Machine Learning and Deep Learning for Autism Spectrum Disorder (ASD) detection

Abstract

Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder characterized by difficulties with social interaction, communication, and behavior. Early identification ensures timely intervention and improves quality of life. The present study proposes a two-modality binary classification approach for ASD detection based on facial image data and movement data. The image dataset consists of facial photos of children labeled as ASD or non-ASD. The movement data, captured using a Kinect v2 sensor, comprises 1,259 features per subject derived from 3D joint positions and gait metrics, with a total of 800 samples. A DenseNet121 convolutional neural network (CNN) was employed for images, achieving 89% accuracy. Grad-CAM was used to provide visual explanations by highlighting important regions in the images. For the movement modality, a Multi-Layer Perceptron (MLP) trained on features learned via an autoencoder achieved 99.38% accuracy, with Shapley Additive exPlanations (SHAP) applied to identify key features influencing model decisions. Finally, a late fusion mechanism combining both models was evaluated, resulting in 88.12% accuracy. Results highlight the effectiveness of unimodal solutions, particularly the MLP with autoencoder, and indicate that multimodal fusion requires further optimization for enhanced overall performance.