Detecting Autism Spectrum Disorders from Resting-State fMRI in Young Children Using Bidirectional Long-Short Term Memory Neural Networks

Autism Spectrum Disorder (ASD) is a neurodevelopmental condition characterized by social communication deficits and repetitive behaviors, with early diagnosis being critical for effective intervention. Traditional diagnostic methods rely on subjective behavioral assessments, often delaying identification. Recent advances in neuroimaging, particularly resting-state functional magnetic resonance imaging (rs-fMRI), offer a promising avenue for objective and early ASD detection by capturing atypical functional connectivity patterns in the brain. This study leverages the temporal dynamics of rs-fMRI data to classify ASD in young children aged 5–10 years using Bidirectional Long Short-Term Memory (BiLSTM) neural networks. The model processes rs-fMRI time-series bidirectionally, capturing both past and future contextual information to identify ASD-related connectivity alterations. Training and testing were conducted on the publicly available Autism Brain Imaging Data Exchange (ABIDE) I dataset, with performance evaluated through stratified 10-fold cross-validation. The proposed BiLSTM model achieved an accuracy of 76.91%, outperforming comparable Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Convolutional Neural Network (CNN) architectures. Sensitivity and specificity were 75.31% and 78.35%, respectively, highlighting the model's balanced performance in identifying both ASD and typically developing controls. These results underscore the importance of temporal modeling in rs-fMRI analysis, as spatial methods like CNNs yielded lower accuracy (71.88%). The study demonstrates the potential of BiLSTMs for pediatric ASD classification while acknowledging the need for further refinement to bridge the gap with higher-performing hybrid models reported in the literature.