Early Stage Prediction of Autism Spectrum Disorder: Analyzing Different Hyperparameter Tuned Machine Learning Classifier

Autism hinders the socio-personal life of individuals and the development of society; so, detecting autism spectrum disorder (ASD) at an early stage is an urge of humanity. There has been a significant amount of research done to detect ASD using Machine Learning (ML) approaches, but there is a considerable research gap in preprocessing the raw datasets. In this study, maintaining well-structured data preprocessing steps, we have performed a comparative analysis between multiple traditional and ensemble ML classifiers and proposed a model for predicting ASD using three (3) ASD datasets of all ages available in the University of California Irvine (UCI) machine learning repository. We have applied six (06) traditional ML classifiers: Logistic Regression (LR), Support Vector Machine (SVM), Naive-Bayes (NB), Random Forest (RF), Decision Tree (DT), and Extra Tree (ET); employed bagging and boosting classifiers using the traditional classifiers as base estimators to improve their performance, evaluated their performance using different statistical evaluation measures. We have also utilized the 10-fold randomized cross-validation technique (RandomizedSearchCV) to find the best hyperparameters. The study concludes with the information that, with adolescent, children, and adult datasets, respectively, DT(Boosting), LR(Traditional), and SVM(Boosting) perform the best, retaining the intended aspects and experiencing reduced bias.