARD, PRISM, CONSORT, and SPIRIT guidelines, respectively. 144 145 146 147 148 149 Since the original, non-AI versions of these guidelines have been widely used by researchers, the AI versions have also become widely recognized by journals. Unlike other guidelines, the DECIDE-AI guideline was designed for early clinical study stage evaluation of any AI modalities (e.g., diagnostic, prognostic, therapeutic), in live clinical settings and importantly does not require any one study design. 149 The DECIDE-AI guideline prioritizes assessing the risk of data shift and reporting of clinical implementation experience as-is to expedite sharing of the usage experience. 149

Having described the design, development, validation, and deployment of personalized ML IVF prognostic models, it may be helpful to return to the broader vision of advancing reproductive medicine and increasing fertility care accessibility. Economic modeling using ML IVF prognostic models can inform the allocation of funding to support fertility care with strategies at the local, regional, or national levels. Most importantly, these local strategies are aligned with a global, scientific and ethical approach adaptable to local fertility centers' clinical care and operational needs. A global collaboration of public, private, research, and operational groups developing validated ML IVF prognostic models help to prioritize women's and couples' family-building goals. By helping more people who are proactively seeking fertility care to become parents, we may also help to mitigate the macro level impact of below replacement fertility currently experienced by more than half of all countries.

In the context of ML-supported provider–patient prognostic counseling, scalability refers to the ability of the ML platform to serve patients through many fertility centers and providers globally with great implementation and model update efficiency at low costs while preserving or improving the quality of the IVF prediction models, the counseling reports, and other supportive services. Scalability is important as it enables the delivery of prognostic information to diverse patient populations. Scalability may be achieved in several ways. For instance, a proprietary, end-to-end platform may be used to support data model pipelines, no-code implementation of customized model and counseling report specs, deployment and usage of models, report generation, multilingual function, and administrative module. However, the platform and related processes must comply with applicable local data privacy laws such as U.S. Health Insurance Portability and Accountability Act (HIPAA) and the European General Data Protection Regulation (EU GDPR). The regulatory framework and pathways governing medical devices comprising AI/ML and software have continued to evolve with the increasing complexity of the devices and needs of patients and providers. 119 Beyond compliance, it is important to conduct 360-degree review with key stakeholders including collaborators, fertility center, clinical and operational leads and teams, internal team, and patients (directly or indirectly via providers) to consider potential unintended consequences of data strategy, product, and life cycle management decisions including efforts to maximize inclusion of diverse patient groups, ensuring that the data are representative of the people served by the resulting model, and ongoing efforts to maximize affordability of and access to fertility treatments. 4 5

Successful application of ML at point-of-care requires the merging of clinical care and data/ML workflow in a streamlined way for the clinical team. In view of staff constraints, it is imperative that the streamlining is operationally and efficiently tailored to each center. 137 Many fertility centers have been scaling up services to meet increasing patients' demands for IVF. In the United States, care teams have been augmented by training advanced practice providers (APPs) and/or general obstetricians and gynecologists. 137 138 139 It is essential that all users are trained on the use of any prognostic tool, but a good tool can support their patient counseling. Additionally, optional integration with fertility centers' electronic health record systems enable largely automated generation of counseling reports.

An IVF prognostic model supports provider–patient relationship at a critical point of the patient's care; therefore, the IVF prognostic information must be clear and easy to understand to both providers and patients. The prognostic information may be presented in an individualized counseling report that includes not only the prognostic information but also the key factors that underpin the prognosis in a graphical format illustrating how the patient compares to other patients treated at the same center. 23

Despite its proven safety and efficacy, ART is vastly underutilized and even in patients for whom ART is appropriate and funded through national reimbursement many patients stop treatment prematurely when they would have still had a good chance of success, if they had continued ART. 6 7 8 9 10 11 12 13 One in six people in the reproductive-age group or 100M+ is estimated to have clinical infertility worldwide. 6 7 In the United States, an estimated 10M+ people have infertility, defined by the American Society for Reproductive Medicine (ASRM) as needing medical care to conceive or have a successful pregnancy. 7 8 9 However, less than 2% of women or couples who could benefit from ART actually used it based on annual reporting by the Society of Assisted Reproductive Technologies (SART). 9 Globally, ∼1M babies are born from ∼4M ART cycles performed annually. 10 11 Realizing the full family- and society-building potential of IVF requires identifying and solving the barriers in patients' navigation of fertility care for one in six women or couples, nontraditional families such as single women and same-sex couples, and families with hereditary genetic diseases. 6 7 8 9 10 11 12 13 14 The causes of the underutilization of ART are complex, but the main barriers cited are emotional stress, uncertainty of treatment success, financial cost, inadequate insurance coverage, and issues arising from limited mechanistic knowledge. 7 14 15 16 17 18 In the United States, while some level of IVF coverage increased from 34% (2015) to 42% (2020) among U.S. employers with 20,000 or more employees, an estimated ∼40% of Americans with employer-funded health insurance did not have IVF coverage. 19 20 Warranting special mention are the racial and socioeconomic disparities in fertility care in the United States. 14 20 Despite higher reporting rates of infertility, Black women are less likely to receive fertility diagnostic testing even when care is sought and less likely to receive fertility treatments including IVF. Black and Hispanic women also have lower fertility treatment success for reasons that are not well understood. 5 14 20 Most countries including Sub-Saharan Africa, low- and middle-income countries in LATAM, South Asia, and East Asia do not have government-funded ART, resulting in low ART utilization. 21 In contrast, many of the industrialized countries already have national support for IVF as a health measure or to attempt to tackle the problem of declining fertility rates or both. 22 Despite such state funding, personal accurate IVF prognostics are typically not available to help patients choose effective treatments and to minimize using ineffective treatments. In fact, ML-based, personalized prognostic counseling may improve efficiency, patient retention, and IVF outcomes while optimizing the use of resources. Many more people would have a family from IVF if they could afford several IVF treatments. A sustainable family-building program—whether paid by third parties or patients—should consider cumulative live birth probability per IVF cycle because this probability directly impacts the number and cost of IVF cycles needed to have a baby. Once a patient starts IVF treatment, the major limitation of achieving an IVF live birth is the high rate of treatment discontinuation or “drop-out” rate. From first-hand data analysis by our team, historically among self-pay patients (e.g., no government or third-party payer), approximately 80% tend not to return after one failed IVF cycle (aka 80% drop-out rate), while even state-funded or state-mandated covered patients may show a drop-out rate of 30 to 50% after one failed IVF cycle. Accurate IVF live birth prediction models can support the pricing of IVF treatments based on the outcome of having “a baby or a partial refund.” Commonly known as “shared risk” program initially popularized in the 1990s, this pricing method simply charges a discounted fee upfront for performing up to two to three IVF treatments, until a baby or a clinical pregnancy is achieved. If there is no live birth after three IVF treatments, then the patient would be paid a “partial refund.” Although patients and fertility centers theoretically benefit from this arrangement, without ML optimization, a substantial percentage of patients may not qualify, or fertility centers set a high upfront fee to protect from financial losses that can be incurred from suboptimal IVF success prediction. In contrast, an ML-driven shared risk program can be offered to the majority of patients and is compliant with the transparency requirements of the Ethics Committee of the American Society of Reproductive Medicine (ASRM). 23 24 Last but not least, information asymmetry currently exists between the business operations of many fertility centers and their providers and patients in which the business operations may have IVF LBP insights informing qualification for shared risk program; yet, those insights may not be known to providers and patients. We advocate for transparency and scientific literacy which are hand and glove and essential to advancing reproductive care and access to care. To patients and payers, IVF prognostic counseling and the cost of having an IVF baby are one and the same conversation, best supported by locally relevant data and ML.