A method and system for obtaining a speech-based measure of the importance of dysphagia symptoms in patients with Parkinson's disease

By combining the GBDT model and the logistic regression model, based on feature extraction of multi-speech tasks, the problems of inefficiency and incomplete analysis of traditional Parkinson's disease evaluation methods are solved, and efficient evaluation of dysphagia symptoms in Parkinson's disease patients are achieved. An interpretable pathological feature analysis is provided to assist doctors in developing treatment strategies.

CN114299925BActive Publication Date: 2024-11-01JIANGSU PROVINCIAL GOVERNMENT HOSPITAL
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111677393.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-12-31
Publication Date
2024-11-01
Estimated Expiration
2041-12-31

AI Technical Summary

Technical Problem

The traditional Parkinson's disease evaluation method is inefficient, the single vowel pronunciation task analysis is incomplete, and the lack of interpretability makes it difficult to effectively evaluate the symptoms of dysphagia in patients with Parkinson's disease.

Method used

Using a speech signal processing method, the GBDT model and the logistic regression model are combined with feature extraction of multi-speech tasks to generate higher-order features for the importance of dysphagia symptoms, so as to achieve the dysphagia symptoms in Parkinson's disease patients. Assessment.

Benefits of technology

It improves the accuracy and efficiency of dysphagia symptoms in patients with Parkinson's disease, provides an interpretable pathological feature analysis report, assists doctors in formulating treatment strategies, and significantly improves the quality of life of patients.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114299925B_ABST
    Figure CN114299925B_ABST
Patent Text Reader

Abstract

The present invention proposes a method for obtaining a speech-based measure of the importance of dysphagia symptoms in patients with Parkinson's disease. By applying multiple speech tasks, a more comprehensive auxiliary analysis of the degree of dysarthria of the subjects is realized. The original features are imported into the GBDT model, and the original features extracted from different speech tasks are combined to obtain high-order features. These high-order features generated by the decision tree are then used as the input of the logistic classification algorithm in machine learning to evaluate dysphagia symptoms, and an importance measure is output to realize early auxiliary analysis of dysphagia symptoms in patients with Parkinson's disease, so as to adopt specific treatment strategies in a timely manner.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the application of machine learning in medicine, and relates to a method and system for obtaining a measurement index of the importance of speech to dysphagia symptoms in Parkinson's disease patients. Background Art

[0002] Parkinson's disease (PD) is a neurological disease that can cause movement and non-motor disorders. Dysphagia is one of the symptoms of Parkinson's disease movement disorders. The clinical manifestations of dysphagia in patients are diverse, including delayed swallowing, aspiration, leakage, residual, incomplete opening of the circumflex muscles, etc. And dysphagia can occur at any stage of the patient, and it can even be the only clinical manifestation of early patients. Therefore, early evaluation of the patient's swallowing function and taking active and effective preventive rehabilitation treatments based on the patient's specific clinical situation have a significant impact on improving the patient's quality of life, reducing mortality, and reducing the burden on the patient's entire family and even society. By studying the incidence, characteristics, and influencing factors of dysphagia in PD patients, it provides a certain basis for timely and effective diagnosis and treatment of dysphagia in patients and improving the quality of life of patients.

[0003] Existing studies have shown that there is a certain pathological connection between Parkinson's disease subtypes and dysphonia. Voice damage may be one of the earliest signs of Parkinson's disease, and its symptoms include slow speech, hoarseness, low volume, and tremor in pronunciation. Voice is mainly produced by the cooperation of the vocal organs in the body, and the cooperation between human organs is controlled by neurons in a unified and coordinated manner. Parkinson's patients cannot stably control the vocal organs due to the loss of related neurons, so that patients all have different degrees of dysphonia. Compared with healthy people, Parkinson's patients often cannot make smooth and accurate sounds. Therefore, speech signals can be used to predict Parkinson's disease early and evaluate the condition of dysphagia symptoms. Compared with traditional Parkinson's disease prediction methods, the use of speech signals to detect Parkinson's disease is economical and efficient, and speech monitoring is non-contact, simple and convenient. For the acquisition of patient speech data, continuous vowels / a / , / i / , / o / , etc., rapid voiced and unvoiced syllable conversion, and situational dialogue reading can be used. Physiologically, the pronunciation of sustained vowels involves the combination of various muscles in the vocal cords and vocal tract, which can be a good measure of the subject's pronunciation ability. Rapid voiceless and unvoiced sound conversion can be a good analysis of the subject's ability to move the alveoli, jaw, tongue and other pronunciation organs. The reading of situational dialogues can analyze the naturalness of the subject's expression ability and whether he can correctly pronounce the intonation of interrogative sentences and declarative sentences.

[0004] When using speech signals to predict Parkinson's disease, it is necessary to use a speech signal processing algorithm to analyze the speech signal to be tested, so as to extract speech feature information that can characterize the pathological characteristics of Parkinson's disease. For different speech tasks, the angle of analysis of articulation disorders is different, and the preprocessing extraction method will also be different, so that the subject's physiological information can be characterized from different dimensions. Based on the above situation, after the corresponding speech features are extracted based on different speech data through the speech signal processing algorithm. The technical model in the field of machine learning can be used to analyze the symptoms of dysphagia. Summary of the invention

[0005] The present invention aims at the problems of low efficiency of traditional Parkinson's disease assessment methods, incomplete analysis of single vowel pronunciation tasks, lack of interpretability, etc., and proposes a method for obtaining a measurement index of the importance of speech to the dysphagia symptoms of Parkinson's disease patients. Feature extraction is performed on the collected speech signals of multiple speech tasks, and the GBDT model is first pre-trained with data to complete the feature combination of the original features extracted from different speech tasks to obtain high-order features. These high-order features generated by the decision tree are then used as the input of the logistic classification algorithm in machine learning to evaluate the dysphagia symptoms, thereby achieving the acquisition of the importance measurement index of the dysphagia symptoms of Parkinson's disease patients, so as to adopt specific treatment strategies in a timely manner.

[0006] In order to achieve the above objectives, the present invention first proposes a method for obtaining a symptom importance measurement index (pathological feature importance measurement) of dysphagia in patients with Parkinson's disease based on speech, which is performed in the following steps:

[0007] Step 1: Collect speech data of Parkinson's disease patients participating in multiple speech tasks (such as sustained vowel pronunciation, voiceless and voiced sound conversion, scene dialogue, etc.), and label the above speech data (whether they suffer from dysphagia);

[0008] Step 2: Perform denoising and enhancement preprocessing on the collected speech data of Parkinson's disease patients, and then perform feature extraction to obtain speech feature vectors that can characterize the data, and construct a data set;

[0009] Step 3: Select N training samples from the data set, train the Gradient Descent Decision Tree (GBDT) model with the Classification And Regression Tree (CART) model as the base classifier, and obtain M base classifiers.

[0010] Step 4: Export the trained M base classifiers from the GBDT model, and use the path from its root node to any child node as a decision feature variable r k ;

[0011] Step 5: The K decision feature variables obtained from the M base classifiers are sent together with the original data feature variables into a binomial logistic regression model for linear fitting;

[0012] Step 6: Derive the importance measures of rule variables and original data feature variables from the trained logistic regression model;

[0013] Step 7: Output the importance measurement result report of the feature variables to assist doctors in making dysphagia symptom assessment.

[0014] Preferably, in step 2: extracting features from the collected speech data of Parkinson's disease patients and constructing a data set, the specific process is:

[0015] 1) Use speech signal processing algorithms to extract speech features. The extracted features include three categories:

[0016] ① Based on the continuous vowel pronunciation task: the fundamental frequency F0 of each vowel, the average fundamental frequency F0_ave, the minimum fundamental frequency F0_min, the maximum fundamental frequency F0_max, the jitter Jitter, the amplitude perturbation entropy APQ, the pitch perturbation quotient PPQ, the pronunciation segment ratio, the characteristics of measuring the change of amplitude over time, the noise harmonic ratio NHR, the harmonic noise ratio HNR, the cycle density entropy RPDE, the trend fluctuation analysis DFA, the gene period entropy PPE; the characteristics of measuring the change of amplitude over time include: Shimmer and its variants, the decibel value of Shimmer, the three-point amplitude perturbation entropy ShimmerAPQ3, the five-point amplitude perturbation entropy ShimmerAPQ5, and the 11-point amplitude perturbation entropy ShimmerAPQ11;

[0017] ② Pronunciation tasks based on rapid voiced-unvoiced conversion: features related to fundamental frequency and their related statistics: average fundamental frequency energy Avg.Energy, standard deviation of energy Std.Energy, as well as voiced segment ratio Voice Rate, silence segment ratio Silence Rate, average duration of silence segment, standard deviation of silence segment duration;

[0018] ③ Pronunciation tasks based on scene dialogues: Features related to fundamental frequency: calculate the fundamental frequency contour, average, standard deviation and maximum value; features related to energy: calculate the average energy, standard deviation and maximum value of energy; features related to duration: vowel segment proportion, average duration of vowel syllable segments, standard deviation of vowel segment duration, average duration of silent segments, standard deviation of silent segment duration;

[0019] 2) Dataset construction: For the selected N sample voices, according to the collected label information, form a sample pair (x n ,y n ), where x n represents the nth sample (n=1,2,···N), represents the h-th dimension feature of the n-th sample (h = 1, 2, ··H), y n is the label of the nth sample: when the label is 0, it means that the subject has no dysphagia symptoms, otherwise it is 1.

[0020] Preferably, in step three: training the integrated model GBDT using the CART model as the base classifier, the specific process is:

[0021] 1) Initialize the first weak learner: Where P(y=1|x) is the total number of training samples (x n ,y n ), that is, the learner is initialized using the prior information of the original probability distribution in the training samples;

[0022] 2) Label prediction value obtained in the current iteration Where F(x n )=F0(x n ) is the output of the learner in this iteration, and the loss function is constructed by the output of the current model:

[0023]

[0024] Among them, y n is the true label value, is the predicted label value;

[0025] 3) The negative gradient r of the loss function obtained in the current iteration m,n for:

[0026]

[0027] where r m,n It will be used as the fitting target of the learner in the next iteration, m represents the number of iterations, n is the nth sample, and this iteration ends;

[0028] 4) A new iteration begins. First, the mth base learner is established. A new CART tree is built to fit the new training data pair (x n ,y n =r m,n ), get the current iteration output F m (x):

[0029] 4-1) Generation of CART-based learner in a new iteration, selecting the optimal segmentation variable h and segmentation point s, and calculating:

[0030]

[0031] Traverse all feature variables h of the data, select one as the split variable, and then traverse the split points s according to its value range, and select the pair (h, s) that minimizes equation (1.3);

[0032] 4-2) After selecting the split variable A and split point S, the data set is divided into D1(h,s)={x|x (h) ≤s} and D2(h,s)={x|x (h) >s} two sub-regions and determine the output value c of the region j :

[0033]

[0034] 4-3) Continue to repeat steps 4-1) and 4-2) for the two sub-regions until the pre-set tree depth is met and the segmentation stops. A new CART base learner is generated, which divides the feature space of the input data into J regions, where J+1 is the number of leaf nodes of the CART model. The best fitting value c for each region is calculated by (1.4) m,j , generates the output F of the current base classifier m (x):

[0035]

[0036] 5) Repeat steps 2) to 4) until M CART trees are obtained after M iterations. The output of the final model for the predicted label of the data is

[0037] in is the integration of the outputs of M base classifiers;

[0038] 6) At this point, the GBDT model training is completed. The model output of this part will only be used to measure the degree of fit of the model to the data set. The M trees generated by the model will serve as the basis for the subsequent construction of rule variables.

[0039] Preferably, in step 4:

[0040] After completing the GBDT model training, each base classifier CART is exported. The path from the root node of the constructed tree model to any child node will be used as a decision feature variable, i.e., a decision rule. The specific process is as follows:

[0041] 1) For each base classifier obtained, the path from the root node to any node in the tree is used as a decision rule:

[0042]

[0043] Where I(·) is an indicator function, which takes the value 1 when the expression in the brackets is true and 0 otherwise. A complete rule variable can be expressed as g m,k (x) = I 节点j (x 切分特征1 <s)*I 节点j+1 (x 切分特征2 >s)I 节点j+· (·), that is, the samples are classified by the split nodes and split point values ​​selected during the classification and regression tree construction process. When the sample falls into the above split data domain, the expression is 1;

[0044] 2) T m is the leaf node set of the classification and regression tree (CART) obtained in the mth iteration, D j,m (h, s) is the segmentation region when the segmentation feature variable h at node j takes the value s; for M trees, the number of nodes in each tree is J+1, and the number of rule variables that can be obtained is

[0045] Preferably, in the step 5:

[0046] The K decision feature variables obtained from the M base classifiers are sent together with the original data feature variables into the logistic regression model for linear fitting. The specific process is as follows:

[0047] 1) Obtain the decision rule from the base classifier as a new variable and import it into the linear function together with the original feature variable to fit the data pair (x n ,y n ), the mathematical expression of the model Q(x) is:

[0048]

[0049] Where e0 represents the fitting deviation, e k represents the kth rule variable g m,k The weight of b h is the original feature x in the linear regression model h The weight of , H represents the number of features;

[0050] 2) The estimation of the model Q(x) coefficients will be obtained by optimizing the loss function in the following formula:

[0051]

[0052] The arg min function is used to obtain the value of the coefficient when the right-hand side term is minimized. λ is used to weigh the weight of the rule variable in the model. The larger the value, the smaller the influence of the rule variable on the model output result.

[0053] 3) The final output of the model is the data x n The label prediction value of

[0054] Preferably, in step six:

[0055] Importance measurement of rule variables and original data feature variables; the specific process is:

[0056] 1) The importance measure of the rule variable is I K :

[0057]

[0058] Among them k Representation rule g m,k The sample ratio, that is, how many samples satisfy rule g m,k , e k represents the kth rule variable g m,k The weight of

[0059] 2) For feature variables, its importance measure I h for:

[0060]

[0061] where std(x j ) refers to The standard deviation of h is the original feature in the linear regression model The weight of

[0062] 3) Based on the calculations in the above two steps, the contribution of each feature variable (original feature, rule feature) to the model output in the model prediction process is obtained, that is, the importance of the feature variable.

[0063] Preferably, the specific process in step seven is: integrating the prediction of the label value by the logistic model and the importance measurement of each feature variable into a disease analysis report of a sample. The doctor observes the distribution range, importance degree and category prediction value of the pathological characteristics through the disease analysis report output by the model, and conducts a comprehensive assessment of the patient's symptoms.

[0064] The present invention also proposes a speech-based importance measurement acquisition system for pathological characteristics of dysphagia symptoms in Parkinson's patients, and the system executes the method.

[0065] Preferably, the system comprises:

[0066] 1) A voice data collection module; used to perform step 1: collect voice data of Parkinson's disease patients participating in multi-speech tasks, as well as characteristic information of dysphagia of Parkinson's disease patients;

[0067] 2) Speech data processing module; used to execute step 2: pre-process the speech data for denoising and enhancement, then perform feature extraction, and convert the denoised and enhanced speech into a speech feature vector as a data set;

[0068] 3) Feature data processing module based on GBDT model; used to execute step 3: select training samples from the data set, train the integrated model GBDT with CART model as the base classifier, obtain M base classifiers, and create a foundation for the subsequent rule variable construction;

[0069] 4) CART tree-based classifier generation module; used to execute step 4: export the trained base classifier, and use the path from its root node to any child node as a decision feature variable to complete the construction of the decision variable;

[0070] 5) A linear fitting module based on a logistic regression model; used to execute step 5: sending the K decision feature variables obtained from the M base classifiers together with the original data feature variables into the logistic regression model for linear fitting to obtain the model's prediction of the category of the test sample;

[0071] 6) Importance measurement export module for rule variables and original data feature variables; used to execute step six: export importance measurement of rule variables and original data feature variables, and generate an analysis report of the test sample together with the prediction of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 Flow chart of the method of the present invention DETAILED DESCRIPTION

[0073] The following experiments were used to verify the beneficial effects of the present invention:

[0074] This experiment uses the Parkinson's disease patient data set collected by the present invention as the research object. The data set collected a total of 212 speech samples from 53 Parkinson's disease patients. After denoising preprocessing, multiple speech features are extracted. Each sample contains 36 speech features, each of which contains 5 statistical values; the sample contains a label to identify the characteristic information of dysphagia, 1 means dysphagia symptoms, and 0 means no dysphagia. According to the attributes in the data set, the Parkinson's disease dysphagia analysis method based on multi-speech tasks is combined with Figure 1 Follow these steps:

[0075] Step 1: Extract speech features from the data collected from the multi-speech task using speech signal processing algorithms:

[0076] The extracted features include:

[0077] ① Based on the continuous vowel pronunciation task: the fundamental frequency F0, average fundamental frequency F0_ave, minimum fundamental frequency F0_min, maximum fundamental frequency F0_max, jitter Jitter, amplitude perturbation entropy APQ, pitch perturbation quotient PPQ and pronunciation proportion of each vowel, the characteristic tremor Shimmer and its variants that measure the amplitude change over time: Shimmer's decibel value, three-point amplitude perturbation entropy ShimmerAPQ3, five-point amplitude perturbation entropy ShimmerAPQ5, 11-point amplitude perturbation entropy ShimmerAPQ11, noise harmonic ratio NHR, harmonic noise ratio HNR, cycle density entropy RPDE, trend fluctuation analysis DFA, gene period entropy PPE;

[0078] ② Pronunciation tasks based on rapid voiced-unvoiced conversion: features related to fundamental frequency and their related statistics - average fundamental frequency energy, standard deviation of energy, percentage of vowel syllable segments, percentage of silent segments, average duration of silent segments, standard deviation of duration of silent segments;

[0079] ③ Pronunciation tasks based on scene dialogues: Features related to fundamental frequency: calculate the fundamental frequency contour, average, standard deviation and maximum value; Features related to energy: calculate the average energy, standard deviation and maximum value of energy; Features related to duration: vowel segment proportion, average duration of vowel syllable segments, standard deviation of vowel segment duration, average duration of silent segments, standard deviation of silent segment duration;

[0080] Step 2: Use the CART model as the base classifier to train the GBDT model and divide the data set into a training set and a test set. Select 70% of the samples as training samples to train the GBDT model, and the remaining 30% of the samples as test samples to test the model performance. The above features are imported into the model pre-training to obtain the cross combination of the original features. The specific implementation is as follows:

[0081] 7) Initialize the first weak learner F0(x): Where P(Y=1|x) is the proportion of label value y=1 in the training samples, that is, the learner is initialized using prior information.

[0082] 8) The learner output obtained in the current iteration is Where i is the current model iteration number, F m (x) is the output of the mth CART tree, and the loss function L(θ) is constructed by the current learner output:

[0083]

[0084] where y i is the true label value, is the predicted label value, Replaced with the learner output, the loss function is rewritten as:

[0085]

[0086] The negative gradient r of the loss function obtained in the current iteration m,j for:

[0087]

[0088] 9) Update the learner by building a new CART tree to fit the data (x i ,r m,i ) Get the current iteration output F m (x):

[0089] 3-1) Generation of CART in a new iteration, randomly select split variable (feature) A, split point S, and calculate:

[0090]

[0091] Get the optimal splitting variable and splitting point, iterate until the stopping condition, and after feature selection is completed, a new CART model is generated; where the data set is divided into D1(A, s) and D2(A, s) according to the splitting point s when the splitting variable A is selected, c1 represents the sample output mean of the data set D1(A, s), and c2 represents the sample output mean of the data set D2(A, s);

[0092] 3-2) The new CART model divides the feature space of the input data into J regions, where J is the number of leaf nodes in the CART model, and calculates the best fitting value C for each region. m,j :

[0093]

[0094] Iterate M times to obtain M CART trees, where R m,j represents the jth region.

[0095] 10) For the trained base classifier, the path from the root node to any child node is used as a decision feature variable, i.e., a decision rule; the specific process is as follows:

[0096] 4-1) For each base classifier, the path from the root node to any node in the tree is used as a decision rule:

[0097]

[0098] Different indicative functions are used according to the different feature types used when splitting nodes; I is the indicative function, which is 1 when the expression in the brackets is true, otherwise it is 0, T m is a leaf node, S j,m is the segmentation area at this node.

[0099] 11) The K decision feature variables obtained from the M base classifiers are sent together with the original data feature variables into the logistic regression model for linear fitting; the specific process is:

[0100] 5-1) The decision rule is obtained from the base classifier as a new variable, and is imported into the linear function together with the original feature variable for fitting. The objective function F(x) is:

[0101]

[0102] Where a0 represents the fitting deviation, a k Represents the kth rule variable r k The weight of b j is the original feature x in the linear regression model j The weight of , J is the number of original features.

[0103] The estimate of the coefficients in the objective function is calculated by minimizing the loss function of the L1 penalty term:

[0104]

[0105] The term on the left is the model parameter value that minimizes the loss on the right. The function of the argmin function is to obtain the parameter value when the cumulative term on the right is the smallest. λ is used to weigh the proportion of rule variables in the model. The larger the weight, the smaller the impact on the model output.

[0106] 12) Importance measurement of rule variables and original data feature variables; the specific process is:

[0107] 6-1) The importance measure of the rule variable is I K :

[0108]

[0109] Where S k Representation rule r k The sample proportion, a k Represents the kth rule variable r k Weight

[0110] 6-2) For feature variables, its importance metric I j for:

[0111] I j =|b j |std(x j ) (1.10)

[0112] where std(x j ) refers to x j The standard deviation of j is the original feature x in the linear regression model j The weight of .

[0113] Using this classification model, the experiment achieved a precision of 86.6% and a recall of 83.6% for dysphagia analysis.

[0114] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention is described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for obtaining a symptom importance measurement index of dysphagia in patients with Parkinson's disease based on speech, characterized in that Follow these steps: Step 1: Collect speech data of Parkinson's disease patients participating in multi-speech tasks, and label the speech data; Step 2: Perform denoising and enhancement preprocessing on the collected speech data of Parkinson's disease patients, and then perform feature extraction to obtain speech feature vectors that can characterize the data, and construct a data set; Step 3: Select N training samples from the data set, train the gradient boosting tree model GBDT with the classification and regression tree model CART as the base classifier, and obtain M base classifiers; Step 4: Export the trained M base classifiers from the GBDT model, and use the path from its root node to any child node as a decision feature variable r k ; After completing the GBDT model training, each base classifier CART is exported. The path from the root node of the constructed tree model to any child node will be used as a decision feature variable, i.e., a decision rule. The specific process is as follows: 1) For each base classifier obtained, the path from the root node to any node in the tree is used as a decision rule: Where I(·) is an indicator function, which takes the value 1 when the expression in the brackets is true and 0 otherwise. A complete rule variable can be expressed as g m,k (x) = I 节点j (x 切分特征1 <s)*I 节点j+1 (x 切分特征2 >s)I 节点j+ ·(·), that is, the samples are classified by the split nodes and split point values ​​selected during the classification and regression tree construction process. When the sample falls into the above split data domain, the expression is 1; 2) T m is the leaf node set of the classification and regression tree (CART) obtained in the mth iteration, D j,m (h, s) is the segmentation region when the segmentation feature variable h at node j takes the value s; for M trees, the number of nodes in each tree is J+1, and the number of rule variables that can be obtained is Step 5: The K decision feature variables obtained from the M base classifiers are sent together with the original data feature variables into the logistic regression model for linear fitting; The K decision feature variables obtained from the M base classifiers are sent together with the original data feature variables into the logistic regression model for linear fitting. The specific process is as follows: 1) Obtain the decision rule from the base classifier as a new variable and import it into the linear function together with the original feature variable to fit the data pair (x n ,y n ), the mathematical expression of the model Q(x) is: Where e0 represents the fitting deviation, e k represents the kth rule variable g m,k The weight of b h is the original feature x in the linear regression model h The weight of , H represents the number of features; 2) The estimation of the model Q(x) coefficients will be obtained by optimizing the loss function in the following formula: The argmin function is used to obtain the value of the coefficient when the right-hand side term is minimized. λ is used to weigh the weight of the rule variable in the model. The larger the value, the smaller the influence of the rule variable on the model output result. The final output of the model is the data x n The label prediction value of Step 6: Derive the importance measures of rule variables and original data feature variables from the trained logistic regression model; Importance measurement of rule variables and original data feature variables; the specific process is: 1) The importance measure of the rule variable is I K : Among them k Representation rule g m,k The sample ratio, that is, how many samples satisfy rule g m,k , e k represents the kth rule variable g m,k The weight of 2) For feature variables, its importance measure I h for: where std(x j ) refers to The standard deviation of h is the original feature in the linear regression model The weight of Step 7: Output the importance measurement result report of the feature variables to assist doctors in making dysphagia symptom assessment.

2. The method according to claim 1, characterized in that In the step 2, feature extraction is performed on the collected speech data of Parkinson's disease patients, and a data set is constructed. The specific process is as follows: 1) Use speech signal processing algorithms to extract speech features. The extracted features include three categories: ① Based on the continuous vowel pronunciation task: the fundamental frequency F0 of each vowel, the average fundamental frequency F0_ave, the minimum fundamental frequency F0_min, the maximum fundamental frequency F0_max, the jitter Jitter, the amplitude perturbation entropy APQ, the pitch perturbation quotient PPQ, the pronunciation segment ratio, the characteristics of measuring the change of amplitude over time, the noise harmonic ratio NHR, the harmonic noise ratio HNR, the cycle density entropy RPDE, the trend fluctuation analysis DFA, the gene period entropy PPE; the characteristics of measuring the change of amplitude over time include: Shimmer and its variants, the decibel value of Shimmer, the three-point amplitude perturbation entropy ShimmerAPQ3, the five-point amplitude perturbation entropy ShimmerAPQ5, and the 11-point amplitude perturbation entropy ShimmerAPQ11; ② Pronunciation tasks based on rapid voiced-unvoiced conversion: features related to fundamental frequency and their related statistics: average fundamental frequency energy Avg.Energy, standard deviation of energy Std.Energy, as well as voiced segment ratio Voice Rate, silence segment ratio Silence Rate, average duration of silence segment, standard deviation of silence segment duration; ③ Pronunciation tasks based on scene dialogues: Features related to fundamental frequency: calculate the fundamental frequency contour, average, standard deviation and maximum value; features related to energy: calculate the average energy, standard deviation and maximum value of energy; features related to duration: vowel segment proportion, average duration of vowel syllable segments, standard deviation of vowel segment duration, average duration of silent segments, standard deviation of silent segment duration; 2) Dataset construction: For the selected N sample voices, according to the collected label information, form a sample pair (x n ,y n ), where x n represents the nth sample (n=1,2,···N), represents the h-th dimension feature of the n-th sample (h = 1, 2, ··H), y n is the label of the nth sample: when the label is 0, it means that the subject has no dysphagia symptoms, otherwise it is 1.

3. The method according to claim 1, characterized in that In step 3: training the integrated model GBDT using the CART model as the base classifier, the specific process is: 1) Initialize the first weak learner: Where P(y=1|x) is the total number of training samples (x n ,y n ), that is, the learner is initialized using the prior information of the original probability distribution in the training samples; 2) Label prediction value obtained in the current iteration Where F(x n )=F0(x n ) is the output of the learner in this iteration, and the loss function is constructed by the output of the current model: Among them, y n is the true label value, is the predicted label value; 3) The negative gradient r of the loss function obtained in the current iteration m,n for: where r m,n It will be used as the fitting target of the learner in the next iteration, m represents the number of iterations, n is the nth sample, and this iteration ends; 4) A new iteration begins. First, the mth base learner is established. A new CART tree is built to fit the new training data pair (x n ,y n =r m,n ), get the current iteration output F m (x): 4-1) Generation of CART-based learner in a new iteration, selecting the optimal segmentation variable h and segmentation point s, and calculating: Traverse all feature variables h of the data, select one as the split variable, and then traverse the split points s according to its value range, and select the pair (h, s) that minimizes equation (1.3); 4-2) After selecting the split variable A and split point S, the data set is divided into D1(h,s)={x|x (h) ≤s} and D2(h,s)={x|x (h) >s} two sub-regions and determine the output value c of the region j : 4-3) Continue to repeat steps 4-1) and 4-2) for the two sub-regions until the pre-set tree depth is met and the segmentation stops. A new CART base learner is generated, which divides the feature space of the input data into J regions, where J+1 is the number of leaf nodes of the CART model. The best fitting value c for each region is calculated by (1.4) m,j , generates the output F of the current base classifier m (x): 5) Repeat steps 2) to 4) until M CART trees are obtained after M iterations. The output of the final model for the predicted label of the data is in is the integration of the outputs of M base classifiers; 6) At this point, the GBDT model training is completed. The model output of this part will only be used to measure the degree of fit of the model to the data set. The M trees generated by the model will serve as the basis for the subsequent construction of rule variables.

4. The method according to claim 1, characterized in that: The specific process in step seven is: integrating the prediction of the label value by the logistic model and the importance measurement of each feature variable into a sample condition analysis report. The doctor observes the distribution range, importance and category prediction value of the pathological characteristics through the condition analysis report output by the model, and conducts a comprehensive assessment of the patient's symptoms.

5. A speech-based system for obtaining the importance measurement of pathological characteristics of dysphagia symptoms in Parkinson's disease patients, characterized by The system executes the method according to any one of claims 1 to 4.

6. The system according to claim 5, characterized in that The system includes: 1) A voice data collection module; used to perform step 1: collect voice data of Parkinson's disease patients participating in multi-speech tasks, as well as characteristic information of dysphagia of Parkinson's disease patients; 2) Speech data processing module; used to execute step 2: pre-process the speech data for denoising and enhancement, then perform feature extraction, and convert the denoised and enhanced speech into a speech feature vector as a data set; 3) Feature data processing module based on GBDT model; used to execute step 3: select training samples from the data set, train the integrated model GBDT with CART model as the base classifier, obtain M base classifiers, and create a foundation for the subsequent rule variable construction; 4) CART tree-based classifier generation module; used to execute step 4: export the trained base classifier, and use the path from its root node to any child node as a decision feature variable to complete the construction of the decision variable; 5) A linear fitting module based on a logistic regression model; used to execute step 5: sending the K decision feature variables obtained from the M base classifiers together with the original data feature variables into the logistic regression model for linear fitting to obtain the model's prediction of the category of the test sample; 6) Importance measurement export module for rule variables and original data feature variables; used to execute step six: export importance measurement of rule variables and original data feature variables, and generate an analysis report of the test sample together with the prediction of the model.