A parkinson's disease dyskinesia identification system based on cross-frequency coupling tensor decomposition

Abstract

The application discloses a Parkinson's disease dyskinesia identification system based on cross-frequency coupling tensor decomposition, comprising: a signal acquisition and processing module, which is used for synchronously collecting local field potential signals of multiple sites of the brain and performing pretreatment; a cross-frequency coupling calculation module, which is used for calculating a cross-frequency coupling atlas of each data segment; a signal-to-noise ratio balance processing module, which is used for generating an average cross-frequency coupling atlas of each subject; a tensor decomposition feature extraction module, which is used for aggregating the average cross-frequency coupling atlas of all subjects, constructing a third-order tensor, and decomposing to obtain three tensor components; and when the average cross-frequency coupling atlas of a to-be-tested subject is input, a weight vector corresponding to the three tensor components is output; and a classification and identification module, which is used for inputting the weight vector of the to-be-tested subject into a pre-trained machine learning classification model, and outputting an identification result of Parkinson's disease dyskinesia. The application can overcome the frequency drift problem existing between different individuals and improve the identification accuracy.

Description

Technical Field

[0001] This invention belongs to the field of electroencephalogram (EEG) signal processing, and in particular relates to a Parkinson's disease dyskinesia recognition system based on cross-frequency coupled tensor decomposition. Background Technology

[0002] Parkinson's disease (PD) is a common neurodegenerative disease, and long-term levodopa (L-Dopa) is currently the gold standard for PD treatment. However, with disease progression and long-term drug use, approximately 50%-90% of patients gradually develop levodopa-induced dyskinesia (LID), manifested as choreiform or dystonia-like involuntary movements, severely impacting their quality of life. In clinical practice, the diagnosis and assessment of LID primarily rely on clinicians' observations and tools such as the Unified Dyskinesia Rating Scale (UDysRS) or the Artificial Involuntary Movement Scale (AIMS). This scale-based assessment method is inherently subjective and typically only allows for "snapshot" assessments during specific outpatient hours, making it difficult to capture the highly fluctuating and intermittent nature of dyskinesia symptoms. Furthermore, it cannot meet the needs of closed-loop deep brain stimulation (DBS) systems for millisecond-level real-time symptom identification and intervention.

[0003] To achieve objective and real-time identification of LID (Low-Intensity Disorder), research on biomarkers based on intracranial neurophysiological signals (such as local field potentials, LFP) has become a hot topic. Early closed-loop deep brain stimulation (DBS) studies mainly focused on power changes in the Beta band (13-30 Hz), which, while effectively characterizing rigidity and bradykinesia symptoms of Parkinson's disease, had poor specificity for dyskinesia. Subsequent studies found that LID episodes are often accompanied by an increase in oscillatory power in the Gamma band (30-100 Hz). However, relying solely on power spectrum characteristics of specific frequency bands is often susceptible to motion artifacts, and the center frequency of Gamma activity varies greatly among different patients, resulting in limitations in the generalization ability and robustness of power spectrum-based detection methods.

[0004] In recent years, increasing evidence suggests that cross-frequency coupling (CFC) in brain networks, particularly the modulation of the phase of low-frequency oscillations (such as theta or alpha rhythms) on the amplitude of high-frequency oscillations (such as gamma rhythms) (i.e., phase-amplitude coupling, PAC), is a key mechanism for revealing abnormal information interaction in neural circuits. In LID (Low-Intensity Disorder) states, significantly enhanced pathological PAC phenomena appear in the striatum (STR) or cortical (M1) regions. Compared to a single power characteristic, PAC reflects higher-order nonlinear interactions between neuronal populations, exhibiting higher specificity and signal-to-noise ratio. Therefore, utilizing the cross-frequency coupling state of electrophysiological signals to identify LID has become an important direction for overcoming current detection bottlenecks.

[0005] Despite the enormous application potential of PAC features, existing analytical methods still face significant challenges in practical applications. Traditional PAC analysis typically employs the "region of interest (ROI) averaging method," which involves manually pre-setting a fixed low-frequency and high-frequency range (e.g., manually specifying 4-8Hz modulation to 60-90Hz) and then calculating the average coupling strength within that region. This method is highly subjective, ignoring the inevitable frequency shift between different subjects, which can easily lead to inaccurate feature extraction or missed detections. Furthermore, simple averaging disrupts the inherent multidimensional structure between "subject-phase frequency-amplitude frequency," failing to effectively separate LID-specific pathological coupling components from complex background EEG activity.

[0006] More critically, existing machine learning-based LID identification research often overlooks the severe bias caused by data imbalance. In actual clinical or animal experimental records, the effective data duration varies greatly among different subjects (ranging from several minutes to several hours). For samples with large datasets, the background noise in the PAC (Packet Accumulation and Canalization) spectra is smoothed by extensive averaging, resulting in extremely low variance; while for samples with small datasets, the background retains strong random noise, exhibiting "blob-like" artifacts. Existing classification algorithms are prone to mistakenly using this "background smoothness" caused by differences in data volume as a feature distinguishing LID from PD (Positive Disorder), rather than based on real neurophysiological signals. This inconsistency in signal-to-noise ratio severely impairs the model's generalization ability, making it difficult for a model trained on one dataset to reproduce on new subjects. Therefore, there is an urgent need for an LID identification method that can automatically extract the intrinsic structural features of the data and effectively overcome the influence of sample size differences. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a Parkinson's disease dyskinesia recognition system based on cross-frequency coupled tensor decomposition. This system can automatically separate specific frequency coupling patterns that are highly correlated with dyskinesia from full-band data, effectively overcoming the frequency drift problem between different individuals and improving recognition accuracy.

[0008] A Parkinson's disease dyskinesia recognition system based on cross-frequency coupled tensor decomposition includes:

[0009] The signal acquisition and processing module is used to simultaneously acquire local field potential signals from multiple sites in the brain and perform signal preprocessing to obtain data segments of fixed duration.

[0010] The cross-frequency coupling calculation module is used to extract the instantaneous phase of the low-frequency oscillation and the instantaneous amplitude of the high-frequency oscillation from each data segment, and to calculate the cross-frequency coupling spectrum of each data segment.

[0011] The signal-to-noise ratio balancing processing module is used to count the number of effective data segments for each subject, determine the minimum baseline number, and generate the average cross-frequency coupling spectrum for each subject using the averaging method.

[0012] The tensor decomposition feature extraction module is used to aggregate the average cross-frequency coupling spectrum of all subjects to construct a third-order tensor, and decompose the tensor to obtain three tensor components; and when the average cross-frequency coupling spectrum of the subject is input, the module outputs the weight vectors corresponding to the three tensor components.

[0013] The classification and recognition module is used to input the weight vector of the test subject into a pre-trained machine learning classification model and output the recognition result of Parkinson's disease dyskinesia.

[0014] The specific process of signal preprocessing in the signal acquisition and processing module is as follows:

[0015] The local field potential signal is filtered by a notch filter to remove power frequency interference and its higher harmonics, and then the signal is divided into multiple data segments with a fixed time length.

[0016] The specific working process of the cross-frequency coupling calculation module is as follows:

[0017] The instantaneous phase of the low-frequency oscillations in each data segment is extracted using bandpass filtering and Hilbert transform. and the instantaneous amplitude of high-frequency oscillations The phase-amplitude modulation index was calculated using the probability density estimation method based on the Gaussian kernel function. The phase frequency and amplitude frequency were used as the horizontal and vertical axes, respectively, to draw a cross-frequency coupling spectrum composed of multiple sets of phase-amplitude modulation indices.

[0018] The formula for calculating the phase-amplitude modulation index is:

[0019] ;

[0020] ;

[0021] in, The phase probability density function is based on kernel density estimation. For phase angle, For the first One sampling point, The number of sampling points. For average amplitude, For bandwidth Gaussian kernel function, For smoothed density The modulation index is obtained by calculating the KL divergence.

[0022] The specific working process of the signal-to-noise ratio balancing processing module is as follows:

[0023] Count the number of valid data segments for all subjects; determine the minimum number of valid data segments. As the minimum baseline; for data segments exceeding... The subjects were randomly selected. Each segment; the average cross-frequency coupling spectrum of each subject is generated by averaging.

[0024] In the tensor decomposition feature extraction module, the average cross-frequency coupling spectra of different subjects are aggregated to construct a third-order tensor with the dimensions of subject × amplitude frequency × phase frequency. A non-negative CP decomposition algorithm with introduced smooth manifold regularization constraints is used to decompose the tensor. The decomposition yields three tensor components.

[0025] A non-negative CP decomposition algorithm with introduced smooth manifold regularization constraints is used to decompose the tensor. The decomposition process is as follows:

[0026] Construct the objective optimization function While minimizing the reconstruction error, a smoothing regularization term for the frequency modes is introduced, and a multiplicative update rule is used for iterative solution to obtain the final smooth non-negative eigenvectors, corresponding to the three tensor components.

[0027] Objective optimization function The formula is:

[0028] ;

[0029] The constraints are: ;in, Denotes the Frobenius norm; The third-order tensor to be decomposed is to be constructed; Let be the rank of the tensor decomposition, i.e., the number of components obtained from the decomposition; , and These are the mode factor matrices corresponding to the sample, amplitude frequency, and phase frequency, respectively; , and They represent the first Factor vectors of components in three modalities; operators Represents the outer product of vectors; and The regularization coefficient is used. and The Laplacian matrix corresponding to the frequency dimension is used to constrain the second-order continuous smoothness of the frequency eigenvectors; the factor matrix... The column vectors are used as cross-frequency coupling feature vectors.

[0030] In the classification and recognition module, the pre-trained machine learning classification model adopts a linear discriminant analysis model.

[0031] A linear discriminant analysis model was trained using cross-frequency coupled feature vectors of Parkinson's disease dyskinesia subjects with known labels. The projection direction that maximizes the ratio of between-class variance to within-class variance was then sought. Leave-one-out cross-validation was used to evaluate the classification accuracy of the model. The identification results of Parkinson's disease dyskinesia for the subjects were then output.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] 1. Solves the instability of traditional histogram calculation under sparse data conditions: Addressing the data reduction issue caused by signal-to-noise ratio balancing, this invention abandons the traditional histogram binning algorithm and innovatively employs a kernel density estimation method based on adaptive bandwidth to calculate the modulation index. By utilizing a Gaussian kernel function to continuously smooth the phase distribution, it effectively overcomes the boundary effects and statistical jitter caused by discrete binning, achieving highly robust coupling strength estimation even with limited sample size.

[0034] 2. Achieved high-purity feature extraction consistent with physiological characteristics: Unlike standard tensor decomposition algorithms, this invention constructs a non-negative CP decomposition model with smooth manifold regularization constraints. By adding a Laplace smoothing regularization term and non-negativity constraints to the objective function, the frequency feature curves obtained from the decomposition are forced to maintain biologically continuous smoothness and non-negativity. This not only automatically eliminates high-frequency noise spikes in the frequency domain but also avoids the generation of meaningless negative values, making the extracted cross-frequency pathological coupling features purer and more interpretable.

[0035] 3. Overcoming the subjective limitations of manual ROI selection: This invention utilizes the blind source separation characteristics of tensor decomposition to automatically separate specific frequency coupling patterns highly correlated with dyskinesia from full-band data, effectively overcoming the frequency drift problem between different individuals and avoiding missed detections caused by subjectively defining regions of interest.

[0036] 4. Eliminates identification bias caused by data imbalance: Addressing the common problem of significant differences in recording duration between clinical and experimental data, this invention introduces a signal-to-noise ratio balancing module. This eliminates artificial statistical differences caused by varying average counts, ensuring that the classifier learns genuine neurophysiological and pathological features, rather than "background smoothness" artifacts from the data processing process, significantly improving the model's generalization ability across different datasets. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a flowchart illustrating the implementation of a Parkinson's disease dyskinesia recognition system based on cross-frequency coupled tensor decomposition according to the present invention.

[0039] Figure 2 The image shows the feature map of component 1 obtained by tensor decomposition and the distribution of individual subjects in this embodiment of the invention.

[0040] Figure 3 The image shows the feature map of component 2 obtained by tensor decomposition and the distribution of the test subjects in this embodiment of the invention.

[0041] Figure 4 The image shows the feature map of component 3 obtained by tensor decomposition and the distribution of individual subjects in this embodiment of the invention.

[0042] Figure 5 This refers to the two-dimensional projection formed by the cross-frequency coupling feature vectors and the corresponding LDA decision boundary in this embodiment of the invention.

[0043] Figure 6 This refers to the three-dimensional space formed by the cross-frequency coupling feature vectors and the corresponding LDA decision plane in this embodiment of the invention.

[0044] Figure 7 This is an example of evaluating the classification performance of the LDA model using leave-one-out cross-validation in this embodiment of the invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] It should be noted that, unless otherwise specified, the features in the following embodiments and implementation methods can be combined with each other.

[0047] In studying the evolution of neurophysiological signals in Parkinson's disease (PD) and its drug-induced dyskinesia (LID), it was found that when transitioning from a PD state to a LID state through drug induction, the phase-amplitude coupling characteristics of specific brain regions (such as the striatum) undergo significant changes, manifested as enhanced modulation of high-frequency amplitude by specific low-frequency phases. Furthermore, the study also found that due to frequency drift among different subjects, the method of calculating the average coupling strength within a manually pre-defined ROI introduces subjective bias, leading to inaccurate feature extraction or missed detections.

[0048] This invention provides a Parkinson's disease dyskinesia recognition system based on cross-frequency coupled tensor decomposition, comprising:

[0049] The signal acquisition and processing module is used to simultaneously acquire local field potential signals from multiple sites in the brain and perform signal preprocessing to obtain data segments of fixed duration.

[0050] The cross-frequency coupling calculation module is used to extract the instantaneous phase of the low-frequency oscillation and the instantaneous amplitude of the high-frequency oscillation from each data segment, and to calculate the cross-frequency coupling spectrum of each data segment.

[0051] The signal-to-noise ratio balancing processing module is used to count the number of effective data segments for each subject, determine the minimum baseline number, and generate the average cross-frequency coupling spectrum for each subject using the averaging method.

[0052] The tensor decomposition feature extraction module is used to aggregate the average cross-frequency coupling spectrum of all subjects to construct a third-order tensor, and decompose the tensor to obtain three tensor components; and when the average cross-frequency coupling spectrum of the subject is input, the module outputs the weight vectors corresponding to the three tensor components.

[0053] The classification and recognition module is used to input the weight vector of the test subject into a pre-trained machine learning classification model and output the recognition result of Parkinson's disease dyskinesia.

[0054] The implementation process of the system of this invention is as follows: Figure 1 As shown, it includes the following steps:

[0055] Step 1: Data Acquisition and Preprocessing.

[0056] In this embodiment, the experimental subjects included Sprague-Dawley (SD) rats, a PD model established by 6-OHDA lesioning, and LID (Least Active Discharge) rats, a LID model established by long-term levodopa induction. Local field potential (LFP) signals from the striatum (STR) and motor cortex (M1) were simultaneously acquired using an implanted microelectrode array. The signal sampling rate was set to 1 kHz.

[0057] Preprocessing steps included: removing power frequency interference and its harmonics using 50 Hz, 150 Hz, and 250 Hz notch filters; and segmenting continuously recorded signals into data segments of fixed duration (e.g., 30 seconds each). After screening, a total of 12 subjects (6 PD models and 6 LID models) were included for subsequent analysis.

[0058] Step 2, cross-frequency coupling spectrum calculation.

[0059] This embodiment first extracts the phase-amplitude coupling feature (i.e., STR phase modulated STR amplitude) of the STR signal, because subsequent comparative experiments show that this pattern has the best recognition performance. Specific steps include:

[0060] (a) Set the phase frequency range to 1-12 Hz (step 0.2 Hz) and the amplitude frequency range to 30-200 Hz (step 4 Hz).

[0061] (b) For each data segment, the instantaneous phase of the STR signal in the above frequency band is extracted using a linear-phase FIR bandpass filter and Hilbert transform. and instantaneous amplitude .

[0062] (c) The coupling spectrum of each segment is calculated using the modulation index formula, and a standardized Z-Map spectrum is generated by permutation test to reduce the influence of random coupling.

[0063] Specifically, to overcome the sample sparsity problem caused by downsampling, this embodiment abandons the traditional histogram binning method and uses a probability density estimation method based on the Gaussian kernel function to calculate the phase-amplitude modulation index. The specific implementation process is as follows:

[0064] (a) Kernel function selection: Gaussian kernel function is selected. Used to smooth the phase distribution.

[0065] (b) Bandwidth parameter selection: The optimal smoothing bandwidth is automatically determined based on Silverman's Rule of Thumb. The calculation formula is:

[0066] ;

[0067] in, The standard deviation of the phase data. This represents the number of sampling points. This ensures that the smoothness adapts to the distribution characteristics of the data.

[0068] (c) Probability density construction: Based on the selected kernel function and bandwidth, construct the amplitude-weighted phase probability density function. :

[0069] ;

[0070] in, This represents the average amplitude within the time window. This step transforms the discrete phase-amplitude points into a continuous, smooth probability distribution curve.

[0071] (d) Modulation index calculation: Calculate the distribution With uniform distribution The Kullback-Leibler (KL) divergence between them is used as the modulation index. The calculation formula is:

[0072] .

[0073] Step 3, signal-to-noise ratio balancing.

[0074] In actual experiments, the effective recording time varies significantly among different subjects, resulting in different numbers of effective data segments. If the cross-frequency coupling spectrum of all segments for each individual is directly averaged, the background noise of individuals with large data volumes will be extremely smoothed, while individuals with relatively small data volumes will retain some random noise in the background. This difference in "background smoothness" caused by sample size can easily be mistaken by the classifier as a physiological characteristic.

[0075] Therefore, this embodiment implements a signal-to-noise ratio balancing strategy: count the number of valid segments for all subjects and determine the minimum baseline number. For all subjects, a random seed algorithm was used to randomly select from the original fragments. Only this segment; calculate only this The average coupling profiles of the segments were processed. After this process, the average coupling profiles of all subjects had consistent statistical confidence and low background noise, ensuring that subsequent classification was based on differences in physiological signal intensity.

[0076] Step 4: Tensor construction and decomposition to extract features.

[0077] The average coupling maps of all subjects after signal-to-noise ratio balancing are aggregated to construct a third-order tensor. Its dimensions are 12 (individual subjects) × 43 (amplitude frequency) × 56 (phase frequency). A non-negative CP decomposition algorithm with smooth manifold regularization constraints is used to decompose the tensor. The specific implementation details are as follows:

[0078] (a) Construction of the Laplace matrix: In order to constrain the smoothness of the frequency features, a Laplace matrix for the amplitude frequency mode is constructed. and the Laplace matrix for phase frequency modes .by For example, first construct the nearest neighbor adjacency matrix. If the frequency point and Adjacent, then Otherwise, it is 0; then calculate the degree matrix. (The elements on the diagonal are) ); ultimately obtained The purpose of this matrix is ​​to penalize drastic changes in weights between adjacent frequency points.

[0079] (b) Objective function construction: The following non-convex optimization objective function, which includes a reconstruction error term and a manifold regularization term, is established:

[0080] ;

[0081] in, It is set to 0.1 to balance data fit and curve smoothness.

[0082] (c) Iterative solution: Since the objective function contains non-negativity constraints ( Since standard gradient descent cannot be used, this embodiment employs Multiplicative Update Rules (MUR) for iterative solution. The amplitude-frequency factor matrix is ​​used as the basis for this solution. For example, its first The update formula for the next iteration is:

[0083] ;

[0084] in, This indicates element-wise multiplication. Let be the expansion matrix of the tensor in mode 2. This is the Khatri-Rao product. It is updated alternately. This continues until the objective function converges, thus obtaining the final smooth non-negative eigenvectors.

[0085] Decomposition results are as follows Figure 2 , Figure 3 , Figure 4 As shown:

[0086] (a) Component 1 (strong pathological coupling component): Its phase frequency peak is located at 3-4 Hz, and its amplitude frequency peak is located at 70-90 Hz. The spatial feature map of this component shows a clear and high-intensity coupling erythema. In the individual distribution map of the subjects, the LID group scored significantly higher on this component than the PD group, indicating that this component captures the most specific cross-frequency pathological coupling features of dyskinesia and is the main contributing factor to the identification of LID.

[0087] (b) Component 2 (secondary pathological coupling component): Its phase frequency peak is located at 4-5 Hz, and its amplitude frequency peak is located at 70-90 Hz. The spatial feature map of this component shows a relatively clear coupling erythema. In the distribution map of the subjects, some individuals in the LID group scored higher on this component than those in the PD group, indicating that this component reflects the dyskinesia-specific cross-frequency pathological coupling to a certain extent and is an effective supplement to component 1.

[0088] (c) Component 3 (background energy component): Its phase frequency peak is located at 0-2 Hz (close to the DC / low frequency edge), and its amplitude frequency peak is located at 70-90 Hz. The spatial characteristic map of this component shows a relatively blurred red spot area at the edge. In the distribution map of the subjects, there is no clear boundary between the LID group and the PD group, indicating that this component reflects the background energy characteristics or edge effect that are prevalent among the subjects.

[0089] Unlike traditional methods that select only a single best feature, this invention, in order to make full use of the multidimensional structural information of the data, uses the weight vectors of the subjects corresponding to the above three components as feature inputs simultaneously to construct a three-dimensional feature vector for each subject.

[0090] Step 5: Feature classification and performance evaluation.

[0091] In this embodiment, the core task of the classification and recognition module is to construct a decision boundary that can distinguish between PD and LID. Based on the three rank-one tensor components obtained from tensor decomposition in step 4, this invention uses Linear Discriminant Analysis (LDA) as the classification model. LDA has extremely high robustness in scenarios with small samples and high-dimensional features. Its core idea is to find an optimal projection direction so that the projected data are as compact as possible among similar classes (minimum intra-class variance) and as far apart as possible among dissimilar classes (maximum inter-class variance). The specific implementation process includes the following four sub-steps:

[0092] (a) Feature vector construction: For the th One subject individual ( In this embodiment The individual factor matrix obtained from tensor decomposition Extract the corresponding weight values ​​and construct the three-dimensional feature vector of the individual. :

[0093] ;

[0094] in, , , Corresponding to components 1, 2, and 3 respectively in the first... Weighted scores on each individual subject.

[0095] (b) LDA model training process: The subjects were divided into a training set, containing PD samples (category 0) and 1 LID sample (Category 1). Model training aims to compute the optimal projection weight vector. The specific calculation steps include:

[0096] Calculate the mean vector. Calculate the mean vectors of the PD group and the LID group in the feature space respectively. and Calculate the within-class scatter matrix. This reflects the degree of dispersion of various types of internal data:

[0097] ;

[0098] Calculate the inter-class scatter matrix This reflects the distance between the two types of centers:

[0099] ;

[0100] Find the optimal projection direction. Maximize the Rayleigh quotient according to Fisher's criterion. By solving the generalized eigenvalue problem, the optimal projection weight vector is obtained. This vector defines the normal direction of the classification hyperplane in the three-dimensional feature space:

[0101] ;

[0102] (c) Decision threshold setting: Project all training samples onto the orientation The above yields a one-dimensional projection value. Typically, the midpoint of the mean of the two projections is taken as the classification decision threshold. :

[0103] ;

[0104] (d) Reasoning and verification process: For a new test subject, first extract its weight vectors on the above three tensor components. Calculate its discrimination score .like If it is determined to be LID; It was determined to be PD.

[0105] To objectively evaluate the model's generalization performance, this embodiment strictly implements leave-one-out cross-validation (LOOCV). That is, one sample is used as the test set each time, and the remaining 11 samples undergo the same training process. and The overall accuracy of the 12 tests was calculated. For example... Figure 5 , Figure 6 As shown, by The defined decision plane clearly segments the feature space in three-dimensional space, and the decision boundary in the two-dimensional projection is also well segmented, verifying that the linear combination of the three features can effectively improve the robustness of the classification boundary. Figure 7 As shown, the confusion matrix obtained by leave-one-out cross-validation indicates that 83.3% of LID samples and all PD samples can be correctly classified and identified, with an overall accuracy of 91.7% and an area under the ROC curve (AUC) of 0.972, indicating that the present invention can meet the requirements of the Parkinson's disease dyskinesia identification task.

[0106] To determine the optimal biomarker source for identifying LID, this embodiment further compared four possible cross-frequency coupling combinations between the STR and M1 brain regions. The analysis procedure was the same as described above, only the source signals for phase and amplitude were changed. The four modes were: STR phase modulated STR amplitude, STR phase modulated M1 amplitude, M1 phase modulated STR amplitude, and M1 phase modulated M1 amplitude. Table 1 shows the classification performance statistics of the four modes under leave-one-out cross-validation.

[0107] Table 1

[0108]

[0109] The results showed that the pathological oscillation network associated with LID exhibited significant directionality and spatial specificity. Among these features, local cross-frequency coupling of STRs was the most sensitive characteristic for identifying LID, followed by M1 regulation of STRs. Therefore, in practical applications, STR signals should be prioritized for acquisition, or higher weights should be assigned to STR-related features in multi-channel analysis.

[0110] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A Parkinson's disease hetero-bromic identification system based on cross-frequency coupling tensor decomposition, characterized in that, include: The signal acquisition and processing module is used to simultaneously acquire local field potential signals from multiple sites in the brain and perform signal preprocessing to obtain data segments of fixed duration. The cross-frequency coupling calculation module is used to extract the instantaneous phase of the low-frequency oscillation and the instantaneous amplitude of the high-frequency oscillation from each data segment, and to calculate the cross-frequency coupling spectrum of each data segment. The signal-to-noise ratio balancing processing module is used to count the number of effective data segments for each subject, determine the minimum baseline number, and generate the average cross-frequency coupling spectrum for each subject using the averaging method. The tensor decomposition feature extraction module is used to aggregate the average cross-frequency coupling spectrum of all subjects to construct a third-order tensor, and decompose the tensor to obtain three tensor components; and when the average cross-frequency coupling spectrum of the subject is input, the module outputs the weight vectors corresponding to the three tensor components. The classification and recognition module is used to input the weight vector of the test subject into a pre-trained machine learning classification model and output the recognition result of Parkinson's disease dyskinesia.

2. The cross-frequency coupling tensor decomposition based Parkinson's disease dyskinesia identification system of claim 1, wherein, The specific process of signal preprocessing in the signal acquisition and processing module is as follows: The local field potential signal is filtered by a notch filter to remove power frequency interference and its higher harmonics, and then the signal is divided into multiple data segments with a fixed time length.

3. The Parkinson's disease dyskinesia recognition system based on cross-frequency coupled tensor decomposition according to claim 1, characterized in that, The specific working process of the cross-frequency coupling calculation module is as follows: The instantaneous phase of the low-frequency oscillations in each data segment is extracted using bandpass filtering and Hilbert transform. and the instantaneous amplitude of high-frequency oscillations The phase-amplitude modulation index was calculated using the probability density estimation method based on the Gaussian kernel function. The phase frequency and amplitude frequency were used as the horizontal and vertical axes, respectively, to draw a cross-frequency coupling spectrum composed of multiple sets of phase-amplitude modulation indices.

4. The Parkinson's disease dyskinesia recognition system based on cross-frequency coupled tensor decomposition according to claim 3, characterized in that, The formula for calculating the phase-amplitude modulation index is: ； ； in, The phase probability density function is based on kernel density estimation. For phase angle, For the first One sampling point, The number of sampling points. For average amplitude, For bandwidth Gaussian kernel function, For based on The modulation index is obtained by calculating the KL divergence.

5. The Parkinson's Disease Dyskinesia Recognition System Based on Cross-Frequency Coupled Tensor Decomposition according to claim 1, characterized in that, The specific working process of the signal-to-noise ratio balancing processing module is as follows: Count the number of valid data segments for all subjects; determine the minimum number of valid data segments. As the minimum baseline; for data segments exceeding... The subjects were randomly selected. Each segment; the average cross-frequency coupling spectrum of each subject is generated by averaging.

6. The Parkinson's disease dyskinesia recognition system based on cross-frequency coupled tensor decomposition according to claim 1, characterized in that, In the tensor decomposition feature extraction module, the average cross-frequency coupling spectra of different subjects are aggregated to construct a third-order tensor with the dimensions of subject × amplitude frequency × phase frequency. ; A non-negative CP decomposition algorithm with introduced smooth manifold regularization constraints is used to decompose the tensor. The decomposition yields three tensor components.

7. The Parkinson's disease dyskinesia recognition system based on cross-frequency coupled tensor decomposition according to claim 6, characterized in that, A non-negative CP decomposition algorithm with introduced smooth manifold regularization constraints is used to decompose the tensor. The decomposition process is as follows: Construct the objective optimization function While minimizing the reconstruction error, a smoothing regularization term for the frequency modes is introduced, and a multiplicative update rule is used for iterative solution to obtain the final smooth non-negative eigenvectors, corresponding to the three tensor components.

8. The Parkinson's disease dyskinesia recognition system based on cross-frequency coupled tensor decomposition according to claim 1, characterized in that, In the classification and recognition module, the pre-trained machine learning classification model adopts a linear discriminant analysis model.

Images