Quantification of Parkinsonian Bradykinesia Based on Axis-Angle Representation and SVM Multiclass Classification Method

Abstract

Accurate evaluation of bradykinesia plays a crucial role in the diagnosis and therapy effect of Parkinson's disease. However, the subjective assessment shows low consistency among different evaluators, and the objective sensor-based methods cannot accurately distinguish patients with different grades of the 5-point clinical bradykinesia ratings. In this paper, an objective scoring method based on axis-angle representation and multi-class support vector machine (SVM) classifier was employed to estimate the bradykinesia severity. To reduce the dimension of attitude data for the better statistical analysis, the axis-angle representation approach was employed to obtain the 1D combined orientation angle to express the 3-D hand-grasping assessment task. The significant bradykinesia features extracted from the 1-D combined orientation angle were used to train the SVM classification algorithm to acquire an objective score on the bradykinesia severity. Clinical experiments with 78 patients and 18 age-matched healthy subjects showed that the classification accuracy of the proposed method was 95.349%, which was superior to other related reports. Hence, the proposed objective scoring method can provide a 5-point bradykinesia score in realtime, and matches the clinical assessment approach of neurologists. Moreover, this method improves both the inter-rater reliability and intra-rater reliability of the bradykinesia assessment.

Keywords

• Parkinsonian bradykinesia
• support vector machine (SVM)
• quantification
• wearable device
• sensor fusion