Abstract: Dyskinesia is the significant feature of Parkinson’s disease patient, therefore the data analysis on parkinsonian gait kinematic parameters can provide a strong basis for disease diagnosis and rehabilitation. The actual number of Parkinson’s disease is far less than the normal human. The traditional machine learning methods are not suitable to classify the imbalanced data directly. This paper proposes a Cost Sensitive Support Vector Machine（CS-SVM） approach to distinguish between healthy human and patients with Parkinson’s disease firstly. The entire subject’s gait data is extracted from real U-shape electronic walkway. The classification performance is also improved by transforming the extracting features into non-dimensional forms. The experimental results indicate that the CS-SVM method can achieve the prediction accuracy of 94.16%, and the F-measure value reaches 87.08%, while it achieves a better performance comparing traditional SVM and the recognition performance is significantly improved by eliminating the influence of height, the prediction accuracy and F-measure value reach 94.81% and 88.66% respectively.

Key words: gait analysis, imbalanced data, electronic walkway system, Parkinson’s disease, Cost Sensitive Support Vector Machine（CS-SVM）

摘要： 运动障碍是帕金森病（PD）患者的重要特征，步态信号分析可以为疾病诊断和康复治疗提供有力依据。现实中PD患者数量远小于正常人群，传统的机器学习方法不适合对正例样本数远多于反例的非平衡数据进行分类。为了准确地区分出PD患者和健康人，使用一种代价敏感支持向量机（CS-SVM）的方法来构建PD患者和健康人之间的步态信号分类模型。所有受试者的步态运动学特征数据是采用真实的U型电子步道系统提取的，并将特征数据转化为无量纲的形式来消除身高对时空属性的影响。实验结果表明使用这种CS-SVM方法得到的预测准确率和F-measure值分别达到了94.16%和87.08%，与传统的SVM方法相比性能更优。同时消除身高对时空属性的影响可以大幅提高识别性能，预测准确率和F-measure值分别达到94.81%和88.66%。

关键词: 步态分析, 非平衡数据, 电子步道系统, 帕金森病, 代价敏感支持向量机