AI-Driven Visual Screening Tool for Early Detection of Autism Using Facial Cues

Abstract

Autism spectrum disorder (ASD) is a neurodevelopmental disorder distinguished by an extensive range of symptoms, including reduced social interaction, communication difficulties and tiresome behaviors. Early detection of ASD is important because it allows for timely intervention, which significantly improves developmental, behavioral, and communicative outcomes in children. Early diagnosis of autism spectrum disorder (ASD) is critical for effective intervention, and this system aims to provide a non-invasive, objective screening method. The proposed project presents an intelligent and automated system for early autism detection using facial image analysis, addressing the growing need for accessible and accurate diagnostic support tools. Traditional diagnostic methods rely heavily on behavioral assessments, which are time-consuming, subjective, and require expert intervention, highlighting the necessity for a reliable AI-based solution. In this work, a transfer learning approach is employed using Residual Neural Network with 50 layers (ResNet50) to extract deep and discriminative facial features from input images. These features are then utilized to train multiple machine learning classifiers, including Gaussian Naïve Bayes (GNB), Decision Tree Classifier (DTC), and the proposed hybrid model DeepRes-FusionRFC, which combines ResNet50-based feature extraction with Random Forest Classifier (RFC) for enhanced performance. The system follows a structured pipeline comprising dataset acquisition, preprocessing, feature extraction, model training, evaluation, and real-time prediction within a secure graphical user interface (GUI) environment with role-based access control. Experimental results demonstrate that while GNB and DTC provide moderate classification performance, the DeepRes-FusionRFC model significantly outperforms them in terms of accuracy, precision, recall, and F1-score, indicating its robustness and effectiveness in capturing complex facial patterns associated with autism. The integration of deep learning with ensemble machine learning not only improves prediction accuracy but also ensures scalability and efficiency. The proposed system offers a reliable, user-friendly, and high-performance solution for early autism detection, contributing to the advancement of AI-driven healthcare diagnostics.