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

Area of Science:

• Computational neuroscience
• Quantum computing applications
• Medical diagnostics

Background:

• Parkinson's disease (PD) diagnosis is challenging due to subtle early symptoms and limitations of current assessment tools like the MDS-UPDRS scale.
• Existing computational approaches for PD detection are often unimodal, failing to integrate diverse data types effectively.

Purpose of the Study:

• To develop and evaluate a hybrid quantum-classical multimodal network for accurate Parkinson's disease classification.
• To leverage both visual-spatial (hand-drawn patterns) and structured clinical data for enhanced diagnostic performance.

Main Methods:

• A dual-track network combining a CNN-based visual encoder (TVSFE) with quantum circuits and a quantum variational circuit for structured data (VQFMN).
• The TVSFE track incorporates advanced attention mechanisms and quantum amplitude embedding, while the VQFMN track uses RY rotation gates and entangling layers.
• Outputs from both tracks are fused and classified using fully connected layers.

Main Results:

• The hybrid model achieved high performance on the HandPD test set, with 97.28% accuracy, 96.60% precision, 96.62% recall, and 96.54% F1-score.
• Consistent results were observed on the NewHandPD dataset (95.45% across metrics) and through five-fold cross-validation (mean accuracy of 96.58%).
• The quantum-classical fusion significantly outperformed unimodal and fully classical models.

Conclusions:

• The proposed quantum-classical network offers a promising, highly accurate approach for Parkinson's disease detection.
• The model's interpretability is enhanced through Grad-CAM and sensitivity analysis, identifying key visual and clinical features.
• This multimodal fusion strategy represents a significant advancement in computational approaches for neurodegenerative disease diagnosis.