Speech Analysis Method and System for Key Feature Parameters of Freezing Gait Symptoms in Parkinson's Disease Based on AdaBoost Algorithm

Abstract

The present invention discloses a voice analysis method for key feature parameters of freezing gait symptoms in Parkinson's disease based on the AdaBoost algorithm. Step 1: Collect continuous and stable vowels of Parkinson's disease patients and record whether the Parkinson's disease patients have freezing gait symptoms; Step 2: Perform denoising preprocessing on the voice signals and remove the silent segments; Step 3: Extract various voice features; Step 4: Use the CART algorithm to perform feature selection on the original features and screen out the key features that can effectively represent the information of freezing gait symptoms; Step 5: Train the AdaBoost model; Step 6: Input the feature vector of the voice to be measured into the model to obtain the key feature parameters of the freezing gait symptoms in Parkinson's disease. The present invention uses the AdaBoost algorithm to analyze the freezing gait symptoms in Parkinson's disease, improves the model accuracy by using ensemble learning, and reduces the cost of early analysis of the freezing gait symptoms in Parkinson's disease.

Description

Technical Field

[0001] The present invention belongs to the application of the machine learning field in medicine, and relates to a voice analysis method and system for key feature parameters of the freezing gait symptom of Parkinson's disease based on the AdaBoost algorithm. Background Art

[0002] Parkinson's disease is the second most common neurodegenerative disease after Alzheimer's disease. Its main symptoms include motor symptoms and non-motor symptoms. Motor symptoms include muscle stiffness, tremors, and some other movement disorders. Non-motor symptoms mainly include hyposmia, constipation, sleep behavior disorders, and depression, etc. These symptoms are caused by the reduction of a large number of dopaminergic neurons.

[0003] Freezing gait is one of the most severe movement disorder symptoms of Parkinson's disease, which refers to the short-term, sudden interruption or significant reduction of the patient's steps when attempting to walk or during the forward movement. The most common symptom manifestation is that the patient hesitates at the start, has difficulty taking a step, and suddenly has difficulty walking. Freezing gait is disabling. Most patients need to use a wheelchair for activities on average five years after the symptom appears, which greatly affects the patient's quality of life.

[0004] Existing research has shown that there is a certain pathological connection between Parkinson's disease and dysarthria. Voice impairment may be one of the earliest signs of Parkinson's disease, and its symptoms include slow speech, hoarseness, low volume, and vocal tremors, etc. These voice impairments are caused by the loss of control of the laryngeal, phonatory, and respiratory muscles in Parkinson's disease patients.

[0005] Voice is mainly produced by the cooperation of the internal vocal organs in the body, and the cooperation between human organs is coordinated and controlled by neurons uniformly. Due to the loss of relevant neurons, Parkinson's disease patients are unable to stably control the vocal organs, resulting in varying degrees of vocal disorders in patients. Compared with healthy people, Parkinson's disease patients often cannot produce a stable and accurate voice. Therefore, voice signals can be used for early symptom analysis of Parkinson's disease. Compared with traditional Parkinson's disease analysis methods, using voice signals to analyze Parkinson's disease is economical and efficient. Voice monitoring is non-contact, simple, and convenient. For pronunciation, sustained vowels / a / , / i / , / o / can be used. The sustained vowel / a / is the easiest to pronounce, and experience has shown that it can convey the most clinically useful information. Physiologically, the vowel / a / involves the combination of various muscles in the vocal cords and vocal tract, so it increases the probability that nerve problems can be identified. When using voice signals to analyze Parkinson's disease, it is necessary to analyze the voice signals to be measured through voice signal processing algorithms to extract voice feature information that can characterize the pathological features of Parkinson's disease.

[0006] Based on the above situation, after extracting speech features through a speech signal processing algorithm, the number of extracted speech features is often large. Among them, there may be irrelevant features, and there may also be correlations between features, which easily leads to the curse of dimensionality. Feature selection can eliminate irrelevant or redundant features and select key features useful for the current classification task, thereby achieving the purpose of improving the model accuracy, reducing the model prediction time, and providing key features with certain interpretability. Then, technical models in the field of machine learning can be used to analyze the freezing gait symptoms of Parkinson's disease. Summary of the Invention

[0007] Aiming at the problem of difficult analysis of freezing gait symptoms in traditional Parkinson's disease, the present invention proposes a speech analysis method and system for key feature parameters of freezing gait symptoms in Parkinson's disease based on AdaBoost. Feature extraction is performed on the collected speech signal, and interpretable feature selection is carried out. Then, combined with the ensemble learning algorithm in machine learning, the freezing gait of Parkinson's disease patients is analyzed to realize the interpretable key feature analysis of the freezing gait symptoms of Parkinson's disease patients, so as to timely adopt subsequent specific treatment strategies.

[0008] To achieve the above objectives, the method of the present invention includes the following steps:

[0009] Step 1: Collect continuous and stable vowels of Parkinson's disease patients and record whether the Parkinson's disease patients have freezing gait symptoms.

[0010] Step 2: Denoise the speech signal and remove the silent segments.

[0011] Step 3: Use a speech signal processing algorithm to extract various speech features.

[0012] Use a speech signal processing algorithm to extract speech features. The extracted features include: fundamental frequency contour F0_contour, average fundamental frequency F0_ave, minimum fundamental frequency F0_min, maximum fundamental frequency F0_max, four features Jitter, RAP, PPQ, DDP for measuring fundamental frequency changes, five features Shimmer, APQ3, APQ5, APQ11, DDA for measuring amplitude changes, noise-to-harmonic ratio NHR, harmonic-to-noise ratio HNR, recurrence period density entropy RPDE, detrended fluctuation analysis DFA, permutation period entropy PPE, and Mel frequency cepstral coefficients MFCC obtained by transforming the features in the Mel cepstral domain.

[0013] Step 4: Use the CART algorithm for feature selection to screen out interpretable key features that can effectively characterize the freezing gait symptom information.

[0014] The process of using the CART algorithm for feature selection is as follows:

[0015] The specific formula for the Gini index Gini(D) of the dataset D is as follows:

[0016]

[0017] where p k represents the probability that a sample point belongs to the k-th class, and K represents K classification problems;

[0018] The Gini index Gini index (D, a) of the feature a is defined as:

[0019]

[0020] where V represents that the feature a has V possible values; therefore, the feature value that makes the Gini index the smallest after partitioning is selected as the optimal partitioning feature a * , that is:

[0021] a * = argmax a∈A Gini index (D, a) (3)

[0022] Step Five: Train the AdaBoost model.

[0023] Step Six: Voice analysis: Input the feature vector of the voice to be measured into the model to obtain the key feature parameters of the freezing gait symptoms of the person to be measured.

[0024] The present invention also discloses a voice analysis system for key feature parameters of Parkinson's disease freezing gait symptoms based on the AdaBoost algorithm. The system includes:

[0025] A voice signal acquisition module, which is used to perform Step One: Acquisition of voice signals: Acquire continuous and stable vowels of Parkinson's disease patients and record whether the Parkinson's disease patients have freezing gait symptoms;

[0026] A voice signal processing module, which is used to perform Step Two: Preprocessing of voice signals: Denoise the voice signals and remove the silent segments;

[0027] A voice feature extraction module, which is used to perform Step Three: Voice feature extraction: Extract various voice features by using a voice signal processing algorithm;

[0028] A voice feature selection module, which is used to perform Step Four: Feature selection: Use the CART algorithm for feature selection to screen out the key features that can characterize the freezing gait symptoms;

[0029] An AdaBoost classification model training module, which is used to perform Step Five: Training the model: Use a decision tree as a base classifier to train the AdaBoost classification model;

[0030] A voice analysis module, configured to perform Step Six, voice analysis: input the feature vector of the voice to be measured into the model to obtain the key feature parameters of the freezing gait symptoms of the person to be measured.

[0031] The method for analyzing Parkinson's disease freezing gait based on the AdaBoost algorithm provided by the present invention has the following beneficial effects:

[0032] 1. It reduces the cost of early analysis of freezing gait in Parkinson's disease patients. Since the present invention extracts features from voice signals and then uses machine learning methods to analyze freezing gait symptoms, it avoids the high cost of going to the hospital for diagnosis. Instead, it only needs to obtain voice signals and gait feature information of Parkinson's disease patients for model establishment, and then early analysis of freezing gait symptoms can be carried out, saving the diagnosis cost.

[0033] 2. It improves the efficiency of analyzing Parkinson's disease freezing gait symptoms. The traditional method for analyzing Parkinson's disease is for doctors to conduct a series of tests on patients in aspects such as movement and tremors and comprehensively analyze based on the patient's performance whether the patient has freezing gait symptoms of Parkinson's disease, the degree of the disease, and the possible disease development trend. The disease analysis means are complex and time-consuming. The present invention is obtained through machine learning methods, and by effectively selecting voice features, the dimension of the input voice feature vector can be reduced, improving the analysis efficiency of the machine learning algorithm program.

[0034] 3. It provides a reliable means for analyzing the key feature parameters of Parkinson's disease freezing gait symptoms. The traditional analysis of freezing gait symptoms is through doctors' tests on patients in multiple aspects such as movement and tremors. Since the motor symptoms in the early stage of Parkinson's disease are not obvious, the analysis accuracy is greatly affected. The present invention utilizes the physiological connection between voice and Parkinson's disease, analyzes the early freezing gait symptoms based on the patient's voice, with little subjective influence, and through CART feature selection, screens out interpretable key voice features as reference values for doctors to analyze and evaluate the condition. Description of the Drawings

[0035] Figure 1 It is the classifier training flowchart of the present invention;

[0036] Figure 2 It is the system module block diagram of the present invention. Detailed Embodiments

[0037] The following experiments are used to verify the beneficial effects of the present invention:

[0038] In this experiment, the Parkinson's disease patient dataset collected by the present invention was selected as the research object. A total of 212 voice samples from 53 Parkinson's disease patients were collected in this dataset. After denoising preprocessing, various voice features were extracted. Each sample contains 59 voice features, and each voice feature contains 7 statistical values; the sample contains a label to identify the gait feature information, where 1 indicates the presence of freezing gait symptoms and 0 indicates the absence of freezing gait. As Figure 1 shown, according to the attributes in the dataset, the voice-based Parkinson's disease freezing gait analysis method is carried out according to the following steps:

[0039] Step 1: Collect the continuous and stable vowels of Parkinson's disease patients and record whether there are freezing gait symptoms in Parkinson's disease patients.

[0040] Step 2: Preprocess the voice signal, including denoising processing and removing silent segments.

[0041] Step 3: Use voice signal processing algorithms to extract various voice features, and each sample contains 59 voice features.

[0042] Step 4: Feature selection: Use the CART algorithm for feature selection to screen out the key features that can characterize freezing gait symptoms.

[0043] The decision tree depth of the CART algorithm is set to 5, the impurity calculation method is the Gini coefficient, and the minimum number of samples required for the branching sub-nodes is set to 3. Run the CART algorithm 10 times to obtain the key features selected on average, including: the 9th and 11th frequency ranges mfcc_mean_9 and mfcc_mean_11 of the Mel Frequency Cepstral Coefficient MFCC, the fundamental frequency contour f0_contour_mean, the fundamental frequency perturbation feature DDP, the detrended fluctuation analysis DFA, Shimmer, etc.

[0044] The dataset after feature selection is divided into a training set and a test set. The five-fold cross-validation method is adopted, and the dataset is randomly divided into five mutually exclusive data subsets of similar sizes. Each time, the union of four subsets is selected as the training set to train the decision tree model, and the remaining one subset is used as the test set to test the model performance. Finally, the mean of the test results of the five groups of training / test sets is used as the performance metric of the model.

[0045] Step 5: Train the model: Use the decision tree as the base classifier to train the AdaBoost classification model.

[0046] Before model training, the training set and test set data are normalized first to map all data into the numerical range of [0, 1]. The functions of the normalization algorithm are as follows: 1. Each attribute in the dataset has an actual physical background, so their units and ranges are different. Normalization can eliminate the influence of units or orders of magnitude, map all data into a pre-defined range, and facilitate subsequent data processing; 2. Normalization can improve the program running speed and accelerate convergence; 3. Singular sample data (sample vectors that are particularly large or small compared to other input samples) may increase the training time and even cause the algorithm to fail to converge. Normalization before training can eliminate the influence of singular sample data on the training process.

[0047] The specific process of constructing the AdaBoost model is as follows:

[0048] 1) Initialize the weight distribution of the training data. Each training sample is initially assigned the same weight: 1 / N;

[0049]

[0050] where w 1i represents the weight of the i-th training sample at the beginning, and N represents the total number of samples;

[0051] 2) Perform M iterations. Each iteration performs the following steps:

[0052] a. Use the training dataset with the weight distribution D n to learn and obtain the base classifier G m (x):

[0053] G m (x): χ → {-1, +1}

[0054] b. Calculate the classification error rate e m of G m on the training dataset:

[0055]

[0056] where G m (x i ) represents the classification result of the base classifier G m (x) on the training data x i , y i represents the true classification of the training data x i , w mi represents the weight of the sample x i at the m-th iteration, and P() represents the probability of a certain event; <() represents that the result is 1 when the event in the parentheses is true, otherwise the result is 0;

[0057] c. Calculate G m Coefficient of (x) to obtain the weight α of the basic classifier in the final classifier m :

[0058]

[0059] where e m represents the classification error rate of G m (x) on the training dataset;

[0060] d. Update the weight distribution of the training dataset:

[0061] D m+1 =(w m+1,1 , w m+1,2 , …, w m+1,i , …, w m+1,N )

[0062]

[0063]

[0064] where w m+1,i represents the updated weight of the +th training sample after m iterations; Z m represents the normalization factor, and exp() represents the exponential function with the natural constant e as the base;

[0065] 3) Combine all the basic classifiers to obtain the final classifier, and the final classification result is obtained by weighted voting of all the basic classifiers:

[0066]

[0067] where f(x) represents the weighted combination of each basic classifier, G(x) represents the final classifier, and sign() represents the sign function.

[0068] Step Six: Speech Analysis: Input the feature vector of the speech to be tested into the model to obtain the key feature parameters of the freezing gait symptoms of the person to be tested. Here, 100 speech samples to be tested are selected for speech analysis, and some key feature parameter values of the freezing gait are shown in the following table.

[0069]

[0070]

[0071] After selecting the key speech features, use these features for early analysis of the freezing gait symptoms, and the test accuracy rate is 87.6%.

[0072] Figure 2The system module block diagram of the present invention is given, and the system includes:

[0073] A voice signal acquisition module; used to execute Step 1, acquisition of voice signals: acquire continuous and stable vowels of Parkinson's disease patients, and record whether Parkinson's disease patients have freezing gait symptoms;

[0074] A voice signal processing module; used to execute Step 2, preprocessing of voice signals: perform noise reduction processing on voice signals and remove silent segments;

[0075] A voice feature extraction module; used to execute Step 3, extraction of voice features: extract various voice features using voice signal processing algorithms;

[0076] A voice feature selection module; used to execute Step 4, feature selection: perform feature selection using the CART algorithm to screen out key features that can characterize freezing gait symptoms;

[0077] An AdaBoost classification model training module; used to execute Step 5, training of the model: use a decision tree as the base classifier to train the AdaBoost classification model;

[0078] A voice analysis module; used to execute Step 6, voice analysis: input the feature vector of the voice to be measured into the model to obtain the key feature parameters of the freezing gait symptoms of the person to be measured.

[0079] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions recorded in the foregoing embodiments or perform equivalent replacements on some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A speech analysis method for key feature parameters of the freezing gait symptom of Parkinson's disease based on the AdaBoost algorithm, characterized in that it is Performed according to the following steps: Step 1, Acquisition of voice signals: Continuously and stably acquire vowels of Parkinson's disease patients, and record whether the Parkinson's disease patients have freezing gait symptoms; Step 2, Preprocessing of voice signals: Denoise the voice signals and remove silent segments; Step 3, Extraction of voice features: Use voice signal processing algorithms to extract various voice features; Step 4, Feature selection: Use the CART algorithm for feature selection to screen out key features that can characterize freezing gait symptoms; The specific process of the said Step 4 is: The specific formula for the Gini index Gini(D) of the dataset D is: where p k represents the probability that the sample point belongs to the k-th class, and K represents K classification problems; The Gini index Gini of feature a index (D, a) is defined as: Where V represents that feature a has V possible values; therefore, select the feature value of feature a that minimizes the Gini index after partitioning as the optimal partitioning feature a * , that is: a * = argmax a∈A Gini index (D, a); Step 5, Training the model: Use a decision tree as the base classifier to train the AdaBoost classification model; The specific process of the said Step 5 is: 1) Initialize the weight distribution of the training data. Each training sample is initially given the same weight: 1 / N; where w 1i represents the weight at the start of the i-th training sample, and N represents the total number of samples; 2) Perform M iterations, and perform the following steps for each iteration: a. Use a training data set with a weight distribution D m to learn and obtain a base classifier G m (x): G m (x): x → {-1, +1} b. Calculate G m (x) The classification error rate e on the training data set m : Among them, G m (x i ) represents the classification result of the base classifier G m (x) on the training data x i , and y i represents the true classification of the training data x i . w mi represents the weight of the sample x i at the m-th iteration. P() represents the probability of an event; I() represents that the result is 1 when the event in the parentheses is true, and 0 otherwise; c. Calculate G m the coefficient of (x) to obtain the weight α of the basic classifier in the final classifier m : where e m represents the classification error rate of G m (x) on the training data set; d. Update the weight distribution of the training dataset: D m+1 = (w m+1,1 , w m+1,2 , …, w m+1,i , …, w m+1,N ) where w m+1,i represents the updated weight after the i-th training sample is iterated m times; Z m represents the normalization factor, and exp() represents the exponential function with the natural constant e as the base; 3) Combine all the base classifiers to obtain the final classifier. The final classification result is obtained by weighted voting of all the base classifiers: where f(x) represents the weighted combination of each base classifier, G(x) represents the final classifier, and sign() represents the sign function; Step 6, Voice analysis: Input the feature vector of the voice to be measured into the model to obtain the key feature parameters of the freezing gait symptoms of the person to be measured.

2. The method according to claim 1, characterized in that, The specific process of the said Step 3 is: Use voice signal processing algorithms to extract various voice features. The extracted features include: fundamental frequency contour F0_contour, average fundamental frequency F0_ave, minimum fundamental frequency F0_min, maximum fundamental frequency F0_max, four features Jitter, RAP, PPQ, DDP that measure fundamental frequency changes, five features Shimmer, Shimmer:APQ3, Shimmer:APQ5, Shimmer:APQ11, Shimmer:DDA that measure amplitude changes, noise-to-harmonic ratio NHR, harmonic-to-noise ratio HNR, recurrence period density entropy RPDE, detrended fluctuation analysis DFA, periodicity entropy PPE; and Mel cepstral coefficients MFCC obtained by converting the voice in the Mel cepstral domain.

3. A speech analysis system for key characteristic parameters of the freezing gait symptom of Parkinson's disease based on the AdaBoost algorithm, characterized in that The system executes the method described in any one of claims 1-2.

4. The system according to claim 3, wherein The system includes: A voice signal acquisition module, used to execute Step 1, Acquisition of voice signals: Continuously and stably acquire vowels of Parkinson's disease patients, and record whether the Parkinson's disease patients have freezing gait symptoms; A voice signal processing module, used to execute Step 2, Preprocessing of voice signals: Denoise the voice signals and remove silent segments; A voice feature extraction module, used to execute Step 3, Extraction of voice features: Use voice signal processing algorithms to extract various voice features; A voice feature selection module, used to execute Step 4, Feature selection: Use the CART algorithm for feature selection to screen out key features that can characterize freezing gait symptoms; An AdaBoost classification model training module, used to execute Step 5, Training the model: Use a decision tree as the base classifier to train the AdaBoost classification model; A voice analysis module, which is used to execute Step Six, voice analysis: input the feature vector of the voice to be measured into the model to obtain the key feature parameters of the freezing gait symptoms of the person to be measured.

Images