Integrating Graph Convolutional Networks (GCNNs) and Long Short-Term Memory (LSTM) for Efficient Diagnosis of Autism

Abstract

Artificial intelligence (AI) has a wide range of practices in biotechnology, specifically for automated diagnosis of behavioural disorders, including autism spectrum disorder (ASD). With the rise and severity of this disorder, machine learning and deep learning methods have proven to provide efficient and less invasive diagnosis for individuals with ASD. Although various deep learning techniques have been employed to achieve a more robust diagnosis for ASD, hybrid graph convolutional neural networks (GCNNs) for ASD diagnosis are not addressed prominently in the literature; GCNNs have received significant adoption in image processing. This paper proposes and evaluates a hybrid deep learning model that combines the power of GCNNs and long short-term memory (LSTM) for ASD diagnosis. The proposed GCNN-LSTM model provides an efficient diagnosis for ASD by identifying the brain functionality between anterior and posterior areas of the brain. The proposed GCNN-LSTM model and other baseline classifiers are trained and tested on various ASD-related fMRI brain images from the Autism Brain Imaging Data Exchange (ABIDE) database. Experimental results show that the proposed model achieved an accuracy of up to 75%, AUC up to 80%, Precision up to 82%, Recall of up to 85%, and F1-score of up to 83%, thus outperforming the baseline classifiers.