AI Scheme for High-Accuracy and Contactless Assessment of Parkinson’s Disease Grades

This paper addresses the significant challenges inaccurately and consistently assessing Parkinson's disease (PD) severity, particularly in elderly patients, where traditional diagnostic methods struggle with quantification and standardization. To overcome these limitations, we propose a novel, AI-based scheme for PD grade assessment that enhances accuracy and standardization while avoiding the complexities associated with wearable sensors. We collected a dataset of 110 videos from volunteers across different PD severity grades. Using MediaPipe, a contactless pose extraction model, we extracted 19 distinct kinematic features from joint movements to generate real-time kinematic data. A rigorous data screening phase was employed to eliminate irrelevant or noisy features, reducing the dimensionality of the dataset. We then evaluated PD grades using five different machine learning algorithms. Among these, the K-Nearest Neighbors (KNN) classifier achieved the highest performance, with an overall accuracy of 96.63% and 100% accuracy in classifying Grade 4 and Grade 5 patients. These results demonstrate the effectiveness of our AI-based scheme in providing accurate, contactless, and standardized PD severity assessments. Notably, the elimination of wearable sensors not only improves diagnostic accuracy but also enhances the feasibility of remote diagnosis, benefiting both physicians and patients. Our study underscores that the proposed AI solution represents a significant advancement in PD diagnosis by offering a reliable, sensor-free, and standardized assessment method, which is highly valuable for clinical use and remote management of PD.