结合残差神经网络和语音诊断的帕金森病识别研究

Research on parkinson’s disease recognition based on residual neural network and voice diagnosis

为降低临床上诊断帕金森病对量表和穿戴设备仪器及医生临床经验的过度依赖,为PD患者在诊断方法上提供新的思路。采用信号处理方法对数据集MDVR-KCL提取包含周期变化、峰值变化及谐波信噪比3大类、12个复杂语音特征形成一维向量数据集。采用传统的决策树和残差神经网络分类方法进行训练和测试,通过对比实验,发现可以有效解决因神经网络加深而准确率下降问题的残差神经网络能够有效的区分PD患者和健康人,取得了97.3%的准确率。

In order to reduce the clinical diagnosis of Parkinson’s disease on the scale and wearable equipment and doctors’clinical experience of excessive dependence,new ideas for PD patients are provided in the diagnosis method.In this paper,signal processing method is used to extract 12 complex speech features from MDVR-KCL data set,including periodic change,peak change and harmonic signal-to-noise ratio.Traditional decision tree and residual neural network are used for training and testing.Through comparative experiments,it is found that residual neural network,which can effectively solve the problem of neural network deepening and accuracy decreasing,can effectively distinguish PD patients and healthy people,and the accuracy rate is up to 97.3%.