Objective Autism Spectrum Disorder (ASD) is a complex neurodevelopmental disorder with dynamic brain dysfunctions that static functional connectivity often fails to capture. To better exploit temporal information in resting-state fMRI (rs-fMRI), this study proposes a novel diagnostic framework based on dynamic functional connectivity (DFC) state transition sequences and deep temporal modeling. Methods Sliding windows are applied to rs-fMRI time series to construct DFC matrices. Each window is represented using fused features from an autoencoder and seven statistical descriptors. These features are clustered into discrete brain states, forming subject-specific state transition sequences. A circular shift-based augmentation strategy is adopted to mitigate data scarcity. Finally, a sequential model based on Mamba—a structured state space model—is employed to capture long-range dependencies and classify the sequences. Results On the Autism Brain Imaging Data Exchange (ABIDE) site dataset, the proposed Mamba-FCN achieves 79.7% accuracy and 81.4% F1-score under 10 × 10-fold cross-validation, outperforming baselines such as Support Vector Machine (SVM), Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM). Ablation studies confirm the contributions of Mamba, feature fusion, and data augmentation. The method also shows strong cross-site generalization across 15 imaging centers, with an average accuracy of 76.6%. Conclusion This work presents an effective ASD diagnostic model that integrates dynamic brain connectivity patterns and advanced sequence modeling. The proposed framework demonstrates strong discriminative and generalization capabilities, even with limited data. Future work will explore generative data augmentation and broader applications to other brain disorders.
Chen et al. (Fri,) studied this question.