Application method of adaptive neural network in individualized neural regulation

Abstract

The invention relates to an application method of an adaptive neural network in individualized neural regulation. Based on multi-source data including electroencephalogram EEG, functional magnetic resonance imaging fMRI, genetic information and psychological evaluation, a nervous system characterization model of an individual is constructed; key features including a neural connection mode and synchronism in an individual nervous system are mined through a machine learning technology including a graph neural network, and a neural activity basic graph for each individual is constructed; according to the constructed individualized neural representation model, a self-adaptive regulation and control path planning method is introduced; calculating a stimulation path for generating a regulation effect on a nervous system by using the variation trend of individual neural activity data in space and time; according to path planning, neural signals are stimulated, time windows, intensity and frequency of stimulation are calculated based on connectivity interaction of functional areas of the brain, and dynamic adjustment of individual neural activities is achieved. And a self-adaptive activation mechanism based on neural feedback is introduced.

Description

Technical Field

[0001] The present invention relates to an application method in individualized neural regulation, and in particular to an application method of an adaptive neural network in individualized neural regulation. Background Art

[0002] At present, although the application methods of adaptive neural networks in individualized neural regulation have made significant progress, they still face many challenges and shortcomings in the actual application process, especially in comparison with the existing technology, such as the individualized neural regulation parameter optimization method based on the tDCS stimulation response model mentioned in the Chinese invention patent 2023110055781. These problems are more obvious. First, in the existing adaptive neural network methods, although the joint modeling of neural activity signals and psychological state data can reflect the individual differences of the nervous system, it is often limited by the heterogeneity and diversity of data in the fusion of multi-source data and the extraction of deep coupling features. Especially when processing complex neural signals such as EEG and fMRI, the network model needs to efficiently encode and decode the spatiotemporal characteristics of these signals, which places extremely high demands on the algorithm. However, the current neural network architecture has not yet been able to fully and effectively integrate these signals and extract accurate features with physiological and psychological significance. Secondly, adaptive neural networks also face limitations in the planning and optimization of neural regulation paths. Traditional path planning methods often rely on the functional connectivity between brain regions to optimize the stimulus distribution, which requires accurate modeling of the functional state of each region.

[0003] However, the differences in the structure of the nervous system between individuals are large, making it impossible for path planning based on the general brain connection model to perfectly adapt to all individuals. This leads to insufficient personalization of the neuromodulation scheme, especially when it comes to the treatment of different patients, and it is unable to effectively cope with the dynamic changes of the nervous system and the differences in neural function. In addition, the existing adaptive adjustment mechanism, especially when dynamically adjusting the neural stimulation path, has the problems of large consumption of computing resources and poor real-time performance. Although adjustment through feedback of real-time data can theoretically improve the treatment effect, in practical applications, the spatiotemporal changes of neural activity are very complex, and the efficiency and accuracy of existing computational models in real-time processing and adjustment of neural stimulation schemes are still low, and it is impossible to accurately capture the slight changes in brain activity in a short time, resulting in a lag in the adjustment of the neural stimulation scheme, which in turn affects the treatment effect. In contrast, the personalized neuromodulation method based on the tDCS stimulation response model can simulate different neuromodulation schemes through finite element modeling, and predict the response of whole-brain neural activity through a large-scale dynamic model, thereby effectively overcoming the ethical constraints and heterogeneity of neuromodulation schemes. This method optimizes the current dosage parameters based on the individual differences of each patient by constructing an individualized tDCS stimulation response model, thereby reducing the uncertainty in treatment and improving the consistency of treatment effects. In contrast, when faced with complex physiological and psychological multi-source data, the existing adaptive neural network methods rely too much on the model's training data in the optimization process, and cannot completely eliminate the heterogeneity and individual differences between samples, and cannot perform large-scale simulations, which limits the applicability and accuracy of the treatment plan. In addition, personalized neuroregulation also faces ethical and safety issues in actual operation, especially in the clinical verification stage.

[0004] Existing neural network models often cannot be widely used in clinical applications without sufficient verification, which not only limits their promotion, but also may bring safety risks to patients. The individualized neuroregulation method based on tDCS can reduce the impact of ethical constraints by simulating different neuroregulation schemes and verifying them using finite element modeling, and conduct safety verification in a virtual environment, providing strong support for subsequent clinical trials. Furthermore, the existing adaptive neural network methods often rely on individual real-time neural activity data for the regulation of the nervous system, but in many cases, the patient's psychological state and emotional fluctuations may have a profound impact on neural activity. Although the present invention attempts to perform deep coupling modeling by introducing psychological data such as emotional fluctuations and stress levels, due to the complexity and variability of emotions and psychological states, how to accurately extract features that match neural activity from these non-physiological signals is still a difficult problem. Summary of the invention

[0005] The purpose of the present invention is to provide an application method of an adaptive neural network in individualized neural regulation, thereby solving some of the drawbacks and shortcomings pointed out in the background technology.

[0006] The present invention solves the above-mentioned technical problems by adopting the following technical solution, which includes the following steps:

[0007] S1. Construction of individualized neural system representation model:

[0008] S1.1. Construct a model of the individual's neural system representation based on multi-source data including EEG, fMRI, genetic information, and psychological assessments;

[0009] S1.1. Use machine learning techniques including graph neural networks to mine key features of individual nervous systems, including neural connection patterns and synchronization, and construct a basic neural activity map for each individual.

[0010] S2. Adaptive control path planning based on neural representation model:

[0011] S2.1. Based on the constructed individualized neural representation model, an adaptive control path planning method is introduced; the stimulation path that produces a regulatory effect on the nervous system is calculated by using the spatial and temporal variation trends of individual neural activity data;

[0012] S2.2, Path planning is the stimulation of neural signals. Based on the connectivity interaction of various functional areas of the brain, the time window, intensity and frequency of stimulation are calculated to achieve dynamic regulation of individual neural activity;

[0013] S3. Adaptive activation mechanism and dynamic regulation of the nervous system:

[0014] S3.1. Introducing an adaptive activation mechanism based on neural feedback; the adaptive activation mechanism monitors individual neural activity in real time and infers potential activation nodes and connections in the neural network through dynamic changes in neural activity;

[0015] S3.2. By adjusting the stimulation parameters in real time, the activation mode of the nervous system is optimized, so that the neural network can adjust itself and adapt to the changes in the external environment at different times;

[0016] S4. Combination of personalized neural regulation and plasticity mechanisms:

[0017] S4.1. Use adaptive neural network models to evaluate the plasticity of individual neural networks in real time, and design regulatory strategies to promote the development of neural plasticity by combining the activation and feedback mechanisms of the nervous system;

[0018] S5. Deep coupling between neural regulation and individual psychological state:

[0019] S5.1. Develop dynamic control methods based on the coupling of neural activity and psychological state; jointly model neural signals with psychological states including mood swings and stress levels through deep learning models including transform autoencoders, and adjust neural control schemes according to real-time changes in individual psychological states.

[0020] Furthermore, the adaptive control path planning method based on the neural representation model includes:

[0021] By combining multi-source physiological data including EEG, fMRI, and genomics with psychological assessment data, a neural activity characteristic model for each individual is constructed; the neural activity characteristic model reflects the structure of the nervous system and describes the dynamic changes of neural functions, including the connectivity and synchronization functional characteristics of the neural network; the individualized neural representation model is expressed as:

[0022]

[0023] in:

[0024] N model (t) represents the output of the individualized neural representation model at time t, which represents the comprehensive performance of individual neural activity; f i (t) is the i-th input feature, which refers to the time series data extracted from different physiological data sources including EEG signals and fMRI image data, reflecting the characteristics of neural activity; w i is the weight coefficient of the i-th feature, indicating the importance of the feature to the output of the neural representation model; b is the bias term, which adjusts the basic output of the model; n is the total number of input features, indicating the dimension of the model input features.

[0025] Furthermore, the adaptive control path planning method based on the neural representation model includes:

[0026] Stimulation paths are planned based on the functional connectivity of different areas of the nervous system, including the cerebral cortex and basal ganglia. The degree of functional connectivity of each area to other areas determines its contribution to the overall neural activity, and stimulation paths are allocated based on connectivity. The formula for spatial optimization is:

[0027]

[0028] in:

[0029] P spatial represents the optimized stimulation path, which reflects the stimulation distribution in space; is the spatial extent of the brain region, indicating the set of regions of the nervous system that need to be stimulated; S j(x, t) is the intensity of neural activity at spatial position x and time t, reflecting the state of neural activity in a specific area; a j is the weight coefficient of the jth region, indicating the contribution weight of the region to the overall neural activity; m is the number of neural regions that need to be regulated, indicating the number of neural regions considered; dx is the spatial integration variable, indicating the integration process of the brain spatial region.

[0030] In terms of time optimization, the stimulation time window is designed based on the time series of individual neural activity. The stimulation timing is selected within the dynamic time window by weighted summing and maximizing the time series characteristics of neural activity. The time optimization formula is:

[0031]

[0032] in:

[0033] P temporal (t) is the optimal stimulation path at time t, indicating the stimulation scheme optimized for the current time period; E i (t) is the time series feature of the i-th neural activity, indicating the fluctuation of neural activity at time t; r i is the weight coefficient of the i-th neural activity feature, indicating the degree of influence of this feature on the time optimization path; n is the number of neural activities involved in the regulation, indicating the dimension of the time series data considered.

[0034] Furthermore, the adaptive control path planning method based on the neural representation model includes:

[0035] Adopt real-time monitoring and dynamic adjustment mechanism; design an adaptive adjustment mechanism based on real-time data to dynamically adjust the stimulation path according to changes in neural activity; the real-time adjustment formula is:

[0036]

[0037] Where: ΔP(t) represents the amount of stimulation path adjustment at time t, indicating the stimulation direction and amplitude that need to be adjusted; γ is the adjustment coefficient, which controls the sensitivity of the adjustment; N model (t) is the neural activity value predicted by the individualized neural representation model, which represents the expected neural system activity state; N current (t) is the current neural activity value obtained by real-time monitoring equipment, indicating the actual activity state of the nervous system; is the gradient operator, which represents the difference N model (t)N current The gradient of (t) is calculated.

[0038] Furthermore, the method for deep coupling of neural regulation and individual psychological state includes:

[0039] First, the neural activity signal and psychological state data are preprocessed by the encoder part of the deep learning model; the encoder part fuses the multimodal data into a low-dimensional representation through nonlinear mapping to form the coupling characteristics of neural activity and psychological state; then the coupling characteristics are restored by the decoder part to generate a control output signal corresponding to the neural activity, and adjust the neural stimulation scheme according to the current psychological state; the coupling mapping expression formula of the transformed autoencoder is:

[0040]

[0041] Where: Z N To represent the low-dimensional coupling representation of neural activity and psychological state; by transforming the encoder part of the autoencoder model, the neural activity signal and psychological state data are mapped to a shared low-dimensional feature space to form a coupling representation of neural and psychological state; X N Input includes EEG, fMRI neural activity signals; X P is the input psychological state data; the state data is obtained through physiological or psychological assessment; Θ E are the weights and biases of the encoder network, which are the parameters adjusted during the network learning process; is a nonlinear mapping function used to transform neural activity and psychological state signals into low-dimensional coupled representations.

[0042] Furthermore, the method for deep coupling of neural regulation and individual psychological state includes:

[0043] By introducing the transform autoencoder, neural activity and psychological state are jointly modeled, and the interaction between the two is adaptively learned and identified under the deep learning framework; the structure of the transform autoencoder includes an encoder, a decoder, and a joint layer; the encoder maps the input neural activity signal and psychological state data to a low-dimensional representation through a nonlinear activation function; the features of neural activity and psychological state in the latent space are learned and fused; and the joint modeling and decoding mapping expression is:

[0044] Z N =W N ·X N +W P ·X P +b

[0045] Among them, Z N is a low-dimensional coupled representation of neural activity and psychological state. After calculation at the encoder layer, the low-dimensional representation contains the comprehensive features of neural activity and psychological state, which is passed as input to the subsequent decoder part. N is the weight matrix of the neural activity signal, which controls the influence of the neural activity signal; W Pis the weight matrix of psychological state data, which controls the contribution of psychological state data in the model; X N is the input neural activity signal; X P is the input psychological state data; b is the bias term used to translate the output.

[0046] Furthermore, the method for deep coupling of neural regulation and individual psychological state includes:

[0047] Real-time mental state monitoring is introduced, and the neural regulation scheme is dynamically adjusted by combining the coupling characteristics of neural activity and mental state extracted by the transform autoencoder. Real-time mental state monitoring includes real-time collection of emotional fluctuations and stress level data. The data is input into the deep learning model, and the neural stimulation adjustment scheme is generated by combining the changing trend of the neural activity signal:

[0048]

[0049] Among them, ΔS(t) is the adjustment amount of the neural stimulation scheme at time t; represents the real-time adjustment of the neural regulation scheme according to the current neural activity and psychological state; S0 is the initial neural stimulation scheme, which is set at the beginning of treatment, including the frequency and intensity of stimulation; P(t) is the psychological state data at time t, including emotional fluctuations and stress levels; Θ A To regulate the parameters of the model, the deep learning model is trained for optimization and the adjustment method of the stimulation scheme is controlled; It is a nonlinear control function used to adjust the neural stimulation scheme according to the coupling characteristics of neural activity and psychological state.

[0050] The application method of the adaptive neural network of the present invention in individualized neural regulation has the following beneficial effects:

[0051] By combining a variety of physiological data (such as EEG, fMRI, genomics data) with psychological state assessment data, a unique neural activity characteristic model for each individual is constructed, which can accurately characterize the structure and function of the nervous system of each patient. Based on these individualized models, the neural regulation pathway can be dynamically adjusted according to specific neural activity patterns, psychological states, and environmental changes, thereby achieving precise treatment.

[0052] An adaptive control mechanism based on real-time neural activity and psychological state data has been introduced. By real-time monitoring of the changing trends of neural activity and changes in psychological states such as mood swings and stress levels, the path, frequency, intensity and other parameters of neural stimulation can be adjusted in a timely manner to better respond to the dynamic changes in the individual's nervous system and psychological state. This real-time response capability improves the flexibility and adaptability of treatment and significantly optimizes the treatment effect.

[0053] By optimizing the activation mechanism and dynamic regulation of the nervous system through adaptive neural networks, the plasticity of the nervous system can be effectively promoted, allowing the brain to self-adjust and adapt to changes in the external environment at different times. This has important clinical significance for rehabilitation treatment, cognitive function improvement, and nerve damage repair, and can help patients gradually recover and enhance their nerve function. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 The present invention is a flowchart of the method for applying the adaptive neural network in individualized neural regulation.

[0055] Figure 2 This is a flow chart of the adaptive control path planning method based on the neural representation model of the present invention.

[0056] Figure 3 This is a flow chart of the deep coupling method of neural regulation and individual psychological state of the present invention. DETAILED DESCRIPTION

[0057] The specific implementation modes of the present invention will be described in detail below in conjunction with the accompanying drawings.

[0058] Combination Figure 1 The flowchart of the adaptive neural network in personalized neural regulation first starts with the construction of a personalized nervous system representation model. The core of this model is to fuse multi-source data to accurately depict the unique nervous system characteristics of each individual. The construction of the personalized nervous system representation model depends on different types of physiological and psychological data, especially electroencephalogram (EEG), functional magnetic resonance imaging (fMRI), genetic information, and psychological assessment data. EEG can capture the brain's electrical activity signals and reflect the characteristics of neural activity in the cerebral cortex; fMRI can provide spatial distribution information of brain activity by measuring changes in brain blood oxygen levels, helping to depict the functional connections and collaborative working modes of different brain regions; genetic information provides the potential genetic basis of individual neural activity, which is important for understanding the differences in individual neural function; and psychological assessment helps collect individual psychological state data in terms of emotions, cognition, behavior, etc. These data comprehensively reflect the individual's neural characteristics and psychological state from multiple dimensions, thereby providing accurate input information for subsequent personalized neural regulation.

[0059] In the process of constructing the neural system representation model, the core technology is to use Graph Neural Networks (GNN) to extract the features of the individual neural system. Graph Neural Networks are a special type of deep learning model that can process graph structure data and are widely used in social networks, recommendation systems and other fields. In this technology, the individual's neural activity data is converted into a graph structure, in which the nodes of the graph represent different areas of the brain (such as the cerebral cortex, basal ganglia, thalamus, etc.), and the edges represent the functional connection relationship between these areas. Through the multi-layer propagation mechanism of the graph neural network, the connection pattern and interaction relationship between these neural regions can be effectively mined, thereby constructing the neural connection network of the individual brain. The node characteristics of each neural region not only include the intensity of the neural activity in the region, but also include the coordinated activity pattern of the region with other regions. During the training process, the graph neural network can extract key information such as the synchronization, functional coupling and dynamic change pattern of neural activity by learning the interaction between different brain regions, and then construct a basic map of individual neural activity.

[0060] This map is not fixed, but dynamically adjusted because neural activity has significant individual differences. By integrating multi-source data such as EEG and fMRI, the graph neural network can fully consider the differences in individual neural activity in space and time, and generate a high-precision, personalized neural system representation model. This model will provide an accurate neural basis map for the optimization of neural regulation schemes, avoiding the drawbacks of the "one-size-fits-all" model used in traditional methods.

[0061] According to the individualized neural representation model constructed in the early stage, the neural stimulation path is designed and adjusted to achieve precise dynamic regulation of individual neural activity. First, the neural representation model provides each individual with a personalized brain neural activity map, including important information such as neural connection patterns, synchronization and interaction of different brain regions. Based on this model, the spatial and temporal trends of individual neural activity can be further analyzed, that is, how the activities of different brain regions fluctuate with the individual's psychological state or environmental changes at a specific time point and spatial position. In this way, neural regulation path planning can flexibly adapt to individual brain activity changes and make precise adjustments. Specifically, the core of path planning lies in the design of the stimulation path, which not only involves the connectivity between different functional areas of the brain, but also takes into account multiple factors such as the timeliness, intensity and frequency of the stimulation. First, based on the neural activity data provided by the individualized neural representation model, the connectivity between the functional areas of the brain is analyzed to identify the synergistic patterns of different regions.

[0062] The connectivity between different brain regions is different. Some regions are the main signal processing centers, while other regions play a supporting role. Based on these connection patterns, the stimulation paths of the nervous system in different spatial regions can be calculated. These paths determine the transmission route and strength of the stimulation signal, ensuring that the stimulation signal can effectively act on the target area and exert the maximum regulatory effect through the synergy between brain regions. Secondly, the stimulation path is not only a spatial planning, but also includes the planning of the time window, intensity and frequency of the stimulation. The time window determines the timing and duration of the application of the stimulation signal, while the intensity and frequency directly affect the strength and continuity of the stimulation effect. For example, some neural regulation methods require stimulation within a specific activity cycle of the brain to maximize its effect, which requires the regulation path to be able to flexibly respond to the temporal fluctuations of neural activity. Therefore, in the adaptive regulation path planning, the system adjusts the stimulation time window according to the real-time changes in neural activity to ensure that the stimulation signal intervenes in the brain area at the most appropriate time. In addition, the adjustment of intensity and frequency changes dynamically according to the intensity of brain activity and the needs of the regulation target.

[0063] By monitoring the individual's neural activity in real time and accurately intervening in the nervous system according to its dynamic changes. This method first introduces an adaptive activation mechanism based on neural feedback, aiming to dynamically infer the potential activation nodes and connections in the neural network by continuously monitoring the changes in the individual's neural activity. The dynamic changes in neural activity include data such as EEG signals and functional magnetic resonance imaging (fMRI) signals, which can reflect the activity state of the brain at a specific time point and situation in real time. Through the real-time feedback of these data, the system can identify the activity changes and potential activation nodes in key areas of the neural network. "Activation nodes" in neural networks refer to areas that show significant neural activity at a certain moment. These nodes are often activated under specific cognitive tasks or emotional states, and their activities play a vital role in the overall neural function of individuals.

[0064] These nodes and their connections can respond quickly to external environments or internal stimuli, thereby changing the overall state of the neural network. By adjusting stimulation parameters such as frequency, intensity, duration, etc. in real time, the system can accurately optimize the activation mode of the nervous system to ensure that the neural network can make efficient and adaptive adjustments when facing different needs and environmental changes. For example, when an individual's emotional state changes, the system can alleviate or enhance specific neural functions by adjusting the degree of activation of specific brain areas, optimizing the individual's emotional regulation or cognitive function. At the same time, the system can also adjust the spatial distribution of the brain and optimize the coordination between different areas of the brain, making the activation mode of the entire nervous system more reasonable and efficient. The optimization and dynamic adjustment of the activation mode is an adaptive process, which means that the nervous system can flexibly adjust to external stimuli and internal needs in different periods and situations.

[0065] By evaluating the plasticity of individual neural networks in real time and combining the activation and feedback mechanisms of the nervous system, we can design regulatory strategies that promote the development of neural plasticity. In this process, we first need to understand that neural plasticity refers to the ability of the brain to change its structure and function after experiencing external stimulation or changes in internal demand, including synaptic plasticity, reorganization of neural circuits, and changes in connections between neurons. Traditional neuroregulatory methods often focus on symptom relief or short-term effects, and lack consideration of the long-term plasticity of the nervous system and its self-adjustment mechanisms.

[0066] The introduction of the adaptive neural network model has made neuroregulation no longer limited to short-term intervention, but can take into account the long-term adaptability and plasticity development of the nervous system. Specifically, the model monitors the activity state of the nervous system in real time, identifies the areas in the neural circuit that need to be adjusted, and designs appropriate stimulation strategies based on the neural activity patterns in these areas. Through dynamic feedback to the neural network, the system can identify the potential plasticity of neural activity and adjust the stimulation parameters at the appropriate time to promote long-term plastic changes in the nervous system. This process is not only an intervention in the current neural activity, but also through the feedback mechanism of the system, it guides the neural network to produce structural and functional adaptive changes after a certain period of stimulation. For example, by adjusting the activation pattern of specific brain areas, stimulating the strengthening or reorganization of synapses, the nervous system forms a new information processing pathway. These adjustments not only improve the short-term neuroregulatory effect, but also promote the long-term healthy regulation of the nervous system.

[0067] Through a deep learning model including a transform autoencoder, neural signals and an individual's psychological state (such as mood swings, stress levels, etc.) are jointly modeled to capture the complex interrelationships between neural activity and psychological state. In this process, neural signals (such as electroencephalogram (EEG), functional magnetic resonance imaging (fMRI), etc.) and psychological state data (such as mood swings, stress levels, emotional labels, etc.) are input into the deep learning model, and through model training, the intrinsic correlation between neural activity and psychological state is automatically extracted. The role of the transform autoencoder is to convert the original complex data (neural signals and psychological state data) into a highly compressed and representative feature space by learning a low-dimensional latent space, thereby achieving deep coupling of neural activity and psychological state.

[0068] This coupling model can accurately capture how neural activity is affected by changes in an individual's mental state, and in turn reveal how mental states are manifested through changes in neural activity. For example, when an individual experiences emotional fluctuations, changes in neural activity are reflected in specific areas of the brain, such as increased or decreased activity in the emotion regulation area, which in turn affects the functional state of the entire nervous system. Through this deep coupling modeling method, not only can we fully understand the interaction between neural activity and mental state, but we can also provide more personalized input for neural regulation. Based on this coupling model, the system can monitor changes in an individual's mental state in real time and adjust the neural regulation scheme based on real-time data of the mental state. Specifically, when an individual's mood fluctuates or stress level changes, the system automatically adjusts the stimulation parameters according to the characteristics of neuro-psychological coupling. This adjustment not only includes the time window, intensity, frequency, etc. of the stimulation, but also involves changes in the target area and stimulation mode of the stimulation.

[0069] Embodiment 1:

[0070] Combination Figure 2 A patient named Xiao Zhang is experiencing a stressful work environment and has mild anxiety and mood swings. In order to improve his nervous system function and alleviate his mood swings, an individualized neural regulation scheme based on adaptive neural networks is designed. The core of this scheme is to accurately describe his neural activity characteristics by constructing an individualized neural representation model, and implement adaptive regulation path planning based on this.

[0071] First, Xiao Zhang's nervous system was comprehensively evaluated through multi-source physiological data. Specifically, his electroencephalogram (EEG), functional magnetic resonance imaging (fMRI), and genomics data were collected. In addition, Xiao Zhang was also given a psychological assessment to quantify his stress level and mood swings. All of these data together helped build an individualized neural representation model. The goal of this model is to combine the structural characteristics of the nervous system as well as its dynamic changes, thereby reflecting the functional characteristics of the nervous system, such as connectivity and synchronization.

[0072] Next, use the following formula to build Xiao Zhang's personalized neural representation model:

[0073]

[0074] Where: N model (t) represents the output of the individualized neural representation model at time t, reflecting the comprehensive performance of Xiao Zhang’s neural activity. i (t) is the i-th input feature, which represents the time series data extracted from physiological data sources such as EEG and fMRI, reflecting the characteristics of neural activity. For example, specific frequency bands in EEG signals (such as α waves, β waves, etc.) can indicate Xiao Zhang’s emotional and cognitive states, while fMRI image data reveals the activation of specific brain regions.

[0075] w i is the weight coefficient of the i-th feature, indicating the importance of this feature to the output of the neural representation model. Assume that after data training, the weight coefficients of these features are obtained. For example, w1 of EEG is 0.4, w2 of fMRI is 0.3, w3 of genomic data is 0.2, and w4 of psychological assessment data is 0.1. b is the bias term, which is used to adjust the basic output of the model to adapt to the baseline neural activity characteristics of Xiao Zhang. n is the total number of input features, which is determined by the dimension of the data. In Xiao Zhang's case, assuming that 4 main input features are used (EEG signals, fMRI data, genomic data, and psychological assessment data), then n=4.

[0076] Example calculation of model output:

[0077] Assume that at time t, Xiao Zhang's EEG signal shows some fluctuations, fMRI images show activation of specific areas (such as the prefrontal lobe), genes related to emotion regulation in genomic data are relatively active, and psychological assessment shows that he is currently at a moderate level of stress. Based on the above features, the calculation process of the model output is as follows:

[0078] N model (t)=w1·f1(t)+w2·f2(t)+w3·f3(t)+w4·f4(t)+b

[0079] Assume that the following values ​​are obtained:

[0080] f1(t) = 0.8 (eigenvalue of EEG signal, reflecting the fluctuation of neural activity)

[0081] f2(t)=1.2 (fMRI data, showing the activation intensity of the prefrontal cortex)

[0082] f3(t) = 0.5 (characteristics of genomics data, activity of related genes)

[0083] f4(t)=0.7 (psychological assessment results, stress level score)

[0084] Substituting the above values ​​and weight coefficients, assuming b = 0.1:

[0085] N model (t)=0.4·0.8+0.3·1.2+0.2·0.5+0.1·0.7+0.1

[0086] N model (t)=0.32+0.36+0.1+0.07+0.1=0.95

[0087] Therefore, based on this output, the model concluded that Xiao Zhang's current neural activity characteristic is 0.95. This value reflects Xiao Zhang's neural activity performance in the current state and provides a basis for subsequent regulation path planning.

[0088] Next, based on this neural activity feature, an adaptive control path planning method is used to design a personalized neural control plan. Based on the spatiotemporal change trend of neural activity and the connectivity interaction between brain functional areas, a stimulation path suitable for Xiao Zhang is calculated. Assume that at this time, the model predicts that Xiao Zhang's mood fluctuates greatly and the activation of the prefrontal lobe is high, and this area needs to be regulated.

[0089] The specific control steps are as follows:

[0090] Time window: According to the timing characteristics of Xiao Zhang's neural activity, it is assumed that neural regulation is performed during the time period when the emotional fluctuations are large (for example, within 10 minutes before and after the peak of the emotional fluctuations), and the time window is 10 minutes.

[0091] Stimulation frequency: By analyzing the spectrum of Xiao Zhang's neural activity, a suitable stimulation frequency, such as 10Hz, is selected to intervene in the activity of the prefrontal area and help relieve anxiety.

[0092] Stimulation intensity: The stimulation intensity is set according to the amplitude of neural activity. For example, by calculating the amplitude of the EEG signal, a moderate stimulation intensity value (assuming 2 mA) is selected to ensure that the stimulation does not have a negative impact on the nervous system.

[0093] In Xiao Zhang's case, because he showed certain mood swings and anxiety reactions, especially strong activity in specific brain areas (such as the prefrontal cortex), it was necessary to plan the stimulation path based on the functional connectivity of different areas such as the cerebral cortex and basal ganglia, and select the appropriate stimulation time through time optimization. These steps will make the neuromodulation program more efficient and able to dynamically adapt to Xiao Zhang's neural activity state.

[0094] Spatial Optimization Path Planning:

[0095] First, consider the spatial distribution of neural activity in different brain regions, especially the connectivity between the cerebral cortex and the basal ganglia. Assume that the neural activity characteristics of these brain regions have been obtained through previous EEG and fMRI data. In spatial optimization, the following formula is used to plan the stimulation path:

[0096]

[0097] Where: P spatial Represents the optimized stimulation path, reflecting the spatial stimulation distribution of the nervous system. is the spatial range of the brain region, indicating the set of regions of the nervous system that need to be stimulated. In Xiao Zhang's case, assuming that the prefrontal lobe and basal ganglia regions of the cerebral cortex need to be stimulated, then It is the combination of these two regions. j (x, t) is the intensity of neural activity at spatial position x and time t, reflecting the state of neural activity in a specific area. These intensity values ​​are obtained through EEG and fMRI data. For example, the neural activity intensity of Xiao Zhang's prefrontal region at a certain moment is 0.8, while the neural activity intensity of the basal ganglia region is 0.6.

[0098] a j is the weight coefficient of the jth region, indicating the contribution weight of the region in the overall neural activity. Assume that according to the spatial connectivity analysis, the weight of the prefrontal region is a1 = 0.7, and the weight of the basal ganglia region is a2 = 0.3. m is the number of neural regions that need to be regulated. Assuming that only the prefrontal lobe and basal ganglia are considered, m = 2. dx is the spatial integration variable, which represents the integration process of the brain spatial region.

[0099] To further specify, the above formula can be applied to Xiao Zhang's case. Assume that the neural activity intensity in the prefrontal region is 0.8, the neural activity intensity in the basal ganglia region is 0.6, and the known weight coefficients are 0.7 and 0.3 respectively. When calculating the optimized stimulation path through the above formula, the contribution and neural activity intensity of each region are taken into account.

[0100]

[0101] This computational result suggests that the optimized stimulation pathway will be balanced between prefrontal and basal ganglia regions, reflecting the higher contribution of prefrontal regions to overall neural activity.

[0102] Time-optimized path planning:

[0103] In addition to spatial optimization, the temporal characteristics of neural activity, especially the timing of stimulation, also need to be considered. In Xiao Zhang's case, his mood swings change over time, so the time window of stimulation needs to be optimized based on the temporal characteristics of neural activity. Use the following formula for time optimization:

[0104]

[0105] in:

[0106] P temporal (t) is the optimal stimulation path at time t, which represents the stimulation scheme optimized for the current time period. i (t) is the time series feature of the i-th neural activity, indicating the fluctuation of neural activity at time t. Assume that through the joint analysis of EEG and fMRI, the time series feature of the prefrontal region E1(t) = 0.7 and the time series feature of the basal ganglia region E2(t) = 0.5 are obtained, reflecting the fluctuation intensity of Xiao Zhang's current emotional fluctuation. i is the weight coefficient of the i-th neural activity feature, indicating the degree of influence of this feature on the time optimization path. Assume that according to the connectivity of the prefrontal lobe and basal ganglia, the weight coefficient of the prefrontal region is r1 = 0.6, and the weight coefficient of the basal ganglia region is r2 = 0.4. n is the number of neural activities involved in the regulation, indicating the dimension of the considered time series data. Assume that only the prefrontal lobe and basal ganglia are considered, so n = 2.

[0107] Apply the above formula to Xiao Zhang's case to calculate the optimal stimulation time path. First, substitute the known time series characteristics and weight coefficients:

[0108]

[0109] By maximizing this weighted summation result, the optimal time window for stimulation is determined, which means that during this time period, stimulation will exert a higher weight on Xiao Zhang's prefrontal area, thereby alleviating his emotional fluctuations.

[0110] Through the above-mentioned spatial and temporal optimization calculations, a personalized neural regulation path was obtained. This path not only allocated space according to Xiao Zhang's neural activity characteristics, but also dynamically adjusted according to his emotional fluctuations and the time series characteristics of neural activity. Finally, based on these calculation results, the neural regulation system can accurately adjust the stimulation timing and stimulation area to achieve the best neural regulation effect.

[0111] In this embodiment, since Xiao Zhang's nervous system will fluctuate during the treatment process, especially when his emotions fluctuate greatly, it is necessary to design an adaptive adjustment mechanism based on real-time data to dynamically adjust the stimulation path according to changes in neural activity. This process relies on the formula:

[0112]

[0113] in:

[0114] ΔP(t) represents the amount of stimulation path adjustment at time t, indicating the stimulation direction and amplitude that need to be adjusted. Through real-time monitoring and adjustment, the stimulation parameters can be fine-tuned to ensure that Xiao Zhang's neural activity remains at an ideal level. γ is the adjustment coefficient, which controls the sensitivity of the adjustment amount. Assume that in Xiao Zhang's personalized regulation plan, γ = 0.8, indicating that it is hoped that a more sensitive adjustment will be made within a certain range to cope with the changes in neural activity caused by Xiao Zhang's emotional fluctuations. N model (t) is the neural activity value predicted by the individualized neural representation model, which represents the expected neural system activity state. Assume that the model predicts that the expected neural activity value of Xiao Zhang at a certain moment is N model (t) = 0.75, representing the intensity of neural activity in the prefrontal area under the expectation state.

[0115] N current (t) is the current neural activity value obtained by the real-time monitoring device, which represents the actual state of nervous system activity. Based on Xiao Zhang’s real-time EEG monitoring data, it is assumed that the current neural activity value is N current (t) = 0.6, indicating that the current neural activity intensity is lower than the expected state. is the gradient operator, which represents the difference N model (t)N current The gradient of the change in (t) is calculated to measure the degree to which the neural activity state deviates from the expected state.

[0116] In Xiao Zhang's case, assume that real-time data is obtained at a certain moment (for example, the 10th minute during the treatment process). According to real-time monitoring, the current neural activity value is N current (10) = 0.6, and according to the prediction of the individualized neural representation model, the expected neural activity value is N model (10)=0.75.

[0117] Calculate the difference:

[0118] ΔN=N model (10)N current (10) = 0.750.6 = 0.15

[0119] Next, calculate the gradient of this difference:

[0120]

[0121] Assume that the rate of change of the difference ΔN is relatively stable within this time window, and assume that the gradient Approximately 1 (i.e. the rate of change per unit time is constant).

[0122] According to the formula:

[0123]

[0124] This means that at the 10th minute, the stimulation path needs to be adjusted by ΔP(10)=0.12, that is, the stimulation intensity or time window is adjusted to make up for the gap between neural activity and expectations.

[0125] Dynamically adjust the stimulation path: According to the above calculation, the adjustment amount ΔP(10) = 0.12 means that the intensity of stimulation needs to be increased or the duration of stimulation needs to be extended to make Xiao Zhang's neural activity closer to the desired state. Assuming that the basic intensity of stimulation is 0.5, after real-time adjustment, the stimulation intensity can be increased to:

[0126] New stimulus intensity = 0.5 + 0.12 = 0.62

[0127] In addition, the stimulation time window will be extended accordingly. Suppose the stimulation time window was originally 10 minutes, and after adjustment, the time window is extended to 11 minutes to better adjust neural activity.

[0128] Throughout the treatment, Xiao Zhang's real-time neural activity will be continuously monitored, and the stimulation path will be dynamically adjusted according to the above formula. For example, in the next few minutes, if Xiao Zhang's neural activity continues to be lower than the expected state, the same adjustment mechanism will continue to be used to update the stimulation intensity and time window in real time. After each adjustment, the stimulation parameters will continue to approach the optimal value, thus ensuring the long-term effect of neural regulation.

[0129] Embodiment 2:

[0130] Combination Figure 3, when Xiao Zhang undergoes neuromodulation treatment, he needs to consider not only his neural activity state, but also his psychological state such as emotions and stress. Traditional neuromodulation methods usually treat neural activity and psychological state as two independent systems, but the method of this embodiment can provide customized regulation for individual changes more accurately by coupling the two.

[0131] First, the neural activity signals and psychological state data are preprocessed by the encoder part of the deep learning model. In Xiao Zhang’s case, the neural activity signals include EEG signals and fMRI image data, while the psychological state data is obtained through physiological or psychological assessments, such as mood swings, stress levels, anxiety, etc.

[0132] Assume that Xiao Zhang's EEG signals show fluctuations in neural activity in the prefrontal region during treatment, and according to psychological assessment, his stress level is 5 / 10 (moderate stress). These signals will be input into the encoder part of the transform autoencoder, nonlinearly mapped, and fused into a low-dimensional representation. This process is represented by the following formula:

[0133]

[0134] in:

[0135] Z N It is a low-dimensional coupled representation of neural activity and psychological state, reflecting the comprehensive characteristics of neural activity and psychological state. Through the encoder part, neural signals and psychological state data are mapped to a shared low-dimensional feature space to form a coupled representation. N is the input neural activity signal, including EEG and fMRI signals. For example, Xiao Zhang’s EEG data shows fluctuations in the prefrontal cortex, where X N An example value of X is [0.65, 0.72, 0.68, 0.71] (reflecting the prefrontal neural activity at different time points). P is the input psychological state data, for example, Xiao Zhang's emotional state is anxiety, stress level is 5, and emotional fluctuation is 3 / 10. Assume X P The value of is [0.5, 0.6, 0.4] (representing emotions, stress, and anxiety). E are the weights and biases of the encoder network, which are optimized during model training to map neural activity and psychological state signals to a low-dimensional coupling space. Assume Θ E The initial values ​​of are randomly set, and after training, the model will automatically optimize these parameters.

[0136] Low-dimensional coupled representation generation:

[0137] By transforming the encoder part of the autoencoder, the neural activity signal XN and psychological state data X P is mapped to a low-dimensional space to form a coupled representation Z N This coupling representation reflects the state of the nervous system at a specific moment, while also incorporating the individual’s psychological state. For example, after being processed by the deep learning model, Xiao Zhang’s coupling representation Z N is a four-dimensional vector:

[0138] Z N =[0.78,0.65,0.80,0.72]

[0139] This vector characterizes the coupling characteristics between neural activity (such as prefrontal fluctuations) and psychological states (such as stress and anxiety).

[0140] Decoding and stimulation control output:

[0141] Next, couple the feature Z N The decoder part of the transform autoencoder is restored to generate a control output signal corresponding to the neural activity. The goal of the decoder part is to represent Z according to the low-dimensional coupling N Generate a corresponding stimulation scheme to effectively regulate the nervous system. Assume that the control signal output by the decoder is:

[0142]

[0143] in:

[0144] S stimulus is the stimulus signal output by the decoder, which indicates the intensity, frequency, and time of the neural stimulation to be applied. Assume that the stimulus signal output by the decoder is S stimulus =[0.6,1.2,0.5], where the first value represents the stimulation intensity, the second value represents the stimulation frequency, and the third value represents the stimulation duration.

[0145] is the nonlinear mapping function of the decoder, which represents Z according to the coupling N and the decoder weights Θ D Map the low-dimensional representation back to the stimulus space. Assume Θ D It is optimized during the training process and can accurately map the low-dimensional representation to the stimulus signal. stimulus Each component of reflects a different regulation of the activity of the Xiaozhang nerve. For example, the stimulation intensity is 0.6, which means that the intensity of the stimulation current is moderate, and the frequency is 1.2Hz, which means that the frequency of the applied neural stimulation signal is 1.2Hz and the duration is 0.5 seconds.

[0146] According to the current psychological state (such as mood swings, stress, etc.), the system will further adjust the neural stimulation scheme. Assuming that Xiao Zhang's stress level rises to 7 / 10 during the treatment and his mood swings increase, the system will adjust the stimulation parameters based on the real-time psychological state. By inputting real-time data into the transform autoencoder model, the coupling characteristics of neural activity and psychological state will be updated, thereby adjusting the output stimulation scheme. For example, when the psychological state changes, the output stimulation intensity, frequency, and duration will be adjusted accordingly to better adapt to Xiao Zhang's current physiological and psychological state.

[0147] In this example, in the personalized neuromodulation treatment process of Xiao Zhang, we deeply explored how to jointly model neural activity signals and psychological state data through the transform autoencoder under the deep learning framework to achieve precise individualized regulation. In this process, Xiao Zhang's neural activity and psychological state will be mapped through the encoder part of the deep neural network, and these data will be learned and integrated in the low-dimensional space through joint modeling, thereby achieving dynamic neural regulation.

[0148] First, by introducing the transform autoencoder, Xiao Zhang’s neural activity and psychological state data are jointly modeled. The encoder part transforms the input neural activity signal X into N and psychological state data X P Mapping to a low-dimensional representation Z N . This low-dimensional representation not only contains the characteristics of neural activity, but also incorporates the characteristics of psychological state, reflecting the interaction between the two. Suppose that during Xiao Zhang's treatment, the input neural activity signals include his EEG waveform data and fMRI image data, while the psychological state data is obtained through psychological assessment. For example, Xiao Zhang's mood swing score is 6 / 10, the anxiety level is 4 / 10, and the stress is 7 / 10. These data will be input into the encoder together with the neural activity signal for joint modeling.

[0149] The joint modeling process of the transform autoencoder is expressed by the following formula:

[0150] Z N =W N ·X N +W P ·X P +b

[0151] in:

[0152] Z N It is a low-dimensional coupling representation of neural activity and psychological state. It is a comprehensive feature of neural activity and psychological state, which represents the state of the nervous system and takes into account the psychological state of the individual. Nis the weight matrix of the neural activity signal, which controls the influence of the neural activity signal in the low-dimensional representation. Assume that W N The value range of W is [0.5, 1.0], which represents the importance of neural activity signals in the model. P is the weight matrix of the psychological state data, which controls the contribution of the psychological state data in the low-dimensional representation. Assume that W P The value range of X is [0.3, 0.7], which represents the influence of psychological state data on the neural regulation model. N is the input neural activity signal. Assume that at a certain moment, Xiao Zhang’s EEG signal X N =[0.65, 0.72, 0.68, 0.70], reflecting the neural activity in the prefrontal region. P is the input psychological state data. Assume that Xiao Zhang’s psychological state data is X P = [0.6, 0.5, 0.7], representing emotion, stress, and anxiety levels, respectively. b is a bias term used to adjust the base value of the output, assuming b = 0.1.

[0153] With these inputs, the transforming autoencoder model can learn and output a low-dimensional coupled representation Z N , thereby integrating Xiao Zhang’s neural activity and psychological state into a comprehensive feature.

[0154] According to the above formula, we can calculate Xiao Zhang’s low-dimensional coupling representation Z N Assume that the weight matrix W of the neural activity signal is N =[0.6, 0.8, 0.7, 0.75] and the weight matrix W of the psychological state data P =[0.5,0.6,0.7], the contribution of each dimension can be calculated: Z N =[0.6 0.65+0.8 0.72+0.7 0.68+0.75 0.70]+[0.5 0.6+0.6 0.5+0.7 0.7]+0.1After calculation, we get:

[0155] Z N =[0.39+0.576+0.476+0.525]+[0.3+0.3+0.49]+0.1=[1.966]+[1.09]+0.1

[0156] =3.156

[0157] Therefore, the low-dimensional coupling representation Z N is 3.156, which represents the comprehensive state of Xiao Zhang’s neural activity and psychological state at that moment. Once the low-dimensional representation Z is obtained N, which will be passed as input to the decoder part of the transform autoencoder. The decoder's task is to output the corresponding neural stimulation scheme based on the low-dimensional representation to regulate Xiao Zhang's neural activity. The decoder will N Generate control signals such as stimulation intensity, frequency and time. Assume that the output of the decoder is:

[0158]

[0159] Among them, S stimulus is the stimulus signal generated by the decoder, which indicates the intensity, frequency, and duration of the neural stimulation to be applied. For example, the stimulus signal output by the decoder is:

[0160] S stimulus =[0.65,1.0,0.8]

[0161] in:

[0162] 0.65 represents the stimulation intensity; 1.0 represents the stimulation frequency (Hz); and 0.8 represents the stimulation duration (seconds).

[0163] In this way, the system is able to generate precise neuromodulation plans based on Xiao Zhang's current neural activity and psychological state.

[0164] During the treatment, Xiao Zhang's neural activity and psychological state will continue to change, and the system needs to dynamically adjust the neural stimulation plan based on real-time data. For example, suppose during the treatment, Xiao Zhang's mood swings and stress levels change, with the stress level rising from 7 / 10 to 8 / 10 and the mood swings falling from 6 / 10 to 5 / 10. At this time, the system will re-input the new psychological state data and adjust the stimulation plan based on the real-time neural activity.

[0165] Assume that the psychological state data obtained through real-time monitoring is:

[0166] X P =[0.65,0.75,0.8]

[0167] At this time, the system will calculate the low-dimensional coupling representation Z again N , and according to the new Z N Output adjusted stimulus signal S stimulus , ensuring that neuromodulation always adapts to Xiao Zhang's actual needs. By jointly modeling neural activity and psychological state through transforming autoencoders, and generating low-dimensional coupling representations based on this, it is possible to accurately provide personalized neurostimulation solutions based on Xiao Zhang's neural and psychological states. This method not only improves the accuracy of neuromodulation, but also can make adaptive adjustments based on real-time changes in psychological states to ensure maximum treatment effects.

[0168] This embodiment introduces real-time psychological state monitoring, and combines the deep learning model to dynamically adjust the coupling characteristics of neural activity and psychological state. During the treatment process, Xiao Zhang's emotional fluctuations, stress levels and other psychological state data will be collected in real time and input into the deep learning model. At the same time, combined with the changing trend of the neural activity signal, a real-time adjustment plan for neural stimulation is generated. This dynamic adjustment process is calculated through formulas and the neural stimulation plan is optimized at each moment to maximize the treatment effect.

[0169] First, Xiao Zhang's mood fluctuations and stress level data are collected through real-time psychological state monitoring equipment (such as emotion monitoring bracelets, pressure sensors, etc.). Assume that at a certain moment, Xiao Zhang's mood fluctuations are 6 / 10 and the stress level is 7 / 10. These data will be combined with neural activity signals (such as EEG data) and passed as input to the deep learning model. Through this model, the coupling characteristics of neural activity and psychological state can be extracted, and the neural regulation scheme can be adjusted in real time based on these characteristics.

[0170] In this process, the adjustment amount ΔS(t) of the neural stimulation scheme is calculated by the following formula:

[0171]

[0172] in:

[0173] ΔS(t) represents the adjustment amount of the neural stimulation scheme at time t, which indicates how the initial stimulation scheme S0 needs to be adjusted according to the current neural activity and psychological state data. N (t) is the coupling feature of neural activity and psychological state, which represents the low-dimensional representation extracted by the transform autoencoder, which combines the interaction between Xiao Zhang's neural activity and psychological state. S0 is the initial neural stimulation scheme, which is usually set at the beginning of treatment and contains parameters such as the frequency and intensity of stimulation. Assume that the initial scheme S0 = [0.7, 1.0, 0.5], where:

[0174] 0.7 represents the stimulus intensity (unit: mA);

[0175] 1.0 represents the stimulation frequency (unit: Hz);

[0176] 0.5 represents the stimulation duration (unit: seconds).

[0177] P(t) is the psychological state data at time t, including emotional fluctuations and stress levels. Assume that at a certain moment, Xiao Zhang’s psychological state data is P(t) = [0.6, 0.7], which represent emotional fluctuations and stress levels respectively. Θ A is a parameter of the control model, which is optimized through the deep learning training process and controls the adjustment of the neural stimulation scheme. Assume ΘA Contains several weight parameters that control the impact of different input features (neural activity and psychological state data) on stimulus adjustment.

[0178] After real-time psychological state monitoring, Xiao Zhang's current psychological state data is P(t) = [0.6, 0.7], and his mood swings and stress levels are 6 / 10 and 7 / 10 respectively. At this time, Xiao Zhang's neural activity signals are collected through the EEG device. Assuming X N (t) = [0.72, 0.68, 0.75, 0.70], representing the neural activity in the prefrontal cortex and other regions. These data are processed by the transform autoencoder to obtain a low-dimensional coupling representation Z N (t).

[0179] Assume that the coupling extracted from the transform autoencoder model is represented as:

[0180] Z N (t) = W N ·X N +W P ·P(t)+b

[0181] Assume W N =[0.6,0.7,0.8,0.9],W P = [0.5, 0.6], b = 0.2, the coupling expression Z can be calculated N (t):

[0182] Z N (t)=[0.6·0.72+0.7·0.68+0.8·0.75+0.9·0.70]+[0.5·0.6+0.6·0.7]+0.2

[0183] Calculated:

[0184] Z N (t)=[0.432+0.476+0.6+0.63]+[0.3+0.42]+0.2=2.138

[0185] This is the low-dimensional coupling representation Z at the current moment N (t) = 2.138.

[0186] Combined with the calculated low-dimensional coupling representation Z N (t), which is input into the control function Make real-time adjustments in:

[0187]

[0188] Assuming the control function According to the parameters Θ obtained from the training of the deep learning modelA =[1.2,0.9,1.1] to adjust, the resulting stimulus adjustment is:

[0189] ΔS(t)=[1.2·2.138+0.9·0.7+1.1·0.6]=[2.5656+0.63+0.66]=3.8556

[0190] By calculating the adjustment amount ΔS(t) of the stimulation scheme, a new neural stimulation scheme can be obtained:

[0191] S new (t) = S0 + ΔS(t)

[0192] Among them, S0 = [0.7, 1.0, 0.5] is the initial stimulus scheme, so the new stimulus scheme is:

[0193] S new (t)=[0.7+3.8556,1.0+3.8556,0.5+3.8556]=[4.5556,4.8556,4.3556]

[0194] This new stimulation scheme will be applied to Xiao Zhang's treatment to ensure that the neuromodulation can dynamically adapt to his changes in mental state. By introducing real-time mental state monitoring and combining the coupling characteristics of neural activity and mental state extracted by the transform autoencoder, the neurostimulation scheme can be adjusted in real time to optimize Xiao Zhang's treatment effect. In this process, the dynamic adjustment mechanism of the deep learning model effectively responded to Xiao Zhang's mood swings and changes in stress levels, ensuring the accuracy and adaptability of individualized neuromodulation.

Claims

1. The application method of adaptive neural network in individualized neural regulation is characterized by The following steps are involved: S1. Construction of individualized neural system representation model: S1.

1. Construct a model of the individual's neural system representation based on multi-source data including EEG, fMRI, genetic information, and psychological assessments; S1.

1. Use machine learning techniques including graph neural networks to mine key features of individual nervous systems, including neural connection patterns and synchronization, and construct a basic neural activity map for each individual. S2. Adaptive control path planning based on neural representation model: S2.

1. Based on the constructed individualized neural representation model, an adaptive control path planning method is introduced; the stimulation path that produces a regulatory effect on the nervous system is calculated by using the spatial and temporal variation trends of individual neural activity data; S2.2, Path planning is the stimulation of neural signals. Based on the connectivity interaction of various functional areas of the brain, the time window, intensity and frequency of stimulation are calculated to achieve dynamic regulation of individual neural activity; S3. Adaptive activation mechanism and dynamic regulation of the nervous system: S3.

1. Introducing an adaptive activation mechanism based on neural feedback; the adaptive activation mechanism monitors individual neural activity in real time and infers potential activation nodes and connections in the neural network through dynamic changes in neural activity; S3.

2. By adjusting the stimulation parameters in real time, the activation mode of the nervous system is optimized, so that the neural network can adjust itself and adapt to the changes in the external environment at different times; S4. Combination of personalized neural regulation and plasticity mechanisms: S4.

1. Use adaptive neural network models to evaluate the plasticity of individual neural networks in real time, and design regulatory strategies to promote the development of neural plasticity by combining the activation and feedback mechanisms of the nervous system; S5. Deep coupling between neural regulation and individual psychological state: S5.

1. Develop dynamic control methods based on the coupling of neural activity and psychological state; jointly model neural signals with psychological states including mood swings and stress levels through deep learning models including transform autoencoders, and adjust neural control schemes according to real-time changes in individual psychological states.

2. The method for applying the adaptive neural network in individualized neural regulation according to claim 1, characterized in that The adaptive control path planning method based on the neural representation model includes: By combining multi-source physiological data including EEG, fMRI, and genomics with psychological assessment data, a neural activity characteristic model for each individual is constructed; the neural activity characteristic model reflects the structure of the nervous system and describes the dynamic changes of neural functions, including the connectivity and synchronization functional characteristics of the neural network; the individualized neural representation model is expressed as: in: N model (t) represents the output of the individualized neural representation model at time t, which represents the comprehensive performance of individual neural activity; f i (t) is the i-th input feature, which refers to the time series data extracted from different physiological data sources including EEG signals and fMRI image data, reflecting the characteristics of neural activity; w i is the weight coefficient of the i-th feature, indicating the importance of the feature to the output of the neural representation model; b is the bias term, which adjusts the basic output of the model; n is the total number of input features, indicating the dimension of the model input features.

3. The method for applying the adaptive neural network in individualized neural regulation according to claim 2, characterized in that The adaptive control path planning method based on the neural representation model includes: Stimulation paths are planned based on the functional connectivity of different areas of the nervous system, including the cerebral cortex and basal ganglia. The degree of functional connectivity of each area to other areas determines its contribution to the overall neural activity, and stimulation paths are allocated based on connectivity. The formula for spatial optimization is: Where: P spatial represents the optimized stimulation path, which reflects the stimulation distribution in space; is the spatial extent of the brain region, indicating the set of regions of the nervous system that need to be stimulated; S j (x, t) is the intensity of neural activity at spatial position x and time t, reflecting the state of neural activity in a specific area; a j is the weight coefficient of the jth region, indicating the contribution weight of the region in the overall neural activity; m is the number of neural regions that need to be regulated, indicating the number of neural regions considered; dx is the spatial integration variable, indicating the integration process of the brain spatial region; In terms of time optimization, a stimulation time window is designed based on the time series of individual neural activity; the stimulation timing is selected within the dynamic time window by weighted summing and maximizing the temporal characteristics of neural activity.

4. The method for applying the adaptive neural network in individualized neural regulation according to claim 3, characterized in that The adaptive control path planning method based on the neural representation model includes: Adopt real-time monitoring and dynamic adjustment mechanism; design an adaptive adjustment mechanism based on real-time data to dynamically adjust the stimulation path according to changes in neural activity; the real-time adjustment formula is: Where: ΔP(t) represents the amount of stimulation path adjustment at time t, indicating the stimulation direction and amplitude that need to be adjusted; γ is the adjustment coefficient, which controls the sensitivity of the adjustment; N model (t) is the neural activity value predicted by the individualized neural representation model, which represents the expected neural system activity state; N current (t) is the current neural activity value obtained by real-time monitoring equipment, indicating the actual activity state of the nervous system; is the gradient operator, which represents the difference N model (t)N current The gradient of (t) is calculated.

5. The method for applying the adaptive neural network in individualized neural regulation according to claim 1, characterized in that The deep coupling method of neural regulation and individual psychological state includes: First, the neural activity signal and psychological state data are preprocessed by the encoder part of the deep learning model; the encoder part fuses the multimodal data into a low-dimensional representation through nonlinear mapping to form the coupling characteristics of neural activity and psychological state; then the coupling characteristics are restored by the decoder part to generate a control output signal corresponding to the neural activity, and adjust the neural stimulation scheme according to the current psychological state; the coupling mapping expression formula of the transformed autoencoder is: Where: Z N To represent the low-dimensional coupling representation of neural activity and psychological state; by transforming the encoder part of the autoencoder model, the neural activity signal and psychological state data are mapped to a shared low-dimensional feature space to form a coupling representation of neural and psychological state; X N Input includes EEG, fMRI neural activity signals; X P is the input psychological state data; the state data is obtained through physiological or psychological assessment; Θ E are the weights and biases of the encoder network, which are the parameters adjusted during the network learning process; It is a nonlinear mapping function used to transform neural activity and psychological state signals into low-dimensional coupled representations.

6. The method for applying the adaptive neural network in individualized neural regulation according to claim 5, characterized in that The deep coupling method of neural regulation and individual psychological state includes: By introducing a transform autoencoder, neural activity and psychological state are jointly modeled; the structure of the transform autoencoder includes an encoder, a decoder, and a joint layer; the encoder maps the input neural activity signal and psychological state data to a low-dimensional representation through a nonlinear activation function; the features of neural activity and psychological state in the latent space are learned and fused.

7. The method for applying the adaptive neural network in individualized neural regulation according to claim 6, characterized in that The deep coupling method of neural regulation and individual psychological state includes: Real-time mental state monitoring is introduced, and the neural regulation scheme is dynamically adjusted by combining the coupling characteristics of neural activity and mental state extracted by the transform autoencoder. Real-time mental state monitoring includes real-time collection of emotional fluctuations and stress level data. The data is input into the deep learning model, and the neural stimulation adjustment scheme is generated by combining the changing trend of the neural activity signal: Among them, ΔS(t) is the adjustment amount of the neural stimulation scheme at time t; represents the real-time adjustment of the neural regulation scheme according to the current neural activity and psychological state; S0 is the initial neural stimulation scheme, which is set at the beginning of treatment, including the frequency and intensity of stimulation; P(t) is the psychological state data at time t, including emotional fluctuations and stress levels; Θ A To regulate the parameters of the model, the deep learning model is trained for optimization and the adjustment method of the stimulation scheme is controlled; It is a nonlinear control function used to adjust the neural stimulation scheme according to the coupling characteristics of neural activity and psychological state.

Images