High accuracy machine learning predictive model for embryo selection in IVF PGT cycles with single embryo transfers

Abstract

Recent publications have demonstrated significant improvement in IVF treatment outcomes by implementing NGS-based preimplantation genetic testing (PGT) while transferring fewer embryos. The objective of this study was to create complex machine learning (ML) model that can accurately predict the reproductive potential of a particular euploid embryo to establish viable pregnancy in IVF PGT cycles with single embryo transfer (SET). A retrospective study of NGS PGT outcome data from blastocysts biopsied on day 5 or day 6 (2013-2017) was conducted using supervised and unsupervised ML algorithms to identify differences in clinical pregnancy rates (PR). 1383 cycles (7120 embryos) of IVF PGT followed by 1108 SETs were included in the study (842 patients). From 320 original features (175 clinical and 145 morph. and kinetic parameters of embryo development), 6470 synthetic features were created (by Weight of Evidence, Encoding of categorical levels, Target Encoding, etc.), tested and 357 features were selected. 73 statistical models were trained and ensembled in the final model. Predictive accuracy was evaluated by 5-fold cross-validation. Clinical PR was defined by the presence of a fetal heartbeat at 6-7 weeks of pregnancy. Analysis of the combined predictions from multiple weak learners (GLM, random forest, gradient boosting, etc.) processed by Generalized Model Stacking produced a predictive performance of AUC = 0.8103, Logloss = 0.5356. The probability of positive clinical outcome was calculated for each euploid embryo and ranged from 0.194 to 0.838 (baseline prediction - 0.645). The variable of highest importance were the morphological characteristics of the blastocysts, the history of previous failed IVF cycles, and time when blastocysts became available for biopsy (0.372, 0.233, and 0.118 respectively). Generalized linear model intercept estimate for embryo morphology was 0.957 (Std. Error = 0.36, Pr (|z|) = 0.008), intercept estimate for the history of previous failed IVF cycles was -0.625 (Std. Error = 0.1, Pr (|z|) = 3.1e-10), and intercept estimate for biopsy day was -0.416 (Std. Error = 0.133, Pr (|z|) = 0.002). These results were confirmed by univariate analysis: the ongoing PR after SET was considerably higher when the transferred euploid embryos were graded as good quality embryos (AA/AB/BA) vs fair (BB) or borderline fair (-B or B-) quality: 68.6% (493/719) vs 56.9% (161/283) and 42.5% (45/106) respectively, χ2=4.31, OR = 0.338, CI = 0.223 - 0.513, p<0.05. Analysis of the data proved that Machine learning algorithms applied to large clinical data sets could predict the outcome of the IVF treatment with high accuracy. In order to achieve high accuracy predictions in IVF PGT cycles factors affecting clinical outcomes were defined, ranked, and evaluated. Ensemble methods of statistical learning offer superior performance over their singleton counterparts and essentially help to transform medical records into medical knowledge.