Abstract

STUDY QUESTION

Is it feasible to identify factors that significantly affect the clinical outcome of IVF-ICSI cycles and use them to reliably design a predictor of implantation?

SUMMARY ANSWER

The Bayesian network (BN) identified top-history embryos, female age and the insemination technique as the most relevant factors for predicting the occurrence of pregnancy (AUC, area under curve, of 0.72). In addition, it could discriminate between no implantation and single or twin implantations in a prognostic model that can be used prospectively.

WHAT IS KNOWN ALREADY

The key requirement for achieving a single live birth in an IVF-ICSI cycle is the capacity to estimate embryo viability in relation to maternal receptivity. Nevertheless, the lack of a strong predictor imposes several restrictions on this strategy.

STUDY DESIGN, SIZE, DURATION

Medical histories, laboratory data and clinical outcomes of all fresh transfer cycles performed at the International Institute for Reproductive Medicine of Lugano, Switzerland, in the period 2006–2008 (n = 388 cycles), were retrospectively evaluated and analyzed.

PARTICIPANTS/MATERIALS, SETTING, METHODS

Patients were unselected for age, sperm parameters or other infertility criteria. Before being admitted to treatment, uterine anomalies were excluded by diagnostic hysteroscopy.

To evaluate the factors possibly related to embryo viability and maternal receptivity, the class variable was categorized as pregnancy versus no pregnancy and the features included: female age, number of previous cycles, insemination technique, sperm of proven fertility, the number of transferred top-history embryos, the number of transferred top-quality embryos, the number of follicles >14 mm and the level of estradiol on the day of HCG administration. To assess the classifier, the indicators of performance were computed by cross-validation. Two statistical models were used: the decision tree and the BN.

MAIN RESULTS AND THE ROLE OF CHOICE

The decision tree identified the number of transferred top-history embryos, female age and the insemination technique as the features discriminating between pregnancy and no pregnancy. The model achieved an accuracy of 81.5% that was significantly higher in comparison with the trivial classifier, but the increase was so modest that the model was clinically useless for predictions of pregnancy. The BN could more reliably predict the occurrence of pregnancy with an AUC of 0.72, and confirmed the importance of top-history embryos, female age and insemination technique in determining implantation. In addition, it could discriminate between no implantation, single implantation and twin implantation with the AUC of 0.72, 0.64 and 0.83, respectively.

LIMITATIONS, REASONS FOR CAUTION

The relatively small sample of the study did not permit the inclusion of more features that could also have a role in determining the clinical outcome. The design of this study was retrospective to identify the relevant features; a prospective study is now needed to verify the validity of the model.

WIDER IMPLICATIONS OF THE FINDINGS

The resulting predictive model can discriminate with reasonable reliability between pregnancy and no pregnancy, and can also predict the occurrence of a single pregnancy or multiple pregnancy. This could represent an effective support for deciding how many embryos and which embryos to transfer for each couple. Due to its flexibility, the number of variables in the predictor can easily be increased to include other features that may affect implantation.

STUDY FUNDING/COMPETING INTERESTS

This study was supported by a grant, CTI Medtech Project Number: 9707.1 PFLS-L, Swiss Confederation. No competing interests are declared.

Introduction

Embryo implantation represents a critical step in the reproductive process during which the blastocyst penetrates the endometrium to establish an interface between the growing fetus and the maternal circulation (Guzeloglu-Kayisli et al., 2009). For a successful implantation to occur a viable embryo must establish a synchronized dialog with a receptive endometrium (Simon et al., 2000).

Owing to the complexity of this process, several factors can determine implantation failure including chromosomal abnormalities, which are especially frequent in female gametes in a manner that is closely related to age (Gianaroli et al., 2005a; Kuliev et al., 2005; Munné et al., 2007; Gianaroli et al., 2010), a poor uterine receptivity, sometimes due to asynchrony between the embryonic development and the endometrium (Norwitz et al., 2001; Margalioth et al., 2006; Swain and Smith, 2011), and factors associated with ART interventions comprising the potential adverse effects of ovarian stimulation (Santos et al., 2010) and of in vitro culture systems (Jones et al., 2001; Horsthemke and Ludwig, 2005; Dumoulin et al., 2010) on embryo development.

Looking at clinical data where single implantation frequently results from the transfer of two or more embryos of comparable morphology, it is clear that a large proportion of failed implantations must be ascribed to the embryo. Despite numerous advances in culture systems and the introduction of more refined tools for the classification of the embryo's developmental capacity (Katz-Jaffe et al., 2006; Brison et al., 2007; Jones et al., 2008a, b; Wong et al., 2010; Meseguer et al., 2011), the lack of a strong predictor of embryo viability imposes severe restrictions to the general implementation of elective single-embryo transfer. In comparative trials between elective single- and elective double-embryo transfers in selected patient groups, the resulting live birth rates are comparable only when the transfer of cryopreserved embryos in subsequent cycles is included (Pandian et al., 2009; Gelbaya et al., 2010).

These considerations lead to the conclusion that the key requirement for achieving a single live birth within an IVF-ICSI cycle is the capacity to estimate the embryo's potential for further development in relation to maternal receptivity, which is a general term including factors associated with the patient characteristics and the treatment cycle. Morphological criteria have been, and still are, routinely used to select embryos for transfer or cryopreservation, and comprise the evaluation of pronucleate oocytes, the occurrence of early cleavage at 25–27 h post-insemination, scoring systems for day 2–3 embryos (including number of cells and synchrony of cleavage as well as the degree and pattern of fragmentation) and grading systems for morulas (progression of compaction) and blastocysts (including the timing of blastocyst formation and the degree of expansion as well as the characteristics of the inner cell mass and trophectoderm cell lineages) (Alpha Scientists in Reproductive Medicine and ESHRE Special Interest Group of Embryology, 2011).

These criteria brought the definition of top-quality embryos. They are currently the most predictive markers of embryo developmental competence and are widely used in most IVF laboratories (Kovacic et al., 2004; Volpes et al., 2004; Gianaroli et al., 2007; Magli et al., 2007; Racowsky et al., 2011). Other studies have shown that the predictive capacity can be further improved by combining the scores obtained by the morphology of the same embryo at different stages (De Placido et al., 2002; Lan et al., 2003; Sjöblom et al., 2006; Loi et al., 2008; Brezinova et al., 2009) as well as by using the information coming from morphokinetic observations (Wong et al., 2010; Meseguer et al., 2011). Nevertheless, reliable prediction of which embryos will implant remains an open problem, as prognostic models generally achieve only limited accuracy (Saith et al., 1998; Beuchat et al., 2008).

Based on the assumption that for a successful implantation to occur a viable embryo must meet a receptive uterus, this study proposes a predictor of pregnancy that was designed by analyzing a series of cycles with known clinical outcomes. This was done to identify the factors that, being specifically related to embryo viability and maternal receptivity, were shown to affect the clinical outcome. The resulting predictor is able to discriminate with reasonable reliability not only between pregnancy and no pregnancy, but also between the occurrence of a single pregnancy or multiple pregnancy. This could represent an effective support to decide for each couple how many embryos and which embryos to transfer.

Materials and Methods

Patients

A total of 388 cycles performed at the International Institute for Reproductive Medicine of Lugano, Switzerland, between 2006 and 2008 were included in the study. Patients were unselected for age, sperm parameters or other infertility criteria. The average maternal age was 36.3 ± 4 years, and the average number of previous IVF-ICSI cycles was 2.2 ± 1.7. Before being admitted to treatment, uterine anomalies were excluded by diagnostic hysteroscopy (Gianaroli et al., 2005b).

Ovarian stimulation was achieved by long down-regulation ovarian simulation protocols (Ferraretti et al., 2004). Oocytes were retrieved transvaginally via ultrasound guidance at 34–36 h after HCG administration, and cultured in HTF (SAGE Cooper Surgical, Inc., Pasadena, CA) supplemented with 5% human serum albumin (HSA, SAGE) at 37°C in a 5.3% CO2 humidified gas atmosphere. Insemination was performed ∼5 h later with conventional IVF in 143 cycles (37%) or ICSI in 245 cycles (63%) depending on sperm indices and the couple's reproductive history.

At 16 h after insemination, oocytes were checked for the presence of pronuclei and polar bodies. Regularly fertilized oocytes were cultured individually and scored daily at 40, 64, 88 and 112 h after insemination. According to the National regulation on IVF, a maximum of three fertilized oocytes per cycle were left in culture. Embryo transfer was performed between Days 2 and 5 depending on the clinical workload: 77 cycles on Days 2, 141 on Days 3, 95 on Day 4 and 75 on Day 5.

A clinical pregnancy was defined as the presence of a gestational sac with fetal heartbeat at 7–8 weeks. All pregnancies were followed-up to term. The implantation rate was calculated by the ratio between the number of gestational sacs with fetal heart beat and the number of transferred embryos.

The study was supported by a grant, CTI Medtech, Project Number: 9707.1 PFLS-L, Swiss Confederation.

Pronucleate oocytes and embryo grading

Pronucleate oocyte morphology was assessed by observing (i) the shape and position of pronuclei, (ii) the distribution and size of nucleoli within pronuclei and (iii) the position of polar bodies with respect to the longitudinal axis of pronuclei. According to previous studies, the configurations having juxtaposed and centrally located pronuclei with a symmetrical distribution of large size nucleoli and polar bodies next to each other are considered as top quality having the highest chances of normal chromosome status, embryo development and implantation (Gianaroli et al., 2003; Gianaroli et al., 2007).

Day 2–3 embryos were assessed by recording the number and morphology of nuclei and blastomeres, the percentage of fragmentation in the perivitelline space and the presence of compaction (Alpha Scientists in Reproductive Medicine and ESHRE Special Interest Group of Embryology, 2011). Day 4 embryos were classified based on the number of cells and grade of compaction and day 5 embryos were assessed according to the grade of expansion associated with inner cell mass and trophectoderm morphology (Alpha Scientists in Reproductive Medicine and ESHRE Special Interest Group of Embryology, 2011).

The definition of top-quality embryos was as follows: Embryos graded as top quality throughout all stages were graded as top history. In addition, day 4–5 embryos were graded as top history also if they had been graded as top quality in all stages but one, provided that the single non-top assessment did not occur on the day of transfer (see later for details).

  • Day 2: embryos with four regular cells and fragmentation <10%;

  • Day 3: embryos with eight regular cells and fragmentation <10%;

  • Day 4: embryos at the morula stage with fully compacted blastomeres;

  • Day 5: embryos at the blastocyst stage with inner cell mass and trophectoderm.

Statistical analysis

Statistical classifiers are mathematical models, which predict the value of a categorical target variable (the class) on the basis of the values of other variables (the features). The class variable was categorized as pregnancy versus no pregnancy; the features included the following variables that normally represent the main characteristics of a treatment cycle: the woman's age, previous IVF-ICSI cycles, insemination technique, quality of sperm, number of transferred top-history embryos, number of transferred top-quality embryos, number of follicles >14 mm and level of estradiol on the day of HCG administration. For each cycle, the classifier estimated the probability of pregnancy and no pregnancy. The most probable outcome was regarded as a prediction and the accuracy of the classifier, namely the proportion of successful predictions, was measured. Since in IVF-ICSI cycles, the occurrence of no pregnancy is two to four times more frequent than pregnancy, a trivial classifier, which always predicts no pregnancy, can achieve high accuracy without conveying any useful information. Therefore, the accuracy of the developed classifiers was checked against that of the trivial classifier.

Given the shortcomings of accuracy, which is very sensitive on the class imbalance and allows the trivial classifier to obtain a good evaluation, the classifiers were also evaluated through the AUC, namely the area under the receiver operating characteristic curve. For example, in the case of two women, of whom only one is destined to become pregnant, the AUC corresponds to the probability that the model will assign a higher probability of pregnancy to that woman compared with the second. The performance of the model is graded as poor if the AUC lies between 0.5 and 0.7, fair if the AUC lies between 0.7 and 0.8, good if the AUC exceeds 0.8, while an AUC of 1 implies perfect discrimination (Leushuis et al., 2009).

As final evaluation, the calibration of the predictions was checked, namely the correspondence between the computed and the observed proportion of pregnancies.

To assess the classifier, the indicators of performance were computed by cross-validation (Witten et al., 2011). The original sample was split into k parts; the model was derived on (k− 1) parts of the data (training data) and the predictions were evaluated on the remaining part (test data). The procedure was repeated several times until each part was used as the test set. The k parts were stratified, namely they contained a similar proportion of pregnancies and no pregnancies.

Decision trees

As a first step, data were analyzed through a decision tree that was expected to provide good accuracy, and to represent at the same time a model easy to be interpreted by IVF-ICSI professionals.

The decision tree tried to discover the patterns, which link the features with the occurrence of pregnancy or no pregnancy. The model expressed the discovered patterns between class and features as a set of understandable rules. Since feature selection was performed as pre-processing step (Witten et al., 2011), not all the features that were initially postulated were included, with the least predictive ones being removed. Feature selection was aimed at making the model better performing and more interpretable.

A Bayesian network model for decision support

Bayesian networks (BNs) are probabilistic models suited for data analysis, representation of expert knowledge and probabilistic modelling (Koller and Friedman, 2009). The BN of this study was developed with the aim of: (i) discriminating with more success than the decision tree not only between no pregnancy and pregnancy, but also between no pregnancy, single pregnancy, twin pregnancy or higher order pregnancy and (ii) serving as a support for the decision of which embryos to transfer, by assessing how the probabilities of having i (i = 0, 1, 2, 3) gestational sacs varies with the number and the type of the transferred embryos.

The BN model assumed a viable embryo and a receptive uterus to be necessary for a successful implantation (Zhou and Weinberg, 1998; Simon et al., 2000; Roberts, 2007). In detail, if the uterus is not receptive, there will be no pregnancy; if the uterus is receptive and there is one viable embryo, a single pregnancy will follow; if there are two viable embryos and a receptive uterus, a twin pregnancy will follow, and so on. Moreover, the BN model assumed that each embryo implants independently of any others.

As shown in Fig. 1, the BN model was represented as a graph, where each node characterizes a variable. Contrary to the decision tree, whose discrimination threshold is learned from data, the BN model requires the variables to be discrete; therefore variables were divided into the categories shown in Table I. An arc connecting two nodes represents a relationship between the two variables; in Fig. 1, the maternal receptivity (intended as the predisposition to pregnancy and including the uterine receptivity) depends on the age of the woman and on the number of her previous IVF-ICSI cycles; the viability of the embryo depends on its score (non-top, top-quality, top history) and on the insemination technique; the pregnancy depends on both the maternal receptivity and viability of each embryo. The structure of the model was selected, among possible alternative layouts, adopting the Bayesian Information Criterion.

Table I

Division of variables for the BN

Variable Categories 
Age (years) <34, 34–40, >40 
Previous IVF-ICSI cycles 0–2, >2 
Grade of each embryo No transfer; non-top; top; top history 
Insemination technique IVF, ICSI 
Maternal/uterine receptivity Receptive; non-receptive 
Embryo viability Viable; non-viable 
Pregnancy No; single; twin; triple 
Variable Categories 
Age (years) <34, 34–40, >40 
Previous IVF-ICSI cycles 0–2, >2 
Grade of each embryo No transfer; non-top; top; top history 
Insemination technique IVF, ICSI 
Maternal/uterine receptivity Receptive; non-receptive 
Embryo viability Viable; non-viable 
Pregnancy No; single; twin; triple 
Figure 1

The BN model. The uterine receptivity (intended as predisposition to pregnancy) depends on the number of previous IVF-ICSI cycles and on age. The viability of each embryo depends on both its grade (non-top, top-quality, top history) and the insemination technique. Pregnancy depends on both maternal receptivity and viability of each embryo.

Figure 1

The BN model. The uterine receptivity (intended as predisposition to pregnancy) depends on the number of previous IVF-ICSI cycles and on age. The viability of each embryo depends on both its grade (non-top, top-quality, top history) and the insemination technique. Pregnancy depends on both maternal receptivity and viability of each embryo.

The BN model required estimating probability distributions for each node, namely the marginal probabilities for all the nodes with no predecessors (e.g. previous cycles, age, score of the embryos) and the conditional probabilities for all other nodes, given all possible combinations of their direct predecessors. For instance, it was necessary to estimate with which probability the embryo would be viable given all the possible combinations of technique (IVF or ICSI) and score (no-transfer, non-top, top-quality, top history); with which probability the woman would be receptive, given all the possible combinations of age (<34, 34–40, >40 years) and previous IVF-ICSI cycles (≤2, >2). The age intervals were decided according to the indications derived from the retrospective analysis of more than 10 000 internal cycles in relation to maternal receptivity.

The main difficulty in making these estimations was that both maternal receptivity and embryo viability were only partially observed. In particular, the uterus viability was observed (as positive) only for the cases with pregnancy; for the cases with no pregnancy, it could not be ascertained whether the uterus was not receptive, the embryos were not viable, or both. Similarly, the viability of the embryos was observed (as positive) only for the cases when all the transferred embryos implanted. To address this problem, the Expectation–Maximization algorithm was used, which is suited to estimate the parameters of a model from a data set with partially observed variables (Zhou and Weinberg, 1998; Roberts, 2007; Roberts et al., 2010).

Once all the conditional probabilities were estimated, the model was ready for predictions; also in this case cross-validation was implemented and the cases used for testing the model predictivity were different from those used for estimating the model parameters. While issuing the prediction for a new case, the BN model was provided with information about embryo grades, woman age, insemination technique and number of previous IVF cycles, while embryo viabilities and uterine receptivity were unknown to the model, as in real clinical practice. On the basis of the available information, the BN model inferred the probability of each embryo being viable and of the uterus being receptive, and consequently estimated the probability of single and multiple pregnancies.

The network of Fig. 1 is handled between one and three embryos to be transferred, as this is the limit imposed by the national regulation on IVF.

Taking decision with expected utility

Based on the variables that were found to be related to pregnancy from the analysis of the 388 cycles, this study provided some examples of the decision, which would have been suggested by applying the criterion of expected utility on top of the probabilities estimated by the BN model. The utility is a score, which represents the desirability of the different outcomes. In reproductive medicine, the most desirable clinical outcome is a singleton pregnancy; therefore, the corresponding utility should be set at the top level having a score of 1, while no pregnancy should be set at the bottom level with a score of 0. The occurrence of multiple pregnancy could be set at any intermediate value, including 0, or even at a negative value like in this study, according to the doctor's experience. As an illustrative example, let us assume the utility of no pregnancy, single pregnancy and multiple pregnancy to be, respectively, 0, 1 and 0.5. Let us also assume the probability of no pregnancy, single pregnancy and multiple pregnancy for a certain category of patient to be, respectively, 0.2, 0.5 and 0.3. The expected utility of embryo transfer is obtained by weighing the utility of the different outcomes by their probability (0 × 0.2 + 1 × 0.5 + 0.5 × 0.3 = 0.65). However, the chances of pregnancy also vary in relation to the embryos selected for the transfer and on their number and morphology. The probabilities computed by the model can be exploited to select the set of embryos whose transfer maximizes the expected utility.

As a first step, it was required to associate a utility score with the possible outcomes of an IVF cycle: no pregnancy, single pregnancy, twin pregnancy or triple pregnancy. The score of no pregnancy (U0) and single pregnancy (U1) were set, respectively, as U0 = 0 and U1 = 1. The utilities U2 and U3 of twin and triple pregnancy, respectively, should reflect the wishes of the patient and/or the recommendation of the doctor. To demonstrate the applicability of the model, we hypothesized the following utilities in relation to the woman's body build that was defined by the body mass index (BMI) according to the WHO obesity classification (World Health Organization, 2000): These utility scores were provided as an example; they could be set at levels different from those set here, including leaving a single implantation as the only positive utility. Depending on the clinical experience, other parameters which have an effect on the utility related to the chances of implantation could have been chosen instead of BMI or in addition to it. The corresponding figures can be adjusted for any single case.

  • Overweight (BMI >25 kg/m2): U2 = 0.5, U3 = −1

  • Normal (BMI 18.5–25 kg/m2): U2 = 0.2, U3 = −2

  • Underweight (BMI <18.5 kg/m2): U2 = −1, U3 = −2.

The expected utility of transferring a certain cohort of embryos was Oi = 3PiUi where Pi was the probability of having i gestational sacs computed by the BN model. In this way, the model could serve as a decision support, suggesting the transfer of the set of embryos resulting in the highest expected utility.

Results

Clinical outcomes

The overall clinical pregnancy rate was 19.8% (77/388), including both single (n = 60) and twin pregnancies (n = 17), while no triple pregnancies were generated. The pregnancy rate was higher for conventional IVF (26.7%) than for ICSI (16.3%, P < 0.01), and was negatively correlated with the number of previous cycles, being, respectively, 21 and 16.5% for women having experienced 0–2 or more than two cycles, although the difference in these pregnancy rates was not significant at a 95% level. A negative correlation was also found with age, going from 28% (25/90) for women aged <34 years to 20% (48/241) for women between 34 and 40 years and 7% (4/57) for women older than 40 years.

The most discriminative variable for predicting a clinical pregnancy was the number of transferred top-history embryos. The pregnancy rate varied from 14% (47/332) to 53% (30/56; P < 0.01) depending on whether the cohort contained 0–1 or 2–3 top-history embryos. The 95% confidence intervals of the pregnancy rate for transfers with two to three top-history embryos was 41–65% with even the lower bound of this confidence interval being about twice the average pregnancy rate in the whole data set.

If embryos were categorized only on the basis of their last grade irrespective of their history, the top-quality label would be assigned to all the top-history embryos and to those, which were top in the last observation without being top history. In this case, the pregnancy rate calculated in the data set would be of, respectively, 11 or 42% for transfers with 0–1 or 2–3 top embryos. Therefore, the top-history score was more discriminative than the usual top-quality score.

No correlation was found between the number of top-history embryos and female age. The percentage of top-history embryos decreased with the day of culture, going from 35% on Day 2 to 10% on Day 5.

Predictions with the decision tree

For the decision tree, the following features were identified as fundamental to discriminate between pregnancy and no pregnancy: (i) the number of transferred top-history embryos, (ii) the female age and (iii) the fertilization technique (IVF or ICSI). The remaining features were discarded probably because of a certain degree of redundancy (e.g. number of transferred top-quality embryos versus number of top-history embryos; semen of proven fertility versus IVF/ICSI) or because of limited predictive power (estradiol, number of follicles >14).

As shown in Fig. 2, the decision tree checked first the number of transferred top-history embryos; if this was 0 or 1, no-pregnancy was predicted. If the number of transferred top-history embryos was 2 or 3, pregnancy was predicted in the case of conventional IVF, while for ICSI pregnancy was predicted in women younger than 37 years. The number of top-history embryos was recognized as the most discriminative feature, being the first to be tested within the tree. Therefore, the tree predicted pregnancy for cycles characterized by two or three top-history embryos and conventional IVF, or by two or three top-history embryos, ICSI and woman age <37 years. Overall, 43 cycles matched this criteria, and the pregnancy rate for these cycles was 60% (26/43), roughly three times the average. This confirmed the ability of the decision trees to identify the conditions favouring pregnancy as well as the importance of top-history embryos in determining implantation.

Figure 2

The decision tree with the features enclosed into rhomboidal nodes. The rectangular nodes indicate a point where the prediction is given, with the numbers representing the cases in which the prediction was correct and incorrect, respectively.

Figure 2

The decision tree with the features enclosed into rhomboidal nodes. The rectangular nodes indicate a point where the prediction is given, with the numbers representing the cases in which the prediction was correct and incorrect, respectively.

The decision tree achieved an accuracy of 81.5%, as measured by cross-validation. This was significantly higher than the 80.2% achieved by the trivial classifier (P < 0.05), but the difference was quite narrow in practical terms. The AUC of the decision tree was 0.65, for which the discrimination ability of the model was quite modest. The performance was even worse when the discrimination between no pregnancy, single pregnancy and twin pregnancy was attempted after re-deriving the model and computing the AUC three times, once for each class (no pregnancy, single pregnancy and twin pregnancy); the AUCs were, respectively, 0.58, 0.50 and 0.70, confirming the limited discrimination ability of the model.

Predictions with the BN

In the task of discriminating between pregnancy and no-pregnancy, the BN model had accuracy equivalent to the decision tree, but a definitively higher AUC (0.72 versus 0.65) suggesting that the BN model can estimate more reliably the probability of pregnancy versus no pregnancy. As shown in Table II, the probabilities estimated by the model were well calibrated. For instance, the model correctly estimated a very low probability of pregnancy (<5%) in a set of 47 cycles in which only 3 clinical pregnancies resulted (6% observed pregnancy rate).

Table II

Calibration of probabilities for discrimination of pregnancy and no pregnancy

Predicted probability of pregnancy (%) 0–5 5–10 10–20 20–30 30–40 40–60 
No. of cycles 47 60 105 78 68 30 
Observed pregnancy rate (%) 10 25 33 49 
95% CI of the pregnancy rate 0–13 2–16 4–16 15–35 22–44 31–67 
Predicted probability of pregnancy (%) 0–5 5–10 10–20 20–30 30–40 40–60 
No. of cycles 47 60 105 78 68 30 
Observed pregnancy rate (%) 10 25 33 49 
95% CI of the pregnancy rate 0–13 2–16 4–16 15–35 22–44 31–67 

The conditions under which the model estimated pregnancy to be more probable than no pregnancy were (i) for IVF, the presence of two or more top-history embryos and age <34 years and (ii) for ICSI, the presence of 3 top-history embryos and age <34 years. Such conditions were quite coherent with those identified by the decision tree, and further confirmed the importance of top-history embryos and the woman's age for the establishment of a pregnancy.

However, the BN model provided an important improvement over the decision tree in discriminating between no pregnancy, single pregnancy and twin pregnancy, as indicated by the AUC measured in cross-validation, and corresponding to 0.72, 0.64 and 0.83, respectively. A crosscheck of the observed and predicted rates of clinical pregnancy for the cohort is given in Table III.

Table III

Observed and predicted rates of clinical pregnancy for cycles transferred with different number of embryos, ranging from 1 to 3

Transferred embryos Total cycles Implanted embryos
 
0
 
1
 
2
 
  Observed (%) Predicted (%) Observed (%) Predicted (%) Observed (%) Predicted (%) 
108 93 90 10 — — 
179 76 80 18 18 
101 73 73 21 20 
Transferred embryos Total cycles Implanted embryos
 
0
 
1
 
2
 
  Observed (%) Predicted (%) Observed (%) Predicted (%) Observed (%) Predicted (%) 
108 93 90 10 — — 
179 76 80 18 18 
101 73 73 21 20 

Additional improvement of the model performance was attempted by introducing further explaining variables. More specifically, maternal receptivity was also made to depend on the occurrence of previous pregnancies, and embryo viability was also made to depend on the use of sperm of proven fertility. Such attempts, despite potentially yielding a more detailed representation of the IVF-ICSI cycles, failed to improve the model's predictivity suggesting that a more complex model is not necessarily a better predictor, especially if the data set contains partially observed variables.

The estimates of maternal receptivity and embryo viability are shown in Tables IV and V, respectively. These probabilities cannot be trivially estimated from data, since for cycles resulting in no pregnancy it is unknown whether the mother was not receptive or the embryos were not viable. Therefore, it was necessary to adopt the Expectation–Maximization algorithm to estimate such probabilities. Age appeared to have a bigger impact on maternal receptivity than the number of previous IVF cycles (Table IV). Regarding embryo viability (Table V), there was an estimated difference of ∼15 points between the viability of top-history and top-quality embryos, and between the viability of top and non-top embryos. Moreover, the estimated viability decreased ∼7 points after ICSI, compared with IVF. The estimates showed that embryo score had a bigger impact on viability than the insemination technique.

Table IV

Probability of the mother being receptive, as a function of the maternal age and the number of previous cycles (95% confidence intervals in parentheses)

Age (years) Previous cycles
 
0–2 >2 
<34 0.73 (0.63–0.83) 0.69 (0.59–0.79) 
34–40 0.60 (0.54–0.66) 0.55 (0.49–0.61) 
>40 0.41 (0.31–0.51) 0.19 (0.09–0.29) 
Age (years) Previous cycles
 
0–2 >2 
<34 0.73 (0.63–0.83) 0.69 (0.59–0.79) 
34–40 0.60 (0.54–0.66) 0.55 (0.49–0.61) 
>40 0.41 (0.31–0.51) 0.19 (0.09–0.29) 
Table V

Probability of embryo implantation (assuming the mother to be receptive), as a function of the score and of the technique (95% confidence interval in parenthesis)

Embryo score Technique
 
IVF ICSI 
Non-top 0.11 (0.08–0.14) 0.07 (0.04–0.10) 
Top quality 0.25 (0.18–0.32) 0.22 (0.15–0.29) 
Top history 0.48 (0.41–0.55) 0.34 (0.27–0.41) 
Embryo score Technique
 
IVF ICSI 
Non-top 0.11 (0.08–0.14) 0.07 (0.04–0.10) 
Top quality 0.25 (0.18–0.32) 0.22 (0.15–0.29) 
Top history 0.48 (0.41–0.55) 0.34 (0.27–0.41) 

Using expected utility

The elaboration of the described estimates (age, previous cycles, embryo quality and insemination technique) permitted the set-up of a prognostic model to be used prospectively when deciding the number and type of embryos to be transferred.

Figure 3 shows, as an example, the expected utilities for different transfer possibilities, based on the probabilities computed by the BN model for a woman with age <34 years. Each row corresponds to a specific combination of her body build, insemination technique and number of previous cycles; once the appropriate row has been selected, the expected utility of transferring a different set of embryos can be read. A white background highlights the set of embryos with the highest expected utility. For example, if three top-history embryos (H) are available, the highest expected utility is attained in most situations by transferring only two of them because of the high risk of multiples. However in the case of overweight and ICSI, transferring all three embryos attains the maximum utility. Conversely, in the case of underweight women, the maximum utility is derived from the transfer of only one embryo (compare the columns ‘HHH’, ‘HH’ and ‘H’). If three non-top embryos (n) are available, the highest utility is attained by transferring three embryos in all situations (compare the columns ‘nnn’, ‘nn’ and ‘n’). Finally, if two top-quality (t) and one non-top embryos are available, the highest utility is obtained by transferring the three of them, apart from the case of underweight women, where the maximum utility is associated with the transfer of two top-quality embryos (compare columns ‘ttn’ and ‘tt’).

Figure 3

Expected utilities of different transfer possibility. The table is constructed for a patient <34 years old, considering all possible embryo type combinations (H = top history, t = top-quality, n = non top). Each row represents a specific combination of body build (underweight, normal and overweight) of the woman, insemination technique (IVF or ICSI) and number of previous cycles. The lighter the green colour, the higher the expected utility; the stronger the red colour, the lower the expected utility. For each row, a white background highlights the configuration with the highest expected utility.

Figure 3

Expected utilities of different transfer possibility. The table is constructed for a patient <34 years old, considering all possible embryo type combinations (H = top history, t = top-quality, n = non top). Each row represents a specific combination of body build (underweight, normal and overweight) of the woman, insemination technique (IVF or ICSI) and number of previous cycles. The lighter the green colour, the higher the expected utility; the stronger the red colour, the lower the expected utility. For each row, a white background highlights the configuration with the highest expected utility.

Probabilities of pregnancy calculated before and after embryo culture

The BN model can be used to provide information on the chances of pregnancy for couples before they start the treatment. In this situation, the quality of the resulting embryos is obviously unknown.

For example, in the case of a couple where the woman is <34 years old at her first stimulation and the male partner has normal semen indices, the BN model would estimate the probability of having 0, 1, 2 or 3 gestational sacs as, respectively, 69, 26, 5 and ∼0%, with a probability of pregnancy corresponding to 31% (26 + 5), compared with the 20% which is the average in the studied sample.

If the information about embryo quality would be added, it becomes evident how the prediction of the probability of pregnancy varies as additional information is collected, and the estimates would be different. Actually, if three top-history embryos would be transferred to the woman described above, the probability of having 0, 1, 2 or 3 gestational sacs would change to 37, 28, 27 and 8%; in the case of two top-history embryos being transferred the probabilities would be 46, 36, 17 and 0%.

Discussion

Since the hallmark of excellence in reproductive medicine is a consistently high pregnancy rate associated with a single implantation, the shift towards single-embryo transfer should be adopted at least in couples at risk of multiple gestations. This policy implies having a reliable method of embryo selection that should be accurate enough as to identify those cases which, based on both embryo and patient characteristics, would possibly be penalized by single-embryo transfer. In this respect, innovative approaches have been proposed to improve the embryo selection process, some of which are invasive (Geraedts et al., 2011; Schoolcraft et al., 2011), others rely on extended culture systems (which are now questioned for the neonatal outcome of the babies born; Källén et al., 2010), and others are dependent on sophisticated devices (Swain and Smith, 2011; Meseguer et al., 2012). The common denominator for all these innovations is the need for validation, as their real efficacy is still largely unproven.

With these concepts in mind, the goal of this study was to design a system that, by combining embryo morphology and the clinical parameters that were found to be relevant in determining implantation, could predict the occurrence of a pregnancy and, more specifically, the chances of single, twin and multiple implantations. The proposed model can easily assist professionals in taking decisions at the time of embryo transfer by means of an objective statistical evaluation that only requires the input of specific data in a pre-set scheme. It is undisputable that, if the correct decision were taken, the chances of pregnancy would be maximized with a concomitant reduction in the incidence of multiple gestations, which is a priority in the field of reproductive medicine.

In consideration of the multifactorial typology of implantation where several variables play a role, designing a prediction model is a complex task. The approach that we adopted was based on the elaboration of previously generated data with known clinical outcomes in order to define the features relating patients' characteristics and embryo morphology, which were determinant in establishing a pregnancy. In this way a model was elaborated, based on cross-validation of data that could be operationally used in every setting to predict not only the event of pregnancy versus no pregnancy, but also the occurrence of single implantation versus twin or multiple gestational sacs.

As a first approach for the prediction of pregnancy, we used a decision tree by which, in agreement with other studies (Leushuis et al., 2009), the accuracy was significantly higher than the trivial classifier (81.5 versus 80.2%, P < 0.05), but at such a narrow level that it was actually useless in practical terms. In addition, the model had poor performance in the more complicated task of discriminating between no pregnancy, pregnancy and multiple pregnancy. To formulate a model that could provide a solid support to IVF professionals, a BN model was proposed that made pregnancy to depend on the features selected as relevant by the training data, namely the age of the woman, the number of her previous IVF-ICSI cycles, the insemination technique and the score of the embryos. This model was also designed to assess how the probability of no pregnancy, single pregnancy and multiple pregnancy varies depending on the number and the type of the transferred embryos.

The relevance of embryo quality in establishing a pregnancy was immediately recognized by both statistical models, the decision tree and the BN. More specifically, it was clear that the probability of pregnancy strongly depended on the number of top-history embryos. In agreement with previous studies (Lan et al., 2003), it was recognized that embryos that were top quality throughout all observations were those with the highest chances of implantation, confirming that synchrony of development plays a crucial role in determining viability (Lundin et al., 2001; Lawler et al., 2007; Magli et al., 2007; Alpha Scientists in Reproductive Medicine and ESHRE Special Interest Group of Embryology, 2011). We set an exception to this rule for day 4 and day 5 embryos, for which the grade of top history was given even if they had been graded as top quality at all stages but one, provided that the single non-top assessment did not occur on the day of transfer. These observations suggest that later stages of embryo growth had higher sensitivity and specificity in the prediction of implantation (Rehman et al., 2007), giving to the assessment of blastocyst morphology an extra value at the time of embryo selection. Interestingly enough, the same characteristics seem to be presented by day 4 embryos (Feil et al., 2008) confirming that overcoming the trauma related to the activation of the embryonic genome is a sign of very good prognosis.

As expected, the concomitant analysis of the other variables indicated the negative association of implantation with age and the number of previous cycles, while a normal sperm sample was positively associated with a favourable outcome (Table V), suggesting that the capacity of activating the oocyte and guiding the first cleavage divisions are crucial events in determining the embryo fate (Heytens et al., 2009; Speyer et al., 2010). For instance, in the last couple of years, many contributions have demonstrated the relevance of sperm indices in affecting implantation due to several defects which are more common in pathological samples (Magli et al., 2009; Speyer et al., 2010; Ajduk et al., 2011; Hammoud et al., 2011).

Altogether these observations lead to the conclusion that, of the variables examined, age, number of previous cycles, insemination technique and embryo quality are the crucial factors affecting maternal receptivity and embryo viability. The derived estimates calculated by the BN model were reasonably coherent with those made by Roberts et al., who reported an average of 0.36 for receptivity and of 0.32 for viability (Roberts et al., 2009).

As an additional advantage, the BN model proposed here was able to discriminate with reasonable reliability among no pregnancy, single pregnancy and twin pregnancy, although the number of twin pregnancies included in the data set was only 17. The corresponding estimates depended on the concept of utility that describes the most desirable clinical outcome for each couple. The weight of utility can be easily modified by IVF professionals at each calculation by keeping into consideration the negative effects related to multiple pregnancies especially in consideration of the woman's physical condition (Fig. 3). We made it depend on her body build, but additional factors could be also included, as the model is extremely flexible.

The benefit of a model that can support the decision at the time of transfer would be especially evident in case of having several embryos in culture. The scheme proposed here was based on three embryos that by law could be left in culture, and this notably decreased the number of observations and the possibility of selection. However, the BN model can be straightforwardly expanded to manage a larger number of embryos; in this case, the calculation of the expected utilities could be of even greater advantage in supporting decisions by offering wide ranges of possible combinations per each cycle. This would be especially important in those situations where the maternal receptivity and the morphological quality of the embryos suggest the transfer of more than one embryo. It is actually clear that the policy of single-embryo transfer is not the best option for all cases, and having an objective support at this stage would be of clear benefit (McLernon et al., 2010; Jonsdottir et al., 2011).

Should specific morphological and developmental details coming from morphokinetic studies be confirmed to be crucial for implantation (Wong et al., 2010; Meseguer et al., 2011), the corresponding features could be included to make the predictor even stronger.

As a further result, the BN model was able to identify cycles having an extremely low chance of implantation (Table II). This opens up to a series of considerations regarding the rationale of cancelling transfers with a very negative prognosis, avoiding, for the patient, the stress of a hopeless transfer with a minimal loss of pregnancies. Although this may be an acceptable approach from a statistical point of view that actually permits verification of the fitness of the utility model, from an ethical perspective it poses great difficulties especially related to the fact that embryos have already been generated.

In conclusion, the present study showed that all the prerequisites are given to construct a model aimed at supporting decisions and making predictions that are specific for every single case. The derived figures are clinically reasonable and have the advantage of being reproducible and objective, as confirmed by several measurements like AUC and the calibration probabilities. The preliminary data derived from its prospective application at the time of transfer (39 cycles so far) confirm its clinical validity resulting in 12 clinical pregnancies (31% clinical pregnancy rate): all singletons (in two cases three embryos were transferred) and one twin (two embryos were transferred). Moreover, the model is flexible enough to make predictions on the chances of pregnancy even before the couple starts a treatment cycle resulting in great acceptance from couples, although the prediction is more robust when the information on embryo quality is added (Smeenk et al., 2000). It is clear that some utility profiles can vary depending on each center's policy and experience, and they can be quickly changed by the IVF professionals to personalize the model represented in Fig. 3. In this way, it is possible to calculate for each couple the expected utility regarding which embryo and how many embryos to transfer for the most desired clinical outcome.

Authors' roles

L.G. formulated the study design, participated in manuscript drafting and critical discussion and approved the final version; M.C.M. participated in the study design, execution, analysis, manuscript drafting and critical discussion and approved the final version; A.G. participated in the study execution and critical discussion and approved the final version; C.G. participated in the study execution and critical discussion and approved the final version; L.M.G. participated in the study design, analysis and critical discussion and approved the final version; G.C. designed and validated the predictive models, participated in the study design, execution, analysis, manuscript drafting and critical discussion and approved the final version.

Funding

This study was supported by a grant, CTI Medtech Project Number: 9707.1 PFLS-L, Swiss Confederation.

Conflict of interest

None declared.

References

Ajduk
A
Ilozue
T
Windsor
S
Yu
Y
Seres
KB
Bomphrey
RJ
Tom
BD
Swann
K
Thomas
A
Graham
C
, et al.  . 
Rhythmic actomyosin-driven contractions induced by sperm entry predict mammalian embryo viability
Nat Commun
 , 
2011
, vol. 
2
 pg. 
417
 
Alpha Scientists in Reproductive Medicine and ESHRE Special Interest Group of Embryology
The Istanbul consensus workshop on embryo assessment: proceedings of an expert meeting
Hum Reprod
 , 
2011
, vol. 
26
 (pg. 
1270
-
1283
)
Beuchat
A
Thevenaz
P
Unser
M
Ebner
T
Senn
A
Urner
F
Germond
M
Sorzano
CO
Quantitative morphometrical characterization of human pronuclear zygotes
Hum Reprod
 , 
2008
, vol. 
23
 (pg. 
1983
-
1992
)
Brezinova
J
Oborna
I
Svobodova
M
Fingerova
H
Evaluation of day one embryo quality and IVF outcome—a comparison of two scoring systems
Reprod Biol Endocrinol
 , 
2009
, vol. 
7
 pg. 
9
 
Brison
DR
Hollywood
K
Arnesen
R
Goodacre
R
Predicting human embryo viability: the road to non-invasive analysis of the secretome using metabolic footprinting
Reprod Biomed Online
 , 
2007
, vol. 
15
 (pg. 
296
-
302
)
De Placido
G
Wilding
M
Strina
I
Alviggi
E
Alviggi
C
Mollo
A
Varicchio
M
Tolino
A
Schiattarella
C
Dale
B
High outcome predictability after IVF using a combined score for zygote and embryo morphology and growth rate
Hum Reprod
 , 
2002
, vol. 
17
 (pg. 
2402
-
2409
)
Dumoulin
JC
Land
JA
Van Montfoort
AP
Nelissen
EC
Coonen
E
Derhaag
JG
Schreurs
IL
Dunselman
GA
Kester
AD
Effect of in vitro culture of human embryos on birthweight of newborns
Hum Reprod
 , 
2010
, vol. 
25
 (pg. 
605
-
612
)
Feil
D
Henshaw
RC
Lane
M
Day 4 embryo selection is equal to Day 5 using a new embryo scoring system validated in single embryo transfers
Hum Reprod
 , 
2008
, vol. 
7
 (pg. 
1505
-
1510
)
Ferraretti
AP
Gianaroli
L
Magli
MC
D'Angelo
A
Farfalli
V
Montanaro
N
Exogenous LH in COH for ART: when and which?
Fertil Steril
 , 
2004
, vol. 
82
 (pg. 
1521
-
1526
)
Gelbaya
TA
Tsoumpou
I
Nardo
LG
The likelihood of live birth and multiple birth after single versus double embryo transfer at the cleavage stage: a systematic review and meta-analysis
Fertil Steril
 , 
2010
, vol. 
94
 (pg. 
936
-
945
)
Geraedts
J
Montag
M
Magli
MC
Repping
S
Handyside
A
Staessen
C
Harper
J
Schmutzler
A
Collins
J
Goossens
V
, et al.  . 
Polar body array CGH for prediction of the status of the corresponding oocyte. Part I: clinical results
Hum Reprod
 , 
2011
, vol. 
26
 (pg. 
3173
-
3180
)
Gianaroli
L
Magli
MC
Ferraretti
AP
Fortini
D
Grieco
N
Pronuclear morphology and chromosomal abnormalities as scoring criteria for embryo selection
Fertil Steril
 , 
2003
, vol. 
80
 (pg. 
341
-
349
)
Gianaroli
L
Magli
MC
Ferraretti
AP
Sperm and blastomere aneuploidy detection in reproductive genetics and medicine
J Histochem Cytochem
 , 
2005a
, vol. 
53
 (pg. 
261
-
268
)
Gianaroli
L
Gordts
S
D'Angelo
A
Magli
MC
Brosens
I
Cetera
C
Campo
R
Ferraretti
AP
Effect of the inner myometrium fibroid on the reproductive outcome after IVF
Reprod Biomed Online
 , 
2005b
, vol. 
10
 (pg. 
473
-
477
)
Gianaroli
L
Magli
MC
Ferraretti
AP
Lappi
M
Borghi
E
Erimini
B
Oocyte euploidy, pronuclear zygote morphology and embryo chromosomal complement
Hum Reprod
 , 
2007
, vol. 
22
 (pg. 
241
-
249
)
Gianaroli
L
Magli
MC
Cavallini
G
Crippa
A
Capoti
A
Resta
S
Robles
F
Ferraretti
AP
Predicting aneuploidy in human oocytes: key factors which affect the meiotic process
Hum Reprod
 , 
2010
, vol. 
25
 (pg. 
2374
-
2386
)
Guzeloglu-Kayisli
O
Kayisli
UA
Taylor
HS
The role of growth factors and cytokines during implantation: endocrine and paracrine interactions
Semin Reprod Med
 , 
2009
, vol. 
27
 (pg. 
62
-
79
)
Hammoud
SS
Nix
DA
Hammoud
AO
Gibson
M
Cairns
BR
Carrell
DT
Genome-wide analysis identifies changes in histone retention and epigenetic modifications at developmental and imprinted gene loci in the sperm of infertile men
Hum Reprod
 , 
2011
, vol. 
26
 (pg. 
2558
-
2569
)
Heytens
E
Parrington
J
Coward
K
Young
C
Lambrecht
S
Yoon
S-Y
Fissore
RA
Hamer
R
Deane
CM
Ruas
M
, et al.  . 
Reduced amounts and abnormal forms of phospholipase C zeta (PLCζ) in spermatozoa from infertile men
Hum Reprod
 , 
2009
, vol. 
24
 (pg. 
2417
-
2428
)
Horsthemke
B
Ludwig
M
Assisted reproduction: the epigenetic perspective
Hum Reprod Update
 , 
2005
, vol. 
11
 (pg. 
473
-
482
)
Jones
GM
Trounson
AO
Vella
PJ
Thouas
GA
Lolatgis
N
Wood
C
Glucose metabolism of human morula and blastocyst-stage embryos and its relationship to viability after transfer
Reprod Biomed Online
 , 
2001
, vol. 
3
 (pg. 
124
-
132
)
Jones
GM
Cram
DS
Song
B
Magli
MC
Gianaroli
L
Lacham-Kaplan
O
Findlay
JK
Jenkin
G
Trounson
AO
Gene expression profiling of human oocytes following in vivo or in vitro maturation
Hum Reprod
 , 
2008a
, vol. 
23
 (pg. 
1138
-
1144
)
Jones
G
Cram
DS
Song
B
Kokkali
G
Pantos
K
Trounson
AO
Novel strategy with potential to identify developmentally competent IVF blastocysts
Hum Reprod
 , 
2008b
, vol. 
23
 (pg. 
1748
-
1759
)
Jonsdottir
I
Lundin
K
Bergh
C
Double embryo transfer gives good pregnancy and live birth rates in poor responders with a modest increase in multiple birth rates: results from an observational study
Acta Obstet Gynecol Scand
 , 
2011
, vol. 
90
 (pg. 
761
-
766
)
Källén
B
Finnström
O
Lindam
A
Nilsson
E
Nygren
KG
Olausson
PO
Blastocyst versus cleavage stage transfer in in vitro fertilization: differences in neonatal outcome?
Fertil Steril
 , 
2010
, vol. 
94
 (pg. 
1680
-
1683
)
Katz-Jaffe
MG
Gardner
DK
Schoolcraft
WB
Proteomic analysis of individual human embryos to identify novel biomarkers of development and viability
Fertil Steril
 , 
2006
, vol. 
85
 (pg. 
101
-
107
)
Koller
D
Friedman
N
Probabilistic Graphical Models
 , 
2009
MIT Press
 
Cambridge, Massachusetts London, England
Kovacic
B
Vlaisavljevic
V
Reljic
M
Cizek-Sajko
M
Developmental capacity of different morphological types of day 5 human morulae and blastocysts
Reprod Biomed Online
 , 
2004
, vol. 
8
 (pg. 
687
-
694
)
Kuliev
A
Cieslak
J
Verlinsky
Y
Frequency and distribution of chromosome abnormalities in human oocytes
Cytogenet Genome Res
 , 
2005
, vol. 
111
 (pg. 
193
-
198
)
Lan
K
Huang
F
Lin
Y
Kung
F
Hsieh
C
Huang
H
Tan
PH
Chang
SY
The predictive value of using a combined Z-score and day 3 embryo morphology score in the assessment of embryo survival on day 5
Hum Reprod
 , 
2003
, vol. 
18
 (pg. 
1299
-
1306
)
Lawler
C
Baker
HWG
Edgar
DH
Relationships between timing of syngamy, female age and implantation potential in human in vitro-fertilised oocytes
Reprod Fertil Dev
 , 
2007
, vol. 
19
 (pg. 
482
-
487
)
Leushuis
E
Van Der Steeg
J
Steures
PB
Eijkemans
M
Van Der Veen
F
Mol
BW
Hompes
PG
Prediction models in reproductive medicine: a critical appraisal
Hum Reprod Update
 , 
2009
, vol. 
15
 (pg. 
537
-
552
)
Loi
K
Prasath
EB
Huang
ZW
Loh
SF
Loh
SKE
A cumulative embryo scoring system for the prediction of pregnancy outcome following intracytoplasmic sperm injection
Singapore Med J
 , 
2008
, vol. 
49
 (pg. 
221
-
227
)
Lundin
K
Bergh
C
Hardarson
T
Early embryo cleavage is a strong indicator of embryo quality in human IVF
Hum Reprod
 , 
2001
, vol. 
16
 (pg. 
2652
-
2657
)
Magli
MC
Gianaroli
L
Ferraretti
AP
Lappi
M
Ruberti
A
Farfalli
V
Embryo morphology and development is dependent on the chromosomal complement
Fertil Steril
 , 
2007
, vol. 
87
 (pg. 
534
-
541
)
Magli
MC
Gianaroli
L
Ferraretti
AP
Gordts
S
Fredericks
V
Crippa
A
Paternal contribution to aneuploidy in preimplantation embryos
Reprod Biomed Online
 , 
2009
, vol. 
18
 (pg. 
536
-
542
)
Margalioth
EJ
Ben-Chetrit
A
Gal
M
Eldar-Geva
T
Investigation and treatment of repeat implantation failure following IVF-ET
Hum Reprod
 , 
2006
, vol. 
21
 (pg. 
3036
-
3043
)
McLernon
DJ
Harrild
K
Bergh
C
Davies
MJ
de Neubourg
D
Dumoulin
JC
Gerris
J
Kremer
JA
Martikainen
H
Mol
BW
, et al.  . 
Clinical effectiveness of elective single versus double embryo transfer: meta-analysis of individual patient data from randomised trials
BMJ
 , 
2010
, vol. 
341
 pg. 
c6945
 
Meseguer
M
Herrero
J
Tejera
A
Hilligsøe
KM
BirgerRamsing
N
Remohı
J
The use of morphokinetics as a predictor of embryo implantation
Hum Reprod
 , 
2011
, vol. 
26
 (pg. 
2658
-
2671
)
Meseguer
M
Kruhne
U
Laursen
S
Full in vitro fertilization laboratory mechanization: toward robotic assisted reproduction?
Fertil Steril
 , 
2012
, vol. 
97
 (pg. 
1277
-
1286
)
Munné
S
Chen
S
Colls
P
Garrisi
J
Zheng
X
Cekleniak
N
Lenzi
M
Hughes
M
Fischer
J
Garrisi
M
, et al.  . 
Maternal age, morphology, development and chromosome abnormalities in over 6000 cleavage-stage embryos
Reprod Biomed Online
 , 
2007
, vol. 
14
 (pg. 
628
-
634
)
Norwitz
ER
Schust
DJ
Fisher
SJ
Implantation and the survival of early pregnancy
N Engl J Med
 , 
2001
, vol. 
345
 (pg. 
1400
-
1408
)
Pandian
Z
Bhattacharya
S
Ozturk
O
Serour
G
Templeton
A
Number of embryos for transfer following in-vitro fertilisation or intra-cytoplasmic sperm injection
Cochrane Database Syst Rev
 , 
2009
, vol. 
2
 pg. 
CD003416
 
Racowsky
C
Stern
JE
Gibbons
WE
Behr
B
Pomeroy
KO
Biggers
JD
National collection of embryo morphology data into Society for Assisted Reproductive Technology Clinic Outcomes Reporting System: associations among day 3 cell number, fragmentation and blastomere asymmetry, and live birth rate
Fertil Steril
 , 
2011
, vol. 
95
 (pg. 
1985
-
1989
)
Rehman
KS
Bukulmez
O
Langley
M
Carr
BR
Nackley
AC
Doody
KM
Doody
KJ
Late stages of embryo progression are a much better predictor of clinical pregnancy than early cleavage in intracytoplasmic sperm injection and in vitro fertilization cycles with blastocyst-stage transfer
Fertil Steril
 , 
2007
, vol. 
87
 (pg. 
1041
-
1052
)
Roberts
S
Models for assisted conception data with embryo-specific covariates
Stat Med
 , 
2007
, vol. 
26
 (pg. 
156
-
170
)
Roberts
S
Fitzgerald
C
Brison
D
Modelling the impact of single embryo transfer in a national health service IVF programme
Hum Reprod
 , 
2009
, vol. 
24
 (pg. 
122
-
131
)
Roberts
SA
Hirst
WM
Brison
DR
Vail
A
Embryo and uterine influences on IVF outcomes: an analysis of a UK multi-centre cohort
Hum Reprod
 , 
2010
, vol. 
25
 (pg. 
2792
-
2802
)
Saith
R
Srinivasan
A
Michie
D
Sargent
IL
Relationships between the developmental potential of human in-vitro fertilization embryos and features describing the embryo, oocyte and follicle
Hum Reprod Update
 , 
1998
, vol. 
4
 (pg. 
121
-
134
)
Santos
M
Kuijk
EW
Macklon
NS
The impact of ovarian stimulation for IVF on the developing embryo
Reprod
 , 
2010
, vol. 
139
 (pg. 
23
-
34
)
Schoolcraft
WB
Treff
NR
Stevens
JM
Ferry
K
Katz-Jaffe
M
Scott
RT
Jr
Live birth outcome with trophectoderm biopsy, blastocyst vitrification, and single-nucleotide polymorphism microarray-based comprehensive chromosome screening in infertile patients
Fertil Steril
 , 
2011
, vol. 
96
 (pg. 
638
-
640
)
Simon
C
Martin
JC
Pellicer
A
Paracrine regulators of implantation
Baillieres Best Pract Res Clin Obstet Gynaecol
 , 
2000
, vol. 
14
 (pg. 
815
-
826
)
Sjöblom
P
Menezes
J
Cummins
L
Mathiyalagan
B
Costello
MF
Prediction of embryo developmental potential and pregnancy based on early stage morphological characteristics
Fertil Steril
 , 
2006
, vol. 
86
 (pg. 
848
-
861
)
Smeenk
JM
Stolwijk
AM
Kremer
JA
Braat
DD
External validation of the templeton model for predicting success after IVF
Hum Reprod
 , 
2000
, vol. 
15
 (pg. 
1065
-
1068
)
Speyer
BE
Pizzey
AR
Ranieri
M
Joshi
R
Delhanty
JD
Serhal
P
Fall in implantation rates following ICSI with sperm with high DNA fragmentation
Hum Reprod
 , 
2010
, vol. 
25
 (pg. 
1609
-
1618
)
Swain
JE
Smith
GD
Advances in embryo culture platforms: novel approaches to improve preimplantation embryo development through modifications of the microenvironment
Hum Reprod Update
 , 
2011
, vol. 
17
 (pg. 
541
-
557
)
Volpes
A
Sammartano
F
Coffaro
F
Mistretta
V
Scaglione
P
Allegra
A
Number of good quality embryos on day 3 is predictive for both pregnancy and implantation rates in in vitro fertilization/intracytoplasmic sperm injection cycles
Fertil Steril
 , 
2004
, vol. 
82
 (pg. 
1330
-
1336
)
Witten
IH
Frank
E
Hal
M
Data Mining: Practical Machine Learning Tools and Techniques
 , 
2011
3rd edn
San Francisco, CA
Morgan Kaufmann Publishers
Wong
CC
Loewke
KE
Bossert
NL
Behr
B
De Jonge
CJ
Baer
TM
Reijo Pera
RA
Non-invasive imaging of human embryos before embryonic genome activation predicts development to the blastocyst stage
Nat Biotechnol
 , 
2010
, vol. 
28
 (pg. 
1115
-
1121
)
World Health Organization
Obesity: preventing and managing the global epidemic
.
 , 
2000
Geneva, Switzerland
World Health Organization
 
Report of a WHO Consultation on Obesity. WHO Technical Report Series 894
Zhou
H
Weinberg
C
Evaluating effects of exposures on embryo viability and uterine receptivity in in vitro fertilization
Stat Med
 , 
1998
, vol. 
17
 (pg. 
1601
-
1612
)