Objective: To introduce an innovative method for distinguishing gait patterns in Parkinson’s Disease (PD) patients with medical or surgical treatment and healthy controls, by exploiting a dynamic- based mathematical tool, named recurrence plots (RPs).

Background: The efficacy of Deep Brain Stimulation (DBS) differs among individuals, and its impact on gait and balance continues to be an area of active investigation. Identifying characteristics through wearable sensors and deep learning (DL) approaches that reflect DBS-specific responses, is of outmost importance for achieving enhanced patient outcomes.

Method: Patients with PD diagnosis, Hoehn & Yahr (H&Y) stages 1-4, were recruited from the Movement Disorders Outpatient Clinics of two hospitals in Greece (University General Hospital of Heraklion and the University General Hospital of Patra). Healthy individuals were enrolled at the Hellenic Mediterranean University, Crete, Greece and served as healthy control group (A). Patients on dopaminergic substitution therapy (non-DBS) were classified as group B, while patients undergone surgical treatment (DBS) represented group C.  All participants executed a segment of the Smart- Insole Gait Assessment Protocol [1]. This study employs a Conditional Deep Convolutional Generative Adversarial Network (DC-GAN) for data augmentation, generating synthetic gait cycles conditioned on patient ID and DBS status to achieve class balance, while a Vision Transformer (ViT) model is employed for classification tasks, using RPs derived from pressure sensor data.

Results: In the study, 174 PD patients (61 women) with a mean age of 65.2±9.5 years were included and the DC-GAN’s ability to generate synthetic gait data was assessed. The results showed a strong resemblance between real and synthetic data, with minimal error metrics. In classification tasks, a ViT model with attention-based feature fusion achieved the highest accuracy (94.61%) in multi-class classification. For binary tasks, cross-attention-based ViT models performed best, especially in B vs. C classification. However, in A vs. C, precision and recall dropped, suggesting gait normalization in DBS patients.

Conclusion: Considering that DBS can profoundly impact motor function, a systematic and data-oriented method for categorizing DBS and non-DBS circumstances may yield critical insights into treatment outcomes, facilitating more tailored management methods for PD patients.

« Back to 2025 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/distinguishing-dbs-and-non-dbs-parkinsons-patients-using-recurrence-plots-and-vision-transformers/