Automated Recognition of Autism Spectrum Disorder from EEG Signals Using a CNN-LSTM Hybrid Model

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental disorder that impacts social interaction and behavior. Accurate early diagnosis is critical for timely intervention, yet current diagnostic methods are often lengthy and influenced by subjective interpretation. This research introduces an automated ASD detection framework based on electroencephalogram (EEG) data analysis using a Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) model. In this approach, the CNN component identifies spatial patterns in the EEG signals, while the LSTM component detects temporal relationships, providing a comprehensive basis for ASD classification. The CNN-LSTM architecture is designed to process raw EEG signals without additional preprocessing, employing distinct temporal and spatial filters focused on motor cortex electrode pairs. A 5-layer CNN paired with a 4-layer max pooling structure optimizes dimensionality reduction, followed by a fully connected (FC) layer for final classification. To mitigate overfitting, dropout and batch normalization techniques are applied. This model was trained and validated using EEG data from 5, 10, 20, and 50 subjects, obtained from the Physionet repository and recorded with a 14-channel Emotiv EPOC EEG device. This study incorporated varying cognitive states such as focused, unfocused, and drowsy. Experimental results indicate a group-level classification accuracy averaging 94.13%, an ROC area of 0.971, and a peak accuracy of 99.41% for the FC3-FC4 electrode pair on a 10-subject dataset. This automated system demonstrates promising support for clinicians in diagnosing ASD, potentially streamlining the diagnostic process and minimizing subjective biases.