Prediction of Placenta Accreta Spectrum by Machine Learning Methods and Determination of Candidate Biomarkers

Abstract

Placenta accreta spectrum (PAS) disorders; Abnormal adhesion of placental villi to the myometrium associated with endometrial trauma or dysplasia. Placenta previa and previous cesarean section operations are two major risk factors for PAS disorders. It is usually diagnosed by ultrasound examinations performed during pregnancy follow-up. After this diagnosis is made, a very careful and strict pregnancy follow-up should be done. If the diagnosis is made during pregnancy, the delivery should be done by cesarean section and the bleeding that the mother will experience should be stopped with an appropriate method. However, no protein candidate to be used in clinical diagnosis has been found so far. The aim of this study is to identify candidate biomarkers that can be used in the diagnosis and follow-up of PAS with machine learning methods.
 In this study, proteomic data obtained from 26 women with and without PAS were used. After using the Lasso method as the variable selection method, machine learning models (XGBoost, Adaboost) were created with 5-fold cross-validation. Accuracy, Balanced accuracy, Sensitivity, Specificity, Positive Predictive Value, Negative Predictive Value, F1-Score, MCC and G-mean metrics were used in the performance evaluation of the models created.
 When the performance metrics of the two models are compared, the best result belongs to the XGBoost machine learning model. Therefore, the Accuracy, Balanced accuracy, Sensitivity, Specificity, Positive Predictive Value, Negative Predictive Value, F1-Score, MCC, and G-mean performance criteria for the XGBoost model are 0.962, 0.950, 1.00, 0.90, 0.94, 1.00, 0.97, 0.92, and 0.97, respectively.
 As a result, considering the experimental results, it can be said that the created machine learning model is quite successful in classifying PAS. In addition, it can be said that KDR and AMH proteins are candidate biomarkers that can be used in the diagnosis and follow-up of PAS according to the significance of the variables related to the model.