Pre-diagnosis for Autism Spectrum Disorder Using Eye-Tracking and Machine Learning Techniques

This study explores the potential of utilizing machine learning in conjunction with gaze-tracking data to facilitate early or pre-diagnosis of ASD which can be cost-effective and beneficial to people with limited access to healthcare resources. A dataset comprising gaze-tracking information mapped onto images to differentiate between control subjects and autistic individuals is utilized and treated as an image classification problem. Two machine learning frameworks were employed for model training and testing: (1) a fast approach using principal component analysis (PCA) on the images followed by conventional machine learning algorithms such as ANN, Decision Tree, and support vector machines (SVM), which yielded an accuracy of 78% and an AUC of 0.82; and (2) a deep learning approach that involved a custom convolutional neural network (CNN) model, achieving an accuracy of 92% and an AUC of 0.96. Several transfer learning models were also evaluated, with the ResNet50 model providing the best results (accuracy: 0.86, AUC: 0.94). These findings demonstrate the viability of these methods for the pre-diagnosis of autism.