Optimal Ensemble - Instance Bootstrap Classifier With Balanced Class for Autism Diagnosis

Abstract

For the better reformation in medical diagnosis, machine learning techniques contribution is
highly recommendable as it has the potential ability to improve prediction accuracy. Autism
Spectrum Disorder (ASD) as neurological based disorder that affects the developmental
milestones of a child at their earlier age. In order to attain the highest precision in
classification, the datasets “Autistic Spectrum Disorder Screening Data for Adult, Toddler,
Child and Adolescent” contains 118, 1054, 509, and 248 instances are used for classification.
The proposed work consists of two phases: i. applies one-hot coding algorithm for the nonnumerical
attributes and implement bruteforce KNN and bootstrap Random forest machine
learning models using balanced class with optimality. It is applied to an extracted features of
all four datasets. It is to classify the new observations for ASD screening process and the
observed results shown that Random forest performed better over KNN. ii.. “Ensemble-
Instance based classifier with Features of optimality” (EI-FO) combines both the
performance of bruteforce KNN and Bootstrap Random Forest algorithm applied to 18
features of importance resulted in 99% of accuracy. EI-OF further examined on 14 optimal
features of Adult dataset: 10 questionnaires (A1-A10), age, Ethnicity and residence with
specific values(United States, Singapore and Malaysia). The experimental outcomes of the
proposed EI-FO classification algorithm can diagnose autism in Adult with an predictable
accuracy of 98.6% (US), 98.2 %( Singapore) and 99.1 %(Malaysia). Overall, the proposed
EI-OF(accuracy 99%) outperformed when compared to KNN(accuracy 96% accuracy) in
terms of time complexity, precision, Recall and F-Score.