23. SCORING METHOD TO ESTIMATE CLINICAL PREGNANCY USING ARTIFICIAL INTELLIGENCE MODEL

Abstract

Introduction Clinical pregnancy represents success outcome in a PGT-a in cycle. It can be defined as the presence of sacs with fetal heart beats at 7-8 weeks. Main purpose of PGT-a study is to identify euploid embryos in order to be replaced to patient endometrium. Nevertheless, PGT-a cycles differs in relation with attributes related to patients as well cycle information. The main goal of the present study is to design an predictive model capable to give a probability of clinical pregnancy before embryo transfer takes place. Material & methods Dataset used included 1190 cycles that belong at New York University Fertility Center between 2012 and 2015. Patient id was encrypted with an alphanumeric code. The classifier implement was a Multilayer Perceptron. Feature engineering and standardization was carried out in the original variables. Algorithm was implemented using Python programing language.Model parameters was searched using grid search methodology. Classification performance was evaluated via 5-fold cross validation and area under the curve (AUC) Results Variables used as a predictor were: maternal age, biopsy specimen method, reason for referral and quantity of euploid and aneuploid embryos generated in the cycle. In order to evaluate normality distribution among the dependent variables selected D'Agostino's K2 test was performed as a goodness-of-fit measure. Results variables showed a non gaussian distribution (p>0.05). Lack of normality was due to presence of outliers. To mitigate it effects we employed a technique called feature engineering applying mathematical function to the original variables. In order to include categorical variables, it were coded using numerical encoding. Model implemented achieves an Area Under the Curve (AUC) of 80%. In other words, the sensitivity-specificity tradeoff in the classifier was of 80%. It is important to note that threshold selected was 0.5. It means that cases with an output after classification with values higher than the cut-off were classified as Pregnant. On the other hand cases with output lower than 0.5 were assigned as non pregnant. Furthermore, Sensitivity (True Positive / True Positive + False Negative) and Specificity (True Negative / True Negative + False Positive) were 82% and 69% respectively. Conclusions We are aware about the presence of an intrinsic bias closely related to the Fertility Center where the PGT-a cycles was performed. In order to overcome that we are developing a new algorithm capable to solve current limitation being capable to be adapted for a specific center. A scoring tool based on artificial intelligence can be helpful in selecting those patients with high probability to achieve clinical pregnancy during PGT-a cycle. Artificial neural networks is a robust estimator that allowed us to overcome multicollinearity between the original variables. Overall performance of the model was high (80%) giving strong evidence that the cycles can be accurately classified.