Predicted probabilities of live birth following assisted reproductive technology using United States national surveillance data from 2016 to 2018.

Using national data from 2016 to 2018, our study provides the most up-to-date estimates of the cumulative probability of live birth based on demographic and reproductive characteristics for new and return autologous oocyte patients and donor oocyte recipient patients undergoing IVF in the United States. Moreover, additional detail is provided on the CLBR after each additional ET cycle within 12 months of oocyte retrieval to allow for further insight into a patient’s expected outcomes. Although a wide variety of clinical prediction models for IVF have been developed, 6 our model has several important distinctions. Most previous models have only estimated live birth rates after a single or multiple fresh ET cycles. 7 – 11 There is growing use of frozen ET cycles, which are now more common than fresh ETs. 1 , 12 Moreover, most existing models were fit using older data—from cycles completed more than 10 years ago—which does not reflect current practices. 7 , 22 For example, there has been increasing use of ICSI, 13 PGT, 14 and elective single ET for the past 2 decades, 15 all of which have influenced IVF success rates. In addition, there are instances where prediction models were developed for use in a specific patient population 23 or clinic. 24 Although this may be advantageous in certain scenarios, our goal was to provide useful information to as many potential IVF patients as possible by capitalizing on the NASS database, which contains approximately 98% of IVF cycles performed in the United States. 25 To date, the most similar clinical model comes from McLernon et al 26 who used data from the Society for Assisted Reproduction and Technology Clinic Outcome Reporting System 2014–2016 to develop a pre- and post-treatment model of CLBR after IVF. Similar to our study, they used a validated database that encompassed almost all of the nonbanking fresh and frozen IVF cycles in the United States during that period. Furthermore, they focused on a similar outcome—CLBR after 3 oocyte retrievals and all transfers from each oocyte retrieval in the following 12 months. However, there were several important differences. First, we used data from more recent years, 2016 to 2018, to develop our models. Furthermore, we did not apply as many exclusion criteria. We included women without infertility as a reason for IVF treatment and women who used PGT and donor sperm. This resulted in a larger sample size for both our new (152,426 vs 88,613) and return (44,490 vs 24,735) patient populations, which is likely representative of a wider scope of patients. Moreover, we included a third patient population, women undergoing donor oocyte IVF (n=25,831), to further enhance the generalizability of our findings to the widest scope of patients. In addition to providing the CLBR after a complete cycle (defined as all ETs within 12 months of the oocyte retrieval), we also estimated CLBR after each ET to provide further granularity on success rates to patients. Scenarios in which this may be useful include patients with too few embryos to support multiple transfers from a single retrieval or patients who only have insurance coverage for 1 transfer. Although the factors included in our models were similar (including age, BMI, parity, and specific infertility diagnoses), we did not include AMH or any posttreatment factors, such as number of oocytes retrieved. This was done to ensure that we could provide estimates of success to as many potential IVF patients as possible, even those who had yet to obtain a full infertility workup and may not know their AMH concentration. Despite these differences, our models achieved similar discriminatory ability and produced similar estimated probabilities of success to the McLernon study. Although our models in new and return patients undergoing autologous IVF performed well, as evidenced by C-statistics ranging from 0.69 to 0.82, the donor oocyte recipient model did not have a good discriminatory ability. This is likely because all of the recipient characteristics included in the model were only weakly associated with CLBR and the sample size was more limited. The CLBRs for donor oocyte IVF patients were high regardless of recipient characteristics. Future models in this population may want to consider adding donor characteristics or evaluating other recipient factors to further enhance the performance. First, because our data were derived from a large database meant for surveillance, we were limited by the accuracy of the data inputted by individual providers and clinics. Moreover, we were not able to include some potentially important predictors of success because either they were not reliably or routinely collected or they had too much missing data. For example, a wealth of literature supports an association between smoking and worse IVF outcomes 27 ; however, we did not find this to be strongly associated with CLBRs in our dataset likely because of underreporting. For the donor oocyte recipient models, we also lacked information on donor characteristics, most notably donor age, which may have improved our discriminatory capabilities. Furthermore, we explicitly decided not to include race and ethnicity in our models because of concerns about the validity of this variable and how results would be interpreted. 28 Markers of ovarian reserve, such as AMH, may also be more relevant for inclusion in future models as clinical and at-home testing for this biomarker become more standardized and routine. 29 Because of the nature of the NASS database, we were unable to identify women who were treated at >1 clinic. Although we expected this to be uncommon, it may have resulted in some patients being counted more than once. Although we included 8 infertility diagnoses in our models, some of these, such as unexplained infertility and infertility because of other reasons, are heterogeneous. Because of the lower number of patients who had 2 or 3 oocyte retrievals, our ability to predict CLBR in these patients was slightly lower. Nevertheless, the C-statistics from these models still indicated good discriminatory ability. Finally, our CLBRs after the second and third ETs may be conservative estimates. The only patients who had a second and third ET within 12 months of their oocyte retrieval are patients whose first ET cycle failed and, thus, may be lower prognosis patients. In addition, these patients would likely be using their second, third, fourth, or higher-choice embryos for these transfers, which may have resulted in lower success rates. Moreover, many patients did not attempt a second or third ET during the timeframe (for various unknown reasons), even though they had supernumerary embryos available. Had they cycled and had we been able to include their outcomes in this analysis, it could have boosted CLBRs within a year; however, this usage pattern reflects the reality of IVF in the United States; therefore, our estimated success rates reflected this. Using a large, national dataset encompassing nearly all IVF cycles in the United States, our study developed clinical models for use in new and return patients undergoing autologous IVF and donor oocyte recipient patients to provide individualized estimates of cumulative probability of live birth after multiple fresh and frozen ETs. These models will be used to update the IVF success estimator tool on the CDC website ( https://www.cdc.gov/art/ivf-success-estimator/index.html ) that can help inform patients and providers of a woman’s probability of having a live birth after IVF.

A total of 196,916 women (152,426 new patients and 44,490 return patients) who underwent autologous ART and 25,831 donor oocyte recipients were included in our analysis ( Table 1 ). Women undergoing autologous ART were most often between 30 and 40 years (74%), were non-Hispanic White (64%), and had no previous pregnancy (50%) or live birth (73%). The most common infertility diagnoses were male factor (32%), DOR (28%), and other reasons (20%). Most women undergoing donor oocyte ART were between 41 and 59 years (59%), were non-Hispanic White (61%), had no previous pregnancy (48%) or live birth (75%), and had a diagnosis of DOR (67%). The autologous ART patients underwent 207,766 autologous ET cycles, and the donor oocyte recipients underwent 36,909 transfer cycles in 2016–2018 ( Table 2 ). Most included cycles were frozen (60% for autologous and 46% for donor) and singe-embryo (59% for autologous and 65% for donor) transfers. ICSI was the most common fertilization method (75% for autologous and 68% donor), and PGT was used in more than a quarter of ET cycles (38% for autologous and 27% for donor). The unadjusted live birth rate per ET cycle was 43% for autologous and 49% for donor oocyte patients. The predicted probability of live birth increased with patient age from 20 to 30 years and steadily declined after 30 years ( Figure 1 , A ). Patient BMI had a nonlinear association with CLBR ( Supplemental Figure 1 , A ). The success rates remained relatively constant for BMIs 17 to 25 kg/m 2 but had a negative relation with CLBR for BMIs >25 kg/m 2 . In descending order of magnitude, new patients diagnosed with ovulatory disorder, unexplained, male factor, and tubal factor infertility had higher odds of live birth, whereas patients with DOR, uterine factor, and other reasons had lower odds of live birth compared with patients without that infertility diagnosis ( Figure 2 ; Supplemental Table 1 ). Parity and gravidity were inconsistently associated with live birth ( Supplemental Table 1 ). Our model predicted that a new patient who is 35 years, with a BMI of 25 kg/m 2 , no previous pregnancy or live birth, and an unexplained infertility diagnosis, has a 48%, 69%, and 80% CLBR after 1, 2, and 3 oocyte retrievals and all ETs (fresh or frozen) in the following year ( Figure 3 ). These CLBRs are 43%, 64%, and 76%, respectively, if an otherwise identical patient has a BMI of 35 kg/m 2 ; 29%, 48%, and 62%, respectively, if this same patient (with a BMI of 25 kg/m 2 ) has an infertility diagnosis of DOR; and 25%, 41%, and 52%, respectively, if this same patient (with a BMI of 25 kg/m 2 and unexplained infertility) was 40 years old. Among return patients, female age also had a nonlinear association with CLBR with success rates peaking around 30 years and steeply declining ( Figure 1 , B ). The negative association between higher BMI and lower CLBR was less consistent in return compared with new patients ( Supplemental Figure 1 , B ). Return patients with one or more previous pregnancies or live births had higher odds of live birth than nulligravid and nulliparous return patients ( Supplemental Table 2 ). The effect of infertility diagnoses on CLBR was consistent in direction and magnitude in new and return patients such that patients with ovulatory disorders had the highest predicted CLBR and patients with DORs had the lowest predicted CLBR ( Supplemental Table 2 ). Figure 3 , B , displays the predicted probabilities of live birth for return patients for various characteristics. Few recipient characteristics were associated with CLBR in new donor oocyte patients. Recipient age and BMI had weak, negative, inconsistent associations with CLBR ( Figure 1 , C ; Supplemental Figure 1 , C ). Tubal factor, uterine factor, and unexplained infertility were the only infertility diagnoses that were associated with lower CLBR after 1 ET ( Supplemental Table 3 ); however, after 2 ETs, only tubal factor, ovulatory disorders, and other reasons were associated with CLBRs. No infertility diagnosis was associated with CLBR after 3 ETs. Higher gravidity was associated with higher CLBR safter 1 ET but not after 2 or 3 transfers. Parity of the donor oocyte recipient was not associated with CLBR in any of the models. As evidenced from Figure 3 , C , the predicted cumulative probability of live birth for donor oocyte recipients was high, between 82% and 88%, after up to 3 ETs, and this did not vary significantly by recipient age, BMI, or infertility diagnosis. The computed C-statistic among new patients ranged from 0.69 to 0.78, indicating a good capacity to discriminate between patients with high and low probabilities of success ( Supplemental Table 4 ). Among return patients, the range of the computed C-statistic values was even higher (0.72–0.82). In contrast, the models among donor oocyte patients yielded C-statistics of 0.54 to 0.57, indicating poor discrimination ability. The calibration plots showed high agreement between the observed and predicted CLBR across ages among all 3 patient populations ( Supplemental Figure 2 ).

We used data from the CDC’s National Assisted Reproductive Technology Surveillance System (NASS), which contains information on approximately 98% of all IVF cycles performed in the United States. 4 To verify the accuracy of reporting, a random sample of fertility clinics is selected annually for data validation. Discrepancy rates are <5% for most fields. 4 All fresh and frozen IVF cycles using autologous or donor oocytes reported to NASS during 2016 to 2018 were eligible for this analysis. Long-term banking cycles where the intent was to cryopreserve embryos for use in ≥12 months were excluded. Epidemiologic research using NASS data is approved by the institutional review board at the CDC. We focused on 3 patient populations to develop the clinical models: (1) new patients, defined as having no history of ovarian stimulation, undergoing autologous ART who had an intended oocyte retrieval in 2016 or 2017; (2) return patients, who had a history of ovarian stimulation before 2016 or 2017, undergoing autologous ART who had at least 1 additional intended oocyte retrieval in 2016 or 2017; and (3) patients who used fresh or frozen donor oocytes for the first time for an ET cycle in 2016 or 2017. New and return patients with an oocyte retrieval in 2018 and donor oocyte patients who had their first transfer in 2018 were excluded as they did not have 12 months of follow-up to calculate CLBR. Demographic, reproductive, and clinical characteristics routinely collected in NASS were considered for inclusion in the models based on the availability of data (eg, frequency of missingness) and their known association with CLBR. The following variables were considered: patient age, body mass index (BMI), parity, gravidity, infertility diagnosis, race and ethnicity, smoking status of the female partner, age of the male partner providing sperm, and most recent antimüllerian hormone (AMH) level. Of the variables included in our final model, only BMI had missing data (18%) and was imputed. Furthermore, we used multiple imputation to account for the missingness of some key descriptive variables, such as race and ethnicity (36%), sperm source (14%), number of oocytes retrieved (8.6%), and use of PGT (10.9%) (see Supplemental Materials and Methods for more details on the multiple imputation), but these variables were not included in the final model. In brief, the variables that were dropped from further consideration had a high amount of missing data, had little to no association with live birth, and/or were parameters that would be unknown to patients who have yet to start their IVF cycle. A more detailed description of the variable inclusion process is provided in the Supplemental Materials and Methods . For all models, we focused on the cycle-level outcome of live birth delivery, which was defined as the birth of one or more live infants. In both new and return autologous oocyte patients, we calculated the CLBR for all fresh or frozen ET cycles that occurred within 12 months of the patient’s oocyte retrieval. In the NASS data, all ET cycles are linked to the specific oocyte retrieval the embryo originated from by oocyte retrieval date and patient identifier, thus allowing us to calculate the 12-month time window from retrieval to transfer. In new donor oocyte patients, we calculated the CLBR for up to 3 ET cycles that occurred within 12 months of their first ET cycle. Among new IVF patients, a cumulative logistic regression or multinomial regression model was used to estimate the probability of live birth after 1, 2, and ≥3 ETs that occurred within 12 months of a woman’s first oocyte retrieval. 20 For our models in new patients, the denominator was all new patients with intended oocyte retrievals, and the outcome had 4 levels: (1) if the first transfer resulted in live birth, (2) if the first transfer did not result in live birth but the second transfer resulted in live birth, (3) if the first and second transfers did not result in live birth but the third transfer resulted in live birth, and (4) if there was no birth for all transfers or there was no transfer within 12 months (reference group). Of note, most patients with no ET within 12 months had oocyte retrievals that were not successful—in other words, their retrieval was canceled because of poor response, they had no oocytes retrieved, or they had no viable embryos created. We did not remove the patients and cycles that did not progress to ET because this would have artificially inflated the CLBR estimates. Assuming that the contribution of transfer to live birth varies but contributions of predictors are similar across transfers, we used the common slope approach to develop our models. The option “cumlogit” was used, which renders the same slope for all covariates and different intercepts for each outcome group. Second and third multinomial regression models were run for all transfers occurring within 12 months of a woman’s second and third oocyte retrieval. To obtain the CLBR after the second or third retrieval and after up to 1, 2, and ≥3 ET cycles, we applied equations proposed by Luke et al (see Supplemental Materials and Methods for more details). 7 In addition, these equations were applied to calculate the CLBR after 3 oocyte retrievals and up to 1, 2, and ≥3 ET cycles. Among return IVF patients, we used a similar methodology. Logistic regression models were used to estimate the probability of live birth after a donor oocyte recipient patient’s first, second, and third ET cycle. To obtain the CLBR after the second and third ET cycles, we applied the equations proposed by Luke et al. 7 To determine the most appropriate way to model patient age, we plotted the empirical live birth rates by age and determined that a quadratic function was likely sufficient to model the nonlinear association between CLBRs and age. Next, we estimated the best-fitting power term using the PROC MODEL function in SAS (SAS Institute, Cary, NC). Patient BMI was modeled as a quadratic function across all models. To reduce the influence of outliers in the multivariable models, we truncated ages 46 years in new patients, 48 in return patients, 50 years in donor oocyte recipients as <1% of women fell outside these cutoffs. Across all models, we truncated BMIs 45 kg/m 2 , for the same reasons as age. Infertility diagnoses, including male factor, tubal factor, endometriosis, ovulatory disorder, diminished ovarian reserve (DOR), uterine factor, other (ie, recurrent pregnancy loss, medical contraindication to pregnancy, use of a gestational carrier, or testing for genetic abnormality in an embryo), and unexplained were modeled as indicator variables as >1 option can be reported. Gravidity and parity were each modeled as 3-level categorical variables (0, 1, and ≥2). To assess the predictive ability of the models, we computed a C-statistic using the SAS macro “MultAUC.” 21 Unlike the logistic model to which only 1 C-statistic is computed, in multinomial regression models, there are multiple C-statistics computed for each level of the outcome vs the reference. As the SAS macro is unable to handle datasets with multiple imputation, we used the average imputed BMI value for model fitting. The SAS macro is not currently able to provide an area under the curve or use the bootstrap method to obtain optimism-corrected C-statistic. We created calibration plots to show the agreement between observed and predicted CLBR for patients after 1, 2, and 3 oocyte retrievals. The observed probability of live birth for each retrieval was obtained in patients by age using Kaplan-Meier estimates, the number of patients with a live birth within 12 months of oocyte retrieval at a given age divided by the total number of patients at a given age. To generate the predicted CLBRs by patient age, first, we estimated each patient’s predicted CLBR using the parameters shown in Supplemental Tables 1 to 3 in combination with the patient’s specific values for each covariate. Next, we calculated the mean predicted CLBRs by patient age. Last, the observed and predicted CLBRs after a second and third oocyte retrieval in autologous oocyte patients and after a second and third ET in donor oocyte recipients were calculated using the Luke formula. 7

Assisted reproductive technologies (ARTs), including in vitro fertilization (IVF) with or without intracytoplasmic sperm injection (ICSI), have become one of the main treatment modalities for couples facing fertility problems. 1 In the United States, as of 2019, IVF contributes to >2% of all infants born. 2 Since 1995, the Centers for Disease Control and Prevention (CDC) has been collecting data on ART procedures performed in fertility clinics in the United States. 3 Annually, the CDC produces several reports (eg, clinic specific, state specific, and national) based on these data, 2 , 4 of which the clinic-specific report gives a potential patient an idea of their average probabilities of success, stratified by age group. However, because many other characteristics may affect the success of IVF, there is interest from patients and providers in developing clinical models that can estimate the individualized cumulative probability of a live birth before starting treatment and after any subsequent treatment cycles. 5 Although previous IVF prediction models have been developed, 6 most models are using older data (before 2010) and predict the probability of a live birth after a fresh embryo transfer (ET). 7 – 11 In addition, the practice of IVF is rapidly changing, including the important contribution of embryo cryopreservation and subsequent treatment cycles to cumulative live birth rates (CLBRs). 1 , 12 Other changes in practice that have changed the profile of ART cycles include the increased use of frozen and thawed ET cycles, ICSI, preimplantation genetic testing (PGT), and elective single ET. 13 – 15 Therefore, it is important for prediction models to reflect the current state of practice. 16 Moreover, donor oocytes are being increasingly used 17 ; however, very few previous studies have identified factors that predict IVF outcomes in donor oocyte recipients. 18 , 19 This analysis aimed to develop models that can predict CLBRs in patients undergoing autologous and donor oocyte IVFs according to demographic and reproductive characteristics using US national surveillance data between 2016 and 2018.