Bioactivity-Enhanced 3D FDG-PET Metabolic Network Modeling With Transformer Architecture for Parkinsonian Syndrome Subtyping

We propose a scalable multi-modal feature fusion framework for analyzing heterogeneous neuroimaging and bioactivity data, enabling hierarchical biomarker modeling across diverse patient cohorts. Designed for clinical decision support, our system integrates multi-scale biological signalsąłincluding metabolic (FDG-PET), molecular (bioactivity), and structural featuresąłthrough transformer-structured learning to uncover coherent, pathology-aligned patient subgroups. The pipeline extracts discriminative network features, such as regional metabolic patterns and ligand-receptor interactions, enhanced via weakly supervised learning to ensure comprehensive disease characterization. A manifold-preserved feature selector filters non-discriminative biological attributes while retaining the intrinsic geometric structure of the neuropathological space, producing refined patient representations optimized for diagnostic precision. These embeddings are projected into a latent neuropathological space, where each patient is modeled as a probabilistic distribution over abstract disease subtypes, enabling fine-grained stratification of Parkinsonian syndromes (e.g., PD vs. atypical parkinsonism). The resulting bioactivity-metabolic network captures both localized dysfunction (e.g., nigrostriatal degeneration) and global system-level perturbations through multi-scale feature fusion.We construct a patient similarity graph from these fused representations, enabling large-scale subgroup detection via spectral clustering on the network manifold. The identified clusters reflect shared pathophysiological dimensions (e.g., dopaminergic vs. cholinergic dominance) and reveal hierarchical progression patterns. This approach demonstrates superior interpretability compared to conventional radiomic or clinical scoring systems. For personalized diagnosis, our classification module integrates individual patient profiles with cluster-level embeddings through attention-based feature fusion. This dual-scale integrationąłcombining localized bioactivity signatures with global network trendsąłenhances diagnostic accuracy while supporting subtype-specific therapeutic planning. Evaluations on 1,365 patients show over 12% higher sensitivity in differentiating PD from mimics compared to single-modality models. The frameworkąŕs robustness is validated across disease stages, with consistent performance (F1-score 0.82) for both early- and late-stage cohorts, confirming its suitability for precision neurology applications requiring fused multi-modal intelligence.