Automatic Classification Of PD Based On Motion Parameters

Parkinson’s disease(PD)has become a very common neurodegenerative disease,which seriously disturbed the patients‘ daily life.With the aging of the world population,the burden of Parkinson’s disease on society and economy is expected to increase year by year.The grading diagnosis of Parkinson’s disease helps the clinic to formulate the treatment plan.According to the symptoms of Parkinson’s disease,proposing the corresponding treatment plan to postpone disability is an important goal in the current and future research.With the aging of population,the burden of Parkinson’s disease on the society and the economy will increase year by year.At present,parkinson’s disease diagnosis is based on Hoehn-Yahr classification and UPDRS scale.However,it depends on the subjective judgement of doctors,which requires a lot of time and manpower.Motor function degradation is the main clinical manifestation of parkinson’s disease,and currently there is no objective evaluation method.The development of wearable motion sensors make it possible for objective and accurate evaluation of motor function of PD patients.The data acquisition system includes 5 wearable sensor nodes worn on the left and right wrist,the waist and the left and right legs.Each node collects 3 axis acceleration and 3 axis angular velocity data.In order to design the experimental movements in accordance with the clinical understanding,this paper designs a set of experimental paradigms basing on clinical PD scale such as UPDRS.The motion data of a total of 71 patients with Parkinson’s Hoehn-Yahrand healthy controls were collected.Each group’s data were examined by two clinical experts for the judgement of Hoehn-Yahr stage.Three methods including support vector machine(SVM),k nearest neighbor algorithm(kNN)and random forest are explored.The original acceleration and angular velocity signals of motion sensors are used to extracts the motion characteristic parameters such as the gait cycle and the wrist angle,and extracts the statistical characteristic parameters,such as the frequency of the main frequency,the deflection and the kurtosis in the time domain.The feature parameters are dimensionality reduced by principal component analysis and other feature engineering methods.Meanwhile,the accuracy of three machine learning classification algorithms for predicting PD classification is compared.The results show that the accuracy of the support vector machine PD classification algorithm is the highest,which is 85.92%.Deep learning method based on convolution neural network is applied to the research of PD classification algorithm.This method uses the original data of sensor nodes related to experimental actions as input,and automatically optimizes the weight parameters of the network by data driven.This paper also compares the effects of different data augmentation methods on the results,and adding noise and flipping data operations at the same time has the best effect.The final classification accuracy of this method is 92.95%.Especially in the early PD population,the classification accuracy of this method is 77%,which is 22% higher than the machine learning algorithm.Deep learning algorithm does not need empirical knowledge to design parameters,and achieve higher classification accuracy.This method has good application prospects for PD classification and early detection.