Quantification and Diagnosis of Mobility Deficits

Background and Research Motivation

Parkinson’s Disease (PD) is a neurodegenerative disorder primarily affecting patients’ motor abilities, leading to tremors, bradykinesia, limb rigidity, and problems with gait and balance. These motor deficits significantly impact patients’ ability to live independently and their quality of life. Statistics predict that by 2030, nearly 1.2 million people in the United States alone will have Parkinson’s Disease, with the global patient count exceeding 10 million. Therefore, accurately assessing and diagnosing the motor deficits in patients is a critical issue that needs urgent resolution.

Current severity assessment methods for PD mainly rely on clinical observations and physician experience, where patients are evaluated based on performing specific tasks in a laboratory or clinical setting. This method is influenced by subjective factors and does not adequately reflect patients’ actual motor conditions in their daily lives. Hence, researchers require a reliable, non-invasive quantitative method to objectively assess the motor deficits in Parkinson’s Disease patients, thereby providing more timely and effective rehabilitation feedback.

Research Source and Publication Information

This paper was authored by Fujian Yan, Jiaqi Gong, Qiang Zhang, and Hongsheng He and has been accepted by IEEE Transactions on Biomedical Engineering. Fujian Yan is affiliated with the School of Computing at Wichita State University, while Jiaqi Gong, Qiang Zhang, and Hongsheng He are affiliated with the Department of Computer Science and Mechanical Engineering at The University of Alabama. This research was supported by NSF grants #2327313 and #2129113.

Research Methods and Experimental Process

This paper proposes a method to collect motion data through non-invasive wearable physiological biosensors and then use deep learning models to analyze and quantify the motor deficits in Parkinson’s Disease patients. This method includes several key steps and technical innovations.

Experimental Process

The main steps of the research are as follows:

1. Data Collection:
   • Laboratory Data: Data were collected from 10 healthy subjects performing six activities (such as walking, squatting, picking up objects, drawing, yoga, and playing with block toys), using sensors including accelerometers and electromyography (EMG) sensors installed on limbs and torso.
   • Clinical Data: Data were collected from 8 PD patients and healthy individuals performing 8 actions.
2. Data Preprocessing:
   • Resampling, gravity compensation, and normalization of data, using a Butterworth high-pass filter to remove gravity components.
   • Data slicing to ensure consistent input shapes for the model.
3. Learning Motion Primitives:
   • Discovering representative motion primitives through multi-layer neural networks and dictionary learning methods.
4. Classification and Quantification:
   • Designing and training a deep convolutional neural network (CNN) model to quantify the severity of motor deficits in patients by establishing a Bag of Motion Primitives (BoMP).

Sample and Processing Details

• Laboratory Data involved healthy subjects performing activities in both normal conditions and conditions simulated to mimic PD by wearing weighted devices. Data included two separate collections for each activity for every subject.
• Clinical Data included PD patients completing a series of specific movements, with data sliced into equal lengths to ensure consistency when input into the model.

Technical Innovation

A significant innovation of this paper is the use of non-invasive sensor data in motion analysis to achieve efficient representation and classification of the original motion time series through dictionary learning and a Bag of Motion Primitives model.

Main Research Results

The results split into laboratory data and clinical data:

• Laboratory Data:
  • Classification accuracy for laboratory data reached 93.95%, significantly higher than linear models (64.86%) and support vector machines (91.35%).
  • Motion primitives obtained through dictionary learning successfully reconstructed and identified differences between healthy and simulated patient states.
• Clinical Data:
  • Classification accuracy reached 99.84%, effectively distinguishing between healthy individuals, mild PD, and severe PD patients.
  • The confusion matrix showed that most mild patient data (97.18%) were correctly classified, though data imbalance led to some misclassifications.

Conclusion and Value

This paper demonstrates the presence of significant motion primitives in simple motion measurement data and shows their effectiveness in quantifying motor deficits. This method performs with higher accuracy when predicting known patient data and offers the potential for real-time diagnosis and quantification for unknown patients. By leveraging large-scale motion tracking data, a user-specific quantification index based on the number of motion primitives can entirely replace observation-based clinical scores, providing real-time reports during rehabilitation, thereby enhancing patients’ positive recovery outcomes. This method not only improves the accuracy of PD symptom classification but also offers new approaches for other conditions requiring motion tracking and assessment.