Neural regulation effect evaluation system based on artificial intelligence

Abstract

The invention discloses a nerve regulation effect evaluation system based on artificial intelligence, which belongs to the technical field of intelligent medical treatment and comprises a regulation data acquisition module, a patient data optimization module, a feature integration module, an optimal feature selection module and a nerve regulation effect evaluation module. According to the method, the Harris eagle hybrid optimization algorithm is adopted to perform feature selection, the global search capability is enhanced, and the balance of global search and local search is adjusted through the adaptive contraction factor, so that the optimal feature combination is efficiently searched, the data redundancy is reduced, and the efficiency and accuracy of neural regulation effect evaluation are improved; a double-attention compact convolution bidirectional long-short-term memory model is adopted for nerve regulation effect evaluation, fine local features are efficiently extracted, time sequence changes of influence of nerve regulation on a patient are deeply captured, the sensitivity of the model to nerve regulation response of the patient is improved, more accurate and reliable intelligent nerve regulation effect evaluation is achieved, and the neural regulation effect evaluation accuracy is improved. And a doctor can be helped to optimize a nerve regulation and control treatment scheme.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent medical technology, and specifically refers to a neural regulation effect evaluation system based on artificial intelligence. Background Art

[0002] Neuroregulation is a therapeutic technology that improves pathological neurological function by regulating the electrical activity of the nervous system through external stimulation, and the neuroregulation effect evaluation system based on artificial intelligence is a system that uses artificial intelligence technology to automatically analyze multi-source data generated during the neuroregulation process and accurately evaluate the effects of neuroregulation. It aims to provide strong data support for doctors' decision-making, help doctors adjust neuroregulation plans according to the neuroregulation effects on different patients, effectively avoid excessive or insufficient treatment, thereby improving the therapeutic effect of neuroregulation and promoting the development of intelligent medicine.

[0003] However, in the existing neuromodulation effect evaluation process, the data involved in the neuromodulation effect evaluation is high in dimension and contains a large amount of redundant information, which affects the calculation efficiency of the model. In addition, the relationship between EEG signals and patient clinical data is complex, and it is difficult to find the most suitable feature combination using traditional feature selection methods, resulting in poor evaluation accuracy. There is also a technical problem that the neuromodulation effect evaluation inevitably requires the use of EEG signals, which have complex spatial and temporal dependencies. The changes in the patient's EEG signals after neuromodulation occur gradually, thus requiring a high level of temporal modeling capability of the model. Summary of the invention

[0004] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides a neural regulation effect evaluation system based on artificial intelligence. In view of the technical problems that in the existing neural regulation effect evaluation process, the data involved in the neural regulation effect evaluation is high in dimension, mixed with a large amount of redundant information, affecting the model calculation efficiency, and the relationship between EEG signals and patient clinical data is complex, and it is difficult to find the most suitable feature combination using traditional feature selection methods, resulting in poor evaluation accuracy, this solution creatively adopts the Harris Eagle hybrid optimization algorithm for feature selection, combines the locust optimization algorithm and the Levy flight distribution, enhances the global search capability, adjusts the balance between global search and local search through an adaptive shrinkage factor, and then efficiently searches for the optimal feature combination, reduces data redundancy, and improves the neural regulation effect evaluation. The efficiency and accuracy of the evaluation are improved; in view of the fact that in the existing neuromodulation effect evaluation process, the neuromodulation effect evaluation inevitably needs to use EEG signals, and EEG signals have complex spatial and temporal dependencies. The changes in the patient's EEG signals after neuromodulation occur gradually, thus requiring a high technical problem in the temporal modeling capability of the model. This solution creatively uses a dual-attention compact convolutional bidirectional long short-term memory model to evaluate the neuromodulation effect. It can efficiently extract more refined local features, deeply capture the temporal changes of the impact of neuromodulation on patients, and enhance the sensitivity of the model to the patient's neuromodulation response through dual attention blocks, thereby achieving more accurate and reliable intelligent neuromodulation effect evaluation, thereby assisting doctors to promptly identify patients with poor treatment effects and optimize treatment plans in a targeted manner.

[0005] The technical solution adopted by the present invention is as follows: the neural regulation effect evaluation system based on artificial intelligence provided by the present invention includes: a regulation data acquisition module, a patient data optimization module, a feature integration module, an optimal feature selection module and a neural regulation effect evaluation module;

[0006] The control data acquisition module is used for data acquisition, and obtains the original patient data and the control effect label through data acquisition, and sends the original patient data to the patient data optimization module, and sends the control effect label to the neural control effect evaluation module;

[0007] The patient data optimization module is used for data preprocessing, obtains patient optimized data through data preprocessing, and sends the patient optimized data to the feature integration module;

[0008] The feature integration module obtains a patient feature set through a frequency domain feature extraction unit, a time-frequency domain feature extraction unit, an EEG signal feature extraction unit and a feature integration unit, and sends the patient feature set to an optimal feature selection module;

[0009] The optimal feature selection module uses the Harris Eagle hybrid optimization algorithm to perform feature selection based on the patient feature set to obtain the patient's optimal feature subset, and sends the patient's optimal feature subset to the neuromodulation effect evaluation module;

[0010] The neural regulation effect evaluation module uses a double-attention compact convolution bidirectional long short-term memory model to evaluate the neural regulation effect based on the optimal feature subset and the regulation effect label to obtain the neural regulation effect level.

[0011] Furthermore, in the control data acquisition module, the data acquisition specifically involves collecting the EEG signals of the patient before and after the neuromodulation, and collecting the patient's clinical data to obtain the patient's original data. The doctor obtains the control effect label through manual annotation based on the psychological scale filled out by the patient;

[0012] The patient original data includes the patient's electroencephalogram signal and the patient's clinical data;

[0013] The regulatory effect labels include significant effect, minor effect and no effect.

[0014] Further, in the patient data optimization module, the data preprocessing is specifically to preprocess the original patient data through data denoising, data cleaning and data normalization to obtain patient optimized data;

[0015] The data cleaning includes outlier deletion, duplicate value filtering and missing value filling operations;

[0016] The data normalization specifically includes normalizing the numerical data in the original patient data and performing a one-hot encoding operation on the type data in the original patient data.

[0017] Furthermore, in the feature integration module, a frequency domain feature extraction unit, a time-frequency domain feature extraction unit, an EEG signal feature extraction unit and a feature integration unit are provided, including the following contents:

[0018] A frequency domain feature extraction unit extracts frequency domain features from the EEG signal by performing Fourier transform on the EEG signal in the patient's optimized data;

[0019] A time-frequency domain feature extraction unit extracts time-frequency domain features from the EEG signal by performing wavelet transform on the EEG signal in the patient optimization data;

[0020] An EEG signal feature extraction unit, which extracts features from the EEG signal in the patient's optimized data through the frequency domain feature extraction unit and the time-frequency domain feature extraction unit to obtain EEG signal features;

[0021] The feature integration unit integrates the EEG signal features with the patient clinical data in the patient optimization data to obtain a patient feature set.

[0022] Furthermore, in the best feature selection module, there are an optimization algorithm initialization unit, a fitness value calculation unit, an improved individual position update unit, a prey position update unit and an optimal feature subset calculation unit, including the following contents:

[0023] The optimization algorithm initialization unit randomly initializes the individual position in the search space and sets the maximum number of iterations. The calculation formula of the individual position is:

[0024] ;

[0025] Where, X i is the i-th individual position, which is used to represent the patient feature subset, i is the individual first index, po i1 is the position component of the ith individual in the first dimension, po i2 is the position component of the ith individual in the second dimension, po id is the position component of the ith individual in the dth dimension, d is the search space dimension, and the search space dimension is equal to the total number of patient features;

[0026] A fitness value calculation unit constructs a fitness function according to the model accuracy, and uses the fitness function to calculate the fitness value of each individual position, and takes the individual position with the largest fitness value as the prey position;

[0027] The calculation formula for constructing the fitness function based on the model accuracy is:

[0028] ;

[0029] Where F(·) is the fitness function, is the model accuracy weight, which is used to control the importance of model accuracy. is the model accuracy, specifically the accuracy of the double-attention compact convolution bidirectional long short-term memory model, is the feature quantity weight, which is used to control the importance of feature selection ratio. is the number of features selected at the i-th individual position;

[0030] Improve the individual position update unit. When the shrinkage factor is ≥ 0.5, perform a global search to update the individual position. Otherwise, perform a local search to update the individual position.

[0031] The calculation formula of the shrinkage factor is:

[0032] ;

[0033] Where c is the shrinkage factor, c max is the maximum value of the shrinkage factor, t is the iteration index, which is used to indicate the current iteration number, c min is the minimum value of the shrinkage factor, and T is the maximum number of iterations;

[0034] The global search is specifically performed using a locust optimization algorithm, and the global search is improved by introducing Levy flight distribution. The calculation formula is:

[0035] ;

[0036] Where, X i (t+1) is the position of the ith individual at the t+1th iteration, j is the second index of the individual, N is the number of individuals, ub is the upper bound of the search space, lb is the lower bound of the search space, s(·) is the social force function used to calculate the interaction force strength between individuals, and dis ij is the distance between the i-th individual position and the j-th individual position, X j (t) is the jth individual position at the tth iteration, X i (t) is the jth individual position at the tth iteration, r LF is a random number that follows the Levy flight distribution, X best (t) is the prey position at the tth iteration;

[0037] The calculation formula of the social force function is:

[0038] :

[0039] In the formula, is the gravitational factor, e is the base of natural logarithms;

[0040] The local search is specifically performed by using the Harris Hawk optimization algorithm. When the fitness value of the individual position is less than 0.5, the Harris Hawk soft encirclement strategy is executed to update the individual position, otherwise the Harris Hawk hard encirclement strategy is executed to update the individual position;

[0041] The calculation formula for executing the Harris Hawk soft encirclement strategy to update the individual position is:

[0042] ;

[0043] In the formula, dis ibest is the distance between the position of the ith individual and the position of the prey, P is the probability of prey escape, and J is the prey jump intensity;

[0044] The calculation formula for executing the Harris Hawk hard encirclement strategy to update the individual position is:

[0045] ;

[0046] In the formula, r is a random number in the range of (0,1), argmax is the maximum value function, which is used to select the candidate position with the largest fitness value from the candidate position set. is a set of candidate positions, A(t) is the first candidate position at the tth iteration, and B(t) is the second candidate position at the tth iteration;

[0047] The calculation formula of the first candidate position is:

[0048] ;

[0049] Where, X rand (t) is the random individual position at the tth iteration;

[0050] The calculation formula of the second candidate position is:

[0051] ;

[0052] Where LF(·) is the Levy flight distribution function;

[0053] A prey position updating unit, which updates the prey position through the fitness value calculation unit;

[0054] The best feature subset calculation unit calculates the prey position as the best feature subset of the patient by repeatedly executing the improved individual position updating unit and the prey position updating unit until a maximum number of iterations is reached.

[0055] Furthermore, in the neuromodulation effect evaluation module, based on the best feature subset and the regulation effect label, a double-attention compact convolution bidirectional long short-term memory model is used to evaluate the neuromodulation effect and obtain the neuromodulation effect level;

[0056] The dual-attention compact convolution bidirectional long short-term memory model includes a compact convolution subnet, a dual-attention block, a bidirectional long short-term memory subnet and an evaluation layer;

[0057] The compact convolutional subnetwork is used to efficiently extract local feature patterns;

[0058] The dual attention block is used to enhance the focus on key feature information and improve the model evaluation effect;

[0059] The bidirectional long short-term memory subnetwork is used to model temporal dependencies and capture the changing trends of patients affected by neuromodulation, so as to more accurately evaluate the effects of neuromodulation;

[0060] The evaluation layer is used to generate and output model evaluation results;

[0061] The neural regulation effect evaluation module is provided with a compact convolution subnet construction unit, a dual attention block construction unit, a bidirectional long short-term memory subnet construction unit, an evaluation layer construction unit, a model training unit and an evaluation result generation unit, including the following contents:

[0062] A compact convolutional subnet construction unit is provided in which a one-dimensional convolutional layer, a depth convolutional layer and a separable convolutional layer are provided, and a regularization block is provided after each convolutional layer, wherein the regularization block includes a batch normalization layer, an average pooling layer and a random dropout layer, and the best feature subset of the patient is processed through the compact convolutional subnet to obtain the convolutional feature of the patient;

[0063] The calculation formula of the one-dimensional convolutional layer is:

[0064] ;

[0065] Where Y tem is the output feature of the one-dimensional convolution layer, Conv1D(·) is a one-dimensional convolution operation, fea tem is the one-dimensional convolutional layer input, specifically the best feature subset, W tem is a one-dimensional convolution kernel, b tem is the one-dimensional convolution bias term;

[0066] The calculation formula of the depth convolution layer is:

[0067] ;

[0068] Where Y dep is the deep convolution output feature, DepConv2D(·) is the deep convolution operation, fea dep is the input of the deep convolutional layer, specifically the one-dimensional convolutional layer output feature after regularization block processing, W dep is the depth convolution kernel, b dep is the depth convolution bias term;

[0069] The calculation formula of the separable convolutional layer is:

[0070] ;

[0071] Where Y sep is the separable convolution output feature, SepConv(·) is the separable convolution operation, fea sep is the input of the separable convolution layer, specifically refers to the deep convolution output feature after being processed by the regularization block, W sep is a separable convolution kernel, b sep is the separable convolutional layer bias;

[0072] A dual attention block construction unit, wherein a spatial attention block and a channel attention block are set in the dual attention block, wherein the channel attention block includes two fully connected layers, and the patient convolution features are processed by the dual attention block to obtain the patient enhanced features;

[0073] The calculation formula of the spatial attention block is:

[0074] ;

[0075] In the formula, att spa is the spatial attention, Sig(·) is the S-type activation function, [·,·] is the feature concatenation operation, and F com is the patient convolution feature, GAP(·) is the global average pooling operation, MP(·) is the maximum pooling operation, and F spac is the patient's spatial attention characteristic, is the element-wise multiplication symbol;

[0076] The calculation formula of the channel attention block is:

[0077] ;

[0078] In the formula, att cha is the channel attention, W2 is the weight of the second fully connected layer, tanh(·) is the tanh activation function, W1 is the weight of the first fully connected layer, F chac It is a patient-enhanced feature;

[0079] A bidirectional long short-term memory subnet construction unit is constructed by stacking M bidirectional long short-term memory units to construct a bidirectional long short-term memory subnet, wherein the bidirectional long short-term memory unit includes a forward long short-term memory unit and a backward long short-term memory unit, and the patient enhancement feature is modeled for temporal dependency through the bidirectional long short-term memory subnet to obtain the patient temporal feature;

[0080] An evaluation layer construction unit is provided in which a fully connected layer and a Softmax activation function are set, and the patient's time series features are processed through the evaluation layer to obtain a model evaluation result;

[0081] A model training unit, which constructs a double-attention compact convolution bidirectional long short-term memory model through the compact convolution subnet construction unit, the double-attention block construction unit, the bidirectional long short-term memory subnet construction unit and the evaluation layer construction unit, and performs model training to obtain a neural regulation effect evaluation model;

[0082] The evaluation result generating unit evaluates the neuromodulation effect through the neuromodulation effect evaluation model to obtain the neuromodulation effect grade.

[0083] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0084] (1) In view of the technical problems in the existing neuromodulation effect evaluation process, the data involved in the neuromodulation effect evaluation is high in dimension and contains a large amount of redundant information, which affects the calculation efficiency of the model. In addition, the relationship between EEG signals and patient clinical data is complex, and it is difficult to find the most suitable feature combination using traditional feature selection methods, resulting in poor evaluation accuracy. This solution creatively uses the Harris Eagle hybrid optimization algorithm for feature selection, combined with the locust optimization algorithm and Levy flight distribution to enhance the global search capability, and adjusts the balance between global search and local search through an adaptive shrinkage factor, thereby efficiently searching for the optimal feature combination, reducing data redundancy, and improving the efficiency and accuracy of neuromodulation effect evaluation.

[0085] (2) In view of the technical problem that in the existing neuromodulation effect evaluation process, the neuromodulation effect evaluation inevitably needs to use EEG signals, which have complex spatial and temporal dependencies. The changes in the EEG signals of patients after neuromodulation occur gradually, thus requiring a high level of temporal modeling capability of the model, this solution creatively adopts a dual-attention compact convolutional bidirectional long short-term memory model to evaluate the neuromodulation effect. It can efficiently extract more detailed local features and deeply capture the temporal changes of the impact of neuromodulation on patients. It also improves the sensitivity of the model to the patient's neuromodulation response through dual attention blocks, achieving more accurate and reliable intelligent neuromodulation effect evaluation, thereby assisting doctors to promptly identify patients with poor treatment effects and optimize treatment plans in a targeted manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 A schematic diagram of the framework of the neural regulation effect evaluation system based on artificial intelligence provided by the present invention;

[0087] Figure 2 It is a flowchart of the feature integration module;

[0088] Figure 3 Schematic diagram of the process of the optimal feature selection module;

[0089] Figure 4 Schematic diagram of the process of the neuromodulation effect evaluation module.

[0090] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0091] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0092] In the description of the present invention, it should be understood that terms such as “upper”, “lower”, “front”, “back”, “left”, “right”, “top”, “bottom”, “inside” and “outside” indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore should not be understood as limiting the present invention.

[0093] Example 1, see Figure 1 , the neural regulation effect evaluation system based on artificial intelligence provided by the present invention includes: a regulation data acquisition module, a patient data optimization module, a feature integration module, an optimal feature selection module and a neural regulation effect evaluation module;

[0094] The control data acquisition module is used for data acquisition, and obtains the original patient data and the control effect label through data acquisition, and sends the original patient data to the patient data optimization module, and sends the control effect label to the neural control effect evaluation module;

[0095] The patient data optimization module is used for data preprocessing, obtains patient optimized data through data preprocessing, and sends the patient optimized data to the feature integration module;

[0096] The feature integration module obtains a patient feature set through a frequency domain feature extraction unit, a time-frequency domain feature extraction unit, an EEG signal feature extraction unit and a feature integration unit, and sends the patient feature set to an optimal feature selection module;

[0097] The optimal feature selection module uses the Harris Eagle hybrid optimization algorithm to perform feature selection based on the patient feature set to obtain the patient's optimal feature subset, and sends the patient's optimal feature subset to the neuromodulation effect evaluation module;

[0098] The neural regulation effect evaluation module uses a double-attention compact convolution bidirectional long short-term memory model to evaluate the neural regulation effect based on the optimal feature subset and the regulation effect label to obtain the neural regulation effect level.

[0099] Example 2, see Figure 1This embodiment is based on the above embodiment. In the control data acquisition module, the data acquisition specifically includes collecting the EEG signals of the patient before and after the neuromodulation, and collecting the patient's clinical data to obtain the patient's original data. The doctor obtains the control effect label through manual annotation based on the psychological scale filled out by the patient;

[0100] The patient original data includes the patient's electroencephalogram signal and the patient's clinical data;

[0101] The patient's clinical data, including gender, age, disease characteristics, and duration of depression;

[0102] The neuromodulation includes transcranial direct current stimulation, transcranial magnetic stimulation and transcutaneous auricular vagus nerve stimulation;

[0103] The psychological scales include the Hamilton Depression Rating Scale and the Positive and Negative Affect Scale;

[0104] The regulatory effect labels include significant effect, minor effect and no effect.

[0105] Example 3, see Figure 1 , this embodiment is based on the above embodiment, in the patient data optimization module, the data preprocessing is specifically to preprocess the original patient data through data denoising, data cleaning and data normalization to obtain patient optimized data;

[0106] The data cleaning includes outlier deletion, duplicate value filtering and missing value filling operations;

[0107] The data normalization specifically includes normalizing the numerical data in the original patient data and performing a one-hot encoding operation on the type data in the original patient data.

[0108] Example 4, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In the feature integration module, a frequency domain feature extraction unit, a time-frequency domain feature extraction unit, an EEG signal feature extraction unit and a feature integration unit are provided, and include the following contents:

[0109] A frequency domain feature extraction unit extracts frequency domain features from the EEG signal by performing Fourier transform on the EEG signal in the patient's optimized data;

[0110] A time-frequency domain feature extraction unit extracts time-frequency domain features from the EEG signal by performing wavelet transform on the EEG signal in the patient optimization data;

[0111] An EEG signal feature extraction unit, which extracts features from the EEG signal in the patient's optimized data through the frequency domain feature extraction unit and the time-frequency domain feature extraction unit to obtain EEG signal features;

[0112] The feature integration unit integrates the EEG signal features with the patient clinical data in the patient optimization data to obtain a patient feature set.

[0113] Example 5, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In the best feature selection module, an optimization algorithm initialization unit, a fitness value calculation unit, an improved individual position update unit, a prey position update unit and an optimal feature subset calculation unit are provided, and the following contents are included:

[0114] The optimization algorithm initialization unit randomly initializes the individual position in the search space and sets the maximum number of iterations. The calculation formula of the individual position is:

[0115] ;

[0116] Where, X i is the i-th individual position, which is used to represent the patient feature subset, i is the individual first index, po i1 is the position component of the ith individual in the first dimension, po i2 is the position component of the ith individual in the second dimension, po id is the position component of the ith individual in the dth dimension, d is the search space dimension, and the search space dimension is equal to the total number of patient features;

[0117] A fitness value calculation unit constructs a fitness function according to the model accuracy, and uses the fitness function to calculate the fitness value of each individual position, and takes the individual position with the largest fitness value as the prey position;

[0118] The calculation formula for constructing the fitness function based on the model accuracy is:

[0119] ;

[0120] Where F(·) is the fitness function, is the model accuracy weight, which is used to control the importance of model accuracy. is the model accuracy, specifically the accuracy of the double-attention compact convolution bidirectional long short-term memory model, is the feature quantity weight, which is used to control the importance of feature selection ratio. is the number of features selected at the i-th individual position;

[0121] Improve the individual position update unit. When the shrinkage factor is ≥ 0.5, perform a global search to update the individual position. Otherwise, perform a local search to update the individual position.

[0122] The calculation formula of the shrinkage factor is:

[0123] ;

[0124] Where c is the shrinkage factor, c max is the maximum value of the shrinkage factor, t is the iteration index, which is used to indicate the current iteration number, c min is the minimum value of the shrinkage factor, and T is the maximum number of iterations;

[0125] The global search is specifically performed using a locust optimization algorithm, and the global search is improved by introducing Levy flight distribution. The calculation formula is:

[0126] ;

[0127] Where, X i (t+1) is the position of the ith individual at the t+1th iteration, j is the second index of the individual, N is the number of individuals, ub is the upper bound of the search space, lb is the lower bound of the search space, s(·) is the social force function used to calculate the interaction force strength between individuals, and dis ij is the distance between the i-th individual position and the j-th individual position, X j (t) is the jth individual position at the tth iteration, X i (t) is the jth individual position at the tth iteration, r LF is a random number that follows the Levy flight distribution, X best (t) is the prey position at the tth iteration;

[0128] The calculation formula of the social force function is:

[0129] :

[0130] In the formula, is the gravitational factor, e is the base of natural logarithms;

[0131] The local search is specifically performed by using the Harris Hawk optimization algorithm. When the fitness value of the individual position is less than 0.5, the Harris Hawk soft encirclement strategy is executed to update the individual position, otherwise the Harris Hawk hard encirclement strategy is executed to update the individual position;

[0132] The calculation formula for executing the Harris Hawk soft encirclement strategy to update the individual position is:

[0133] ;

[0134] In the formula, dis ibest is the distance between the position of the ith individual and the position of the prey, P is the probability of prey escape, and J is the prey jump intensity;

[0135] The calculation formula for executing the Harris Hawk hard encirclement strategy to update the individual position is:

[0136] ;

[0137] In the formula, r is a random number in the range of (0,1), argmax is the maximum value function, which is used to select the candidate position with the largest fitness value from the candidate position set. is a set of candidate positions, A(t) is the first candidate position at the tth iteration, and B(t) is the second candidate position at the tth iteration;

[0138] The calculation formula of the first candidate position is:

[0139] ;

[0140] Where, X rand (t) is the random individual position at the tth iteration;

[0141] The calculation formula of the second candidate position is:

[0142] ;

[0143] Where LF(·) is the Levy flight distribution function;

[0144] A prey position updating unit, which updates the prey position through the fitness value calculation unit;

[0145] An optimal feature subset calculation unit, which calculates the prey position as the patient's optimal feature subset by repeatedly executing the improved individual position update unit and the prey position update unit until a maximum number of iterations is reached;

[0146] By performing the above operations, in order to solve the technical problems in the existing neuromodulation effect evaluation process, such as the high dimension of the data involved in the neuromodulation effect evaluation, a large amount of redundant information, which affects the calculation efficiency of the model, and the complex relationship between EEG signals and patient clinical data, it is difficult to find the most suitable feature combination using traditional feature selection methods, resulting in poor evaluation accuracy, this solution creatively adopts the Harris Eagle hybrid optimization algorithm for feature selection, combines the locust optimization algorithm and Levy flight distribution, enhances the global search capability, adjusts the balance between global search and local search through an adaptive shrinkage factor, and then efficiently searches for the optimal feature combination, reduces data redundancy, and improves the efficiency and accuracy of neuromodulation effect evaluation.

[0147] Example 6, see Figure 1 and Figure 4, based on the above embodiment, in the neural regulation effect evaluation module, according to the best feature subset and the regulation effect label, a double attention compact convolution bidirectional long short-term memory model is used to evaluate the neural regulation effect and obtain the neural regulation effect level;

[0148] The dual-attention compact convolution bidirectional long short-term memory model includes a compact convolution subnet, a dual-attention block, a bidirectional long short-term memory subnet and an evaluation layer;

[0149] The compact convolutional subnetwork is used to efficiently extract local feature patterns;

[0150] The dual attention block is used to enhance the focus on key feature information and improve the model evaluation effect;

[0151] The bidirectional long short-term memory subnetwork is used to model temporal dependencies and capture the changing trends of patients affected by neuromodulation, so as to more accurately evaluate the effects of neuromodulation;

[0152] The evaluation layer is used to generate and output model evaluation results;

[0153] The neural regulation effect evaluation module is provided with a compact convolution subnet construction unit, a dual attention block construction unit, a bidirectional long short-term memory subnet construction unit, an evaluation layer construction unit, a model training unit and an evaluation result generation unit, including the following contents:

[0154] A compact convolutional subnet construction unit is provided in which a one-dimensional convolutional layer, a depth convolutional layer and a separable convolutional layer are provided, and a regularization block is provided after each convolutional layer, wherein the regularization block includes a batch normalization layer, an average pooling layer and a random dropout layer, and the best feature subset of the patient is processed through the compact convolutional subnet to obtain the convolutional feature of the patient;

[0155] The calculation formula of the one-dimensional convolutional layer is:

[0156] ;

[0157] Where Y tem is the output feature of the one-dimensional convolution layer, Conv1D(·) is a one-dimensional convolution operation, fea tem is the one-dimensional convolutional layer input, specifically the best feature subset, W tem is a one-dimensional convolution kernel, b tem is the one-dimensional convolution bias term;

[0158] The calculation formula of the depth convolution layer is:

[0159] ;

[0160] Where Y depis the deep convolution output feature, DepConv2D(·) is the deep convolution operation, fea dep is the input of the deep convolutional layer, specifically the one-dimensional convolutional layer output feature after regularization block processing, W dep is the depth convolution kernel, b dep is the depth convolution bias term;

[0161] The calculation formula of the separable convolutional layer is:

[0162] ;

[0163] Where Y sep is the separable convolution output feature, SepConv(·) is the separable convolution operation, fea sep is the input of the separable convolution layer, specifically refers to the deep convolution output feature after being processed by the regularization block, W sep is a separable convolution kernel, b sep is the separable convolutional layer bias;

[0164] A dual attention block construction unit, wherein a spatial attention block and a channel attention block are set in the dual attention block, wherein the channel attention block includes two fully connected layers, and the patient convolution features are processed by the dual attention block to obtain the patient enhanced features;

[0165] The calculation formula of the spatial attention block is:

[0166] ;

[0167] In the formula, att spa is the spatial attention, Sig(·) is the S-type activation function, [·,·] is the feature concatenation operation, and F com is the patient convolution feature, GAP(·) is the global average pooling operation, MP(·) is the maximum pooling operation, and F spac is the patient's spatial attention characteristic, is the element-wise multiplication symbol;

[0168] The calculation formula of the channel attention block is:

[0169] ;

[0170] In the formula, att cha is the channel attention, W2 is the weight of the second fully connected layer, tanh(·) is the tanh activation function, W1 is the weight of the first fully connected layer, F chac It is a patient-enhanced feature;

[0171] A bidirectional long short-term memory subnet construction unit is constructed by stacking M bidirectional long short-term memory units to construct a bidirectional long short-term memory subnet, wherein the bidirectional long short-term memory unit includes a forward long short-term memory unit and a backward long short-term memory unit, and the patient enhancement feature is modeled for temporal dependency through the bidirectional long short-term memory subnet to obtain the patient temporal feature;

[0172] An evaluation layer construction unit is provided in which a fully connected layer and a Softmax activation function are set, and the patient's time series features are processed through the evaluation layer to obtain a model evaluation result;

[0173] A model training unit, which constructs a double-attention compact convolution bidirectional long short-term memory model through the compact convolution subnet construction unit, the double-attention block construction unit, the bidirectional long short-term memory subnet construction unit and the evaluation layer construction unit, and performs model training to obtain a neural regulation effect evaluation model;

[0174] An evaluation result generating unit, which evaluates the neuromodulation effect through a neuromodulation effect evaluation model to obtain a neuromodulation effect grade;

[0175] By performing the above operations, in view of the technical problem that in the existing neuromodulation effect evaluation process, the neuromodulation effect evaluation inevitably needs to use EEG signals, and EEG signals have complex spatial and temporal dependencies. The changes in the patient's EEG signals after neuromodulation occur gradually, and therefore require a high level of temporal modeling capability of the model, this solution creatively uses a dual-attention compact convolution bidirectional long short-term memory model to evaluate the neuromodulation effect. It can efficiently extract more refined local features, deeply capture the temporal changes of the impact of neuromodulation on patients, and enhance the sensitivity of the model to the patient's neuromodulation response through dual attention blocks, thereby achieving more accurate and reliable intelligent neuromodulation effect evaluation, thereby assisting doctors to promptly identify patients with poor treatment effects and optimize treatment plans in a targeted manner.

[0176] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0177] While the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that many changes, modifications, substitutions and variations can be made to the embodiments without departing from the principles and spirit of the invention.

[0178] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. A neural regulation effect evaluation system based on artificial intelligence, characterized by: It includes a control data acquisition module, a patient data optimization module, a feature integration module, an optimal feature selection module and a neuromodulation effect evaluation module; The control data acquisition module is used for data acquisition, and obtains the original patient data and the control effect label through data acquisition, and sends the original patient data to the patient data optimization module, and sends the control effect label to the neural control effect evaluation module; The patient data optimization module is used for data preprocessing, obtains patient optimized data through data preprocessing, and sends the patient optimized data to the feature integration module; The feature integration module obtains a patient feature set through a frequency domain feature extraction unit, a time-frequency domain feature extraction unit, an EEG signal feature extraction unit and a feature integration unit, and sends the patient feature set to an optimal feature selection module; The optimal feature selection module uses the Harris Eagle hybrid optimization algorithm to perform feature selection based on the patient feature set to obtain the patient's optimal feature subset, and sends the patient's optimal feature subset to the neuromodulation effect evaluation module; The neural regulation effect evaluation module uses a double-attention compact convolution bidirectional long short-term memory model to evaluate the neural regulation effect based on the optimal feature subset and the regulation effect label to obtain the neural regulation effect level.

2. The neural regulation effect evaluation system based on artificial intelligence according to claim 1, characterized in that: In the best feature selection module, there are an optimization algorithm initialization unit, a fitness value calculation unit, an improved individual position update unit, a prey position update unit and an optimal feature subset calculation unit, including the following contents: The optimization algorithm initialization unit randomly initializes the individual position in the search space and sets the maximum number of iterations. The calculation formula of the individual position is: ； Where, X i is the i-th individual position, which is used to represent the patient feature subset, i is the individual first index, po i1 is the position component of the ith individual in the first dimension, po i2 is the position component of the ith individual in the second dimension, po id is the position component of the ith individual in the dth dimension, d is the search space dimension, and the search space dimension is equal to the total number of patient features; A fitness value calculation unit constructs a fitness function according to the model accuracy, and uses the fitness function to calculate the fitness value of each individual position, and takes the individual position with the largest fitness value as the prey position; The calculation formula for constructing the fitness function based on the model accuracy is: ； Where F(·) is the fitness function, is the model accuracy weight, which is used to control the importance of model accuracy. is the model accuracy, specifically the accuracy of the double-attention compact convolution bidirectional long short-term memory model, is the feature quantity weight, which is used to control the importance of feature selection ratio. is the number of features selected at the i-th individual position; Improve the individual position update unit. When the shrinkage factor is ≥ 0.5, perform a global search to update the individual position. Otherwise, perform a local search to update the individual position. A prey position updating unit, which updates the prey position through the fitness value calculation unit; The best feature subset calculation unit calculates the prey position as the best feature subset of the patient by repeatedly executing the improved individual position updating unit and the prey position updating unit until a maximum number of iterations is reached.

3. The artificial intelligence-based neural regulation effect evaluation system according to claim 2, characterized in that: In the improved individual position update unit, the calculation formula of the shrinkage factor is: ； Where c is the shrinkage factor, c max is the maximum value of the shrinkage factor, t is the iteration index, which is used to indicate the current iteration number, c min is the minimum value of the shrinkage factor, and T is the maximum number of iterations; The global search is specifically performed using a locust optimization algorithm, and the global search is improved by introducing Levy flight distribution. The calculation formula is: ； Where, X i (t+1) is the position of the ith individual at the t+1th iteration, j is the second index of the individual, N is the number of individuals, ub is the upper bound of the search space, lb is the lower bound of the search space, s(·) is the social force function used to calculate the interaction force strength between individuals, and dis ij is the distance between the i-th individual position and the j-th individual position, X j (t) is the jth individual position at the tth iteration, X i (t) is the jth individual position at the tth iteration, r LF is a random number that follows the Levy flight distribution, X best (t) is the prey position at the tth iteration; The calculation formula of the social force function is: ： In the formula, is the gravitational factor, e is the base of natural logarithms; The local search is specifically performed by using the Harris Hawk optimization algorithm. When the fitness value of the individual position is less than 0.5, the Harris Hawk soft encirclement strategy is executed to update the individual position, otherwise the Harris Hawk hard encirclement strategy is executed to update the individual position; The calculation formula for executing the Harris Hawk soft encirclement strategy to update the individual position is: ； In the formula, dis ibest is the distance between the position of the ith individual and the position of the prey, P is the probability of prey escape, and J is the prey jump intensity; The calculation formula for executing the Harris Hawk hard encirclement strategy to update the individual position is: ； In the formula, r is a random number in the range of (0,1), argmax is the maximum value function, which is used to select the candidate position with the largest fitness value from the candidate position set. is a set of candidate positions, A(t) is the first candidate position at the tth iteration, and B(t) is the second candidate position at the tth iteration; The calculation formula of the first candidate position is: ； Where, X rand (t) is the random individual position at the tth iteration; The calculation formula of the second candidate position is: ； Where LF(·) is the Levy flight distribution function.

4. The neural regulation effect evaluation system based on artificial intelligence according to claim 3 is characterized in that: In the neuromodulation effect evaluation module, based on the best feature subset and the regulation effect label, a double-attention compact convolution bidirectional long short-term memory model is used to evaluate the neuromodulation effect and obtain the neuromodulation effect level. The dual-attention compact convolution bidirectional long short-term memory model includes a compact convolution subnet, a dual-attention block, a bidirectional long short-term memory subnet and an evaluation layer; The compact convolutional subnetwork is used to efficiently extract local feature patterns; The dual attention block is used to enhance the focus on key feature information and improve the model evaluation effect; The bidirectional long short-term memory subnetwork is used to model temporal dependencies and capture the changing trends of patients affected by neural regulation; The evaluation layer is used to generate and output model evaluation results; The neural regulation effect evaluation module is provided with a compact convolution subnet construction unit, a dual attention block construction unit, a bidirectional long short-term memory subnet construction unit, an evaluation layer construction unit, a model training unit and an evaluation result generation unit, including the following contents: A compact convolutional subnet construction unit is provided in which a one-dimensional convolutional layer, a depth convolutional layer and a separable convolutional layer are provided, and a regularization block is provided after each convolutional layer, wherein the regularization block includes a batch normalization layer, an average pooling layer and a random dropout layer, and the best feature subset of the patient is processed through the compact convolutional subnet to obtain the convolutional feature of the patient; The calculation formula of the one-dimensional convolutional layer is: ； Where Y tem is the one-dimensional convolution layer output feature, Conv1D(·) is a one-dimensional convolution operation, fea tem is the one-dimensional convolutional layer input, specifically the best feature subset, W tem is a one-dimensional convolution kernel, b tem is the one-dimensional convolution bias term; The calculation formula of the depth convolution layer is: ； Where Y dep is the deep convolution output feature, DepConv2D(·) is the deep convolution operation, fea dep is the input of the deep convolutional layer, specifically the one-dimensional convolutional layer output feature after regularization block processing, W dep is the depth convolution kernel, b dep is the depth convolution bias term; The calculation formula of the separable convolutional layer is: ； Where Y sep is the separable convolution output feature, SepConv(·) is the separable convolution operation, fea sep is the input of the separable convolution layer, specifically refers to the deep convolution output feature after being processed by the regularization block, W sep is a separable convolution kernel, b sep is the separable convolutional layer bias; A dual attention block construction unit, wherein a spatial attention block and a channel attention block are set in the dual attention block, wherein the channel attention block includes two fully connected layers, and the patient convolution features are processed by the dual attention block to obtain the patient enhanced features; The calculation formula of the spatial attention block is: ； In the formula, att spa is the spatial attention, Sig(·) is the S-type activation function, [·,·] is the feature concatenation operation, and F com is the patient convolution feature, GAP(·) is the global average pooling operation, MP(·) is the maximum pooling operation, and F spac is the patient's spatial attention characteristic, is the element-wise multiplication symbol; The calculation formula of the channel attention block is: ； In the formula, att cha is the channel attention, W2 is the weight of the second fully connected layer, tanh(·) is the tanh activation function, W1 is the weight of the first fully connected layer, F chac It is a patient-enhanced feature; A bidirectional long short-term memory subnet construction unit is constructed by stacking M bidirectional long short-term memory units to construct a bidirectional long short-term memory subnet, wherein the bidirectional long short-term memory unit includes a forward long short-term memory unit and a backward long short-term memory unit, and the patient enhancement feature is modeled for temporal dependency through the bidirectional long short-term memory subnet to obtain the patient temporal feature; An evaluation layer construction unit is provided in which a fully connected layer and a Softmax activation function are set, and the patient's time series features are processed through the evaluation layer to obtain a model evaluation result; A model training unit, which constructs a double-attention compact convolution bidirectional long short-term memory model through the compact convolution subnet construction unit, the double-attention block construction unit, the bidirectional long short-term memory subnet construction unit and the evaluation layer construction unit, and performs model training to obtain a neural regulation effect evaluation model; The evaluation result generating unit evaluates the neuromodulation effect through the neuromodulation effect evaluation model to obtain the neuromodulation effect level.

5. The neural regulation effect evaluation system based on artificial intelligence according to claim 4 is characterized in that: In the feature integration module, there are frequency domain feature extraction unit, time-frequency domain feature extraction unit, EEG signal feature extraction unit and feature integration unit, including the following contents: A frequency domain feature extraction unit extracts frequency domain features from the EEG signal by performing Fourier transform on the EEG signal in the patient's optimized data; A time-frequency domain feature extraction unit extracts time-frequency domain features from the EEG signal by performing wavelet transform on the EEG signal in the patient optimization data; An EEG signal feature extraction unit, which extracts features from the EEG signal in the patient's optimized data through the frequency domain feature extraction unit and the time-frequency domain feature extraction unit to obtain EEG signal features; The feature integration unit integrates the EEG signal features with the patient clinical data in the patient optimization data to obtain a patient feature set.

6. The neural regulation effect evaluation system based on artificial intelligence according to claim 5, characterized in that: In the patient data optimization module, the data preprocessing is specifically to preprocess the original patient data through data denoising, data cleaning and data normalization to obtain patient optimized data; The data cleaning includes outlier deletion, duplicate value filtering and missing value filling operations; The data normalization specifically includes normalizing the numerical data in the original patient data and performing a one-hot encoding operation on the type data in the original patient data.

7. The neural regulation effect evaluation system based on artificial intelligence according to claim 6 is characterized in that: In the control data acquisition module, the data acquisition specifically involves collecting the EEG signals of the patient before and after the neuromodulation, and collecting the patient's clinical data to obtain the patient's original data. The doctor obtains the control effect label through manual annotation based on the psychological scale filled out by the patient; The patient original data includes the patient's electroencephalogram signal and the patient's clinical data; The regulatory effect labels include significant effect, minor effect and no effect.

Images