Quantum-Inspired Variational Optimization of neuromorphic self-supervised heterogeneous graph transformers for multimodal Parkinson’s disease representation learning

Abstract

Early diagnosis of Parkinson’s disease (PD) remains challenging due to the heterogeneous nature of multimodal clinical data, limited annotated datasets, suboptimal multimodal fusion strategies, and the absence of uncertainty-aware decision-making in existing deep learning models. This paper proposes NeuroQ-MGST, a Neuromorphic Self-Supervised Multimodal Graph Transformer with Quantum-Inspired Variational Optimization, for robust and interpretable PD diagnosis. The proposed framework integrates a Spiking Neural Network (SNN) encoder to capture temporally sparse modality-specific representations, a self-supervised contrastive and masked reconstruction objective to exploit unlabeled data, and a Dynamic Heterogeneous Graph Construction (DHGC) with a Hierarchical Multimodal Graph Transformer (HMGT) to model complex cross-modal interactions among speech, gait, neuroimaging, and clinical biomarkers. To improve optimization in highly non-convex multimodal learning spaces, a Quantum-Inspired Variational Optimization (QIVO) module performs structured global representation refinement, while a Bayesian Monte Carlo Dropout layer provides calibrated predictive uncertainty for confidence-aware clinical decision support. Experimental evaluation on the PMMD 2.0 and UCI Parkinson’s Disease datasets demonstrates the effectiveness of the proposed framework, achieving 98.7% accuracy and 98.9% AUC-ROC, outperforming recent state-of-the-art multimodal approaches. Ablation studies validate the contribution of each architectural component, and cross-cohort evaluation on an independent European dataset achieves 95.8% zero-shot accuracy, confirming strong generalization capability. The proposed NeuroQ-MGST framework provides an accurate, interpretable, and uncertainty-aware solution for next-generation AI-assisted Parkinson’s disease diagnosis.