The study converts pressure-sensor gait data into recurrence plots and applies Vision Transformer fusion models with DC-GAN augmentation to classify PD patients across treatment states and controls, achieving up to 94.58% multi-class accuracy.

Offers a noninvasive, data-driven gait biomarker and analytic pipeline that could help monitor and stratify patient treatment responses in PD, but provides little mechanistic or therapeutic target information for drug discovery.

Goal: This study aims to develop an innovative gait analysis framework using recurrence plots (RPs) to differentiate gait patterns between Parkinson's disease (PD) patients under varying treatment regimes and healthy individuals. Methods: Pressure sensor data were transformed into RPs and analyzed using a Vision Transformer (ViT) model with multiple fusion strategies. To address class imbalance, a conditional Deep Convolutional Generative Adversarial Network (DC-GAN) was employed to generate synthetic gait data. Four ViT-based fusion architectures were investigated and evaluated across multi-class and binary classification tasks. Results: The dual ViT stream with late fusion achieved the highest accuracy in multi-class classification (94.58%), while the cross-attention fusion model outperformed others in binary classification tasks. Conclusions: The findings indicate that gait characteristics captured via RPs can effectively distinguish between PD patients under different treatments and healthy controls. This approach provides a data-driven pathway for objective and individualized assessment of PD therapies, potentially supporting improved clinical decision-making.