Research And Application Of MMSE Score Prediction Model Based On Spontaneous Speech

Alzheimer’s disease（AD）,commonly known as dementia,is a neurodegenerative and irreversible disease.It will continue to deteriorate over time,leading to a decline in people’s cognition,behavior,and social skills,seriously destroying a person’s ability to survive and take care of themselves.Early screening of AD is helpful for early detection,early intervention,and effectively delaying the progression of the disease.In the process of cognitive degradation,language impairment is a key sign,including difficulty in naming,finding words,repetition,and inappropriate use of pronouns,which makes it possible to use phonetics to assess AD progress.The Mini-mental State Examination（MMSE）can comprehensively and quickly assess the cognitive impairment of subjects from five cognitive dimensions of orientation,attention,calculation,recall and language.Although simple,it will also take up a lot of doctors’ time and bring psychological burden to patients.The use of automated methods to predict MMSE scores from open communication enables more frequent and objective assessments of individuals through an easily accessible automated assessment mechanism,thereby facilitating early screening for AD.Based on the ADReSS 2020 challenge dataset,this paper studies an automated model for predicting the MMSE score of subjects using their spontaneous speech.It is mainly divided into three parts:The first part uses openSMILE to extract acoustic features in speech,and the features with high correlation are ComParE16,emobase,eGeMAPS and Is09-13.Putting the extracted acoustic features into the Multilayer Perceptron（MLP）,the Root Mean Square Error（RMSE）obtained is 5.49,which is 10%lower than the ADRe SS 2020 challenge acoustic baseline.The second part is the processing of linguistic features,which mainly considers two aspects.On the one hand,Lexical Features are combined with Sentiment Scores and put into the Support Vector Regression（SVR）model.On the other hand,using the pre-training model BERT,the sentence vector of each subject is obtained and put into the Ridge regression model（Ridge）.The final RMSE is 4.5 1,which is 13%lower than the linguistic baseline.The third part comprehensively considers the acoustic and linguistic models.The impact of the linguistic model on the final result is greater than that of the acoustic model.Therefore,it is considered to increase the proportion of the impact of the linguistic model on the final result and reduce the proportion of the impact of the acoustic model on the final result.Combining the acoustic feature ComParE16 and the linguistic feature Lexical+sentiment yields the best result,with an RMSE of 4.18.Based on the above research,this paper designs and implements the MMSE score prediction system based on spontaneous speech.It is a WEB program,which can be used to automatically predict the MMSE score and keep abreast of one’s own health.