基于传统方法和机器学习对帕金森病语音信号早期预警模型的对比研究

Based on Traditional Methods and Machine Learning for Parkinson’s Disease Speech Signals: A Comparative Study of Early Warning Models

近年来,研究发现大约90%以上的帕金森病(PD)患者在疾病的早期阶段表现出某种形式的声音障碍。本文从UCI机器学习存储库中选取经最先进的语音信号处理技术提取252名受试者的语音信号特征数据,运用Logistic回归和决策树、Bagging、随机森林建立PD语音信号早期预警模型。研究发现,基线特征、梅尔频率倒谱系数和小波变换与PD有明显的显著关系。经对比均方误差发现,无论是长期预测还是短期预测决策树预测模型的误判率约为0.10,能高精度分类出受试者是否患有帕金森病。Logistic模型的预测效果都显著不如机器学习预测,主要原因与数据的特殊性和复杂性相关,而非只在于模型本身。利用统计的方法在受试者监测早期进行预警分析,可提前做好预防准备,减轻频繁就医的麻烦。

In recent years, studies have found that about 90% of people with Parkinson’s Disease (PD) exhibit some forms of sound disorder in the early stages of the disease. In this paper, the speech signal feature data of 252 subjects were extracted from the most advanced speech signal processing technology from the UCI machine learning repository, and the PD voice signal early warning model was established by using Logistic Regression, Decision Tree, Bagging, and Random Forest. It was found that the Baseline characteristics, FMCC and wavelet transform had obvious significant relationships with PD. After comparing the mean squared error, it is found that the false positive rate of the Decision tree prediction model is about 0.10, whether it is a long-term prediction or a short-term prediction, which can classify whether the subject has Parkinson’s Disease with high accuracy. The prediction effect of Logistic models is significantly inferior to that of machine learning, mainly because of the particularity and complexity of the data, not just the model itself. The use of statistical methods to carry out early warning analysis in the early stage of subject monitoring can prepare for prevention in advance and reduce the trouble of frequent medical treatment.