Discover ComputingSource publication:
Loading content…
Parkinson's disease diagnosis still rests on symptoms and expert assessment, while multimodal clinical data are heterogeneous, labels are scarce, and deep models rarely report calibrated confidence. NeuroQ-MGST encodes voice, gait, neuroimaging and clinical variables with spiking neural networks, aligns them through self-supervised contrastive and masked reconstruction objectives, reasons over a dynamically rebuilt heterogeneous graph with a hierarchical graph transformer, and adds quantum-inspired variational refinement plus Monte Carlo dropout uncertainty. On the PMMD 2.0 test set it reaches 98.7% accuracy and 0.989 AUC-ROC, and 95.8% zero-shot accuracy on an external European cohort without fine-tuning.
Multimodal Parkinson's diagnosis becomes one end-to-end pipeline that returns a diagnostic class together with a calibrated predictive uncertainty, rather than a class label alone. Earlier multimodal frameworks typically concatenate features before a classifier, which captures only additive contributions of each modality and misses cross-modal second-order dependencies. On the PMMD 2.0 test set the pipeline reaches 98.7% accuracy and 0.989 AUC-ROC against 14 re-implemented baselines; the strongest comparator, QIML-PD, reaches 95.4%, with McNemar p<0.001.
Each modality is first encoded into a sparse temporal representation by a spiking neural network, then aligned across modalities on unlabeled data through contrastive and masked reconstruction objectives. Limited labeled data is a standing constraint in multimodal Parkinson's modeling, and the self-supervised objective lets unlabeled subjects contribute to representation learning. Removing contrastive pretraining lowers accuracy by 2.1% and removing the spiking encoder by 1.2%; the spiking encoder uses about 31% less inference energy than a CNN encoder.
The aligned representations form a dynamic heterogeneous graph of subject, modality, feature and concept nodes, and a hierarchical graph transformer performs typed attention over it. Static graphs or plain concatenation cannot rebuild the relational structure per patient input, whereas dynamic construction makes relations a function of each sample. Replacing the hierarchical graph transformer with MLP fusion lowers accuracy by 3.4%, and substituting a static k-nearest-neighbor graph drops it to 96.1%.
Quantum-inspired variational optimization refines the representation on a fully classical computation graph through phase modulation and Hadamard-structured mixing layers, while Monte Carlo dropout supplies predictive variance. First-order gradient descent is confined to directions set by the local gradient, and the mixing layer lets updates leave that local descent trajectory. Removing the optimization layer lowers accuracy by 1.0%; it converges 16 epochs faster than a 4-layer MLP projection (43 versus 59), and calibration error ECE is 0.024 versus 0.089 for an uncalibrated transformer.
Results cover 1,847 subjects in PMMD 2.0, leave-one-subject-out cross-validation on 31 subjects in the UCI Parkinson's voice dataset, and zero-shot evaluation on an external European cohort of 312 subjects. The intended reader is a researcher working on multimodal clinical decision support or graph neural networks in medicine. All reported results are retrospective and dataset-dependent, and the authors state that multicenter prospective validation with ethics approval is required before clinical deployment.
The authors note that 1,847 subjects in PMMD 2.0 remain limited for high-capacity multimodal architectures and that larger, longitudinal, multi-institution data are needed. The quantum-inspired optimization layer adds trainable parameters and hyperparameters, and rotation-angle initialization may affect convergence stability, so its convergence rate and scalability still need systematic analysis. Missing-modality experiments provide only initial evidence, and systematic modality absence still requires prospective study or domain adaptation.