A deep learning model for diagnosing autism using brain time series.

The early identification of autism is especially critical as it can significantly enhance the effectiveness of intervention strategies. However, the recognition task remains challenging due to the subtle differences between ASD patients and neurotypical individuals. The presented approach leverages a hybrid model that combines Long Short-Term Memory (LSTM) networks with an Attention mechanism, enabling the extraction of both long-term and short-term features for more accurate autism diagnosis. Additionally, we integrate a residual block with channel Attention to enhance feature fusion and mitigate the risk of network degradation. Innovatively, we introduce a sliding window-based data preprocessing method alongside a voting strategy and validate the framework using subject-level 5-fold cross-validation to ensure generalizability across data splits. Our model was evaluated on the Region of Interest (ROI) time series dataset from the Autism Brain Imaging Data Exchange (ABIDE), achieving 73.1 % accuracy on the DOS brain atlas and 81.1 % on the HO brain atlas-outperforming baseline models. Moreover, we constructed brain functional connectivity topological structures for both ASD patients and healthy individuals, providing valuable resources for future autism research.