Loading...
Integrating motion physics knowledge into deep learning for accurate Parkinson's disease classification
Kishore Ponnada, Ajay
Kishore Ponnada, Ajay
Files
Loading...
Kishore Ponnada_2025.pdf
Adobe PDF, 25.35 KB
Authors
Other Names
Location
Time Period
Advisors
Original Date
Digitization Date
Issue Date
2025-04-11
Type
Abstract
Genre
Keywords
Subjects (LCSH)
Citation
Kishore Ponnada, A. 2025. Integrating motion physics knowledge into deep learning for accurate Parkinson's disease classification. -- In Proceedings: 21st Annual Symposium on Graduate Research and Scholarly Projects. Wichita, KS: Wichita State University
Abstract
Parkinson’s Disease (PD) is a progressive neurological disorder characterized by a variety of motor and non-motor symptoms, significantly impacting patients’ quality of life. Early and accurate detection of PD is crucial for timely interventions and improved patient outcomes. Recent advancements have highlighted the potential of time-series data collected from Inertial Measurement Unit (IMU) sensors during patient-performed activities as a valuable source for PD detection. Previous studies have employed diverse data collection methodologies, ranging from single-device setups attached to the patient’s waist to multi-sensor configurations attached to various body segments. However, these approaches often face limitations due to missing modalities or complex sensor setups. To address these challenges, this research proposes a novel hybrid framework for PD detection and classification. The designed Physics-Informed Neural Network (PINN) model integrates kinematics formulas with neural networks to approximate kinematic parameters such as relative rotational angles between body segments. Then, the proposed Bayesian Neural Network (BNN) model classifies the severities of PD affected levels. The proposed method also focuses on data collected using IMU sensors strategically placed on the upper body—specifically the lower and upper arms on both sides, as well as the chest—to achieve simplified yet efficient data acquisition. Given the limited availability of clinical datasets, this study incorporates an extensive data augmentation pipeline. Techniques such as noise injection, Fourier-based transformations, time-warping, and scaling are applied to enrich the dataset, improving the generalization and robustness of the neural network models. By leveraging IMU sensor data captured during multiple activity performances, this work developed a scalable and practical solution for PD detection that reduces dependency on complex multi-modal systems. The proposed hybrid framework not only improves classification accuracy but also addresses critical challenges in current research, paving the way for more accessible and reliable PD monitoring solutions.
Table of Contents
Description
Presented to the 21st Annual Symposium on Graduate Research and Scholarly Projects (GRASP) held at the Rhatigan Student Center, Wichita State University, April 11, 2025.
Research completed in the School of Computing, College of Engineering.
Research completed in the School of Computing, College of Engineering.
Publisher
Wichita State University
Journal
Book Title
Series
GRASP
v. 21
v. 21