A data mining approach for the prediction of pregnancy during IVF treatment

Introduction: One of the most relevant clinical aspects during IVF-ICSI treatments is to decide if endometrium is receptive for embryo implantation, in order to perform embryo transfer or to postpone it in another cycle. This might increase both patients’ convenience and the cost-effectiveness of ART procedure. Different strategies have been developed to evaluate non-invasive endometrial receptivity mainly based on ultrasound examination of endometrial features including thickness, pattern, volume and blood flow. To the best of our knowledge, almost all the published works analyze all these features singularly, without taking into account their possible interaction. The aim of this work is to approach the problem by a data mining technique able to consider an optimal combination of features for designing a system that may predict pregnancy during IVF-ICSI treatment. Material & methods: A total of 62 cycles of ICSI, each related to a different patient, resulting in 27 pregnancy and 35 failures were included in this study. All the ultrasound scans and measurements were performed on the day of hCG administration by the same observer with the digital platform VOLUSON i System using a standard acquisition procedure. Ten features were measured in each cycle: age of the woman (AGE), endometrial vascularization index (E-VI), endometrial flow index (E-FI), endometrial vascularization/flow index (E-VFI), sub-endometrial vascularization index (S-VI), sub-endometrial flow index (S-FI), sub-endometrial vascularization/flow index (S-VFI), endometrial thickness (E-T), endometrial volume (E-V), sub-endometrial volume (S-V). In order to evaluate if each feature is significant for this prediction problem the area under the ROC curve (AUC) was calculated. Then, a data mining approach based on support vector machines (SVM) was designed, combining a set of features to give a prediction score to the IVF-ICSI treatment. The proposed system needs a training phase for the optimization of internal SVM parameters, that is performed according to the leave-one-out testing protocol, which involves the use of a single cycle from the dataset as testing data (each once), and the remaining ones as training. Results: The analysis of the AUC for each feature revealed that the three best features are: S-VI (0.703), E-VI (0.685) and E-VFI (0.677). Using a SVM system trained by the best three features above cited the performance does not appreciably change, probably due to the fact that these features contain similar information. If a feature selection approach is used to choose the three most suited features to be combined, the performance increases to 0.78. This very encouraging result has been obtained training a SVM by the combination of the following three features: S-VI, E-FI and AGE. In our experiments the well-known Sequential Forward Floating Selection (SFFS) algorithm for selecting a subset of features was used. SFFS is based on a search strategy that checks whether, at each sequential step, removing any feature in the selected set and adding a new one can improve the resultant set. Conclusions: Our experiments confirm that vascularization data, measured from both the endometrium and the subendometrium in the day of hCG administration, are significant for the problem of predicting pregnancy. More interestingly our main practical finding is that it is possible to build and use a reliable data mining technique, with the optimal combination of a small set of non invasive measurable features, to predict pregnancy during IVF treatment.