A Bayesian network model for predicting pregnancy after in vitro fertilization.

We present a Bayesian network model for predicting the outcome of in vitro fertilization (IVF). The problem is characterized by a particular missingness process; we propose a simple but effective averaging approach which improves parameter estimates compared to the traditional MAP estimation. We present results with generated data and the analysis of a real data set. Moreover, we assess by means of a simulation study the effectiveness of the model in supporting the selection of the embryos to be transferred.

Objective way to support embryo transfer: a probabilistic decision.

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.