Research Of Auxiliary Diagnosis Of Parkinson’s Disease Based On Speech Time-frequency Feature Fusion

Parkinson’s disease(PD)is a common neurodegenerative disease,which is not completely curable and can only be alleviated with timely treatment.Dysarthria is one of the early symptoms of Parkinson’s disease,and research of auxiliary diagnosis of Parkinson’s disease based on speech signals and machine learning methods has the advantages of non-invasive,easy acquisition and lower cost,which improves the efficiency of the diagnosis and treatment stage.Therefore,research of accurate and efficient speech auxiliary diagnosis for Parkinson’s disease has important theoretical significance and social value.The following problems exist in current research on Parkinson’s disease speech auxiliary diagnosis:(1)most of the studies is based on the combination of relevant speech features at the Mel Scale and deep neural network models;however,the existing methods cannot adequately focus on the global timing information of the speech signal;(2)speech features at the Mel Scale have limited effectiveness in accurately characterizing the pathological information of Parkinson’s disease;(3)speech-based studies on Parkinson’s disease assessment focus on a single assessment metrics and often ignore the correlation between different assessment tasks.To address the above problems,this thesis proposes a method for Parkinson’s disease auxiliary diagnosis based on speech time-frequency feature fusion from the perspective of speech feature fusion and multi-task learning,which works as follows:(1)To address the problem of characterizing the speech information of Parkinson’s disease,this thesis proposes a method for Parkinson’s disease detection based on speech time-frequency feature fusion.First,the Conformer encoder module is introduced in the existing S-vectors model to extract time-domain global features of speech signals.Then,the frequency-domain global features related to Parkinson’s disease speech detection are embedded into the time-domain features for time-frequency information fusion to achieve Parkinson’s disease speech detection.Finally,this thesis conducts several comparative experiments on a public Parkinson’s disease speech dataset and a self-collected speech dataset to verify the effectiveness of the method.(2)To address the problem of Parkinson’s disease severity assessment,this thesis proposes a multi-task learning-based method for Parkinson’s disease severity assessment.First,based on the time-frequency feature fusion model described in the research work(1),a shared-private feature separation method is used to capture the shared features and task-related features from the time-frequency fused features separately.Then,a dynamic weighted average method is used to balance the learning speed of multiple tasks.Finally,this thesis conducts comparison experiments with existing methods on the public Parkinson’s disease speech dataset to verify the effectiveness of the method.(3)A speech-based Parkinson’s disease auxiliary diagnosis system is developed in this thesis.The system uses an Android client as the front-end and consists of user module,speech recording module and diagnosis reports modules,which mainly implements the functions of recording speech and auxiliary diagnosis,diagnosis reports,and doctor-patient communication.After testing,the system can be used as a tool platform Parkinson’s disease auxiliary medical diagnosis.