Study On Parkinson's Speech Diagnosis Method Based On Instance Distribution Learning And Collaborative Learning

Parkinson's Disease(PD)is an irreversible neurodegenerative disease which is seriously harmful to the physical and mental health of patients with PD,and there is still no effective treatment.Existing methods based on biochemistry,imaging,and scales are effective,but they rely heavily on medical equipment and medical personnel.Due to the slowness and concealed incidence of PD,the condition of PD patients will gradually increase with time goes by,so it is very important to study an efficient,convenient and objective diagnostic method.The diagnosis method using speech data is a new method for PD diagnosis that has emerged in recent years,which realizes detection and diagnosis through data mining of the speech data of subjects.It has the advantages of non-invasive,fast and low-cost,and has attracted extensive attention at home and abroad.Although the current public research on PD speech detection has made significant progress,there are still some problems,such as: 1)The mixed noise in the sample acquisition process leads to the poor quality of the sample,which forms an aliasing region in the sample space and greatly affects the classification accuracy of the model;2)The study found that the sensitive features of different corpora were not consistent,while the existing studies extracted the same features for different speech tasks,resulting in the acquired samples having redundant features and poor sample classification ability.In view of the above problems,the research work is conducted from sample distribution learning and collaborative learning in this thesis.(1)Aiming at the serious problem of sample aliasing,a partition bagging ensemble learning classification algorithm for PD speech data is proposed in this thesis.Firstly,the metrics ratio of sample centroid distance is used to measure the aliasing degree between a sample and the heterogeneous samples,and the training set is divided into multiple subsets according to the aliasing degree.Secondly,each subset is trained by a classification model,and the subset is self-tested.The misclassified samples are passed to the next subset after enhancement,and the weights of the sub-classifiers corresponding to each subset are calculated.Finally,the prediction results of each subset are weighted and fused to obtain the final results.The influence of the samples in the aliasing region on other samples is reduced by dividing the train set into subsets,and the utilization of samples is improved because of the transmission of misclassified samples,while the influence of the samples in the aliasing region on the model is further weakened due to the use of weights in the final integration process,thus improving the classification accuracy(2)Aiming at the inconsistency of sensitive features of different corpus,a diagnosis algorithm based on sample and feature collaborative learning for PD speech data is proposed in this thesis.First,multiple speech samples of each subject are combined into one sample by means of sample merging learning;Secondly,Sequential Forward Selection algorithm is used,in conjunction with evaluation criteria such as Pearson Correlation coefficient,distance separability,Minimum Redundancy Maximum Correlation,for feature selection and sequencing.Finally,the top features are selected in turn for classification,and the feature set with the best classification performance is selected for classification diagnosis,and the final predicted result is obtained.The differences between different corpus samples and the feature redundancy of different samples are fully considered by this method,and the subset of speech features closely related to each corpus sample and higher classification accuracy can be obtained through sample and feature collaborative learning.New ideas for improving the accuracy of speech-based PD diagnosis is proposed in this thesis,which provides reference value for corpus design and feature extraction in the research of PD speech diagnosis to some extent,and has important theoretical value and practical significance for promoting the clinical and practical development of PD speech classification diagnosis methods.