Using fMRI Time Series and Functional Connectivity for Autism Classification: Integrating Mamba and KAN in...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Functional Magnetic Resonance Imaging (fMRI) analysis for autism classification is often hindered by domain biases.
• These biases negatively impact the accuracy and reliability of diagnostic models.
• Existing methods may not sufficiently address the challenge of domain shift in neuroimaging data.

Purpose of the Study:

• To develop a novel pipeline for autism classification using fMRI data that mitigates domain-induced biases.
• To extract domain-invariant features that are still informative for classification.
• To enhance the robustness and generalizability of autism diagnostic models.

Main Methods:

• A Domain Adversarial Neural Network (DANN) architecture was employed, integrating Mamba for fMRI time series and Kolmogorov-Arnold Network (KAN) for functional connectivity.
• The DANN framework included an extractor, a domain classifier, and a label classifier, trained adversarially.
• Parallel processing paths within the extractor utilized Mamba and KAN, with features concatenated for classification.

Main Results:

• The proposed method achieved an accuracy of 72.56% and an Area Under the Curve (AUC) of 72.46%.
• Experimental results demonstrated comparability to state-of-the-art methods that do not utilize phenotype information.
• Adversarial training successfully ensured domain invariance of the extracted features.

Conclusions:

• The novel pipeline effectively reduces domain-induced biases in fMRI-based autism classification.
• The integration of Mamba and KAN within a DANN framework offers a robust solution.
• This approach shows significant promise for improving the clinical relevance and accuracy of autism diagnosis using neuroimaging data.