A quantum-classical dual-track deep learning network for explainable Parkinson's disease classification

Parkinson's disease (PD) is a progressive neurodegenerative disorder caused by the loss of dopaminergic neurons in the substantia nigra, presenting with motor and non-motor symptoms in roughly 2-3% of the global population above age 60. Early detection is difficult because symptoms are subtle and often indistinguishable from normal aging. The MDS-UPDRS rating scale, the current clinical standard, is time-intensive, subjective, and requires experienced clinicians. Most computational approaches are unimodal and do not use image and structured clinical data in combination.

We propose a hybrid quantum-classical dual-track multimodal network that classifies PD patients and healthy controls from hand-drawn spiral and meander patterns. The first track, the Topological Visual-Spatial Feature Encoder Network (TVSFE), uses a ghost module-based CNN with Cross-Dimensional Attention Bottleneck (CDAB) blocks incorporating coordinate attention, squeeze-and-excitation, and triplet attention, followed by a quantum variational circuit with amplitude embedding. The second track, the Variational Quantum Feature Mapping Network (VQFMN), encodes structured clinical and demographic data through RY rotation gates and strongly entangling layers. Outputs from both tracks are concatenated and passed through fully connected layers for classification.

On the HandPD test set, the model achieved 97.28% accuracy, 96.60% precision, 96.62% recall, and 96.54% F1-score, outperforming all CNN, transformer, and ML-based baselines compared. Five-fold cross-validation produced a mean accuracy of 96.58%. On the NewHandPD dataset, accuracy, precision, recall, and F1-score were all 95.45%.

The quantum-classical fusion outperforms both single-modality and fully classical variants. Grad-CAM localizes the spatial image regions driving classification and the perturbation-based sensitivity analysis identifies Root Mean Square (RMS) and age as the most influential structured features. Both together make the model's reasoning traceable at the modality level, which is important for decision-making.