Reservoir Computing with Spatial Filtering and Manifold Learning for fMRI Classification

Artem Badarin, Nikita Kulagin, Andrey Andreev, Kurkin Semyon, Alexander E. Hramov

Keywords: Reservoir Computing, Common Spatial Patterns, UMAP, fMRI, Classification, Interpretability

Abstract: We introduce a parametric framework that couples discriminative spatial filtering with reservoir computing to distinguish spatiotemporal structure in resting-state fMRI in two classes. Temporal dependencies are encoded in a reservoir, while supervised spatial filtering on reservoir states isolates condition-specific patterns; parametric Uniform Manifold Approximation and Projection (UMAP) then yields compact nonlinear embeddings fit on training data and evaluated with cross-subject validation. On 163 participants (97 healthy controls, 66 major depressive disorder), the method reaches 87\% accuracy, outperforming network-feature pipelines using LDA, SVM, kNN, and GNN. The framework also generalizes to autism spectrum disorder classification, achieving competitive accuracy on the ABIDE (NYU) benchmark and ranking among top state-of-the-art methods. Interpretability combines spatial-pattern maps with Shapley-value attribution, providing coherent, region-level explanations that consistently implicate cortical and subcortical areas associated with both major depressive disorder and autism spectrum disorder. The framework offers an interpretable route to modeling spatiotemporal organization in clinical and cognitive fMRI.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 14280