Implementation of Vision Transformer for Early Detection of Autism Based on EEG Signal Heatmap Visualization

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental disorder characterized by difficulties in social interaction, communication, and repetitive behavioral patterns. Early detection of ASD is crucial for improving the quality of life of affected individuals and alleviating the burden on their families. This study proposes a computer-aided diagnostic system for ASD by applying a pre-trained Vision Transformer (ViT-B/16) architecture to EEG signal data obtained from King Abdul Aziz University. The dataset comprises EEG recordings from 16 subjects (8 normal and 8 ASD) that have undergone preprocessing—including filtering using the Discrete Wavelet Transform (DWT), segmentation (windowing), and conversion into heatmap representations—and were subsequently partitioned into training, validation, and testing subsets. The ViT model was trained for 100 epochs with a batch size of 16, using the AdamW optimizer and the CrossEntropy loss function, while two learning rate configurations (0.0001 and 0.00001) were evaluated; the best-performing weights were selected based on the lowest validation loss. Test results indicate that the model trained with a learning rate of 0.00001 achieved a testing accuracy of 99.53%, accompanied by excellent precision, specificity, recall, and f1-score, thereby demonstrating strong generalization capabilities and minimal overfitting. Future research is recommended to incorporate locally sourced datasets and to further customize the ViT architecture through comprehensive hyperparameter tuning, with the aim of developing a mobile application to support clinical ASD diagnosis.