Movement Disorder Evaluation of Parkinson’s Disease Severity Based on Deep Neural Network Models

Abstract

Parkinson’s disease has been extensively studied using artificial intelligence for severity assessment. Previous research has achieved excellent results using wearable solutions. However, this approach has significant cost disadvantages, especially in the post-COVID-19 era. Recently, researchers are investigating vision-based alternatives for severity assessment. Unfortunately, the outcomes of these attempts often fall short when compared to wearable solutions. In this paper, we propose a novel feature extraction method suitable for temporal models, called ‘Multitemporal Feature Average Pooling.’ This method effectively captures crucial features from temporal data. Additionally, we employ interpretable artificial intelligence techniques to analyze the models, demonstrating their importance in severity assessment. We trained four severity classification models based on MDS-UPDRS. In the multi-class task of finger tapping, we achieved accuracies of 77% for the left hand and 85% for the right hand. These results represent an improvement of 4.6% and 12.6% compared to Shao et al.’s Three-Stream CNN model. In the binary classification, both left- and right-hand tasks achieved 92% accuracy, showing an increase of 11.4% and 14% compared to Chang et al. Furthermore, we achieved accuracies ranging from 60% to 80% for multi-class tasks involving Hand Movements, Rapid Alternating Movements of Hands, and Leg Agility. In binary classification, these tasks reached an even higher accuracy of 80% to 90%. This research is crucial for advancing vision-based severity assessment in clinical settings, providing more accurate evaluations for Parkinson’s disease patients. Through the application of artificial intelligence technology, we anticipate achieving substantial breakthroughs and advancements in clinical practice.

Keywords

• Deep learning
• LSTM
• MDS-UPDRS
• Parkinson’s disease
• severity classification