Research On Speech-based Identification Of Parkinson’s Disease Subtypes

Parkinson’s disease is a common neurological disease that seriously affects patients’ daily life.With the accelerating trend of population aging in my country,the number of patients with Parkinson’s disease is also increasing,and the proportion of doctors and patients is decreasing.The traditional detection of Parkinson’s disease usually adopts manual evaluation,and the time and labor costs are high,so it is very necessary to use convenient and quick detection methods.The Parkinson’s disease detection method based on machine learning technology can realize the auxiliary diagnosis of Parkinson’s disease by analyzing the speech signals of Parkinson’s disease patients.However,the individual differences of Parkinson’s disease patients are large,and the problem of the large number of Parkinson’s disease types has not been fully studied in the field of speech-based auxiliary diagnosis of Parkinson’s disease.At present,how to use speech to distinguish the type of Parkinson’s disease patients has practical significance.In view of this gap in the research field,the paper first discusses how to use speech to simultaneously realize the distinction of Parkinson’s disease subtypes.Parkinson’s disease patients often have one or more Parkinson’s disease subtypes.According to this phenomenon,the paper abstracts how to distinguish the types of Parkinson’s disease patients into a multi-label classification problem in machine learning.The proposed method is modeled in the order of speech feature extraction,feature selection,and multi-label classification,and fuses information from three different speech signals to classify three Parkinson’s disease subtypes: tremor,frozen gait,and dysphagia at the same time.By comparing different feature selection algorithms and multi-label classification algorithms on a real Parkinson’s disease speech dataset,the experimental results show that the combination of graph-margin based multi-label feature selection algorithm and multi-label k nearest neighbour classification algorithm achieves the best performance.When more data of Parkinson’s disease subtypes is added to the training,the high aliasing and small sample characteristics of the Parkinson’s disease dataset will lead to a decrease in classification accuracy.In response to this problem,the paper further proposes a multi-label classification method based on the prototype network model.In this method,the multi-label data is regarded as the superposition of the class prototypes of each label,and the fully connected network and the selfattention encoder are combined as the feature extraction network.When constructing the task set,the mean of the sample set without a certain label is calculated.As a negative prototype,the mean of the sample set with this label is used as a positive prototype,and the loss function includes not only the classification prediction loss,but also the contrastive loss of positive and negative prototypes.Through comparative experiments on five Mulan multi-label classification datasets and Parkinson’s disease speech datasets.The results show that the proposed method shows better performance on multiple indicators,and improves the model’s discriminative ability on small dataset.