基于Lasso-Logistic回归的帕金森疾病声学特征诊断研究

Diagnostic Study of Acoustic Features in Parkinson’s Disease Based on Lasso-Logistic Regression

摘要: 本文基于埃斯特雷马杜拉(西班牙)帕金森病区域协会的80名欧洲受试者声学特征数据，结合lasso回归提出两阶段变量选择法对44个声学特征因子筛选，最后得到6个显著的声学特征因子：Gender、Shim_loc、MFCC3、HNR35、PPE、GNE。将上述因素通过多因素logistic回归构建患PD疾病风险的列线图模型，并从多个角度验证该模型的有效性和校准性。结果表明，早期PD患者基底神经节运动调节功能异常，声学数据中MFCC3和HNR35数值偏低，PPE、GNE和Shim_loc数值偏高，临床表现为发音时声带振动的最大频率降低，声音低沉，进一步说明所构建的列线图模型可以根据不同的声学特征较好地诊断研究对象患有PD疾病的风险高低。今后声学特征有望成为早期PD诊断的重要生物标记物，为疾病远程筛查提供辅助手段。

Abstract: In this paper, based on the acoustic profile data of 80 European subjects from the Regional Association of Parkinson’s Disease in Extremadura (Spain), a two-stage variable selection method was proposed to screen 44 acoustic profile factors in conjunction with lasso regression, and finally six significant acoustic profile factors were obtained: gender, Shim_loc, MFCC3, HNR35, PPE, and GNE. The above factors were combined to construct a column plot model of the risk of developing PD disease by multifactorial logistic regression to construct a column-line graph model of the risk of developing PD disease, and validate the validity and calibration of the model from multiple perspectives. The results showed that early PD patients had abnormal motor regulation of the basal ganglia, low values of MFCC3 and HNR35, high values of PPE, GNE and Shim_loc in the acoustic data, and reduced maximal frequency of vocal fold vibration during articulation and muffled voices, which indicated that the constructed columnar plot model could diagnose the risk of PD in the subjects according to the different acoustic features. In the future, acoustic features are expected to become important biomarkers for early PD diagnosis and provide an aid for remote screening of the disease.