Pretreatment prediction for IVF outcomes: generalized applicable model or centre-specific model?

IVF is an effective treatment for patients with unresolved fertility problems ( Carson and Kallen, 2021 ), but not without cost and risks. Therefore, a consult regarding the chance of success is desirable to assist the expectation management and treatment decision ( Ratna et al. , 2020 ). Much effort has been made to establish models that predict the prognosis of IVF ( Ratna et al. , 2020 ). Among them, several pretreatment consulting tools based on large-scale data have been developed ( Luke et al. , 2014 ; McLernon et al. , 2016 , 2022 ). These models allow patients and clinicians to estimate the cumulative probability of live birth in a cycle before the treatment is given. However, they are mostly based on the population of Europe or the USA, and their performance and applicability in the countries and regions beyond these areas are largely unknown. National-wide registry data and models based on them are not always available, especially for low- and middle-income countries where cost and efficacy are of significant concern ( Chiware et al. , 2021 ). For regions where national-wide data were not available, prediction models based on local data including only one or a few clinics have been established ( Cai et al. , 2011 ; Qiu et al. , 2019 ; Vogiatzi et al. , 2019 ; Tarin et al. , 2020 ; Zhu et al. , 2021 ; Chen et al. , 2022 ; Qu et al. , 2022 ; Wang et al. , 2022 ). Although it is recommended that individual clinics first validate and recalibrate national-wide data-based models using their datasets before developing their clinic-specific prediction models ( McLernon et al. , 2022 ), the centre-specific models are more specific to their patients and may include potentially important predictors that are not recorded in a national registry database. Several single-centre studies reported centre-specific models with notable discriminatory power ( Cai et al. , 2011 ; Qiu et al. , 2019 ; Vogiatzi et al. , 2019 ; Zhu et al. , 2021 ; Chen et al. , 2022 ). However, the performance of the generalized applicable models based on national-wide data and centre-specific models is yet to be compared in the same population. The present study aimed to validate the generalized applicable pre-treatment models based on Human Fertilisation and Embryology Authority (HFEA) and Society for Assisted Reproductive Technology (SART) datasets in a Chinese population and compare their performance with a prediction model derived from local data.

A total of 26 382 OPUs were included in the study following the initial inclusion of 28 743 cycles ( Supplementary Fig. S1 ). The main reason for exclusion was incomplete cycles (n = 2223), followed by OPU cancellation other than fertility treatment (n = 138). The characteristics of the cycles are shown in Table 1 . The characteristics of patients with incomplete cycles are shown in Supplementary Table S1 . Patients’ characteristics and outcomes. PCOS, polycystic ovary syndrome; AFC, antral follicle count; OPU, ovum pick-up; ET, embryo transfer; FET, frozen-thawed embryo transfer. The dataset was split into a development dataset including 24 086 cycles and a temporal validation data set including 2296 cycles. The importance of predictors in the Xgboost models is shown in Supplementary Fig. S3 . The coefficients of the Lasso models and the Glm models are shown in Supplementary Tables S2 and S3 . The AUCs of the ROC curves from the models are compared in Supplementary Fig. S4 and Supplementary Table S3 . For each type of model (algorithm), including the female endocrine profile significantly improved the predictive power of the model. In the development dataset, Xgboost models appeared to have the highest predictive power, whereas the AUCs of Lasso models and Glm models were similar. In the validation dataset, however, the AUCs of the Xgboost models decreased while the models with lasso and Glm gave relatively stable AUCs ( Supplementary Table S4 ). Supplementary Figure S5 demonstrates the calibration curves of the models in the validation dataset. All models demonstrated a calibration curve close to the diagonal of the plots, indicating a good fit in the large. The slopes for the Xgboost model, the Xgboost model with endocrine profile, the Lasso model, the Lasso model with endocrine profile, the Glm model, and the Glm model with endocrine profile were 0.99, 0.98, 0.99, 0.96 0.97, and 0.95, respectively, and the intercepts were 0.02, 0.04, 0.03, 0.05, 0.04, and 0.06, respectively. Because the Lasso models which were less complex than the Xgboost models and required fewer coefficients than all-in Glm models gave acceptable AUCs across development and validation datasets, we considered the Lasso models as the final models and compared the performance of the models with external validation models. The details for using the model for validation are shown in Supplementary Data File S1 . Because female age was shown to be the most important predictor in our dataset ( Supplementary Fig. S3 ), simple Glm models with female age as the only predictor were also established to show what additional information would be provided by more complex models. Because female age with RCS transformation provided a better fit ( Supplementary Fig. S6 ), the Glm models with RCS-transformed female age were used for subsequent comparison. The AUCs of ROC curves and calibration of the external models which were validated in the dataset are shown in Figs 1 and 2 . In general, all the external models demonstrated a similar discriminatory power in our dataset with that in the original dataset according to the AUCs. The magnitude of the AUCs was similar between the four models validated, but the AUC of the Dhillon model appeared to be higher than the others ( Fig. 1 ). The AUCs of the external models were all lower than the centre-specific models using the Lasso algorithm. Notably, the centre-specific model with female age only also demonstrated a considerable AUC which was similar to external models but still lower than the centre-specific models with a full set of predictors. Discriminatory power of the models. ( A ) Receiver operator characteristic curves (ROC) curves of the models. ( B ) A comparison of AUCs, bars indicate 95% CI. Calibration curves of the models. ( A ) Centre-specific model without endocrine profiles, ( B ) centre-specific model with endocrine profiles, ( C ) McLernon 2016 model, ( D ) Luke model, ( E ) Dhillon model, and ( F ) McLernon 2022 model. Green lines indicate linear fits with the original models. Blue lines indicate smooth fits with original models. Red lines indicate smooth fits with updated models. Original models were updated by rescaling according to intercept and slope. The grey histogram indicates sparse data before rescaling. The blue histogram indicates sparse data after rescaling. For each model, individual prediction values were fitted against an observed outcome. The calibration curves suggested that all external models systematically underestimate live birth ( Fig. 2 ). The slopes of the McLernon 2016 model, the Luke model, the Dhillon model, and the McLernon 2022 model were 1.73, 1.70, 2.66, and 1.12, respectively. The intercepts of the models were −0.003, −0.07, −0.15, and 0.06, respectively. After rescaling according to the slopes and intercepts, the McLernon 2016 model and the Luke model overestimated the live births at the higher end of the prediction. The Dhillon model underestimated live births at the lower end and overestimated at the higher end. The McLernon 2022 model demonstrated the most optimal calibration among all external models. AUCs and calibration of subgroups were shown in Table 2 and Supplementary Figs S7 and S8 . The discriminatory power of the models as the form of AUCs dramatically decreased in the younger patients ages <35 years and increased in the patients with the freeze-all strategy. The calibration curves suggested that the external models and centre-specific models all had a suboptimal calibration in the free-all subgroup. Descriptive information on subgroups can be found in Supplementary Table S5 . Performance of the models in subgroups of patients. ‘*’ indicates accuracy, sensitivity, and specificity are based on a cut-off of 0.5. Since the age categories may affect the performance of the prediction, we also plotted the crude live birth rates and individual predictions in each age category to show often the personalized predictions deferred from the overall live birth rates ( Supplementary Fig. S9 ). It appeared that the predictions of centre-specific models were closer to the crude live birth in younger patients but the McLernon 2016 model may provide a prediction that was closer to the crude live birth rates and patients with advanced age.

The data were acquired from the Reproductive Medicine Centre of Xiamen University affiliated Chenggong Hospital. All patients who received ovum pick-up (OPU) between January 2013 and December 2020 were reviewed for potential inclusion. The OPU cycles were linked to their subsequent frozen–thawed embryo transfer (FET) cycles. The cumulative live birth per OPU was defined as live birth episodes per OPU over 2 years ( Maheshwari et al. , 2015 ) and only the first delivery event from the same OPU is taken into account when calculating the cumulative live birth rate. The patients who cancelled their OPU due to reasons other than fertility treatment (e.g. personal affairs, a change of schedule, or an emergency) were excluded. The patients who did not achieve live births following OPU but still had frozen embryos were considered to have undergone an uncompleted cycle. They were also excluded due to uncertainty of the outcome. On the other hand, the patients who cancelled OUP due to unsatisfying follicular growth or patients who had no embryo for transfer were considered to have a completed cycle. The detailed criteria are shown in Supplementary Fig. S1 . The study was approved by the Institutional Review Board of Xiamen University affiliated Chenggong Hospital. Of all the cycles reviewed (n = 28 743), there were 1412 (4.9%) cycles with one or more missing values in patients’ BMI or endocrine profiles due to the mistakes in data input. We used multiple imputations by chained equations based on an R package ( van Buuren, 2011 ) to fill in the missing values. We externally validated four pre-treatment models based on HFEA, SART, and Centres for Assisted Reproduction (CARE), respectively. McLernon et al. published two pre-treatment models based on HFEA (McLernon 2016 model ( McLernon et al. , 2016 ) and SART data (McLernon 2022 model ( McLernon et al. , 2022 )), respectively. The McLernon 2016 model includes 10 variables, including cycle number, female age, duration of infertility, previous pregnancies, tubal infertility, anovulation, male factor infertility, unexplained infertility, insemination protocol, and year of treatment. The McLernon 2022 model included the following variables: cycle number, female age, female BMI, previous full-term birth, male factor infertility, polycystic ovary syndrome (PCOS), uterine factor infertility, anovulation, and unexplained infertility. McLernon et al. also established a pre-treatment model including AMH to provide additional predictive power. Because an AMH test is not mandatory in our clinic during the period of study and only a part of the patients received an AMH test, we did not validate the model with AMH. Luke et al. (2014) established their model based on an earlier SART dataset (Luke model) and Dhillon et al. (2016) reported a model based on the CARE dataset (Dhillon model). Although both models lack the FET data and may not be suitable to predict cumulative live births following a completed cycle, they aimed to provide pre-treatment counselling and represented an attempt at IVF outcome prediction under different technique backgrounds. Therefore, we also validated the performance of these models. The Luke model included the following variables: female age, female BMI, previous full-term birth, male factor infertility, anovulation, diminished ovarian reserve, tubal hydrosalpinx, other tubal infertility, uterine factor infertility, other infertility factors, unexplained infertility, and the number of diagnoses. The Dhillon model includes the variables of female age, female BMI, tubal infertility, anovulation, male factor infertility, unexplained infertility, other infertility factors, ethnicity, previous live birth, previous miscarriage, antral follicle count (AFC), and duration of infertility. We did not validate several other well-known models, such as the Templeton ( Smith et al. , 2015 ) model and the Nelson model ( Nelson and Lawlor, 2011 ), because they either require parameters that do not exist in our dataset ( Smith et al. , 2015 ) (e.g. previous cycle with pregnancy but no live birth) or include post-treatment parameters ( Nelson and Lawlor, 2011 ; van Loendersloot et al. , 2013 ). For all models, the diagnoses of our dataset were reformatted to meet the requirements of different models. Continuous variables were categorized or transformed as required by the models. For the Luke model, data were split according to the number of cycles, because different coefficients were used in Cycles 1 and 2 or more according to the model. To develop the centre-specific prediction models, we split our dataset into a development set (2013–2019) and a validation set (2020) according to the year of treatment. The following variables were listed as candidate predictors of the models: patient characteristics (age, BMI, weight, height, and smoking status of the couple), reproductive history (previous OPU cycles, duration of infertility, previous live birth, recurrent miscarriage, and ectopic pregnancy history), female infertility diagnoses (uterine problem, tubal problem, endometriosis, and PCOS), male infertility diagnoses (azoospermia, ejaculation disorders, and poor semen quality), and menstruation status ( Munro et al. , 2012 ) (menarche age, polymenorrhea, oligomenorrhea, amenorrhea, dysmenorrhea, and the amount and duration of menstruation bleeding). Because female smoking is very rare in our dataset, we only considered the smoking status of the male counterparts. Female endocrine profiles (basal FSH, LH, prolactin (PRL), and AFC) are also considered. All continuous predictors were plotted against live birth with either linear fit or smooth fit to identify the linearity of association ( Supplementary Fig. S2 ). For the predictors whose linear fitted curve and smooth fitted curve graphically diverged, the values underwent restricted cubic spline transformation using the RMS packages. We used three strategies to develop the centre-specific prediction models with the aforementioned predictors. First, an Extreme Gradient Boosting (Xgboost) algorithm ( Chen, 2022 ) was used to establish gradient boosting trees with the features using the following parameters: max depth = 8, min_child_weight = 4, learning_rate = 0.05, gamma = 0.01, subsample= 0.8, colsample_bytree = 0.8. The number of rounds (iterations) used in the final models was determined according to 1000-round 5-fold cross-validation in search of the minimal root mean squared error (rmse). Second, a Least Absolute Shrinkage and Selection Operator (Lasso) model ( Friedman et al. , 2010 ) was used for feature selection (Lasso model) by penalizing the size of the coefficients and reducing some of them to zero. The amount of penalty (lambda) was determined by 10-fold cross-validation. The Lambda that gave the minimum mean squared error of the model was selected for the final model. Finally, a conventional generalized linear model (Glm) including all predictors was used as a control. To evaluate the contribution of the endocrine profile to the predictive power, all models were established either with or without female endocrine profile profiles. For all models, the AUCs of the Receiver Operator Characteristic (ROC) curves were compared between the development and validation datasets. Calibration curves in the validation dataset were also evaluated. The calibration curve was used as the measure of the agreement between predictions from the models and observed data. We plotted the predicted chance of live birth against the observed chance of live birth and fit them linearly and smoothly. The smooth-fitted curves demonstrate the goodness of fit in the large and the linear-fitted curves yield slopes and intercepts. The predicted values could be rescaled according to the slopes and intercepts to obtain a better fit. AUC and calibration were also evaluated in subgroups to examine the performance of the models in different scenarios. The subgroups included patients with their first attempt and patients with multiple attempts, patients aged <35 years and patients aged ≥35 years, patients with fresh transfer or no transfer, and patients with freeze-all strategy. Accuracy was evaluated, according to a fixed cut-off of 0.5. Accuracy was defined as the proportion of correctly predicted cases (true positives + true negatives) in all cases based on the cut-off. A Youden index-based cut-off was not used because the prediction is often provided as the form of predicted live birth rate to patients in previously published tools and optimal cut-offs were not given. Bier scores provided another way to measure the accuracy. All analyses were performed using R Statistical Software ( R Core Team, 2022 ).

In summary, our study suggested that models derived from a national registry like HFEA and SART could also provide a meaningful prediction for patients from a different region with modest discriminatory power. However, recalibration or developing a local data-based prediction model based may still increase the discriminatory power and give a better calibration.

The study validated four well-known pretreatment models for IVF outcomes derived from the HFEA ( McLernon et al. , 2016 ), SART ( Luke et al. , 2014 ; McLernon et al. , 2022 ), and CARE ( Dhillon et al. , 2016 ) datasets in a Chinese cohort, finding that the discriminatory power of the models remained stable despite the difference in the region, ethnicity, preference of the clinicians and patients may exist between the development cohorts of the models and the validation cohort. On the other hand, however, the calibration suggested an underestimation of the models, and additional rescaling is needed. In addition, the study also developed centre-specific models based on the cohort. The discriminatory power and calibration with the centre-specific models appeared to be improved compared with the external models. Our study may be the first study to validate the generalized prediction models derived from HFEA and SART in a Chinese population. The sample size of the dataset also fortifies the validation. However, the data are limited by the single-centre design and might not have enough representativeness of the Chinese population. Our clinic preferred an agonist ovarian stimulation protocol ( Ren et al. , 2014 ) and fresh embryo transfer and the study cohort was of a younger female age and lower BMI than the previously reported national data ( Luke et al. , 2014 ; McLernon et al. , 2016 , 2022 ). All the factors may contribute to bias and heterogeneity between populations. Due to the limitation of the dataset, we did not validate all the components of the mentioned models. The study failed to validate the McLernon models with AMH included because a significant proportion of our cycles lacked AMH levels. The McLernon models not only evaluated the cumulative live birth of a complete cycle but also cumulative live birth over three complete cycles. The McLernon models were evaluated based on the cumulative live birth of up to three completed cycles in their original population. However, the present study focused on the complete cycle of live birth. Unlike the regions where insurance may cover several treatment cycles for ART treatment, patients in China pay for each cycle of the treatment ( Chiware et al. , 2021 ) and the per-complete cycle of live birth may be a priority concern for the patients and clinicians. Nevertheless, it is still possible to predict the per-cycle live birth using the McLernon models. The per-complete cycle analysis also provides a similar framework to compare different models in the sample population. For instance, the Dhillon model accounts for only a single complete cycle, and the Luke model considers only fresh transfer. The per-cycle manner also avoids discrepancies in the patient population initiating and continuing treatment. There is also a challenge to adapting the infertility indications and diagnoses from other data sources to our dataset. The diagnoses were made by clinicians for individual patients with different preferences. A categorized variable may not fully describe the extent and severity of a disease. Also, the ambiguous nature of some diagnoses, such as male factor infertility, may further contribute to the heterogeneity of the diagnosis and thus affect the accuracy of the validation. Among the models validated, the McLernon 2016 model and the Luke model have been externally validated. Leijdekkers et al. (2018) validated the McLernon 2016 model in a randomized controlled trial dataset from the Netherlands and an AUC of pretreatment prediction of 0.62 (95% CI 0.59–0.64) was reported. With a modification of the model by including AMH, AFC, and body weight, the AUC improved to 0.66 (95% CI 0.64–0.68) ( Leijdekkers et al. , 2018 ). Dubaut et al. validated the Luke model in a single-centre dataset from Oklahoma City, reporting an AUC of 0.62 (95% CI 0.56–0.68) in this external validation ( Dubaut et al. , 2018 ). By validating the models in a larger cohort with different regions and ethnicities, our study may contribute to a better understanding of these models. All validated models demonstrated reasonable AUCs in our cohort. However, the McLernon models ( McLernon et al. , 2016 , 2022 ), which reported AUCs higher than 0.7 in their original cohorts suffered a decrease in discriminatory power. It could be explained by the model fitting, the population difference, and the different study designs. Nevertheless, the discriminatory power of the McLernon models in our cohort appeared to be comparable to that in the Netherlands cohort ( Leijdekkers et al. , 2018 ). It might suggest that the predictors and their coefficients provided by the McLernon 2016 model which is derived from the HFEA database are as useful for our patients as they are for the patients in the Netherlands. The Dhillon model is the only model adjusted for ethnicity factor among the models validated. Theoretically, it might make it more easily adapted to other populations with a different ethnicity distribution. The Dhillon model demonstrated a trend towards higher discriminatory power than other validated models, especially for the patients who received their first cycles, who are the target population for the model. As commented by McLernon et al. (2022) , ethnicity is one of the potentially important predictors but poorly recorded. However, the heterogeneous nature of the race predictor also hampered its use in the prediction models. The well-constructed models derived from the multiethnic data might extend their generalizability by a future inclusion of ethnic predictors. Despite that, the discriminatory power of all the validated models fell in a close range, the calibration varied among models. While all calibration plots suggested a need for re-calibration, the McLernon 2022 model showed better calibration than other external models. It might be partially explained by the date of the development data. The McLernon 2022 models were based on an SART dataset between 2014 and 2015, whereas all other three models were based on the datasets before 2013. It is well known that the date of the data might be associated with the change in clinical practice over time. Also, the establishment of widely accepted consensus among clinicians in the most recent decade might narrow the difference in practice between regions. While predictive models derived from the national registry represent an average clinical practice of a given area and time, the prediction models based on local datasets are supposed to be more specific to the local patients. Our study also established ‘centre-specific models’ which were based on a single-centre dataset. While several IVF prediction models have been constructed for a similar purpose ( Cai et al. , 2011 ; Qiu et al. , 2019 ; Vogiatzi et al. , 2019 ; Tarin et al. , 2020 ; Chen et al. , 2022 ; Qu et al. , 2022 ; Wang et al. , 2022 ), many of them included treatment parameters and intermediate outcomes such as oocyte yield and therefore could not predict the patients’ prognosis before treatment ( Cai et al. 2011 ; Vogiatzi et al. , 2019 ; Chen et al. , 2022 ; Wang et al. , 2022 ). A few recent studies reported pretreatment models comparable to ours. Qiu et al. (2019) reported an Xgoobst-based predictive model for the live birth following the first IVF cycle, showing an AUC of 0.73 in the validation cohort. Qu et al. (2022) reported an AUC of 0.676 based on a larger cohort of 26 689 patients. Both studies are based on validations of random split data, rather than validations in a more recent cohort or other clinics/countries. The work of Tarin et al. (2020) included the so-called ‘high-assisted-fecundity women’ who achieved live birth following their first attempt and ‘low-assisted-fecundity women’ who failed to have a live birth after completing three treatment cycles. The model reported an AUC of 0.718 and included a temporal validation. However, the AUC decreased to 0.649 in the validation cohort, possibly due to the small sample size. Although the pretreatment models are supposed to give a prediction of patients’ prognosis before treatment, the extract treatment the patients received still affects the chance of live birth they might achieve. The patients’ pretreatment characteristics also lead to different clinical scenarios and therefore affect the selection of the treatment protocols. For instance, the age of patients might affect the ovarian stimulation protocol they received, the gonadotropin dosage they took, and the selection of transferred embryos ( Nakagawa et al. , 2019 ; Ubaldi et al. , 2019 ) and a free-all strategy due to ovarian hyper-response may lead to a series of subsequent treatments, including cryopreservation and endometrial preparation. Therefore, we attempted to evaluate the models in several different subgroups which represent common clinical scenarios that a clinic may encounter. The discriminatory power of the models did not significantly change among patients with advanced age, however, the discriminatory power dramatically decreased in the patients younger than 35 years. It may be explained that the same model could fail to discriminate the patients in a more homogenous population ( Alba et al. , 2017 ). As our data show, female age was the most important predictor in our centre-specific models and a model including age alone could yield an AUC of 0.68 in our patients. The flatter age-live birth association in younger patients ( Supplementary Fig. S1A ) may also support the explanation. The age of patients also biased the distribution of aetiologies. While the common theories may say that both unexplained infertility and age-related infertility increase with patients’ age ( ESHRE Capri Workshop Group, 2017 ), our data suggest that young patients had a greater incidence of primary infertility, PCOS diagnosis, and male-related diagnosis ( Supplementary Fig. S10 ). The young primary infertility patients might be a mix of women who suffer from serious ‘true’ unexplained infertility (such as genetic-related disorders) and merely ‘unlucky’ women with normal fertility potential. While the current clinical information cannot distinguish the two kinds of patients when diagnoses such as unexplained fertility of male fertility were given, the latter is expected to have a high live birth rate following treatment. The prognosis of the ‘unlucky’ women is more likely to be determined by treatment-related factors that were not present in the models, rather than the characteristics of the patients. It explains the miscalibrations of external models, as they are established in the clinical practice on a national basis which are less specific to the centre-specific data. For freeze-all patients, however, the models all showed a suboptimal calibration, which may suggest the models need to be refitted in those patients. Notably, the previously published models have different policies for including freeze-all patients or not. For instance, the McLernon 2016 models excluded ‘women whose first treatment was a frozen treatment’. However, in the McLernon 2022 model, these patients were no longer excluded. Because there is an increasing trend towards freeze-all cycles ( Adamson et al. , 2018 ), the models developed with previous data might be poorly fitted to future cycles, and models specific to freeze-all cycles are warranted. The IVF prediction might be the statistic of greatest interest to most patients, as the models are often published in the form of publicly accessible online calculators and the results of the prediction may affect the patient’s view on the treatment. Many regions lack a comprehensive national dataset like HFEA and SART and therefore lack the generally applicable prediction model based on their own country. Our data may be meaningful for the patients and clinicians who would use a prediction model derived from other countries to predict their outcome, as it found the discriminatory power of the models remained reasonable despite the different regions and ethnicities. Caution should also be taken because of the need for recalibration. If a Chinese couple accesses a Web tool based on the aforementioned Models ( Luke et al. , 2014 ; Dhillon et al. , 2016 ), they may find the prediction is informative but underestimates the live rate they might have. Previous discrete choice experiment (DCE) studies have suggested that the expected live birth rate is the most important attribute that affects the patients’ preference for ART treatment ( Abdulrahim et al. , 2021 ; Cornelisse et al. , 2022 ). Therefore, an underestimated live birth rate may affect the patient’s decision-making. It may be necessary to execute further DCE studies to investigate the real value of the models in clinical practice. Although there have been various models that were designed to predict IVF outcomes, there is still limited data regarding the comparison of models serving a similar purpose in the same population. Patients and clinicians may be interested in the difference in the predictive tools on their prediction. However, a direct comparison between published works is often invalid, simply because different populations were reported ( Kragh and Karstoft, 2021 ). Also, if predictive models report their results on the same population, the sample size of the test sample matters because it greatly affects the certainty of a comparison. We reported and compared the performance of four published pretreatment models in our population, due to their wide representativeness and potentially significant implication. While there is only a marginal difference in terms of discriminatory power between the models, we recommended the use of the latest McLernon model when recalibration is not available, because the model is based on a more recent dataset and may provide a better calibration for the incoming treatment. Finally, our data also suggested that although the predictive models for IVF outcomes have been abundant and to use of an existing high-quality model may save the resource and time, developing a ‘local model’ specific to the clinic or region is still meaningful. Also, an update of the existing models with more recent datasets may be desired, as the clinical practice may significantly change since the development of the models.

Click here for additional data file. Click here for additional data file. Click here for additional data file. Click here for additional data file. Click here for additional data file. Click here for additional data file. Click here for additional data file. Click here for additional data file. Click here for additional data file. Click here for additional data file. Click here for additional data file. Click here for additional data file. Click here for additional data file. Click here for additional data file. Click here for additional data file. Click here for additional data file.