Fusion of Convolutional and Recurrent Networks for Autism Detection from EEG Signals

Abstract

Autism Spectrum Disorder (ASD) is a multifaceted neurodevelopmental condition that affects communication, behavior, and social interaction. Early and accurate diagnosis is essential for effective intervention, yet existing clinical assessments are often time-consuming and subject to human interpretation. This study presents a novel deep learning framework for automated ASD recognition using electroencephalogram (EEG) signals, integrating a hybrid Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM) model. The proposed architecture leverages CNNs to extract spatial patterns from EEG data and LSTM networks to capture temporal dynamics, enabling robust end-to-end classification without extensive preprocessing. The model architecture features a 5-layer CNN followed by a 4-layer max-pooling structure for feature abstraction, concluding with a fully connected layer for final decision-making. To enhance generalization, dropout regularization and batch normalization are employed. EEG data collected via a 14-channel Emotiv EPOC device, encompassing cognitive states such as focused, unfocused, and drowsy, were sourced from the PhysioNet repository. Datasets involving 5, 10, 20, and 50 subjects were used to evaluate model performance. Results reveal an average classification accuracy of 94.13%, an ROC-AUC of 0.971, and a peak accuracy of 99.41% using the FC3-FC4 electrode pair on the 10-subject dataset. These findings highlight the model's potential to augment clinical workflows by enabling rapid, objective ASD screening using non-invasive EEG recordings.