A FRAMEWORK OF MACHINE LEARNING FOR EARLYSTAGE DETECTION OF AUTISM SPECTRUM DISORDERS

Abstract

The project focuses on proposing an effective framework for early detection of Autism Spectrum Disorder
(ASD) using Machine Learning (ML) techniques, recognizing the challenges in completely eradicating the disorder
but aiming to mitigate its severity through early interventions. The proposed framework employs four Feature
Scaling (FS) strategies (Quantile Transformer, Power Transformer, Normalizer, Max Abs Scaler) and evaluates their
impact on four standard ASD datasets representing different age groups (Toddlers, Adolescents, Children, and
Adults). ML algorithms (Ada Boost, Random Forest, Decision Tree, K-Nearest Neighbors, Gaussian Naïve Bayes,
Logistic Regression, Support Vector Machine, Linear Discriminant Analysis) are applied to feature-scaled datasets.
The classification outcomes are compared using various statistical measures, revealing the best-performing
classifiers and FS techniques for each age group. The experimental results highlight significant accuracy
achievements, with Voting classifier predicting ASD with the highest accuracy for Toddlers and for Children, while
Voting classifier achieves the highest accuracy of for Adolescents and for Adults. The project includes a detailed
feature importance analysis using four Feature Selection Techniques ,emphasizing the role of fine-tuning ML
methods in predicting ASD across different age groups and suggesting that the feature analysis can guide healthcare
practitioners in decision-making during ASD screenings. The proposed framework demonstrates promising results
compared to existing approaches for early ASD detection.The proposed algorithms has to enhance the robustness
and accuracy of ASD detection, an ensemble method using a Voting Classifier with Random Forest (RF) and
AdaBoost was applied, achieving a remarkable 100% accuracy.