Autism Disorder Diagnosis Enhancement Using Adaptive Ranking Features and Machine Learning Classifiers

Autism is a disorder of brain function that can appear in children younger than two years and affects their communication and learning. The incidence of Autism in children has been increasing, and at the same time, numerous schemes have been proposed for its diagnosis. However, the researchers continue to face challenges in the early and accurate diagnosis of Autism. Early and accurate detection of autism allows doctors to determine the severity of the disease and begin appropriate treatment protocols to develop communication skills in children. This paper proposes feature ranking methods integrated with four classifiers for the accurate diagnosis of autism. Multiple experiments are conducted by integrating ranking features methods (chi-square, Anova-F, and chi2-yates) with Decision Tree (DT), Multi-layer perceptron (MLP), NaiveBayes Bernoulli (NB), and Support Vector Machine –Radial Basis Function (SVM-RBF) classifiers to achieve accurate early detection. The feature ranking is applied with five sets (5, 10, 20, 40, and 46), and each set is individually evaluated to assess its contribution to autism diagnosis. The best diagnosis model (chi2-yates-based SVM-RBF) has achieved 0.971611, 0.94797, and 0.997335 for the accuracy, F1-score, and ROC-AUC, respectively. The proposed framework would be effective in helping therapists in early detection and selecting suitable treatment for autism.