Optimization method for Parkinson's disease signal processing

Abstract

The invention provides an optimization method for Parkinson's disease signal processing, and relates to the technical field of signal processing, and the method specifically comprises the steps: collecting a data set, and carrying out the preprocessing; calculating instantaneous energy distribution and compression mapping to form energy time sequence characteristics; the multi-scale time module extracts and aligns time features in parallel, and captures long-time and short-time dynamic changes; the frequency spectrum reconstruction module introduces a learnable phase modulation mechanism into a frequency domain, and performs joint analysis and feature extraction on frequency domain features; a brain region aggregation module and a graph attention layer execute partition integration and association enhancement on the spatial features, and represent functional relationships between different brain regions; the region self-attention module and the classification module carry out self-adaptive fusion and discriminant optimization, and output a detection result; meanwhile, a Parkinson's disease signal processing model is provided, detection is completed by processing Parkinson's disease signals, and the judgment accuracy is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of signal processing, and particularly relates to an optimization method for Parkinson's disease signal processing. BACKGROUND

[0002] Parkinson's disease is a common central nervous system degenerative disease, and its main pathological feature is the gradual loss of dopaminergic neurons in the substantia nigra, which is manifested as clinical symptoms such as tremor, muscle rigidity and bradykinesia. In recent years, neural electrophysiological signals, as a non-invasive physiological information that can reflect the central nervous activity state in real time, have been widely used in the early diagnosis and pathological mechanism research of Parkinson's disease. However, Parkinson's disease signals have significant non-stationarity and multi-dimensional complexity, and are greatly affected by individual differences, channel distribution and environmental noise, so that the effective pathological features contained in the Parkinson's disease signals are difficult to directly extract and analyze. Traditional Parkinson's disease signal processing methods mainly rely on experience-based filtering, time-frequency transformation or fixed parameter models for feature extraction, and cannot effectively represent the dynamic change rule of Parkinson's disease signals in time, frequency and spatial dimensions. In addition, the process of artificial feature selection is subjective and limited, which leads to insufficient adaptability to the changes of Parkinson's disease signals, and the results are easily disturbed by noise and individual differences, affecting the stability and accuracy of detection.

[0003] With the development of deep learning and graph neural network technology, the automatic feature extraction and pattern recognition method for Parkinson's disease signals has gradually become a research hotspot. However, most of the existing methods only focus on single domain features or local spatial features, ignoring the interaction between Parkinson's disease signals in multi-scale time structure and brain function connection level, and it is difficult to consider both global rhythm features and local dynamic changes. In addition, when facing multi-channel high-dimensional signals, there are problems such as complex structure, large parameters, unstable training, etc., which limit the real-time performance and interpretability in clinical scenarios. Therefore, there is an urgent need for an optimization processing method that can comprehensively utilize the time, frequency and spatial features of Parkinson's disease signals, to realize high-precision and strong-robustness detection and intelligent analysis of Parkinson's disease signals under the premise of ensuring computational efficiency and model stability. SUMMARY

[0004] The application provides an optimization method for Parkinson's disease signal processing, aiming to propose a Parkinson's disease signal processing model, construct a feature extraction system from the time domain to the frequency domain and then to the spatial domain, and improve the integrity of feature expression and the accuracy of discrimination; calculate the instantaneous energy distribution and compression mapping to form an energy time sequence feature; a multi-scale time module extracts and aligns the time features in parallel to capture long and short time dynamic changes; a spectrum reconstruction module introduces a learnable phase modulation mechanism in the frequency domain to jointly analyze and extract features from the frequency domain features; a brain region aggregation module and a graph attention layer perform partition integration and correlation enhancement on the spatial features to represent the functional relationship between different brain regions; a regional self-attention module and a classification module perform adaptive fusion and discriminative optimization to output the detection result of Parkinson's disease; through the Parkinson's disease signal processing model, the detection of Parkinson's disease signal is realized, and the accuracy of Parkinson's disease detection is improved.

[0005] In order to achieve the above purpose, the application provides the following technical scheme: an optimization method for Parkinson's disease signal processing, comprising the following steps: S1, collecting Parkinson's disease signal data set and performing pretreatment operation; S2, calculating the instantaneous energy distribution of the Parkinson's disease signal to form an energy time sequence primary feature, and then performing nonlinear compression processing to generate an energy time sequence feature; S3, constructing a multi-scale time module, performing parallel extraction and feature alignment on the energy time sequence feature in different time scales to obtain a multi-scale time feature; S4, using an adaptive fusion module to fuse the multi-scale time feature, and using a channel attention module to nonlinearly adjust the weight of each channel feature to generate a channel weighted feature; S5, constructing a spectrum reconstruction module, mapping the channel weighted feature to the frequency domain to obtain a frequency domain feature, introducing a phase modulation parameter, modulating the phase modulation component, and obtaining a frequency spectrum steady-state feature; S6, according to the brain region division index, regionally aggregating the frequency spectrum steady-state feature to form a regionally aggregated feature, and then inputting the regionally aggregated feature into a graph attention layer to model the spatial dependence relationship between brain regions and obtain a graph attention feature; S7, inputting the graph attention feature into a regional self-attention module, calculating the correlation matrix between brain regions, obtaining a global fusion feature, and outputting a classification result through a classification module; S8, constructing a Parkinson's disease signal processing model, the Parkinson's disease signal data set sequentially passes through S2 step and S7 step and is trained, the Parkinson's disease signal to be detected is input into the trained model, and the processed result is obtained.

[0006] Preferably, in step S1, a bandpass filter of 0.5Hz to 12Hz is performed on the Parkinson's disease signal dataset to retain low-frequency oscillations and rhythmic activity components and suppress high-frequency noise. In the channel selection stage, common electrode channels are retained and non-common electrode channels are removed to ensure that the number of channels in all data files is consistent, thereby obtaining the preprocessed Parkinson's disease signal.

[0007] Preferably, in step S2, the Parkinson's disease signal is first segmented according to a time window; then, the instantaneous energy distribution is calculated by performing a squaring operation on the Parkinson's disease signal to obtain the instantaneous energy value of each sampling point. The squaring operation is achieved by squaring the amplitude of the Parkinson's disease signal point by point to form an energy distribution matrix; a sliding window average pooling operation is performed on the energy distribution matrix. Average pooling calculates the local energy average value in the time dimension to form a time-continuous energy time series. The energy time series maintains the original number of channels, but the number of sampling points is reduced, constituting the primary energy time series feature; in order to suppress the difference in energy magnitude between different channels, nonlinear compression processing is adopted. The compression function is a logarithmic function. The primary energy time series feature is operated on element by element, and then normalized and mapped to the [0, 1] interval to obtain the energy time series feature.

[0008] Preferably, in step S2, by extracting and nonlinearly compressing the instantaneous energy distribution of the Parkinson's disease signal, a structured model of the energy characteristics of the Parkinson's disease signal is achieved. First, the preprocessed Parkinson's disease signal is segmented according to a fixed time window. Without destroying the original channel structure, the Parkinson's disease signal is divided into several stable time periods, so that each segment corresponds to an independently analyzable time segment. Then, a squaring operation is performed point-by-point on each channel to obtain the instantaneous energy value of each sampling point and form an energy distribution matrix, thereby transforming the noise-sensitive Parkinson's disease signal into an energy trajectory that better reflects the oscillation intensity variation. On this basis, a sliding window average pooling operation is introduced to calculate the local energy average value, forming a temporally continuous and stable energy distribution matrix. The length-compressed energy time series significantly reduces the number of sampling points and alleviates the subsequent computational burden, while effectively smoothing high-frequency random fluctuations and occasional spikes, making the differences in low-frequency oscillation intensity and rhythm stability between Parkinson's disease patients and healthy controls more clearly discernible. Finally, to address the differences in energy amplitude between different electrode channels and individuals, a nonlinear compression strategy based on a logarithmic function is adopted, combined with normalization mapping to unify the amplitude to the [0,1] interval, thereby obtaining energy time series features with a consistent amplitude scale. The processing in step S2 can suppress the adverse effects of noise and energy scale differences while preserving the correspondence between Parkinson's disease signal channels and time windows, enhancing the discernibility of Parkinson's disease energy fluctuation patterns, and improving accuracy and robustness.

[0009] Preferably, in step S3, the input is energy time series features, the multi-scale time module consists of multiple parallel time extraction units, the time extraction units use different time convolution kernel scales to extract features, the convolution kernel scale is set proportionally according to the sampling rate, and the output of the time extraction unit is channel mapped and aligned with the time dimension to generate multi-scale time features.

[0010] Preferably, in step S3, a multi-scale temporal module is constructed to extract temporal features from energy temporal features. The multi-scale temporal module includes multiple parallel temporal extraction units, each of which sequentially includes a two-dimensional convolutional layer, an average pooling layer, and a batch normalization layer to extract the temporal variation features of Parkinson's disease signals at different time scales. The three temporal extraction units correspond to three time scales: long time window, medium time window, and short time window. The lengths of the convolutional kernels of the three units are converted into the number of sampling points according to the sampling rate at ratios of 1 / 2, 1 / 4, and 1 / 8, respectively, and then rounded to obtain feature responses at different time resolutions. The energy temporal features are input into the three temporal extraction units. Each temporal extraction unit performs a two-dimensional convolutional operation to extract local temporal patterns, which are then compressed in the time dimension by an average pooling layer, and the feature distribution is stabilized by a batch normalization layer. The output of each temporal extraction unit is uniformly mapped in the channel dimension by a convolutional layer, and the features at different time scales are aligned to the same length in the time dimension by an adaptive average pooling layer to generate multi-scale temporal features containing multi-time scale information.

[0011] Preferably, in step S3, the energy temporal features of the Parkinson's disease signal are extracted and aligned in parallel at different time scales, achieving a fine characterization of the multi-layered temporal variation patterns of the Parkinson's disease signal. The energy temporal features are input into time extraction units at three time scales: long time window, medium time window, and short time window. Each time extraction unit includes a two-dimensional convolutional layer, an average pooling layer, and a batch normalization layer, which can simultaneously capture the overall rhythmic changes and local energy fluctuations of the Parkinson's disease signal at different time resolutions, thus taking into account both long-term trends and short-term abrupt changes. The output of the time extraction unit is convolved... The layer unifies the channel dimension and aligns the temporal dimension through adaptive average pooling, ensuring that features at different scales are consistent in spatial and statistical distribution. This guarantees the coordination and stability of multi-scale information fusion. The multi-scale structure can effectively enhance the temporal sensitivity of non-stationary Parkinson's disease signals and exhibits stronger characterization capabilities in capturing rhythm abnormalities, energy fluctuations, and oscillation imbalances. The processing in step S3 can simultaneously analyze the dynamic features of Parkinson's disease patient signals at multiple temporal granularities, improve the ability to identify pathological rhythm disorders and energy pattern disturbances, and enhance the accuracy and robustness of disease detection.

[0012] Preferably, in step S4, the input is multi-scale temporal features. The adaptive fusion module generates scale weights based on the global statistical information of the multi-scale temporal features, and fuses the multi-scale temporal features to obtain scale fusion features. These are then input into the channel attention module, which generates channel weight coefficients through the weight generation network, and performs nonlinear weighting on the scale fusion features to obtain channel weighted features.

[0013] Preferably, in step S4, an adaptive fusion module and a channel attention module are constructed to perform weighted fusion of multi-scale temporal features and generate channel-weighted features. The input is multi-scale temporal features. First, a global averaging operation is performed to obtain the global scale features at each scale. Then, a fusion unit consisting of one-dimensional convolution, ReLU activation function, one-dimensional convolution and Softmax function normalization is constructed. The global scale features are input into the fusion unit to generate scale weights. The multi-scale temporal features are weighted and summed along the scale dimension according to the scale weights to obtain the scale fusion features. The channel attention module receives the scale fusion features as input. First, global average pooling is performed in the time dimension to obtain the channel description vector. Then, the channel description vector is input into a weight generation network consisting of two fully connected layers. The first layer reduces the channel dimension and uses the ReLU activation function to extract compact features. The second layer restores the original channel dimension and uses the Sigmoid function to constrain the output range [0,1] to obtain the channel weight coefficients. The channel features in the scale fusion features are weighted channel by channel according to the channel weight coefficients to generate channel-weighted features.

[0014] Preferably, in step S4, hierarchical weighting and saliency enhancement are performed on the multi-scale temporal features to achieve dynamic evaluation of the contribution of features at different time scales and adaptive focusing on key pathological channels, solving the problems of uneven feature distribution, significant differences in information weights, and inconsistent contributions of each channel in Parkinson's disease signals at different time scales; the adaptive fusion module calculates global scale features from the multi-scale temporal features, generates scale weights, and performs nonlinear weighted fusion of features at different time scales, adaptively adjusting the feature proportions according to the rhythmic complexity and temporal resolution of the Parkinson's disease signal, avoiding information loss caused by a fixed fusion method, thereby achieving dynamic balance between time scales; channel attention The module introduces a learnable channel weight generation mechanism based on the fused features. It recalibrates the importance of channels according to the global feature response distribution of each channel, which can highlight signals carrying typical Parkinson's disease pathological activities and suppress interference from irrelevant noise channels, thereby improving the discrimination ability of spatial features. The processing in step S4 achieves the saliency extraction of spatial features while fusing multi-scale temporal information. It ensures the comprehensiveness of temporal information and highlights the keyness of spatial channels. It achieves adaptive optimization in both temporal scale and brain region channel dimensions, enhances the ability to focus on the abnormal rhythm and energy perturbation features of Parkinson's disease signals, and improves the overall recognition accuracy, stability and physiological interpretability.

[0015] Preferably, in step S5, the input is the channel weighted feature. The spectrum reconstruction module performs a frequency domain transformation on the channel weighted feature to obtain a complex form of spectrum feature containing energy mapping component and phase modulation component. Adaptive energy normalization is performed on the frequency domain energy mapping component to obtain the spectral domain energy feature quantity. The phase modulation parameter is introduced to modulate the phase modulation component to obtain the phase modulation structure quantity. The spectral domain energy feature quantity and the phase modulation structure quantity are combined to form a spectrum reconstruction characterization. Then, after inverse Fourier transform, the steady-state spectrum feature is generated.

[0016] Preferably, in step S5, a spectrum reconstruction module is constructed to perform frequency domain transformation, amplitude normalization, and phase modulation processing, with the input being channel weighted features. Perform a Fast Fourier Transform on the input to obtain the frequency domain features in complex form. , ,in For Fourier transform operators, For energy mapping components, For phase modulation components, It is the imaginary unit; the mean and standard deviation of each channel are calculated, and dynamic normalization is performed on the energy mapping components to keep the spectral amplitude distribution of different channels and samples within a uniform statistical range. Adaptive energy standardization is then performed to generate spectral domain energy features. , ,in, and These represent the mean and standard deviation of the energy mapping components for each channel. To prevent division by zero of constant terms.

[0017] Preferably, in step S5, learnable phase modulation parameters are introduced. The phase modulation component is modulated to obtain the phase modulation structure quantity. This allows for the establishment of a learnable correlation between amplitude distribution and phase structure in the frequency domain, enabling adaptive correction of phase shifts in characteristic rhythms. ,in, The modulation intensity coefficient, Let be a hyperbolic tangent function, such that its range is limited to [−π,π]. The phase modulation network generates and jointly optimizes the phase modulation parameters during the training phase. The phase modulation network takes the phase modulation components as input, processes them through one-dimensional convolutional layers, batch normalization layers, and a hyperbolic tangent activation function, and outputs the phase modulation parameters. The spectral domain energy features and phase modulation structure quantities are then reconstructed into a complex spectral reconstruction representation. , Then, the steady-state spectral characteristics are obtained through inverse Fourier transform.

[0018] Preferably, in step S5, frequency domain transformation and phase modulation modeling are performed on the channel weighted features to achieve adaptive reconstruction of the spectral structure of the Parkinson's disease signal and extraction of rhythmic steady-state features. This characterizes the pathological rhythmic abnormalities and energy disturbances at the frequency domain level. The design of step S5 addresses the strong non-stationarity and complex rhythmic overlap issues exhibited by the Parkinson's disease signal in the time domain. By introducing a frequency domain analysis mechanism, it is possible to identify characteristic components related to Parkinson's disease in a relatively stable spectral space. The spectrum reconstruction module first performs a fast Fourier transform on the channel weighted features to map the Parkinson's disease signal from the time domain to the frequency domain, obtaining the complex form of the spectral features containing energy mapping components and phase modulation components. Subsequently, the energy mapping components are dynamically normalized to maintain the stability of different channels and frequencies. The energy amplitudes between samples are within a uniform statistical range, thus eliminating the influence of channel gain differences on the spectral distribution. Subsequently, learnable phase modulation parameters are introduced to nonlinearly modulate the phase modulation components, achieving adaptive correction of phase drift and rhythm offset, and capturing the abnormal performance of Parkinson's disease signals in rhythm synchronization and spectral coupling characteristics. Step S5 transforms the channel weighted features into spectral steady-state features that combine energy steady-state characteristics and phase modulation correlation. This preserves the rhythmic structure of Parkinson's disease signals while enhancing the differential expression capability of frequency domain features. It realizes adaptive reconstruction and steady-state feature extraction of Parkinson's disease pathological rhythms in the frequency domain, effectively improving the sensitivity and discrimination accuracy of abnormal oscillation activities, and providing more interpretable and robust support for the diagnosis of Parkinson's disease signals.

[0019] Preferably, in step S6, the input is the steady-state spectral features. The brain region aggregation module performs region aggregation operations on the steady-state spectral features based on the brain region functional division index to form region aggregation features, which are then sent to the graph attention layer. By constructing the connection relationship between brain regions, attention weighting calculation is performed on the region aggregation features to learn the spatial dependency relationship and interaction pattern between brain regions and generate graph attention features.

[0020] Preferably, in step S6, the steady-state characteristics of the spectrum are analyzed. Brain region aggregation processing is performed and a graph attention representation is constructed, in which... For channel indexing, For time sampling points; firstly, based on the brain region division index table, each channel is grouped according to the corresponding brain region function to form a brain region set. , For the first There are several brain regions, each containing several corresponding channel numbers. A region aggregation operation is performed on the steady-state spectral features within each brain region. An averaging aggregation method along the channel dimension is used to average the features of all channels within the brain region element-wise, resulting in a brain region feature vector. , , For the first Individual brain regions are then stacked in order to obtain region aggregation features. , For the first Brain region feature vectors of each brain region.

[0021] Preferably, in step S6, the region aggregation features are... As input to the graph attention layer, the feature vectors of each brain region are first transformed linearly to obtain the node embedding vectors. , The weight matrix is ​​a linear transformation; for any pair of adjacent brain regions and Attention coefficients are obtained through combination operations of node embedding vectors. Attention coefficients are converted into attention weights by the LeakyReLU activation function and Softmax normalization. Then, based on the attention weights, a weighted summation is performed on the node embedding vectors of adjacent nodes to obtain the updated attention-related brain region features. ,in, To be related to brain regions A set of connected neighboring brain regions, after being processed by a graph attention layer, is then stacked sequentially to generate graph attention features. , For the first Attentional characteristics of individual brain regions.

[0022] Preferably, in step S6, adaptive learning of the structured expression of the spatial distribution characteristics of Parkinson's disease signals and the correlation between brain regions is achieved, thereby characterizing the interaction patterns and functional imbalances of pathological neural activity at the spatial level. The design of step S6 solves the problem of significant spatial heterogeneity of Parkinson's disease signals between different electrode channels. It can not only identify local rhythmic changes in a single channel, but also capture synchronization dyssynchrony and coupling abnormalities across brain regions. The brain region aggregation module, based on the brain region functional division index, groups the spectral steady-state features according to the neuroscience 10-20 electrode system, aggregates the features of multiple channels into brain region feature vectors, and extracts the overall energy and rhythmic features of the brain region through average aggregation operation along the channel dimension, thereby reducing channel noise differences and enhancing the statistical stability of regional aggregation features. This process elevates the representation from channel-level signals to region-level representations. Subsequently, the graph attention layer constructs connections between brain regions and performs attention-weighted calculations based on feature similarity and association weights between nodes. It adaptively learns the interaction strength and information flow between different brain regions, highlighting functional connectivity features closely related to Parkinson's disease pathological activities and suppressing redundant information in irrelevant regions. The overall design captures both local features and global dependencies in the spatial dimension, forming a more hierarchical and interpretable brain functional network representation. This enables high-precision modeling of abnormal synchronization, rhythm coupling disorder, and energy distribution imbalance in the brain regions of Parkinson's disease patients, significantly improving the ability to identify pathological neural network dysfunctions and the interpretability of spatial features. This provides support for graph structure analysis and spatial pattern recognition of Parkinson's disease signals.

[0023] Preferably, in step S7, the input is graph attention features. The region self-attention module calculates the correlation matrix of brain regions based on the graph attention features and performs weighted fusion to generate global fusion features. These features are then fed into a classification module consisting of a residual structure and a linear classification layer. The residual structure performs inter-layer connections and feature mapping on the input features. The linear classification layer calculates the category probability based on the features output by the residual structure. Finally, the classification results for Parkinson's disease patients and healthy controls are output.

[0024] Preferably, in step S7, the input is graph attention features. The region self-attention module first applies query mapping, key mapping, and value mapping to the input graph attention features respectively, forming three feature representations with different modes of action in the feature space of the same dimension. Then, the matching score between any two brain regions is calculated based on the dot product between the query vector and the key vector, and the numerical amplification effect caused by the increase in scale is suppressed by dividing by the square root of the feature dimension. On this basis, Softmax normalization is applied to the matching score of each brain region along the same direction to construct a correlation matrix that reflects the inter-brain region dependency. Then, the correlation matrix is ​​weighted and fused with the value-mapped brain region features by matrix multiplication to form a global fused feature.

[0025] Preferably, in step S7, the global fusion features are flattened into a single vector in the order of brain region dimension and feature dimension as input to the classification module. The classification module first applies layer normalization to the input to reduce training instability caused by scale differences between different samples and different brain regions. The normalized feature vector is then input into a residual structure composed of two fully connected mapping layers. The first layer maps the input dimension to the hidden dimension and applies the ReLU activation function to introduce stronger representation ability. Then, the random deactivation layer is discarded to improve the generalization effect of the model. The second layer remaps the hidden dimension back to the original dimension and adds it to the input of the residual block through a short-circuit connection to form a residual output with inter-layer connection and feature mapping function. On this basis, the residual output is further fed into the linear classification layer. The linear classification layer calculates the original classification score of each category based on the weighted fusion features. In the inference stage, the original classification score is converted into category probability through the Softmax function, and finally the judgment result of Parkinson's disease patient or healthy control is output.

[0026] Preferably, in step S7, adaptive fusion of graph attention features and intelligent discrimination of pathological patterns are realized, establishing a correlation mapping relationship from high-dimensional spatial features to Parkinson's disease diagnosis. The design of step S7 aims to solve the problems of complex inter-brain region feature correlation and significant nonlinear coupling of pathological patterns in EEG signals of Parkinson's disease, realizing dynamic dependency modeling and category discrimination at the global feature level. The regional self-attention mechanism learns the correlation weights between features of different brain regions, adaptively determining the importance of each brain region in the global decision-making process, enhancing the response ability to abnormal activity in key brain regions. On this basis, the classification module uses layer normalization and residual connections to stabilize feature distribution, and further utilizes a linear classification layer to realize the mapping from feature space to category space, so that the differences in EEG of different individuals can be clearly distinguished at the feature level. It not only has strong generalization and robustness, but also maintains good interpretability. The overall design realizes the unified fusion of features between brain regions and the accurate representation of pattern differences at the global level, effectively improving the recognition accuracy of abnormal EEG rhythms and functional imbalances in Parkinson's disease patients, and significantly enhancing robustness and intelligent diagnostic capabilities.

[0027] Preferably, in step S8, a Parkinson's disease signal processing model is constructed, integrating the processing structures from steps S2 to S7 into a unified end-to-end detection framework for training and detecting Parkinson's disease signals. During the training phase, the Parkinson's disease signal dataset is used as input to calculate instantaneous energy distribution and compression mapping, forming energy temporal features. These features are then extracted using a multi-scale time module to extract long- and short-term dynamic change information, forming multi-scale time features. After weighted integration by an adaptive fusion module and a channel attention module, channel-weighted features are obtained. These are then input into a spectrum reconstruction module to obtain spectrum steady-state features that combine energy steady-state characteristics and phase modulation correlation. These features are then converted into graph attention features reflecting the functional connectivity of brain regions by a brain region aggregation module and a graph attention layer, ultimately... The region self-attention module and classification module generate global fusion features and output corresponding category judgments. During iterative training, parameters are updated in reverse based on the real labels of Parkinson's disease patients and healthy controls, and the Parkinson's disease signal processing model converges to a stable discriminative ability. In the detection phase, the Parkinson's disease signal to be detected is passed through the above modules in the same process as in the training phase, and the trained Parkinson's disease signal processing model directly outputs the Parkinson's disease detection result. Through the S8 steps, the processing mechanism is uniformly encapsulated within the same model framework, realizing an automated processing flow of Parkinson's disease signals from input to diagnostic results. This avoids the subjective uncertainty caused by manual feature design and human interpretation in traditional methods, significantly improving the accuracy and efficiency of Parkinson's disease detection.

[0028] Compared with the prior art, the present invention has the following technical effects:

[0029] This invention provides an optimized method for Parkinson's disease signal processing, aiming to propose a Parkinson's disease signal processing model. By constructing a feature extraction system from the time domain to the frequency domain and then to the spatial domain, the completeness of feature representation and the accuracy of discrimination are improved. Instantaneous energy distribution and compression mapping are calculated to form energy temporal features. A multi-scale time module extracts and aligns temporal features in parallel, capturing long- and short-term dynamic changes. A spectrum reconstruction module introduces a learnable phase modulation mechanism in the frequency domain for joint analysis and feature extraction of frequency domain features. A brain region aggregation module and a graph attention layer perform partitioning integration and association enhancement of spatial features, representing the functional connections between different brain regions. A region self-attention module and a classification module perform adaptive fusion and discrimination optimization, outputting the detection results of Parkinson's disease. Through the Parkinson's disease signal processing model, the detection of Parkinson's disease signals is achieved, improving the accuracy of Parkinson's disease detection. Attached Figure Description

[0030] Figure 1 This is a flowchart of the Parkinson's disease signal processing optimization method provided by the present invention.

[0031] Figure 2 This is a flowchart of the generation energy timing characteristics provided by the present invention.

[0032] Figure 3 This is a flowchart of the multi-scale time module provided by the present invention.

[0033] Figure 4 This is a flowchart of the generation channel weighted features provided by the present invention.

[0034] Figure 5 This is a flowchart of the spectrum reconstruction module provided by the present invention.

[0035] Figure 6 This is a flowchart of obtaining graph attention features provided by the present invention.

[0036] Figure 7 This is a flowchart of the output classification results provided by the present invention.

[0037] Figure 8 This is a comparison chart of the classification results of different models on the Parkinson's disease signal dataset provided by this invention.

[0038] Figure 9 This is the confusion matrix diagram of the Parkinson's disease signal processing model on the test set. Detailed Implementation

[0039] This invention provides an optimized method for Parkinson's disease signal processing. It proposes a Parkinson's disease signal processing model, which improves the completeness of feature representation and the accuracy of discrimination by constructing a feature extraction system from the time domain to the frequency domain and then to the spatial domain. It calculates instantaneous energy distribution and compression mapping to form energy temporal features; a multi-scale temporal module extracts and aligns temporal features in parallel, capturing long- and short-term dynamic changes; a spectrum reconstruction module introduces a learnable phase modulation mechanism in the frequency domain for joint analysis and feature extraction of frequency domain features; a brain region aggregation module and a graph attention layer perform partitioning integration and association enhancement of spatial features, representing the functional connections between different brain regions; a region self-attention module and a classification module perform adaptive fusion and discrimination optimization, outputting the detection results of Parkinson's disease; through the Parkinson's disease signal processing model, the detection of Parkinson's disease signals is achieved, improving the accuracy of Parkinson's disease detection.

[0040] Please see Figure 1 As shown in the embodiment of this application, an optimized method for signal processing in Parkinson's disease is presented.

[0041] S1. Collect Parkinson's disease signal dataset and perform preprocessing operations.

[0042] Furthermore, in step S1, in this embodiment, a publicly available dataset is selected as the input data. The University of New Mexico dataset is the selected publicly available dataset, containing 27 Parkinson's disease patients and 27 healthy controls. 64-channel EEG signals were collected from the subjects in a completely relaxed state with their eyes open and closed at a sampling frequency of 500Hz. The Parkinson's group and the control group were statistically matched in terms of gender, age, and education level to ensure consistency of sample sources. Bandpass filtering of 0.5Hz to 12Hz was applied to the EEG signals from the University of New Mexico dataset to retain low-frequency oscillations and rhythmic activity components while suppressing... To suppress high-frequency noise, during the channel selection stage, 62 effective electrode channels were retained, while CPz and Pz channels were removed to ensure that the number of channels was consistent across all data files, thus obtaining the preprocessed Parkinson's disease signal. The preprocessed Parkinson's disease signal was then divided into a training set and a test set at a ratio of 9:1. Five-fold cross-validation was used for training on the training set, with 50 training epochs, a batch size of 64, and a learning rate of 0.001. Adam was used as the optimizer, and five-fold cross-validation was employed for a total of five training epochs. Each epoch was tested on the test set, and finally, the average of the five epochs' metrics was taken as the result.

[0043] S2. Calculate the instantaneous energy distribution of the Parkinson's disease signal to form the primary energy time series feature, and then perform nonlinear compression processing to generate the energy time series feature.

[0044] Further, step S2 includes: inputting a Parkinson's disease signal, calculating the instantaneous energy distribution of the Parkinson's disease signal to obtain the energy distribution matrix, then performing a sliding window average pooling operation to form a primary energy temporal feature in the time axis direction, and finally performing nonlinear compression processing and mapping the compressed Parkinson's disease signal to form an energy temporal feature.

[0045] Furthermore, in step S2, the process is as follows: Figure 2 As shown, the Parkinson's disease signal is first segmented. The Parkinson's disease signal is multi-channel time series data, denoted as N channels. Each channel consists of continuous sampling points. In this embodiment, the number of channels N is 62, the length of the time window is 1 second, and the corresponding number of sampling points is 500. The Parkinson's disease signal is segmented according to the set time window. The segmented data constitutes a three-dimensional tensor, with the dimensions corresponding to the number of samples, the number of channels, and the number of sampling points, respectively. Then, the instantaneous energy distribution is calculated. The Parkinson's disease signal is squared to obtain the instantaneous energy value of each sampling point. The square operation is achieved by squaring the amplitude of the Parkinson's disease signal point by point to form an energy distribution matrix.

[0046] Furthermore, in step S2, a sliding window average pooling operation is performed on the energy distribution matrix. The sliding window length is 32 sampling points, and the step size is 8 sampling points. Average pooling calculates the local energy average value in the time dimension, forming a temporally continuous energy time series. The energy time series maintains the original number of channels, but the number of sampling points is reduced, constituting the primary energy time series feature. In order to suppress the energy magnitude difference between different channels, nonlinear compression processing is adopted. The compression function is a logarithmic function. The primary energy time series feature is operated element-wise, and then normalized and mapped to the [0, 1] interval to obtain the energy time series feature. This maintains the correspondence between the channels and time windows of the Parkinson's disease signal and has a uniform amplitude scale.

[0047] S3. Construct a multi-scale time module to extract and align energy time series features in parallel at different time scales to obtain multi-scale time features.

[0048] Furthermore, step S3 includes: the input is energy temporal features, the multi-scale temporal module consists of multiple parallel temporal extraction units, the temporal extraction units use different temporal convolutional kernel scales to extract features, the convolutional kernel scale is set proportionally according to the sampling rate, and the output of the temporal extraction units is channel-mapped and aligned with the temporal dimension to generate multi-scale temporal features.

[0049] Furthermore, in step S3, a multi-scale time module is constructed to extract time-domain features from energy temporal features, as follows: Figure 3As shown, the multi-scale temporal module includes multiple parallel temporal extraction units. Each temporal extraction unit sequentially includes a two-dimensional convolutional layer, an average pooling layer, and a batch normalization layer to extract the temporal variation features of Parkinson's disease signals at different time scales. In this embodiment, the three temporal extraction units correspond to three time scales: long time window, medium time window, and short time window. The lengths of the convolutional kernels of the three units are converted into the number of sampling points according to the sampling rate at ratios of 1 / 2, 1 / 4, and 1 / 8, respectively, and then rounded down to obtain the feature responses at different temporal resolutions.

[0050] Furthermore, in step S3, in this embodiment, the kernel lengths of the three time extraction units are converted and set according to the sampling rate ratios of 1 / 2, 1 / 4, and 1 / 8. This is designed based on the physiological characteristics of Parkinson's disease signals exhibiting multi-layered rhythmic features and energy change patterns at different time scales. Since Parkinson's disease signals contain both low-frequency slow-wave components reflecting the overall rhythm of neural activity and high-frequency transient tremor components characterizing abnormal motor fluctuations, there are significant differences in the temporal variation characteristics of Parkinson's disease signals at different frequency bands. Using a fixed-scale convolution kernel can easily lead to feature extraction bias towards a single frequency band, resulting in local... If rhythmic information is lost or global trend is not adequately expressed, the kernel length can be dynamically set based on the sampling rate. Long-window convolution can capture low-frequency energy changes and global trends, medium-window convolution focuses on mid-frequency oscillation modes, and short-window convolution emphasizes high-frequency local dynamics. This enables hierarchical perception and complementary fusion of features with different temporal resolutions. The proportional design of the kernel length not only ensures the balance and complementarity of the multi-scale temporal modules in terms of feature coverage, but also adaptively adjusts the time window length according to the sampling rate, improving robustness and transferability under different experimental conditions. This results in more comprehensive and stable multi-scale capture of rhythmic abnormalities in Parkinson's disease signals.

[0051] Furthermore, in step S3, the energy temporal features are input into three temporal extraction units. Each temporal extraction unit performs a two-dimensional convolution operation to extract local temporal patterns. Subsequently, the local features are compressed in the temporal dimension through an average pooling layer, and the feature distribution is stabilized through a batch normalization layer. The output of each temporal extraction unit is uniformly mapped in the channel dimension through a 1×1 convolutional layer, and the features of different time scales are aligned to the same length in the temporal dimension through an adaptive average pooling layer, generating multi-scale temporal features containing information of multiple time scales.

[0052] S4. An adaptive fusion module is used to fuse multi-scale temporal features, and a channel attention module is used to nonlinearly adjust the weights of each channel feature to generate channel-weighted features.

[0053] Further, step S4 includes: the input is multi-scale temporal features, the adaptive fusion module generates scale weights based on the global statistical information of the multi-scale temporal features, and fuses the multi-scale temporal features to obtain scale fusion features, which are then input into the channel attention module, pass through the weight generation network to generate channel weight coefficients, and perform nonlinear weighting processing on the scale fusion features to obtain channel weighted features.

[0054] Furthermore, in step S4, an adaptive fusion module and a channel attention module are constructed to perform weighted fusion of multi-scale temporal features and generate channel-weighted features, as follows: Figure 4 As shown; the input is multi-scale temporal features. First, a global averaging operation is performed to obtain the global scale features at each scale. Then, a fusion unit is constructed, which consists of one-dimensional convolution, ReLU activation function, one-dimensional convolution and Softmax function normalization in sequence. The global scale features are input into the fusion unit to generate scale weights. The multi-scale temporal features are weighted and summed along the scale dimension according to the scale weights to obtain the scale fusion features.

[0055] Further, in step S4, the channel attention module receives the scale fusion features as input; first, global average pooling is performed in the time dimension to obtain the channel description vector; then the channel description vector is input into a weight generation network consisting of two fully connected layers. The first layer reduces the channel dimension and uses the ReLU activation function to extract compact features, and the second layer restores the original channel dimension and uses the Sigmoid function to constrain the output range [0, 1] to obtain the channel weight coefficients. Based on the channel weight coefficients, the channel features in the scale fusion features are weighted channel by channel to generate channel-weighted features.

[0056] S5. Construct a spectrum reconstruction module to map the channel weighted features to the frequency domain to obtain frequency domain features. Introduce phase modulation parameters to modulate the phase modulation components and obtain steady-state spectrum features.

[0057] Further, step S5 includes: the input is channel weighted features; the spectrum reconstruction module performs frequency domain transformation on the channel weighted features to obtain complex spectral features containing energy mapping components and phase modulation components; adaptive energy normalization is performed on the energy mapping components to obtain spectral domain energy features; phase modulation parameters are introduced to modulate the phase modulation components to obtain phase modulation structure quantities; the spectral domain energy features and phase modulation structure quantities are combined to form a spectrum reconstruction representation; and then, through inverse Fourier transform, a steady-state spectral feature is generated.

[0058] Furthermore, in step S5, a spectrum reconstruction module is constructed to perform frequency domain transformation, amplitude normalization, and phase modulation processing, as follows: Figure 5 As shown, the input is channel-weighted features. Perform a Fast Fourier Transform on the input to obtain the frequency domain features in complex form. , ,in For Fourier transform operators, For energy mapping components, For phase modulation components, It is the imaginary unit; the mean and standard deviation of each channel are calculated, and dynamic normalization is performed on the energy mapping components to keep the spectral amplitude distribution of different channels and samples within a uniform statistical range. Adaptive energy standardization is then performed to generate spectral domain energy features. , ,in, and These represent the mean and standard deviation of the energy mapping components for each channel. To prevent the constant term from being divided by zero, The value is .

[0059] Furthermore, in step S5, learnable phase modulation parameters are introduced. The phase modulation component is modulated to obtain the phase modulation structure quantity. This allows for the establishment of a learnable correlation between amplitude distribution and phase structure in the frequency domain, enabling adaptive correction of phase shifts in characteristic rhythms. ,in, The modulation intensity coefficient, The value is , Let be a hyperbolic tangent function, such that its range is limited to [−π,π]. The phase modulation network generates and jointly optimizes the phase modulation parameters during the training phase. The phase modulation network takes the phase modulation components as input, processes them through one-dimensional convolutional layers, batch normalization layers, and a hyperbolic tangent activation function, and outputs the phase modulation parameters. The spectral domain energy features and phase modulation structure quantities are then reconstructed into a complex spectral reconstruction representation. , Then, the steady-state spectral characteristics are obtained through inverse Fourier transform.

[0060] Furthermore, in step S5, the phase modulation network is designed based on the characteristics of Parkinson's disease signals, which exhibit significant individual variability in phase, severe phase entanglement, and sensitivity to noise. On the one hand, the pathological rhythms associated with Parkinson's disease often manifest as changes in phase synchronicity and phase shift patterns between different frequency bands. Simply normalizing the energy spectrum is insufficient to characterize the phase information closely related to Parkinson's disease. By introducing learnable phase modulation parameters and incorporating labels for reverse updates during training, the network can adaptively correct phase drift introduced by differences in electrode placement, conduction paths, and noise interference, making the spectral structure more consistent with Parkinson's disease. The study investigates the real differences in rhythm synchronicity between Parkinson's disease patients and healthy controls. Furthermore, by employing a hyperbolic tangent function as a nonlinear mapping and combining it with bounded modulation intensity coefficients, the study ensures that the phase modulation components are confined to a finite interval while avoiding spectral distortion and gradient explosion caused by excessive phase abrupt changes. This improves the numerical stability and network convergence performance during training. Through the design of the phase modulation network, both steady-state energy characteristics and phase coupling characteristics can be preserved in the frequency domain, making it more sensitive to rhythmic abnormalities and phase synchronization imbalances related to Parkinson's disease. This significantly enhances the discriminative ability and overall detection performance of Parkinson's disease signal spectral reconstruction.

[0061] S6. Based on the brain region division index, the spectral steady-state features are aggregated into regions to form region aggregated features, which are then sent to the graph attention layer to model the spatial dependencies between brain regions and obtain graph attention features.

[0062] Furthermore, step S6 includes: the input is the steady-state spectral features; the brain region aggregation module performs region aggregation operations on the steady-state spectral features based on the brain region functional division index to form region aggregation features, which are then sent to the graph attention layer. By constructing the connection relationship between brain regions, attention weighting calculation is performed on the region aggregation features to learn the spatial dependency relationship and interaction pattern between brain regions and generate graph attention features.

[0063] Furthermore, in step S6, the steady-state characteristics of the spectrum are analyzed. Brain region aggregation processing is performed and a graph attention representation is constructed, in which... For channel indexing, For time sampling points, the process is as follows: Figure 6 As shown; firstly, based on the brain region division index table, each channel is grouped according to the function of the corresponding brain region, forming a brain region set. , For the first There are several brain regions, each containing several corresponding channel numbers. A region aggregation operation is performed on the steady-state spectral features within each brain region. An averaging aggregation method along the channel dimension is used to average the features of all channels within the brain region element-wise, resulting in a brain region feature vector. , , For the first Individual brain regions are then stacked in order to obtain region aggregation features. , For the first Brain region feature vectors of each brain region.

[0064] Furthermore, in step S6, based on the neuroscience 10-20 electrode system, in this embodiment, the brain is divided into 12 regions. =12, with specific brain region groupings as follows: ['Fp1', 'Fp2'], ['AF3', 'AF4', 'AF7', 'AF8', 'AFz'], ['F1', 'F2', 'F3', 'F4', 'F5', 'F6', 'F7', 'F8', 'Fz'], ['FT7', 'FT8', 'FT9', 'FT10'], ['FC1', 'FC2', 'FC3', 'FC4', 'FC5', 'FC6', 'FCz'], ['C1', 'C2', 'C3', 'C4', 'C5', 'C6', 'Cz'], ['T7', 'T8'], ['TP7', 'TP8', 'TP9', 'TP10'], ['CP1', 'CP2', 'CP3', 'CP4', 'CP5', 'CP6'], ['P1', 'P2', 'P3', 'P4', 'P5', 'P6', 'P7', 'P8'], ['PO3', 'PO4', 'PO7', 'PO8', 'POz'], ['O1', 'O2', 'Oz'].

[0065] Furthermore, in step S6, the channels are divided into 12 brain regions based on the neuroscience 10-20 electrode system. The design fully considers the anatomical location of the electrodes on the scalp, functional zoning, and the statistical characteristics of Parkinson's disease signals. Among these, ['Fp1', 'Fp2'] correspond to the frontal pole region, reflecting prefrontal activity related to the prefrontal cortex; ['AF3', 'AF4', 'AF7', 'AF8', 'AFz'] and ['F1', 'F2', 'F3', 'F4', 'F5', 'F6', 'F7', 'F8', 'Fz'] together cover the suprafrontal and midfrontal regions, facilitating the aggregation of signals related to higher motor control, executive function, and disease-related prefrontal rhythm changes; ['FT7', 'FT8', 'FT9', 'FT10'], ['T7', 'T8'], and ['TP7', [TP8, TP9, TP10] correspond to the anterior temporal region and the temporoparietal transition zone, respectively, facilitating the characterization of activity features related to hearing, the limbic system, and multimodal integration; [FC1, FC2, FC3, FC4, FC5, FC6, FCz] and [C1, C2, C3, C4, C5, C6, Cz] cover the frontocentral and central regions, closely related to the motor cortex and sensorimotor circuits, and are key areas for characterizing the pathodynamic rhythms associated with motor symptoms in Parkinson's disease; [CP1, CP2, CP3, CP4, CP5, CP6] and [P1, P2, P3, P4, P5, P6] [P7, P8] correspond to the paraparietal and parietal lobe areas, which helps to integrate rhythmic information related to somatosensory perception, spatial perception, and posture regulation; [PO3, PO4, PO7, PO8, POz] and [O1, O2, Oz] cover the parietal and occipital lobe areas, which are used to aggregate background rhythms related to vision and their modulation effects on motor networks; by grouping brain regions with electrodes that are functionally similar, geographically adjacent, and bilaterally symmetrical, high-dimensional channel features are compressed into physiologically meaningful regional features in space, reducing redundant channel noise and improving the statistical stability of features. On the other hand, it ensures that the brain region division is consistent with the region division system commonly used in clinical and neuroscience, so that the subsequent learning of brain region connectivity by graph attention has clear anatomical and functional orientation. Thus, while controlling the complexity of the model, it enhances the discriminability and interpretability of abnormal functional connectivity and rhythm imbalance patterns related to Parkinson's disease.

[0066] Furthermore, in step S6, the regional aggregation features are... As input to the graph attention layer, the feature vectors of each brain region are first transformed linearly to obtain the node embedding vectors. , The weight matrix is ​​a linear transformation; for any pair of adjacent brain regions and Attention coefficients are obtained through combination operations of node embedding vectors. Attention coefficients are converted into attention weights by the LeakyReLU activation function and Softmax normalization. Then, based on the attention weights, a weighted summation is performed on the node embedding vectors of adjacent nodes to obtain the updated attention-related brain region features. ,in, To be related to brain regions A set of connected neighboring brain regions, after being processed by a graph attention layer, is then stacked sequentially to generate graph attention features. , For the first Attentional characteristics of individual brain regions.

[0067] S7. Input the graph attention features into the region self-attention module, calculate the correlation matrix between brain regions, obtain global fusion features, and output the classification results through the classification module.

[0068] Further, step S7 includes: the input is graph attention features, the region self-attention module calculates the correlation matrix of brain regions based on the graph attention features and performs weighted fusion to generate global fusion features, which are then fed into a classification module composed of residual structures and linear classification layers. The residual structures perform inter-layer connections and feature mapping on the input features, and the linear classification layer calculates the class probability based on the features output by the residual structures. Finally, the classification results of Parkinson's disease patients and healthy controls are output.

[0069] Furthermore, in step S7, the process is as follows: Figure 7 As shown, the input is graph attention features. The region self-attention module first applies query mapping, key mapping, and value mapping to the input graph attention features, forming three different feature representations in the same feature space. Then, it calculates the matching score between any two brain regions based on the dot product between the query vector and the key vector, and suppresses the numerical amplification effect caused by scale increase by dividing by the square root of the feature dimension. On this basis, Softmax normalization is applied to the matching score of each brain region along the same direction to construct a correlation matrix that reflects the inter-brain region dependency. Then, it is weighted and fused with the value-mapped brain region features by matrix multiplication to form a global fused feature.

[0070] Furthermore, in step S7, the globally fused features are flattened into a single vector in the order of brain region dimension and feature dimension as input to the classification module. The classification module first applies layer normalization to the input to reduce training instability caused by scale differences between different samples and different brain regions. The normalized feature vector is then input into a residual structure composed of two fully connected mapping layers. The first layer maps the input dimension to the preset hidden dimension 128 and applies the ReLU activation function to introduce stronger representation ability. Then, the random deactivation layer is discarded to improve the generalization effect of the model. The second layer remaps the hidden dimension back to the original dimension and adds it to the input of the residual block through a short-circuit connection, thereby forming a residual output with inter-layer connection and feature mapping function. On this basis, the residual output is further fed into the linear classification layer. The linear classification layer calculates the original classification score of each category based on the weighted fused features. In the inference stage, the original classification score is converted into category probability through the Softmax function, and finally the judgment result of Parkinson's disease patient or healthy control is output.

[0071] S8. Construct a Parkinson's disease signal processing model. The Parkinson's disease signal dataset is sequentially processed from step S2 through step S7 and the model is trained. After obtaining the Parkinson's disease signals to be detected, the data is input into the trained model to obtain the processed results.

[0072] Further, in step S8, a Parkinson's disease signal processing model is constructed. The processing structures from steps S2 to S7 are sequentially integrated into a unified end-to-end detection framework for training and detecting Parkinson's disease signals. During the training phase, the Parkinson's disease signal dataset is used as input to calculate the instantaneous energy distribution and compression mapping, forming energy temporal features. These features are then extracted using a multi-scale time module to extract long- and short-term dynamic change information, forming multi-scale time features. After weighted integration by an adaptive fusion module and a channel attention module, channel-weighted features are obtained. These are then input into a spectrum reconstruction module to obtain spectral steady-state features that combine energy steady-state characteristics and phase modulation correlation. These features are then converted into graph attention features reflecting the functional connectivity of brain regions by a brain region aggregation module and a graph attention layer. Finally, these features are processed by the region... The domain self-attention module and classification module generate global fusion features and output corresponding category judgments. During iterative training, parameters are updated in reverse based on the real labels of Parkinson's disease patients and healthy controls, and the Parkinson's disease signal processing model converges to a stable discriminative ability. In the detection phase, the Parkinson's disease signals to be detected are passed through the above modules in the same process as in the training phase, and the trained Parkinson's disease signal processing model directly outputs the Parkinson's disease detection results. Through the setting of step S8, the processing mechanism is uniformly encapsulated within the same model framework, realizing an automated processing flow of Parkinson's disease signals from input to diagnostic results. This avoids the subjective uncertainty caused by manual feature design and manual interpretation in traditional methods, and significantly improves the accuracy and efficiency of Parkinson's disease detection.

[0073] To verify the effectiveness of the Parkinson's disease signal processing model, PyCharm was used to write the code, and PyTorch was used as the framework. The Parkinson's disease signal dataset was input into the model for training and testing, and the processed results were obtained.

[0074] like Figure 8 As shown, the Parkinson's disease signal processing model ASLNet proposed in this invention achieves an accuracy of 75.79% and an F1 score of 75.35%, which are significantly improved compared with traditional deep convolutional models (DeepConvNet, EEGNet) and temporal structure models (TSeception, Conformer). This indicates that the Parkinson's disease signal processing model significantly outperforms the comparative models in classification performance, possessing high detection accuracy and strong discriminative ability. Figure 9 As shown, the Parkinson's disease signal processing model achieved an accuracy rate of 0.86 for healthy controls and 0.65 for Parkinson's disease patients. While maintaining high accuracy, it also demonstrated strong recognition capabilities for Parkinson's disease signals. This verifies that the Parkinson's disease signal processing model of this invention can effectively capture the feature distribution related to rhythm abnormalities and functional imbalances in Parkinson's disease signals, enabling accurate detection and classification of individuals with Parkinson's disease. This provides high-performance technical support for intelligent diagnosis of Parkinson's disease based on neural electrical signals.

[0075] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. An optimized method for signal processing in Parkinson's disease, characterized in that, Includes the following steps: S1. Collect Parkinson's disease signal dataset and perform preprocessing operations; S2. Calculate the instantaneous energy distribution of the Parkinson's disease signal to form the primary energy time series feature, and then perform nonlinear compression processing to generate the energy time series feature; S3. Construct a multi-scale time module to extract and align energy time series features in parallel at different time scales to obtain multi-scale time features; S4. An adaptive fusion module is used to fuse multi-scale temporal features, and a channel attention module is used to non-linearly adjust the weights of each channel feature to generate channel-weighted features. S5. Construct a spectrum reconstruction module to map the channel weighted features to the frequency domain to obtain frequency domain features, introduce phase modulation parameters, modulate the phase modulation components, and obtain steady-state spectrum features. S6. Based on the brain region division index, the spectral steady-state features are aggregated into regions to form region aggregated features, which are then sent to the graph attention layer to model the spatial dependencies between brain regions and obtain graph attention features. S7. Input the graph attention features into the region self-attention module, calculate the correlation matrix between brain regions, obtain global fusion features, and output the classification results through the classification module. S8. Construct a Parkinson's disease signal processing model. The Parkinson's disease signal dataset is sequentially passed from S2 through S7 and trained. After obtaining the Parkinson's disease signals to be detected, they are input into the trained model to obtain the processed results.

2. The optimized method for signal processing in Parkinson's disease according to claim 1, characterized in that, Step S2 includes: inputting a Parkinson's disease signal, calculating the instantaneous energy distribution of the Parkinson's disease signal to obtain the energy distribution matrix, then performing a sliding window average pooling operation to form a primary energy temporal feature in the time axis direction, and finally performing nonlinear compression processing and mapping the compressed Parkinson's disease signal to form an energy temporal feature.

3. The optimized method for signal processing in Parkinson's disease according to claim 2, characterized in that, Step S3 includes: the input is energy time series features, the multi-scale time module consists of multiple parallel time extraction units, the time extraction units use different time convolution kernel scales to extract features, the convolution kernel scale is set proportionally according to the sampling rate, and the output of the time extraction unit is channel mapped and aligned with the time dimension to generate multi-scale time features.

4. The optimization method for signal processing in Parkinson's disease according to claim 3, characterized in that, Step S4 includes: the input is multi-scale temporal features, the adaptive fusion module generates scale weights based on the global statistical information of the multi-scale temporal features, and fuses the multi-scale temporal features to obtain scale fusion features, which are then input into the channel attention module, pass through the weight generation network to generate channel weight coefficients, and perform nonlinear weighting processing on the scale fusion features to obtain channel weighted features.

5. An optimized method for signal processing in Parkinson's disease according to claim 4, characterized in that, Step S5 includes: the input is channel weighted features; the spectrum reconstruction module performs frequency domain transformation on the channel weighted features to obtain complex spectral features containing energy mapping components and phase modulation components; adaptive energy normalization is performed on the energy mapping components to obtain spectral domain energy features; phase modulation parameters are introduced to modulate the phase modulation components to obtain phase modulation structure quantities; the spectral domain energy features and phase modulation structure quantities are combined to form a spectrum reconstruction representation; and then, through inverse Fourier transform, a steady-state spectral feature is generated.

6. An optimization method for signal processing in Parkinson's disease according to claim 5, characterized in that, Step S6 includes: The input is the steady-state spectral features. The brain region aggregation module performs region aggregation operations on the steady-state spectral features based on the brain region functional division index to form region aggregation features. These features are then sent to the graph attention layer. By constructing the connection relationships between brain regions, attention-weighted calculations are performed on the region aggregation features to learn the spatial dependencies and interaction patterns between brain regions and generate graph attention features.

7. An optimized method for signal processing in Parkinson's disease according to claim 6, characterized in that, Step S7 includes: the input is graph attention features, the region self-attention module calculates the correlation matrix of brain regions based on the graph attention features and performs weighted fusion to generate global fusion features, which are then fed into the classification module composed of residual structure and linear classification layer. The residual structure performs inter-layer connections and feature mapping on the input features, and the linear classification layer calculates the class probability based on the features output by the residual structure. Finally, the classification results of Parkinson's disease patients and healthy controls are output.

Images