stonelight816/Interpretable-GAT-aided-rs-fMRI-Analysis-and-LLM-based-Multi-Modal-Classification-for-Schizophrenia: v1.0.0: Schizophrenia Multimodal Dataset and Code
Overview This release provides the complete open-source codebase and curated multimodal neuroimaging, eye-tracking, and cognitive assessment dataset supporting the multimodal diagnostic framework for schizophrenia detection described in the associated research paper. All The graph network models and LLM multimodal fusion modules are fully reproducible. Dataset Assets Curated multimodal cohort data from 63 participants: 32 schizophrenia (SZ) patients and 31 healthy control (HC) subjects Functional connectivity of whole-brain based on resting-state functional MRI (rs-fMRI) scans Recordings from three independent eye-tracking experimental paradigms Standardized MCCB cognitive battery quantitative scores Core Model Implementations 2.1 Single-modal brain network model: Improved Graph Attention Network (I-GAT) Custom I-GAT architecture optimized for encoding fMRI functional connectivity and spectral brain features Supports localization of disease-relevant connections across visual, salience, and limbic brain subnetworks Benchmark reproduction script: single-modal classification reaches 87.38% accuracy, exceeding mainstream baseline graph neural networks 2.2 Multimodal fusion framework: DeepSeek-MMC End-to-end multimodal integration pipeline built upon pre-trained large language models (LLMs) Embedded token projection modules to unify heterogeneous brain, eye movement, and cognitive score feature spaces Full multimodal model classification accuracy: 93.8% on the released patient-control dataset