KSNet: Advancing autism prediction via KAN-based graph convolution and multi-source data fusion.

Integrating functional brain networks and Graph Convolutional Networks (GCN) architecture reveal promising diagnostic capabilities for brain disorders. Prevailing approaches remain constrained when capturing dynamic nonlinearities, characterizing intricate connectomes, and fusing heterogeneous data sources. To address these limitations, we present KSNet, a novel architectural grounded in graph convolutional networks. Our design synergistically combines KAN-derived graph convolutions (KANGCN) with subject-relational convolutions (SRGCN). The KANGCN module employs B-spline functions to enable tractable optimization of high-dimension nonlinear neurography manifolds. We further implement selective node retention to amplify biologically relevant regions and attenuate noisy connections. SRGCN integrates phenotypic features, such as age and gender, with functional connectivity representations derived from KANGCN to construct an inter-subject relationship graph. The model employs a trainable hierarchical architecture that adaptively aggregates multi-scale features, thereby enhancing its discriminative power. Experimental results using the AAL116 and CC200 brain atlases demonstrate notable improvements across key performance metrics. Specifically, KSNet achieved a 13.8% increase in classification accuracy and a 12.4% improvement in the Area Under the Receiver Operating Characteristic Curve (AUC). Remarkably, the biomarkers identified by KSNet align closely with those reported in existing ASD research literature, offering new biologically interpretable support for clinical ASD diagnosis.