Prediction and Evaluation of Autism Spectrum Disorder using AI-enabled Convolutional Neural Network and Transfer Learning : An Ensemble Approach

Recent research in Artificial Intelligence (AI) in diagnosing Autism Spectrum Disorders (ASD) helps autistic people to improve their social, communicational and emotional skills. Diagnosis of ASD through enhanced models provides a reality-based therapy for promising improvements in autism disorders. This research paper proposes a novel approach for autism detection using ensemble learning with AI-enabled convolutional neural networks (CNNs). The proposed approach focuses on evaluating risk factors and disorders through an ensemble approach with transfer learning schemes. Our method leverages transfer learning from pre-trained CNN models, including EfficientNet B5, MobileNet, and InceptionV3, trained on the ImageNet dataset. We fine-tune these models for autism detection by improvising their architectures, freezing convolutional layers, and adding new fully connected layers for binary classification problems. Through extensive experimentation, the demonstration was made with a keen focus on each CNN model, which then improved accuracy in identifying autistic traits from image data. Furthermore, by combining the predictions of these models using a soft voting ensemble technique, they achieve superior performance with an accuracy of about 91% respectively. Our results indicate the effectiveness of ensemble learning in improving autism detection accuracy, showcasing the potential of deep learning approaches in aiding clinical diagnosis and treatment planning for autism spectrum disorders (ASDs).