How much information about embryo implantation potential is included in morphokinetic data? A prediction model based on artificial neural networks and principal component analysis

Abstract

PurposeThe aim of this study was to answer the question of how much information about embryo implantation potential can be obtained from morphokinetic parameters through the creation a predictive model based on morphokinetic information and using advanced data-mining and artificial intelligence methods. Materials and methodsTime-lapse recordings of 610 embryos were included in the analysis. For each embryo, absolute (t2, t3, t4, t5) and relative (cc2 and s2) morphokinetic parameters were collected. Additionally, fragmentation levels assessed at t2, t3, t4 and t5 time-points, presence of multinucleation, evenness of blastomeres after the first and second division and women's age were included in the analysis. ResultsThe created predictive model of implantation was constructed on the basis of two advanced data-mining methods: principal component analysis (PCA) and artificial neural networks (ANN). The receiver operating characteristic (ROC) curve constructed for the created model demonstrated its good quality. The area under the ROC curve was AUC=0.75 with a 95% confidence interval (0.70, 0.80). Finally, the model was verified on a validation set and the results were also good, although slightly weaker: AUC=0.71, with a 95% confidence interval (0.59, 0.84). ConclusionsThe combination of two data-mining algorithms: PCA and ANN may be considered a method which can extract virtually all the available information from data. This methodology is indeed efficient, but models presented in the literature are also effective and close to the limit of the maximal information which can be extracted from morphokinetic data. The limit can be determined at the level of AUC value marginally above 0.7.