Abstract
Brain functional connectivity has shown promise for developing objective biomarkers for autism spectrum disorder (ASD). Although many imaging studies have demonstrated its potential, most have focused on static measurements. In this study, we explored the dynamic changes in functional connectivity over time to uncover potential temporal dependencies. These dynamic patterns were abstracted into high-level representations and used as predictors to identify individuals at risk of ASD. To achieve this, we employed a deep learning framework that combines attention mechanism with long short-term memory (LSTM) neural network. Experiments were conducted using heterogeneous resting-state functional magnetic resonance imaging (rs-fMRI) data from the Autism Brain Imaging Data Exchange (ABIDE) database. The resulting classification achieved an accuracy of 74.9% and precision of 75.5% under intra-site cross-validation, outperforming traditional classifiers such as support vector machines (SVM), random forests (RF), and single LSTM network. Further analyses demonstrated the robustness and generalizability of our model, with classification performance less affected by subjects' gender or age. The optimal model's weights revealed atypical temporal dependencies in the brain functional connectivity of individuals with ASD, highlighting the potential for these patterns to serve as biomarkers. Our findings underscore the importance of dynamic functional connectivity in understanding ASD and suggest that our deep learning framework could aid in the development of more accurate and reliable diagnostic tools for this disorder.