Standardizing Video Capture and Clinical Workflows for Reliable Pose-Based Assessment of Parkinson’s Disease

Abstract

Parkinson’s disease is a movement disorder characterized by resting tremor, bradykinesia, rigidity, and postural instability. This disease impacts a large portion of the global population, specifically the elderly, and poses a significant burden to patients, families, and caregivers. The current diagnostic landscape relies heavily on the Movement Disorder Society-Sponsored Revision of the Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) which uses subjective clinical assessment to diagnose patients, resulting in suboptimal diagnostic accuracy, with only 76% of clinically diagnosed cases of Parkinson’s disease meeting the criteria for the disease upon autopsy (Tolosa et al., 2006). Current efforts to improve biomarker identification are often unreliable, invasive, and/or expensive, with the exception of computer vision based biomarker identification, which shows promise. There are a few significant barriers to research into computer vision based biomarkers in the present day research landscape. These barriers include lack of standardized capture conditions for viable video for analysis and a lack of a streamlined, clinically integratable video capture and de-identification process that medical providers can navigate in a timely manner. This thesis evaluates capture conditions for reliability, and streamlines this capture and de-identification procedure for clinical deployment. By using MediaPipe, a machine vision approach for markerless pose tracking, I was able to assess the impact of a variety of capture conditions on accuracy of keypoint placement. Capture conditions include brightness, camera stability, patient clothing worn, recording background, use of assistive mobility device, and reliability of capture when patient movement is atypical. A synchronized capture system was developed to coordinate video and wearable sensor data using a single board computer, reducing clinician burden and minimizing human error. Additionally, a reliable de-identification procedure was developed and validated. Overall, clinicians must ensure that their videos are well lit, the cameras they are using are stable, and the relevant body regions remain in frame and unobstructed in order to yield the highest quality video footage for reliable analysis. Background, patient clothing, and assistive mobility devices were not observed to significantly impact video viability. This work lays the foundation for the creation of standardized datasets, allowing for scalable clinical research that can one day lead to objective biomarker identification for Parkinson’s disease. Although this thesis uses Parkinson’s disease as a case study, the guidelines and workflows can also be implemented for other movement disorders.