An Artificial Intelligence Video-Based Assessment of Upper-limb Bradykinesia in Parkinson’s Disease

Abstract

As there is a lack of an objective biomarker for Parkinson’s disease (PD), the clinical practice relies on serial neurologic assessments to track and characterize disease progression. These clinical scales lack the resolution to detect subtle changes, classify the movements into broad categories, require significant expertise for proper application, and are not sensible enough for early diagnosis; imaging markers are objective and sensible for early diagnosis but are costly and not commonly available. The video-based movement measures estimated using AI algorithms for motion tracking represent a novel alternative for diagnosing and assessing PD. Video-based measures are objective, sensitive to small changes, inexpensive, and easily accessible. This paper explores the correlation between objective kinematics measures derived from videos of the Finger Tapping Test using AI-based algorithms for markerless motion tracking with well-established clinical and imaging markers of disease progression in PD. The results demonstrate that AI-based algorithms for markerless motion tracking accurately track hand movements during the Finger Tapping Test. Different movement aspects, such as speed, duration, amplitude, and rhythm alterations, can be quantified from the videos. Our analysis showed that these video-based measures were significantly correlated with clinical rating scales, including the Movement Disorder Society – Parkinson’s Disease Rating Scale Part III and its Finger Tapping subscore. Additionally, some measures significantly correlated with imaging markers, such as Free Water in the posterior Substantia Nigra. Furthermore, video-based measures could detect the presence of PD from the Finger Tapping Test videos with high accuracy. The best classification model achieved an average accuracy of 95% and an area under the receiver operating characteristic curve of 0.87. Finally, we introduce The Hand Tracking Tool, an easy-to-use interface for analyzing Finger Tapping Test videos using AI-based algorithms for markerless motion tracking and obtaining objective movement measures.