A Parkinson's disease sample feature classification method based on local field potential, feature classification system and sample classification model training method

Abstract

The present invention discloses a characteristic classification method for Parkinson's disease samples based on local field potential, a characteristic classification system and a sample classification model training method. The classification method comprises: calculating the power spectral density of the local field potential Beta frequency band signal of the sample to obtain an average power spectral density curve and performing peak analysis on the curve to obtain a peak power spectral density; modeling the periodic component and the non-periodic component of the average power spectrum respectively, extracting the periodic power spectrum and the non-periodic component; calculating the phase-amplitude coupling characteristics of the local field potential Beta frequency band signal of the sample; performing feature cluster analysis using the peak power spectral density, periodic power spectrum, non-periodic component and phase-amplitude coupling characteristics, and outputting the sample classification results. The present invention constructs a multidimensional data model by extracting and integrating advanced neurophysiological features such as peak power spectral density, periodic oscillation characteristics, non-periodic background characteristics and phase-amplitude coupling, thereby improving the accuracy and reliability of classification.

Description

Technical Field

[0001] The present invention belongs to the field of neuroscience, and more specifically, relates to a Parkinson's disease sample feature typing method based on local field potential, a feature typing system, and a sample classification model training method. Background Art

[0002] Parkinson's disease (PD) is a common neurodegenerative disease that primarily affects middle-aged and elderly people. Clinically, PD is characterized by its typical motor symptoms—tremor, rigidity, bradykinesia, and postural instability—as well as non-motor symptoms such as cognitive dysfunction, depression, and sleep disorders. Despite similar clinical manifestations, PD exhibits significant heterogeneity among different individuals, making early diagnosis, disease classification, and personalized treatment a challenge in medical research. The aging of the global population has led to an increasing prevalence of PD, placing a heavy burden on society and families. Therefore, exploring the different subtypes of PD and achieving more effective intervention and treatment through precise classification has become an important research direction in the current field of neuroscience.

[0003] Currently, the diagnosis of Parkinson's disease mainly relies on clinical assessment tools. However, the clinical manifestations of Parkinson's disease are highly individual, and patients may present with completely different combinations of symptoms at different stages of the disease, making classification and prediction based solely on clinical symptoms very difficult. Studies have shown that Beta-band local field potentials (LFPs) are closely related to motor symptoms, especially during motor control and motor planning. Changes in Beta wave activity can reflect the patient's motor function status. Although EEG signals have made some progress in Parkinson's disease research, most existing studies focus on the correlation analysis between EEG features and single clinical indicators, lacking comprehensive modeling of multiple clinical characteristics and neural activity data.

[0004] The prior art invention patent with publication number CN116458898A proposes a method and system for extracting features of Parkinson's disease depression, including: acquiring EEG signals, processing the EEG signals to obtain local field potentials, and preprocessing the local field potentials to obtain local field potentials to be processed; performing continuous wavelet transform on the local field potentials to be processed, retaining the alpha, lowbeta, highbeta, and beta frequency bands; extracting burst features of the above-mentioned frequency bands, processing the local field potentials using continuous wavelet transform, extracting the burst characteristics of the alpha, lowbeta, highbeta, and beta frequency bands, and using the Parkinson's disease depression feature extraction system to achieve high-precision classification of Parkinson's depression degree based on the extracted burst features. This solution only uses wavelet transform to extract burst features, and the feature extraction is not comprehensive. Summary of the Invention

[0005] In order to overcome the shortcomings of the existing technology, including the failure to comprehensively model multiple clinical characteristics and neural activity data, and the problem of incomplete feature extraction, the present invention provides a Parkinson's disease sample feature typing method based on local field potential, a feature typing system and a sample classification model training method.

[0006] The primary purpose of the present invention is to solve the above technical problems, and the technical solutions of the present invention are as follows:

[0007] A first aspect of the present invention provides a method for characterizing Parkinson's disease samples based on local field potentials, comprising the following steps:

[0008] The power spectrum density of the Beta frequency band signal of the local field potential of the sample is calculated to obtain the average power spectrum density curve;

[0009] Perform peak analysis on the average power spectrum density curve to obtain the peak power spectrum density;

[0010] Model the periodic component and non-periodic component of the average power spectrum density curve separately, and extract the periodic power spectrum and non-periodic component;

[0011] Calculate the phase-amplitude coupling characteristics of the Beta-band signal of the sample's local field potential;

[0012] The peak power spectrum density, periodic power spectrum, non-periodic component and phase-amplitude coupling characteristics are used to perform feature clustering analysis and output the sample classification results.

[0013] Furthermore, the power spectral density of the Beta-band signal of the local field potential of the sample is calculated using the Welch method, which specifically includes the following steps:

[0014] The original local field potential signal x[n] is framed and the windowed signal is calculated using the Hanning window function w[n]. The expression is as follows:

[0015] x i [n] = x[n+i·N]·w[n]

[0016] Where n represents the index of the sample within the frame, n∈[1,N-1], x i [n] represents the windowed signal of the i-th frame, x[n] represents the original signal, w[n] represents the Hanning window, and N is the length of the frame;

[0017] Apply discrete Fourier transform to each frame of windowed signal and calculate the frequency domain signal X i [f], the expression is as follows:

[0018] X i [f] = FFT(xi [n])

[0019] Where f represents frequency, FFT represents discrete Fourier transform, and x i [n] is the windowed signal of the i-th frame;

[0020] The power spectrum density P of the i-th frame is obtained by taking the modulus square of the frequency domain signal of each frame. i [f], the expression is as follows:

[0021] P i [f]=|X i [f]| 2

[0022] The power spectra of all frames are averaged to obtain the average power spectrum density curve P avg [f], the expression is as follows:

[0023]

[0024] Where N is the number of frames.

[0025] Furthermore, the periodic component and the non-periodic component of the average power spectrum density curve are modeled separately to extract the periodic power spectrum and the non-periodic component, which specifically includes the following steps:

[0026] The periodic oscillation components of the power spectrum are fitted using Gaussian functions to obtain the periodic power spectrum density P periodic (f), the expression is as follows:

[0027]

[0028] Where f represents frequency, A represents the amplitude of oscillation, CF represents the center frequency of oscillation, and BW represents the bandwidth of oscillation;

[0029] The 1 / f-like function is used to fit the non-periodic component of the power spectrum to obtain the non-periodic power spectrum density P aperiodic (f), the expression is as follows:

[0030]

[0031] Where f represents frequency, offset represents the offset of the non-periodic component, and exponent represents the exponent of the non-periodic component;

[0032] Combining the periodic power spectrum density and the non-periodic power spectrum density, a complete power spectrum density model is constructed. The model parameters are optimized by the least square method to minimize the mean square error and obtain the spectrum curve of the fitted power spectrum density.

[0033] The non-periodic power spectrum density is subtracted from the fitted power spectrum to obtain the periodic power spectrum, and the non-periodic component is extracted from the non-periodic fitting curve.

[0034] Furthermore, the phase-amplitude coupling characteristics of the local field potential Beta signal are calculated, specifically including the following steps:

[0035] The instantaneous phase of the low-frequency band of the Beta signal is obtained by Hilbert transform; the instantaneous amplitude of the high-frequency band of the Beta signal is obtained by Hilbert transform;

[0036] The relationship between the probability distribution of high-frequency amplitude and low-frequency phase is established using a preset method to obtain the phase-amplitude coupling characteristics.

[0037] Furthermore, a preset method for establishing a relationship between the probability distribution of high-frequency amplitude and low-frequency phase specifically includes the following steps:

[0038] Divide the low-frequency phase into n equal-width intervals;

[0039] In each phase interval, the average value of the high-frequency signal amplitude is calculated, the distribution histogram of the amplitude in the phase is constructed, and the distribution histogram is normalized. The expression is as follows:

[0040]

[0041] Among them, P norm,i is the normalized probability distribution of the ith phase interval, P i is the amplitude histogram frequency in the i-th interval, ∑ j P j is the sum of the amplitude histograms in all phase bins;

[0042] The Kullback-Leibler distance is used to calculate the degree to which the distribution deviates from the uniform distribution. KL , the expression is as follows:

[0043]

[0044] Where n represents the number of intervals and U represents the uniform distribution probability;

[0045] The phase-amplitude coupling characteristic MI is calculated by KL distance normalization, and the expression is as follows:

[0046]

[0047] Furthermore, feature cluster analysis is performed using peak power spectrum density, periodic power spectrum, non-periodic component, and phase-amplitude coupling features, specifically including the following steps:

[0048] Perform dimensionality reduction on the extracted peak power spectrum density, periodic power spectrum, non-periodic component and phase-amplitude coupling features to obtain the reduced-dimensional features;

[0049] Input the reduced-dimensional features into n clustering models and output the sample typing results;

[0050] Calculate the quality indicators of sample typing results, evaluate clustering quality, and select the best clustering scheme.

[0051] A second aspect of the present invention provides a Parkinson's disease sample characteristic typing system based on local field potentials, comprising a memory and a processor, wherein the memory comprises a Parkinson's disease sample characteristic typing method program based on local field potentials, and when the Parkinson's disease sample characteristic typing method program based on local field potentials is executed by the processor, steps of a Parkinson's disease sample characteristic typing method based on local field potentials are implemented.

[0052] A third aspect of the present invention provides a method for training a Parkinson's disease sample classification model, which uses a classification result obtained by a Parkinson's disease sample feature classification method based on local field potential for training, comprising the following steps:

[0053] Using the sample rating scale as input features, m classification models are trained separately, samples are classified according to the classification results, and the optimal model configuration is searched through hyperparameter optimization;

[0054] The leave-one-out cross-validation method was used to evaluate the model performance, output the best classification model and verify its interpretability.

[0055] Furthermore, the hyperparameter optimization search method is Bayesian optimization.

[0056] Furthermore, the method for verifying the interpretability of the model includes the following steps:

[0057] Use the SHAP method to calculate the impact of each input feature on the model classification results and output the SHAP importance ranking graph;

[0058] The SHAP importance ranking diagram was verified by combining statistical analysis, and the inter-group difference significance test was used to evaluate the impact of features on the classification results to ensure the interpretability of the model.

[0059] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0060] This paper proposes a comprehensive analysis method that integrates beta-band EEG local field potential characteristics with multidimensional clinical features to accurately classify Parkinson's disease samples. By extracting and integrating advanced neurophysiological features such as peak power spectral density, periodic oscillation characteristics, aperiodic background characteristics, and phase-amplitude coupling (PAC), a multidimensional data model is constructed to improve the accuracy and reliability of classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to make the purpose and technical solution of the present invention clearer, the present invention provides the following drawings and descriptions:

[0062] Figure 1 A flow chart of the typing method provided in an embodiment of the present invention;

[0063] Figure 2 A schematic diagram of peak power spectrum extraction provided by an embodiment of the present invention;

[0064] Figure 3 A schematic diagram of extracting periodic power spectrum and non-periodic components provided by an embodiment of the present invention;

[0065] Figure 4 A comparison chart of indicators of different clustering methods provided by the embodiment of the present invention;

[0066] Figure 5 A comparison chart of clustering results provided by an embodiment of the present invention;

[0067] Figure 6 A comparison chart of machine learning prediction results provided by an embodiment of the present invention;

[0068] Figure 7 The SHAP model provided in the embodiment of the present invention can interpret the analysis diagram. DETAILED DESCRIPTION

[0069] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0070] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0071] Example 1:

[0072] The present invention provides a method for characterizing Parkinson's disease samples based on local field potentials, such as Figure 1The figure shows a flow chart of a method for characterizing Parkinson's disease samples based on local field potentials. The specific steps are as follows:

[0073] S1: Calculate the power spectral density of the Beta frequency band signal of the local field potential (LFP) of the sample to obtain the average power spectral density curve.

[0074] More specifically, the method for calculating the power spectral density of the sample's local field potential (LFP) Beta frequency band signal is the Welch method, and the specific process is as follows:

[0075] The original local field potential signal x[n] is framed, with each frame length of 1 second. The windowed signal is calculated using the Hanning window function w[n], and the expression is as follows:

[0076] x i [n] = x[n+i·N]·w[n]

[0077] Where n represents the index of the sample within the frame, n∈[1,N-1], x i [n] represents the windowed signal of the i-th frame, x[n] represents the original signal, w[n] represents the Hanning window, and N is the length of the frame;

[0078] Apply discrete Fourier transform (FFT) to each frame of windowed signal to calculate the frequency domain signal X i [f], the expression is as follows:

[0079] X i [f] = FFT(x i [n])

[0080] Where f represents frequency, x i [n] is the windowed signal of the i-th frame;

[0081] The power spectrum density P of the i-th frame is obtained by taking the modulus square of the frequency domain signal of each frame. i [f], the expression is as follows:

[0082] P i [f]=|X i [f]| 2

[0083] The power spectra of all frames are averaged to obtain the average power spectrum density curve P avg [f], the expression is as follows:

[0084]

[0085] Where N is the number of frames.

[0086] S2: Perform peak analysis on the average power spectral density curve to obtain the peak power spectral density.

[0087] In this embodiment, the peak of the Beta frequency band is first searched, and then the frequency band around the peak is calculated by 5 Hz, and the peak power spectrum density is calculated. The reason for choosing this indicator is that it is a parameter that is tracked in the long term in a set of sensor systems and has potential applications in adaptive DBS. The schematic diagram is shown in FIG. Figure 2 shown.

[0088] S3: Model the periodic and non-periodic components of the average power spectrum density curve separately, and extract the periodic power spectrum and non-periodic components.

[0089] The specific process is:

[0090] The periodic oscillation components of the power spectrum are fitted using Gaussian functions to obtain the periodic power spectrum density P periodic (f), the expression is as follows:

[0091]

[0092] Where f represents frequency, A represents the amplitude of oscillation, CF represents the center frequency of oscillation, and BW represents the bandwidth of oscillation;

[0093] The 1 / f-like function is used to fit the non-periodic component of the power spectrum to obtain the non-periodic power spectrum density P aperiodic (f), the expression is as follows:

[0094]

[0095] Where f represents frequency, offset represents the offset of the non-periodic component, and exponent represents the exponent of the non-periodic component.

[0096] The periodic power spectral density and the non-periodic power spectral density are combined to construct a complete power spectral density model. The model parameters are optimized by the least square method to minimize the mean square error and obtain the spectrum curve of the fitted power spectral density. The expression is as follows:

[0097]

[0098] Among them, P i is the actual power spectrum, P(f i ) is the preliminary fitting power spectrum of the model, and N is the number of frequency points;

[0099] Then the feature extraction of power spectrum parameterization is performed. After parameterizing the power spectrum, the spectrum curve of the fitted power spectrum density can be obtained, and the non-periodic fitting curve can also be drawn, such as Figure 3As shown, the periodic power spectrum (the orange shaded area in the figure) is obtained by subtracting the aperiodic power spectral density (the area under the aperiodic fitting curve) from the fitted power spectrum, and the aperiodic components (offset and index) are extracted from the aperiodic fitting curve.

[0100] S3: Calculate the phase-amplitude coupling characteristics of the Beta frequency band signal of the sample's local field potential (LFP);

[0101] The specific process is:

[0102] The instantaneous phase φ(t) obtained by Hilbert transform in the low-frequency band of the Beta signal is expressed as follows:

[0103] φ(t)=arg[Hilbert(low-frequency Beta signal)]

[0104] The Hilbert transform is used to obtain the instantaneous amplitude of the high frequency band of the Beta signal (300-400Hz, 400-500Hz). h (t)|, the expression is as follows:

[0105] |A h (t)|=|Hilbert (high frequency HFO signal)|

[0106] The relationship between the probability distribution of high-frequency amplitude and low-frequency phase is established using a preset method to obtain the phase-amplitude coupling (PAC) feature, which specifically includes the following steps:

[0107] The low-frequency phase is divided into n equal-width intervals, 18 in this embodiment, corresponding to every 20° in [0, 2π]. In each phase interval, the average value of the high-frequency signal amplitude is calculated, and a distribution histogram of the amplitude in the phase is constructed and normalized. The expression is as follows:

[0108]

[0109] Among them, P norm,i is the normalized probability distribution of the ith phase interval, P i is the amplitude histogram frequency in the i-th interval, ∑ j P j is the sum of the amplitude histograms in all phase bins;

[0110] The Kullback-Leibler (KL) distance is used to calculate the degree to which the distribution deviates from the uniform distribution. KL , the expression is as follows:

[0111]

[0112] Where n represents the number of intervals, and U represents the uniform distribution probability, which is usually 1 / n;

[0113] The phase-amplitude coupling characteristic MI is calculated by KL distance normalization, and the expression is as follows:

[0114]

[0115] S4: Perform feature clustering analysis using peak power spectrum density, periodic power spectrum, non-periodic components, and phase-amplitude coupling features to output sample classification results.

[0116] In this embodiment, the specific process is:

[0117] The peak power spectral density, periodic power spectrum, non-periodic component and phase-amplitude coupling characteristics of the extracted 55 patient samples were subjected to principal component analysis (PCA) dimensionality reduction to obtain the reduced-dimensional features.

[0118] The reduced features are input into n clustering models respectively. In this embodiment, four clustering methods are considered, including hierarchical clustering (HAC), K-Means, partitioned K-Means (BiKMeans), and hierarchical DBSCAN (HDBSCAN). The number of clusters k is set to 2, and the sample typing results are output;

[0119] The silhouette coefficient and Davies-Bouldin index of the sample typing results were calculated to evaluate the clustering quality and select the best clustering scheme.

[0120] Silhouette measures the quality of a cluster by calculating how similar a point is to its own cluster, and then averages the result over the entire dataset. A score close to -1 indicates a meaningless grouping, while a score close to +1 indicates a completely independent and compact grouping, as shown below:

[0121]

[0122] Where a(i) is the average distance between the i-th data point and other points in the same cluster, b(i) is the minimum average distance between the i-th data point and other points in different clusters, and N is the total number of points.

[0123] The Davies-Bouldin index (DB index) is also an indicator for measuring clustering performance. It reflects the balance between intra-class compactness and inter-class separation. The expression is as follows:

[0124]

[0125] Where k is the number of clusters, σ i , σj are the average intra-class distances of cluster i and cluster j, respectively, indicating the average distance from the sample to the centroid in the cluster; d ij is the Euclidean distance between the centroids of cluster i and cluster j. The closeness and separation between clusters are evaluated by calculating the maximum similarity index between each cluster and the other clusters and taking the average. The smaller the DB index, the closer the samples within a cluster, the greater the separation between clusters, and the better the clustering effect. The larger the DB index, the worse the clustering performance.

[0126] By comparing the clustering effects of various methods, the best clustering method is selected. The clustering index results of the present invention are as follows: Figure 4 As shown in Figure 2, K-Means performs best overall, and the final clustering result is as follows: Figure 5 As shown, one cluster is named CL0 (Cluster0) and the other cluster is named CL1 (Cluster1).

[0127] Example 2:

[0128] This embodiment provides a system for characteristic typing of Parkinson's disease samples based on local field potentials, comprising a memory and a processor. The memory comprises a program for characteristic typing of Parkinson's disease samples based on local field potentials. When the program for characteristic typing of Parkinson's disease samples based on local field potentials is executed by the processor, the steps of the method for characteristic typing of Parkinson's disease samples based on local field potentials as described in Example 1 are implemented.

[0129] Example 3:

[0130] This embodiment provides a method for training a Parkinson's disease sample classification model, which uses the classification results obtained by the Parkinson's disease sample feature classification method based on local field potentials described in Example 1 for training, including the following steps:

[0131] Using the first, second, third, and fourth parts of the Unified Parkinson's Disease Rating Scale (MDS-UPDRS), the Montreal Cognitive Assessment (MMSE), and the Hamilton Depression Rating Scale (HAMD-24) as input features, different classification models were trained, including logistic regression, AdaBoost, SVM (Support Vector Machine), Naive Bayes, XGBoost (Extreme Gradient Boosting), CatBoost (Category Boosting), KNN (K-Nearest Neighbors), and MLP (Multi-Layer Perceptron). Samples were classified according to the newly described clustering types, CL1 and CL0. To ensure optimal performance for each model, Bayesian optimization was used to search for the optimal hyperparameters for each classifier, fully utilizing the model's predictive power.

[0132] The leave-one-out cross-validation method is used to evaluate model performance, calculate accuracy, recall, and F1 score, output the best performing classification model, and verify its interpretability. This validation method is particularly effective when dealing with small medical datasets because it ensures that the model is tested on completely unseen sample data, thereby providing a more realistic and reliable evaluation of the model's generative ability. The evaluation results of multiple models are as follows: Figure 6 shown.

[0133] The model with the best overall effect, Naive Bayes, was selected, with accuracy, precision, recall score, and F1 score reaching 87%.

[0134] More specifically, the method for verifying the model interpretability includes the following steps:

[0135] The SHAP method was used to calculate the impact of each input feature on the model classification results. The input features included the World Movement Disorder Society Unified Parkinson's Disease Rating Scale (MDS-UPDRS), Mini-Mental State Examination (MMSE), Hamilton Depression Rating Scale 24 (HAMD-24), Non-Motor Symptoms in Daily Life (MDS-UPDRSI), Motor Symptoms in Daily Life (MDS-UPDRSII), Motor Function Examination (MDS-UPDRSIII), Motor Complications (MDS-UPDRSIV), and Montreal Cognitive Assessment (MoCA). The SHAP importance ranking diagram was output to quantify the contribution of different features.

[0136] Combined with statistical analysis to verify the SHAP importance ranking diagram, the SHAP analysis results are as follows Figure 7 As shown in the figure, the contribution of each feature in the model is clearly demonstrated. Finally, statistical analysis of all clinical features within the two clusters was performed to further explore the relationship between the clustering results and clinical variables. Intergroup differences in each feature were tested, revealing significant differences in multiple clinical features between the two groups. These differences are highly consistent with the feature importance rankings derived from the previous SHAP analysis. Table 1 shows the intergroup differences in clinical scores within the new classification.

[0137] Table 1

[0138] ClinicFeatures CLO CL1 P value MDS-UPDRSIII 38.703±11.813 54.167±18.573 <0.001*** MDS-UPDRSIV 3.649±3.482 8.222±3.813 <0.001*** HAMD-24 8.054±4.66 12.167±6.373 0.009** MMSE 25.322±2.224 26.782±1.957 0022* MDS-UPDRSII 18.622±8.281 20.444±6.07 0.410 MDS-UPDRSI 12.676±5.647 13.056±6.159 0.821

[0139] Specifically, features that showed high importance in the SHAP analysis, such as MDS-UPDRS III and MDS-UPDRS IV, also showed highly significant differences between the groups in the statistical analysis. This suggests that the key variables that the model focused on in making predictions were mutually validated by differences in actual clinical characteristics. In addition, the differences between the HAMA and MMSE groups also reached statistical significance, further supporting the contribution of these variables to the model's classification predictions.

[0140] Obviously, the above embodiments of the present invention are merely examples for the purpose of clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. A method for characterizing Parkinson's disease samples based on local field potentials, characterized in that: The steps include: The power spectrum density of the Beta frequency band signal of the local field potential of the sample is calculated to obtain the average power spectrum density curve; Perform peak analysis on the average power spectrum density curve to obtain the peak power spectrum density; Model the periodic component and non-periodic component of the average power spectrum density curve separately, and extract the periodic power spectrum and non-periodic component; The instantaneous phase of the low-frequency band of the Beta frequency band signal of the local field potential of the sample is obtained by Hilbert transform; the instantaneous amplitude of the high-frequency band of the Beta signal is obtained by Hilbert transform; The relationship between the probability distribution of high-frequency amplitude and low-frequency phase is established using a preset method to obtain the phase-amplitude coupling characteristics; Perform dimensionality reduction on the extracted peak power spectrum density, periodic power spectrum, non-periodic component and phase-amplitude coupling features to obtain the reduced-dimensional features; Input the reduced-dimensional features into n clustering models and output the sample typing results; Calculate the quality indicators of sample typing results, including silhouette coefficient and Davies-Bouldin index, use the quality indicators to evaluate clustering quality, select the best clustering scheme, and output the final sample typing results.

2. The method for characteristic classification of Parkinson's disease samples based on local field potential according to claim 1, characterized in that: The method for calculating the power spectral density of the Beta frequency band signal of the local field potential of the sample is the Welch method, which includes the following steps: The original local field potential signal x[n] is framed and the windowed signal is calculated using the Hanning window function w[n]. The expression is as follows: x i [n]=x[n+i·N]·w[n] Where n represents the index of the sample within the frame, n∈[1,N-1], x i [n] represents the windowed signal of the i-th frame, x[n] represents the original signal, w[n] represents the Hanning window, and N is the length of the frame; Apply discrete Fourier transform to each frame of windowed signal and calculate the frequency domain signal X i [f], the expression is as follows: X i [f]=FFT(x i [n]) Where f represents frequency, FFT represents discrete Fourier transform, and x i [n] is the windowed signal of the i-th frame; The power spectrum density P of the i-th frame is obtained by taking the modulus square of the frequency domain signal of each frame. i [f], the expression is as follows: P i [f]=|X i [f]| 2 The power spectra of all frames are averaged to obtain the average power spectrum density curve P avg [f], the expression is as follows: Where N is the number of frames.

3. The method for characteristic classification of Parkinson's disease samples based on local field potential according to claim 1, characterized in that: Modeling the periodic component and non-periodic component of the average power spectrum density curve separately, extracting the periodic power spectrum and non-periodic component, including the following steps: The periodic oscillation components of the power spectrum are fitted using Gaussian functions to obtain the periodic power spectrum density P periodic (f), the expression is as follows: Where f represents frequency, A represents the amplitude of oscillation, CF represents the center frequency of oscillation, and BW represents the bandwidth of oscillation; The 1 / f-like function is used to fit the non-periodic component of the power spectrum to obtain the non-periodic power spectrum density P aperiodic (f), the expression is as follows: Where f represents frequency, offset represents the offset of the non-periodic component, and exponent represents the exponent of the non-periodic component; Combining the periodic power spectrum density and the non-periodic power spectrum density, a complete power spectrum density model is constructed. The model parameters are optimized by the least square method to minimize the mean square error and obtain the spectrum curve of the fitted power spectrum density. The non-periodic power spectrum density is subtracted from the fitted power spectrum to obtain the periodic power spectrum, and the non-periodic component is extracted from the non-periodic fitting curve.

4. The method for characteristic classification of Parkinson's disease samples based on local field potential according to claim 1, characterized in that: The preset method for establishing a relationship between the probability distribution of high-frequency amplitude and low-frequency phase includes the following steps: Divide the low-frequency phase into n equal-width intervals; In each phase interval, the average value of the high-frequency signal amplitude is calculated, the distribution histogram of the amplitude in the phase is constructed, and the distribution histogram is normalized. The expression is as follows: Among them, P norm,i is the normalized probability distribution of the ith phase interval, P i is the amplitude histogram frequency in the i-th interval, ∑ j P j is the sum of the amplitude histograms in all phase bins; The Kullback-Leibler distance is used to calculate the degree to which the distribution deviates from the uniform distribution. KL , the expression is as follows: Where n represents the number of intervals and U represents the uniform distribution probability; The phase-amplitude coupling characteristic MI is calculated by KL distance normalization, and the expression is as follows:

5. A Parkinson's disease sample characteristic classification system based on local field potential, characterized in that: The system includes: a memory and a processor, wherein the memory includes a program for a method for characteristic typing of Parkinson's disease samples based on local field potentials. When the program is executed by the processor, the steps of a method for characteristic typing of Parkinson's disease samples based on local field potentials are implemented as described in any one of claims 1 to 4.

6. A method for training a Parkinson's disease sample classification model based on local field potentials, using the classification results of the Parkinson's disease sample feature classification method based on local field potentials according to any one of claims 1 to 4 for training, characterized in that: The following steps are involved: Using the sample rating scale as input features, m classification models are trained to classify the samples according to the classification results, and the optimal hyperparameters are searched; The leave-one-out cross-validation method was used to evaluate the model performance, output the best classification model and verify its interpretability.

7. The method for training a Parkinson's disease sample classification model based on local field potential according to claim 6, characterized in that: The method for verifying the interpretability of the model includes the following steps: Use the SHAP method to calculate the impact of each input feature on the model classification results and output the SHAP importance ranking graph; The SHAP importance ranking diagram was verified by combining statistical analysis, and the inter-group difference significance test was used to evaluate the impact of features on the classification results to ensure the interpretability of the model.

Images