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

Domain differences in fMRI analysis often cause biases that negatively affect Autism classification. To address this, we propose a novel pipeline leveraging a Domain Adversarial Neural Network (DANN) architecture to extract domain-invariant yet classification-informative features by integrating Mamba and Kolmogorov-Arnold Network (KAN) models. The DANN framework consists of an extractor, a domain classifier, and a label classifier, trained in an adversarial way to reduce domain bias while maintaining classification accuracy. The extractor employs two parallel paths: one processes fMRI time series with the Mamba model, and the other analyzes functional connectivity data using the KAN. The extracted features are concatenated and utilized by KAN-based domain and label classifiers. Adversarial training ensures the domain classifier cannot distinguish domain labels, confirming the domain invariance of the features. Experimental results show that this method achieves an accuracy of 72.56% and an AUC of 72.46%, demonstrating comparability to state-of-the-art methods which rely solely on fMRI data without utilizing phenotype information. Source code and implementation details are available at https://github.com/fatemehghanadi/fMRI-Based-Autism-Classification-Mamba-KAN-with-DANN.Clinical Relevance- The proposed approach effectively mitigates domain-induced biases and offers a robust solution for fMRI-based Autism classification tasks.