An Automatic Turner Syndrome Identification System with Facial Images

Abstract

Gestational weight is an essential parameter for the pediatrician to clinically evaluate the health of both neonate and the mother. During the last several decades, an increase in the prevalence of Large for Gestational Age (LGA) neonate is reported and several researchers engaged themselves to discover the cause. Most of them conducted observational or retrospective studies that used simple statistical test (i.e. univariate/multivariate logistic regression etc.,). However, machine learning schemes are rarely been employed to discover the cause. In this research, one proposed expert-driven and seven automated feature selection schemes with five well-known machine learning classifiers using (10 & 30)-fold cross-validations are employed for the establishment of an efficient and accurate LGA classification model. Accuracy, precision, and AUC scores are selected for the evaluation of the proposed scheme. Wilcoxon signed rank, friedman, and bonferroni-dunn tests are used to observe the variations among (10 & 30)-fold cross validation results and to rank various feature selection and classification schemes. Two baseline methods are also used to compare the results of the proposed expert-driven feature selection scheme. The top 20 features selected by the proposed expert-driven feature selection scheme outperformed among seven automated feature selection schemes. A comparison analysis is also performed between expert-driven and data-driven feature subsets. Furthermore, with the intersection of proposed expert-driven and data-driven feature subsets, it is foreseen that out of 20 features, 11 features are found similar, which authenticates the proposed scheme. The classification performance of the 11 extracted features is almost similar to the proposed expert-driven feature selection scheme. Ensemble technique is also exploited to build the better and effective LGA classification model.