A central nervous system disease acupuncture response prediction method based on clinical characteristics

Abstract

The present application relates to the technical field of acupuncture response prediction, and discloses a central nervous system disease acupuncture response prediction method based on clinical characteristics; wherein, a spectrum embedding network module captures frequency changes at different levels through a multi-level feature embedding method, then introduces dynamic spectrum disturbance to each feature, evaluates the contribution of the feature in the frequency domain, and dynamically adjusts the learning weight of the feature in the model, finally extracts key frequency components in combination with spectrum disturbance analysis; an adaptive decoding module of the spectrum feature constructs a weighting matrix through disturbance response energy, inputs the weighted spectrum components to an adaptive spectrum decoding mechanism, realizes band-level decoupling transformation and aggregation, and through a spectrum partition weighted fusion mechanism, fuses low-frequency and high-frequency subbands according to the energy proportion dominated by the disturbance response, so as to realize efficient analysis and integration of multi-frequency information, enhance the discrimination ability of the model to acupuncture response, and solve the problem of predicting the acupuncture response of central nervous system diseases.

Description

Technical Field

[0001] This invention relates to the field of acupuncture response prediction technology, specifically to a method for predicting acupuncture responses in central nervous system diseases based on clinical characteristics. Background Technology

[0002] Central nervous system diseases are a class of neurological dysfunctions that seriously affect patients' cognition, motor function, and quality of life, such as stroke, Parkinson's disease, and Alzheimer's disease. In recent years, acupuncture, as an important traditional Chinese medicine treatment, has received widespread attention for its auxiliary therapeutic effects in the rehabilitation of central nervous system diseases. Clinical studies have shown that there are significant differences in the response of different individuals to acupuncture stimulation, and its efficacy is affected by multiple factors such as the patient's clinical characteristics, physiological state, and disease type. Therefore, establishing a method for predicting acupuncture response based on clinical characteristics can not only help in the formulation of individualized treatment plans, but also improve the accuracy and safety of acupuncture intervention, which is of great significance in the modernization of traditional Chinese medicine and the development of precision medicine.

[0003] Currently, researchers both domestically and internationally have begun to explore the use of machine learning methods to predict and model the efficacy of acupuncture. Common techniques include support vector machines, random forests, artificial neural networks, and ensemble learning methods. These methods model clinical data such as patients' basic physical signs, medical history, and laboratory indicators to identify groups that are sensitive or have low responses to acupuncture. Some studies have also integrated neural activity data such as electroencephalogram (EEG) signals and functional magnetic resonance imaging (fMRI) to construct multimodal predictive models, thereby improving the explanatory power of acupuncture mechanisms and the accuracy of response assessment.

[0004] However, existing research still has many limitations. First, some methods rely excessively on high-dimensional but difficult-to-obtain neuroimaging data, which limits their application in primary care settings. Second, most current models are black-box structures, lacking interpretability of the contribution of key features, making it difficult to guide clinical practice. In addition, existing methods fail to fully consider the dynamic changes in acupuncture responses and their interaction with individual patient characteristics, resulting in poor stability and generalization ability of prediction results. Therefore, there is an urgent need for an acupuncture response modeling method based on clinically readily available features, which combines interpretability, adaptability, and predictive ability to promote the precision development of acupuncture treatment for central nervous system diseases. Summary of the Invention

[0005] The main objective of this invention is to provide a method for predicting acupuncture responses in central nervous system diseases based on clinical features. The core of this method lies in designing a spectral embedding network module and an adaptive decoding module for spectral features to enhance the model's feature perception and prediction capabilities. The spectral embedding network module captures frequency changes at different levels through multi-level feature embedding, introduces dynamic spectral perturbations into each feature, assesses its importance in the frequency domain, and dynamically adjusts the model's learning weights for the features. Finally, it extracts key frequency components by combining spectral perturbation analysis. The adaptive decoding module constructs a weighted matrix using perturbation response energy, weights the spectral components channel-by-channel, and inputs them into the adaptive spectral decoding mechanism. This achieves decoupling transformation and aggregation at the frequency band level. Furthermore, through a spectral partitioning weighted fusion mechanism, it fuses low-frequency and high-frequency sub-bands based on the energy ratio dominated by the perturbation response, thereby achieving efficient analysis and integration of multi-frequency information and significantly enhancing the central nervous system disease acupuncture response prediction model's ability to understand and discriminate the structure of acupuncture responses.

[0006] To achieve the above objectives, the technical solution of the present invention is: a method for predicting acupuncture responses in central nervous system diseases based on clinical characteristics, the method comprising:

[0007] S1. Collect relevant information on acupuncture responses to central nervous system diseases, including: patient-related demographic information, clinical signs and neurological function assessment, and acupuncture treatment response assessment, which together constitute a dataset of acupuncture responses to central nervous system diseases.

[0008] S2. Preprocessing of the acupuncture response dataset for central nervous system diseases, including methods such as missing value imputation, outlier detection and removal, data normalization, and label consistency verification.

[0009] S3. A spectrum embedding network module is proposed. This module first captures frequency variations at different levels through multi-level feature embedding. Second, each feature undergoes dynamic spectrum perturbation. The values ​​obtained through the perturbation mechanism are used to evaluate the feature's fit for prediction, dynamically adjusting the model's learning weights for the features. Finally, through spectrum perturbation analysis, based on... Values ​​are used to filter the frequency components of the disturbance response;

[0010] S4. An adaptive decoding module for spectral features is proposed. A weighting matrix is ​​constructed using the energy of the disturbance response, and each spectral component is dynamically weighted channel by channel. The weighted spectral features are then input into the adaptive spectral decoding mechanism, which performs hierarchical analysis of multi-frequency information through decoupling transformation and aggregation at the frequency band level. Finally, the low-frequency and high-frequency sub-bands are fused according to the energy ratio dominated by the disturbance response through a spectral partitioning weighted fusion mechanism.

[0011] S5. Construct a central nervous system disease acupuncture response prediction model, which includes data preprocessing, a spectrum embedding network module, an adaptive decoding module for spectrum features, and loss function optimization; wherein, loss function optimization is the training phase of the prediction model. After the prediction model is trained, the test set is input into the central nervous system disease acupuncture response prediction model, and finally the central nervous system disease acupuncture response prediction model outputs the type of acupuncture response.

[0012] Furthermore, in step S3, for the training set... ,in For the first One sample, The number of features for each sample; each feature vector Mapped to the frequency domain, this yields a multi-level spectral representation, with each feature... After being converted into frequency components by the Discrete Fourier Transform, its mathematical model is as follows:

[0013] ;

[0014] In the formula, For the first In the nth sample The value of each feature, For the first In the nth sample The nth frequency component of the feature For the first In the nth sample The frequency components of each feature, For the index of frequency components, The embedding frequency dimension; the component of each feature in the frequency domain reflects its contribution to the frequency distribution in the time domain. To capture low-frequency and high-frequency information, the frequency representation is divided into multiple frequency bands, and the mathematical model of these multiple frequency bands is as follows:

[0015] ;

[0016] In the formula, M represents the number of frequency bands, and each frequency band corresponds to a feature representation at a different scale;

[0017] To evaluate the fit of each feature to the prediction across different frequency bands, an adaptive spectral perturbation analysis mechanism is introduced. Perturbations are introduced into each frequency component, and their impact on the model prediction is quantified by calculating the energy change after the perturbation. Firstly, for the... Frequency components of each feature Introducing a perturbation vector Its mathematical model is:

[0018] ;

[0019] in, The disturbance amplitude controls the degree to which the disturbance affects the spectrum. The direction of the disturbance. The spectrum representation after perturbation is given; then the energy change of the spectrum after perturbation is calculated to obtain the perturbation response energy, the mathematical model of which is:

[0020] ;

[0021] In the formula, Let be the perturbation energy response of the r-th feature, reflecting the impact of the perturbation on the prediction results. The energy of the original spectrum, The energy of the perturbed spectrum reflects the overall impact of the perturbation on the frequency components.

[0022] Through spectral perturbation analysis, the frequency components that have the most impact on the model are selected; perturbation energy response. Features are used to evaluate the fit of features to the prediction, and features with strong perturbation responses are selected as feature inputs. Based on the perturbation response energy, features are selected, and the mathematical model is as follows:

[0023] ;

[0024] In the formula, , A threshold is used to select features with larger perturbation responses. This is the feature set after perturbation filtering.

[0025] Furthermore, in step S3, the spectrum embedding network module captures potential frequency change patterns in the acupuncture response data of central nervous system diseases through multi-level feature embedding, and then introduces a dynamic spectrum perturbation mechanism to evaluate the response sensitivity of each feature in the frequency domain, thereby quantifying its importance in the prediction task; finally, the key frequency components in the acupuncture response data of central nervous system diseases are screened by the intensity of spectrum perturbation response, so that the model focuses on the information dimension with high contribution; the spectrum embedding network module enhances the ability of the central nervous system disease acupuncture response prediction model to identify nonlinear changes, improves the ability to model the hidden structure in multidimensional complex features, and realizes the synergistic optimization of feature selection and representation learning.

[0026] Furthermore, in step S4, the mathematical model is first derived from the weighted coefficients after normalization of the disturbance energy response:

[0027] ;

[0028] in, Pay attention to the weights of the normalized spectrum of the disturbance response; then calculate the spectral weighting matrix. , for features Each feature is weighted. The mathematical model is as follows:

[0029] ;

[0030] In the formula, This is an operation that uses a one-dimensional vector as its diagonal to generate a diagonal matrix. This is a spectral feature weighting matrix, corresponding to the sample This is used to adjust the amplitude of each characteristic spectral component;

[0031] Using a weighted matrix For input features We perform weighting to obtain the weighted spectral feature representation. , The mathematical model is as follows:

[0032] Subsequently, a frequency division threshold is set. This is used to divide frequencies into low-frequency and high-frequency bands, where the low-frequency band index is... The high-frequency segment index is For indexes Based on the dynamic selection of spectral energy distribution, this application sets 70% of the frequency as the upper bound of the low frequency range and 30% of the frequency range as the lower bound of the high frequency range. Subsequently, low-frequency and high-frequency spectral sub-bands are constructed, and the perturbation-weighted spectral components are extracted in segments. The sub-vector of the low-frequency band is: The sub-vectors of the high-frequency band are: Finally, the weighted spectral features after band splitting are obtained. Its mathematical model is:

[0033] ;

[0034] In the formula, , These are the low-frequency and high-frequency characteristic weighting coefficients, used to control the importance of the two frequency bands and satisfy... ;

[0035] The weighted spectral features are input into the structure response decoder of the frequency auto-cointegration transform. In the context of the structure response decoder of the frequency self-cointegration transform, To dynamically adjust the contribution of each frequency band through an adaptive mechanism, spectral cointegration residual transformation is first required for the sample. First, calculate any two dimensions of its spectral characteristics. and The linear residuals have the following mathematical model:

[0036] ;

[0037] In the formula, For the first Each spectral feature For the first Each spectral feature The cointegration coefficients are obtained by fitting the minimum residuals; subsequently, cointegration structure constraints are constructed, combining all cointegration residuals into a structure matrix, the mathematical model of which is:

[0038] ;

[0039] In the formula, For the sample The cointegration residual combination matrix contains the cointegration deviation between all spectral features and is used to measure the stable dependency structure across frequency bands. Subsequently, a nonlinear cointegration response function is introduced, taking the cointegration residual combination matrix as input, to extract its structural response features. Its mathematical model is as follows:

[0040] ;

[0041] In the formula, , This is a trainable parameter matrix used to adjust the spectral cointegration combination. To enhance the nonlinear response to disturbances, Structural response characteristics; final frequency auto-cointegration transform structural response decoder Output final value Its mathematical model is:

[0042] ;

[0043] In the formula, To output the weight vector, For bias terms, For encoder For the sample The output result;

[0044] Furthermore, in step S4, the adaptive decoding module of the spectrum features constructs a weighted matrix through the perturbation response energy to model and dynamically weight the importance of each spectrum channel, thereby enhancing the expressive power of key frequency components. Subsequently, the adaptive decoding module of the spectrum features performs band-level decoupling and aggregation on the weighted spectrum, and hierarchically analyzes the multi-frequency information structure such as low-frequency trends and high-frequency abrupt changes, thereby improving the ability to identify complex spectrum patterns. Finally, through the spectrum partitioning weighted fusion mechanism, the low-frequency and high-frequency information is optimized and fused according to the proportion of perturbation dominant energy, effectively improving the frequency domain expression accuracy and structural sensitivity of the prediction results of the adaptive decoding module of the spectrum features.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] This invention proposes a method for predicting acupuncture responses in central nervous system diseases based on clinical features. The method comprises a spectrum embedding network module and an adaptive decoding module for spectrum features. The spectrum embedding network module captures frequency changes at different levels through multi-level feature embedding, introduces dynamic spectrum perturbations into each feature, assesses its importance in the frequency domain, and dynamically adjusts the model's learning weights for the features. Finally, it extracts key frequency components by combining spectrum perturbation analysis. The adaptive decoding module constructs a weighted matrix using the perturbation response energy, weights the spectrum components channel by channel, and inputs them into the adaptive spectrum decoding mechanism to achieve decoupling transformation and aggregation at the frequency band level. Through a spectrum partitioning weighted fusion mechanism, it fuses low-frequency and high-frequency sub-bands based on the energy ratio dominated by the perturbation response, thereby achieving efficient analysis and integration of multi-frequency information and significantly enhancing the model's ability to discriminate acupuncture responses, thus solving the problem of predicting acupuncture responses in central nervous system diseases. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating the steps of a method for predicting acupuncture responses in central nervous system diseases based on clinical characteristics.

[0048] Figure 2 This is a structural diagram of the spectrum embedding network module.

[0049] Figure 3 This is a structural diagram of the adaptive decoding module based on spectral features.

[0050] Figure 4 This is a training structure diagram for a prediction model of acupuncture response to central nervous system diseases.

[0051] Figure 5 This is a training diagram for a prediction model of acupuncture response to central nervous system diseases.

[0052] Figure 6 This figure shows a comparison between the predicted and actual values ​​of the acupuncture response prediction model for central nervous system diseases. Detailed Implementation

[0053] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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.

[0054] Please see Figures 1-6The present invention provides a technical solution: a method for predicting acupuncture responses of central nervous system diseases based on clinical characteristics. The method includes the following steps: collecting relevant information on acupuncture responses of central nervous system diseases, preprocessing acupuncture response datasets of central nervous system diseases, constructing a spectrum embedding network module, constructing an adaptive decoding module for spectrum features, and constructing a prediction model for acupuncture responses of central nervous system diseases.

[0055] Please refer to Figure 1 As shown in the embodiments of this application, a method for predicting acupuncture responses to central nervous system diseases based on clinical characteristics includes the following specific steps:

[0056] 0021.S1. Collect relevant information on acupuncture responses to central nervous system diseases, including: patient-related demographic information, clinical signs and neurological function assessments, and acupuncture treatment response assessments, which together constitute a central nervous system disease acupuncture response dataset. Among them, patient-related demographic information includes: gender, age, height, weight, body mass index, disease type, duration of disease, first onset or recurrence, and previous acupuncture history. Clinical signs and neurological function assessment data include: blood pressure, heart rate, respiratory rate, NIHSS, UPDRS, MMSE, Barthel index, VAS score, and temperature and pain threshold test results. Acupuncture treatment response assessment results are used as label columns in the dataset, with label types including: high response on the day of treatment, low response within 1 hour after treatment, and no response 1 week after treatment.

[0057] S2. Preprocessing the acupuncture response dataset for central nervous system diseases, including steps such as missing value imputation, outlier detection and removal, data normalization, and label consistency verification. Missing values ​​are rare in the acupuncture response dataset for central nervous system diseases; therefore, this application removes data with missing values. Subsequently, [the data is used...]. The method performs outlier detection on acupuncture response datasets for central nervous system diseases and analyzes feature variables. Standardization, calculating the value of each sample Fractions, whose mathematical model is:

[0058] ;

[0059] In the formula, and They are respectively The mean and standard deviation, For characteristic variables The Each sample value for Sample values Fractions, if If the value is an outlier, it is considered an outlier and the sample is removed. The value is 3; This applies to datasets after outlier handling. The normalization process has the following mathematical model:

[0060] ;

[0061] In the formula, For characteristic variables The minimum value in, For characteristic variables The maximum value in, For characteristic variables The Middle Normalize the sample values, and normalize all variable values ​​so that all feature values ​​are mapped to the interval. ;

[0062] Finally, the labels were validated for consistency. Samples with high response were labeled as 0, samples with low response were labeled as 1, and samples with no response were labeled as 2, forming a central nervous system disease acupuncture response prediction dataset. The central nervous system disease acupuncture response prediction dataset was divided into a training set and a test set in an 8:2 ratio, with 16,000 samples in the training set and 4,000 samples in the test set.

[0063] S3. Construct a spectrum embedding network module, such as Figure 2 As shown, firstly, frequency variations at different levels are captured through multi-level feature embedding; secondly, each feature undergoes dynamic spectral perturbation, and the importance of the feature is evaluated through the perturbation mechanism, dynamically adjusting the model's learning weights for the feature; finally, important frequency components are selected based on response intensity through spectral perturbation analysis; the specific steps are as follows:

[0064] S301, Regarding the training set ,in For the first One sample, The number of features for the sample, with a value of 19; each feature vector Mapped to the frequency domain, this yields a multi-level spectral representation, with each feature... After being converted into frequency components by the Discrete Fourier Transform, its mathematical model is as follows:

[0065] ;

[0066] In the formula, For the first In the nth sample The value of each feature, For the first In the nth sample The nth frequency component of the feature For the first In the nth sample The frequency components of each feature, For the index of frequency components, values ​​range from 0 to... , The embedding frequency dimension is set to 512; each feature component in the frequency domain reflects its contribution to the frequency distribution in the time domain. To capture low-frequency and high-frequency information, the frequency representation is divided into multiple frequency bands, and the mathematical model for these multiple frequency bands is as follows:

[0067] ;

[0068] In the formula, M represents the number of frequency bands, and each frequency band corresponds to a feature representation at a different scale;

[0069] S302. To assess the importance of each feature across different frequency bands, an adaptive spectral perturbation analysis mechanism is introduced. Perturbations are introduced for each frequency component, and their impact on model predictions is quantified by calculating the energy change after the perturbation. First, for the... Frequency components of each feature Introducing a perturbation vector Its mathematical model is:

[0070] ;

[0071] in, The value is 0.3, representing the disturbance amplitude and controlling the degree of impact of the disturbance on the spectrum. The direction of the disturbance. The spectrum representation after perturbation is given; then the energy change of the spectrum after perturbation is calculated to obtain the perturbation response energy, the mathematical model of which is:

[0072] ;

[0073] In the formula, Let be the perturbation energy response of the r-th feature, reflecting the impact of the perturbation on the prediction results. The energy of the original spectrum, The energy of the perturbed spectrum reflects the overall impact of the perturbation on the frequency components.

[0074] S303. Through spectral perturbation analysis, select the features that have the most impact on the model; perturbation energy response. Features are used to evaluate the fit of the features to the prediction, and features with strong perturbation responses are selected as feature inputs; based on the perturbation response energy, the selected features have the following mathematical model:

[0075] ;

[0076] In the formula, , A threshold value of 0.8 is set to select features with larger perturbation responses. This is the feature set after perturbation filtering.

[0077] S4. Construct an adaptive decoding module for spectral features, such as Figure 3 As shown, a weighting matrix is ​​constructed using the perturbation response energy, and each spectral component is dynamically weighted channel by channel. The weighted spectral features are then input into an adaptive spectrum decoding mechanism, which performs hierarchical analysis of multi-frequency information through band-level decoupling transformation and aggregation. Finally, a spectrum partitioning weighted fusion mechanism is used to fuse low-frequency and high-frequency sub-bands according to the energy ratio dominated by the perturbation response. The specific steps are as follows:

[0078] S401. First, based on the weighted coefficients after normalization of the disturbance energy response, the mathematical model is as follows:

[0079] ;

[0080] in, Pay attention to the weights of the normalized spectrum of the disturbance response; then calculate the spectral weighting matrix. For feature sets Each feature is weighted. The mathematical model is as follows:

[0081] ;

[0082] In the formula, This is an operation that uses a one-dimensional vector as its diagonal to generate a diagonal matrix. This is a spectral feature weighting matrix, corresponding to the sample It is A diagonal matrix is ​​used to adjust the amplitude of each feature spectral component;

[0083] S402, Using a weighted matrix For input features We perform weighting to obtain the weighted spectral feature representation. , The mathematical model is as follows:

[0084] Subsequently, a frequency division threshold is set. This is used to divide frequencies into low-frequency and high-frequency bands, where the low-frequency band index is... The high-frequency segment index is For indexes Based on the dynamic selection of spectral energy distribution, this application sets 70% of the frequency as the upper bound of the low frequency range and 30% of the frequency range as the lower bound of the high frequency range. Subsequently, low-frequency and high-frequency spectral sub-bands are constructed, and the perturbation-weighted spectral components are extracted in segments. The sub-vector of the low-frequency band is: The sub-vectors of the high-frequency band are: Finally, the weighted spectral features after band splitting are obtained. Its mathematical model is:

[0085] ;

[0086] In the formula, , These are the low-frequency and high-frequency characteristic weighting coefficients, used to control the importance of the two frequency bands and satisfy... ;

[0087] S403. Input the weighted spectral features into the structure response decoder of the frequency auto-cointegration transform. In the context of the structure response decoder of the frequency self-cointegration transform, To dynamically adjust the contribution of each frequency band through an adaptive mechanism, spectral cointegration residual transformation is first required for the sample. First, calculate any two dimensions of its spectral characteristics. and The linear residuals have the following mathematical model:

[0088] ;

[0089] In the formula, For the first Each spectral feature For the first Each spectral feature The cointegration coefficients obtained by fitting the minimum residuals are initialized to 1 and iteratively optimized. Then, cointegration structure constraints are constructed, combining all cointegration residuals into a structure matrix. The mathematical model is as follows:

[0090] ;

[0091] In the formula, For the sample The cointegration residual combination matrix contains the cointegration deviation between all spectral features and is used to measure the stable dependency structure across frequency bands. Subsequently, a nonlinear cointegration response function is introduced, taking the cointegration residual combination matrix as input, to extract its structural response features. Its mathematical model is as follows:

[0092] ;

[0093] In the formula, , This is a trainable parameter matrix used to adjust the spectral cointegration combination. To enhance the nonlinear response to disturbances, Structural response characteristics; final frequency auto-cointegration transform structural response decoder Output final value Its mathematical model is:

[0094] ;

[0095] In the formula, To output the weight vector, For bias terms, For encoder For the sample The output result.

[0096] S5. Construct a prediction model for acupuncture responses to central nervous system diseases. This model is built from multiple aspects, including data preprocessing, a spectral embedding network module, an adaptive decoding module for spectral features, and loss function optimization. The model is built on a Linux system and implemented using the PyTorch deep learning framework. During the model training phase, a dynamic feedback adjustment mechanism and a joint loss function calculation mechanism are set to ensure that the model automatically adjusts its learning strategy at each stage to improve prediction accuracy. The training structure is as follows: Figure 4 As shown, the specific implementation method is as follows:

[0097] S501. During training, each sample in the acupuncture response dataset for central nervous system diseases is normalized and then input into the model. A spectral embedding network module is used to perform frequency domain mapping on the input samples, and the learning weights of each frequency band feature are dynamically adjusted through frequency perturbation analysis. Furthermore, an adaptive decoding module for spectral features is used to achieve multi-level analysis of low-frequency and high-frequency features through frequency band-level decoupling and fusion operations. For loss function optimization, the learning rate is set to 0.0001, the batch size to 64, the optimizer to Adam, the weight decay coefficient to 1e-4, the gradient clipping threshold to 5, and the ReLU activation function to be used. The dropout ratio is set to 0.3 in the feature decoding stage, the training epochs to 200, and the batch size to 128. In each epoch, the training set is loaded for forward propagation to calculate the loss and backpropagation to update the model parameters. A joint loss function is constructed, which utilizes the standard prediction error term and introduces an additional adjustable structure regularization term to enhance the controllability of the model output. The mathematical model of the loss function is:

[0098] ;

[0099] In the formula, For the first The true needle prick response label value of each sample. For the acupuncture response prediction model of central nervous system diseases, the first The predicted value for each sample, This is the regularization strength parameter, with a value of 0.1, representing the regularization term. To Perform regularization processing. The value of the loss function; such as Figure 5 The figure shows the training loss values ​​of the acupuncture response prediction model for central nervous system diseases. As the number of epochs increases, The continuously decreasing value indicates that the model accuracy is constantly improving and the training effect is good;

[0100] S502. After training the model using the training set in step S501, the trained model is validated by inputting the test set into the trained central nervous system disease acupuncture response prediction model to obtain the model's predicted output for acupuncture response types; such as Figure 6 As shown in the figure, the horizontal axis represents the sample number, the vertical axis represents the predicted value of acupuncture response, the dashed line represents the actual response value, and the solid line represents the model prediction value. It can be seen from the figure that the prediction results are consistent with the actual response values, with small deviations. The experimental results show that the constructed acupuncture response prediction model for central nervous system diseases can accurately predict the response type of acupuncture treatment for central nervous system diseases and has good application effects.

[0101] The specific embodiments described above further illustrate the purpose, technical solution, 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 scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, performs the following steps: S1. Preprocessing the dataset, including imputing missing data, identifying and removing outliers, and normalizing the dataset; the dataset is constructed by collecting acupuncture response information from patients with central nervous system diseases; systematic preprocessing of the dataset includes imputing missing data, identifying and removing outliers, normalizing numerical values, and verifying the consistency and completeness of predicted labels; the dataset is constructed by collecting acupuncture response information from patients with central nervous system diseases, which includes patient demographic data, clinical signs, neurological function assessment results, and response results; S2. Construct a spectrum embedding network module. Extract feature components at different frequency levels through multi-layer feature embedding. Combined with a dynamic spectrum perturbation mechanism, measure the fit of each feature to the response and adjust the feature learning weights. Finally, based on... The characteristics of the disturbance response are selected for subsequent analysis; S3. Design an adaptive decoding module for spectral features. First, construct a channel weighting matrix based on the disturbance energy to assign weights to the spectral features channel by channel. Then, realize the hierarchical reconstruction of multi-frequency information through frequency band-level structural decoupling and fusion operations. Finally, weighted fusion of low-frequency and high-frequency features based on the proportion of disturbance dominant energy. In step S2, an adaptive spectral perturbation mechanism is introduced to add a small perturbation in a specific direction to each spectral component. This small perturbation is obtained by introducing a perturbation vector to obtain the spectral representation after the perturbation. The influence of the frequency component on the model output is measured by calculating the change in energy before and after the perturbation. Specifically, for each frequency component, the difference between its original energy and the energy after the perturbation is calculated, and this energy difference is used as the perturbation response intensity in the frequency dimension. The mathematical model is as follows: ； In the formula, Let be the perturbation energy response of the r-th feature, reflecting the impact of the perturbation on the prediction results. The energy of the original spectrum, The energy of the perturbed spectrum reflects the overall impact of the perturbation on the frequency components. The disturbance amplitude controls the degree to which the disturbance affects the spectrum. This is the spectral representation after perturbation; after calculating the spectral perturbation response energy, the model uses the response intensity of each frequency component as a metric. Subsequently, it determines whether the energy change of the perturbation response exceeds a set threshold. Each feature in the spectrum is filtered, and the filtered features together constitute the final set of spectrum features. , used as input for subsequent models; In S3, the adaptive decoding module for the spectral characteristics includes: firstly, based on the perturbation response intensity of each frequency component... The frequency bands are normalized to obtain their weights; subsequently, a diagonal weighted matrix is ​​constructed. Used for spectral features Dynamic weighting at the channel level yields the weighted spectral feature representation. , The mathematical model is as follows: Subsequently, a frequency division threshold is set. This is used to divide frequencies into low-frequency and high-frequency bands, where the low-frequency band index is... The high-frequency segment index is For indexes Based on the dynamic selection of spectral energy distribution, low-frequency and high-frequency spectral sub-bands are then constructed. The perturbation-weighted spectral components are then extracted in segments. The sub-vector of the low-frequency band is: The sub-vectors of the high-frequency band are: The model extracts sub-vector representations of low and high frequencies respectively, and combines them with weights according to a set fusion coefficient to finally form the feature representation after spectral fusion. ; The weighted spectral features are input into the structure response decoder of the frequency auto-cointegration transform. In the context of the structure response decoder of the frequency self-cointegration transform, To dynamically adjust the contribution of each frequency band through an adaptive mechanism, spectral cointegration residual transformation is first required for the sample. First, calculate any two dimensions of its spectral characteristics. and The linear residuals, whose mathematical model is: ； In the formula, For the first Each spectral feature For the first Each spectral feature The cointegration coefficients are obtained by fitting the minimum residuals; subsequently, cointegration structure constraints are constructed, combining all cointegration residuals into a structure matrix, the mathematical model of which is: ； In the formula, For the sample The cointegration residual combination matrix, which contains the cointegration deviation between all spectral features, is used to measure the stable dependency structure across frequency bands. Subsequently, a nonlinear cointegration response function, consisting of a set of trainable linear transformation matrices and nonlinear activation functions, is introduced. Taking the cointegration residual combination matrix as input, its structural response features are extracted. Finally, the extracted structural response features are input to the output layer of the decoder, and samples are generated through linear weighting and bias term correction. Prediction results This enables accurate prediction of acupuncture response based on spectral cointegration structure.

2. The computer-readable storage medium according to claim 1, characterized in that, In step S2, for each original feature of each sample in the dataset, its corresponding numerical sequence is first mapped to the frequency domain to extract the structural contribution of different frequency components to the original feature. This is achieved through Fourier transform, and its mathematical model is as follows: ； In the formula, For the first In the nth sample The value of each feature, For the first In the nth sample The nth frequency component of the feature For the first In the nth sample The frequency components of each feature, For the index of frequency components, The embedding frequency dimension is used; the sequence of each feature value in the time domain is converted into a complex representation in the frequency domain, thereby capturing its response intensity across multiple frequency components; the component of each feature in the frequency domain reflects its contribution to the frequency distribution in the time domain, and the frequency representation is divided into multiple frequency bands to capture low-frequency and high-frequency information. The mathematical model for these multiple frequency bands is as follows: ； In the formula, M represents the number of frequency bands, and each frequency band corresponds to a feature representation at a different scale.

Images