Intelligent assessment method for mental disordersby integrating EEG and MRI multimodalhigh-dimensional data

To overcome the inherent limitations of existing diagnostic frameworks in multimodal clinical analysis, we propose a highcapacity, cross-modal diagnostic system grounded in a novel dual-space neural architecture. This system introduces two coreinnovations: the Hierarchically-Aligned Dual-Space Transformer (HADST) and a strategic training protocol termed ContextuallyRegularized Reverse Consistency Training (CRRCT). HADST is designed to significantly enhance cross-modality fusioncapabilities by independently encoding EEG time-series data and MRI spatial structures through a dual attention mechanism.This mechanism enables a rich and dynamic interaction between syntactic hierarchy and semantic representation layers,allowing the model to capture nuanced spatiotemporal dependencies across modalities. As a result, the system preservesspatial integrity and temporal coherence, both of which are essential for identifying subtle yet clinically meaningful latentbiomarkers associated with complex mental health conditions such as depression, schizophrenia, and bipolar disorder. Inparallel, CRRCT serves as a robust learning strategy that strengthens the model’s generalization ability and resilience tocommon data issues in clinical domains. By incorporating bidirectional consistency constraints and domain-aware trainingroutines, CRRCT addresses challenges such as exposure bias, inter-modality noise perturbation, and distributional shiftsacross patient demographics or acquisition protocols. Furthermore, the reverse consistency component ensures semanticalignment between forward and backward reconstructions of modality-specific representations, enhancing interpretability andstability. Together, HADST and CRRCT form a comprehensive and adaptable framework tailored to the high variability andheterogeneity characteristic of clinical neuroimaging data, paving the way for more precise, interpretable, and generalizablediagnostics in computational psychiatry.