Closed-loop regulated transcranial alternating current stimulation method and system

Abstract

The present application relates to the technical field of neuromodulation, and particularly relates to a closed-loop controlled transcranial alternating current stimulation method and system, comprising calculating a global brain network efficiency value at the current time according to a blood oxygen level dependent signal in real time; predicting a global brain network efficiency prediction value at the next sampling time through a nonlinear prediction model according to a time sequence of historical global brain network efficiency values and current and historical stimulation parameters; comparing the global brain network efficiency prediction value with a preset individualized target brain network efficiency, and calculating a prediction efficiency deviation; and obtaining a group of advanced stimulation parameters according to the prediction efficiency deviation mapping, the present application realizes advanced closed-loop control of generating and issuing stimulation instructions in advance for the predicted future brain network state deviation, thereby overcoming the mismatch problem between the control instructions and the real-time state of the brain caused by the inherent hemodynamic delay of functional magnetic resonance signals.

Description

Technical Field

[0001] This invention relates to the field of neuromodulation technology, and in particular to a closed-loop controlled transcranial alternating current stimulation method and system. Background Technology

[0002] In closed-loop transcranial alternating current stimulation (TCC) based on functional magnetic resonance imaging (fMRI), the oxygen-dependent signal from fMRI inherently exhibits a hemodynamic delay, typically 2 to 6 seconds. This means that the currently acquired oxygen-dependent signal reflects the brain's neural activity state from several seconds prior. To achieve real-time modulation, existing technologies generally employ two strategies: first, directly calculating stimulation parameters using the current oxygen-dependent signal characteristics—essentially a delayed modulation, where stimulation acts on a past brain state; second, attempting to linearly extrapolate the oxygen-dependent signal or functional connectivity strength to predict future signal values, thereby guiding current stimulation. The second approach, prediction and compensation, has been explored in existing technologies. Its typical process can be summarized as follows: real-time acquisition of the oxygen-dependent signal, calculation of the instantaneous functional connectivity strength in a specific brain region, extrapolation of the next moment's strength value using linear regression or moving average methods based on the recent trend of this connectivity strength, and finally adjustment of stimulation parameters based on the extrapolated value. This method, to some extent, targets the modulation towards the future.

[0003] However, the brain is a highly nonlinear and complex dynamic system, and the evolution of its functional network state is not a simple linear process. In particular, when closed-loop transcranial alternating current stimulation is applied as an intervention, the resulting brain network response is complex and non-stationary. Existing prediction methods based on linear extrapolation have limited ability to represent the overall functional state of the network by predicting objects (such as individual connection strengths), and the extrapolation model is difficult to characterize the nonlinear dynamics under intervention. This leads to insufficient accuracy of prediction results, resulting in insufficient precision and reliability of proactive regulation, and failing to solve the problem of regulatory mismatch caused by delay. Summary of the Invention

[0004] The main objective of this invention is to provide a closed-loop controlled transcranial alternating current stimulation method and system, aiming to solve the technical problems mentioned in the background art.

[0005] This invention proposes a closed-loop controlled transcranial alternating current stimulation method, comprising: Real-time acquisition of blood oxygen level-dependent signals of the target brain region, and real-time calculation of the global brain network efficiency value at the current moment based on the blood oxygen level-dependent signals; The time series of historical global brain network efficiency values ​​and current and historical stimulus parameters are obtained to predict the global brain network efficiency value at the next sampling time through a nonlinear prediction model. The predicted global brain network efficiency is compared with the preset individualized target brain network efficiency, and the predicted efficiency deviation is calculated. Based on the predicted efficiency deviation, the pre-stored stimulation parameter-network efficiency response surface for the target brain region is queried, and a set of advanced stimulation parameters is obtained by mapping. The advanced stimulation parameters are used to issue before the neural activity corresponding to the next sampling time occurs. A transcranial alternating current stimulation signal is generated based on the aforementioned pre-stimulation parameters and applied to the target brain region.

[0006] The present invention is further configured such that the step of acquiring the blood oxygen level-dependent signal of the target brain region in real time and calculating the global brain network efficiency value at the current moment based on the blood oxygen level-dependent signal includes: Based on a preset time window, extract the time series of multiple brain regions in the blood oxygen level dependent signal at the current and nearby times; Based on the time series of each brain region, the Pearson correlation coefficient between every two brain regions is calculated to construct the dynamic functional connectivity matrix at the current moment; The dynamic functional connectivity matrix is ​​thresholded to obtain a binarized brain functional connectivity map. The global brain network efficiency value at the current moment is obtained by calculating the inverse average of the shortest path lengths between all node pairs in the brain functional connectivity graph.

[0007] The present invention is further configured such that the step of obtaining the time series of historical global brain network efficiency values ​​and current and historical stimulus parameters, and predicting the global brain network efficiency value at the next sampling time using a nonlinear prediction model, includes: Efficiency time series vector consisting of global brain network efficiency values ​​at historical moments is obtained, as well as stimulation history sequence consisting of stimulation parameters at the current moment and historical moments, wherein the number of historical moments corresponding to the stimulation parameters is determined according to hemodynamic delay time, and the stimulation parameters include stimulation frequency and stimulation current amplitude. The stimulus history sequence is input into a temporal convolutional network to extract stimulus history feature vectors; The efficiency time series vector is concatenated with the stimulus history feature vector to obtain the fused feature vector.

[0008] The fused feature vector is input into a multilayer perceptron, which contains at least one hidden layer, and the output is used for the next... Global brain network efficiency prediction value

[0009] The present invention is further configured such that the temporal convolutional network extracts temporal features from the stimulus history sequence through causal convolution operations, wherein the causal convolution operations satisfy: Its computational output depends only on the stimulus parameters at the current and historical moments; The temporal receptive field length of the convolutional kernel used for feature extraction does not exceed the length of the stimulus history sequence; The weight and bias parameters in the causal convolution operation are obtained by training the nonlinear prediction model with the goal of minimizing the network efficiency prediction error.

[0010] The present invention is further configured such that the preset individualized target brain network efficiency is determined through the following steps: Acquire functional magnetic resonance imaging data of subjects in a resting state to calculate their baseline brain network efficiency; Obtain the subjects' clinical diagnosis or cognitive task performance to determine the expected direction and magnitude coefficient of network efficiency modulation; Individualized target brain network efficiency is obtained based on the baseline brain network efficiency and the network efficiency adjustment direction and amplitude coefficients.

[0011] The present invention is further configured such that the step of querying a pre-stored stimulation parameter-network efficiency response surface for the target brain region based on the prediction efficiency deviation, and mapping to obtain a set of advanced stimulation parameters includes: Obtain a pre-constructed stimulation parameter-network efficiency response surface for the target brain region, wherein the stimulation parameter-network efficiency response surface takes stimulation frequency and stimulation current amplitude as input and the induced network efficiency change as output. Using the predicted efficiency bias as the desired output, search for the cost function on the stimulus parameter-network efficiency response surface. Stimulation current amplitude; The found stimulation frequency and stimulation current amplitude are used as the advanced stimulation parameters.

[0012] Preferably, after the step of generating a transcranial alternating current stimulation signal based on the pre-stimulation parameters and applying it to the target brain region, the method further includes: Continue to acquire blood oxygen level-dependent signals in real time and calculate the actual global brain network efficiency value generated; The actual global brain network efficiency value is compared with the predicted global brain network efficiency value, and the prediction error is calculated. If the prediction error exceeds a preset threshold, at least one of the following will be triggered: The parameters of the nonlinear prediction model are fine-tuned online. Local refitting of the stimulus parameter-network efficiency response surface.

[0013] The present invention also provides a closed-loop controlled transcranial alternating current stimulation system, comprising: The acquisition and calculation module is used to acquire the blood oxygen level dependent signal of the target brain region in real time, and calculate the global brain network efficiency value at the current moment based on the blood oxygen level dependent signal. The efficiency prediction module is used to obtain the time series of historical global brain network efficiency values ​​and current and historical stimulus parameters, so as to predict the global brain network efficiency value at the next sampling time through a nonlinear prediction model. The comparison calculation module is used to compare the predicted global brain network efficiency with the preset individualized target brain network efficiency and calculate the prediction efficiency deviation. The parameter mapping module is used to query the pre-stored stimulation parameter-network efficiency response surface for the target brain region based on the prediction efficiency deviation, and map to obtain a set of advanced stimulation parameters. The advanced stimulation parameters are used to be issued before the neural activity corresponding to the next sampling time occurs. The synchronization control module is used to generate a transcranial alternating current stimulation signal based on the advanced stimulation parameters and apply it to the target brain region.

[0014] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of a closed-loop controlled transcranial alternating current stimulation method.

[0015] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of a closed-loop controlled transcranial alternating current stimulation method.

[0016] The beneficial effects of this invention are as follows: This invention uses global brain network efficiency as the core regulatory indicator and constructs a nonlinear prediction model that integrates historical brain network state and complete stimulation history to characterize the non-stationary dynamic evolution law of brain functional network under transcranial alternating current stimulation intervention. Based on this prediction, the system realizes advanced closed-loop regulation that generates and issues stimulation instructions in advance for the predicted deviation of future brain network state, thereby effectively overcoming the mismatch between regulatory instructions and real-time brain state caused by the inherent hemodynamic delay of functional magnetic resonance imaging signals. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of a method flow according to an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of the system structure according to an embodiment of this application.

[0019] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0021] like Figure 1 As shown, this application provides a closed-loop controlled transcranial alternating current stimulation method, comprising: S1, acquire the blood oxygen level dependent signal of the target brain region in real time, and calculate the global brain network efficiency value at the current moment based on the blood oxygen level dependent signal. S2, obtain the time series of historical global brain network efficiency values ​​and current and historical stimulus parameters, so as to predict the global brain network efficiency value at the next sampling time through a nonlinear prediction model; S3, compare the predicted global brain network efficiency with the preset individualized target brain network efficiency, and calculate the predicted efficiency deviation; S4. Based on the predicted efficiency deviation, query the pre-stored stimulation parameter-network efficiency response surface for the target brain region and map it to obtain a set of advanced stimulation parameters. The advanced stimulation parameters are used to issue before the neural activity corresponding to the next sampling time occurs. S5. Generate a transcranial alternating current stimulation signal based on the aforementioned pre-stimulation parameters and apply it to the target brain region.

[0022] For example, upon obtaining the pre-stimulation parameters, the system immediately sends this digital instruction to the transcranial alternating current stimulation generation unit. The core of this unit is a high-precision digital-to-analog converter module, responsible for converting the digital frequency and amplitude instructions into analog voltage control signals of the corresponding frequency. This control signal is then input to a voltage-controlled constant current source circuit, which maintains a constant output current despite changes in load impedance, thereby generating a sinusoidal alternating current stimulation signal with a specific stimulation frequency and a constant stimulation current amplitude. The generated stimulation signal is guided via a multi-channel switching and output interface to a multi-electrode array pre-placed on the subject's scalp according to the target brain region. To ensure safety, throughout the entire stimulation signal output process, an independent safety protection circuit monitors multiple physical parameters in real time, including the contact impedance between each channel electrode and the scalp, and the actual output current and voltage waveforms. This safety module has strict thresholds set, such as an upper limit for contact impedance and upper and lower limits for current output. Once any parameter is detected to exceed the safety range, the circuit will directly cut off the stimulation output within milliseconds and trigger an audiovisual alarm.

[0023] In one embodiment of the present invention, the step of acquiring the blood oxygen level-dependent signal of the target brain region in real time and calculating the global brain network efficiency value at the current moment based on the blood oxygen level-dependent signal includes: S11, according to a preset time window, extract the time sequence of multiple brain regions in the blood oxygen level dependent signal at the current and nearby times; For example, the preset time window length is set to 1 to 2 seconds, specifically adjusted adaptively according to the sampling frequency of functional magnetic resonance imaging (fMRI) (for example, when the sampling frequency is 1 Hz, the time window length is set to 1 second, corresponding to 1 sampling point; when the sampling frequency is 2 Hz, the time window length is set to 1 second, corresponding to 2 sampling points). The design basis for this value range is: when it is less than 1 second, the noise ratio of the blood oxygen level dependent signal is too high, resulting in unstable calculation of the functional connectivity matrix; when it is greater than 2 seconds, the time resolution is insufficient, and it is impossible to capture the dynamic changes in brain network efficiency, thus balancing the core requirements of time resolution and computational stability. S12, Based on the time series of each brain region, calculate the Pearson correlation coefficient between every two brain regions to construct the dynamic functional connectivity matrix at the current moment; S13, Threshold the dynamic functional connectivity matrix to obtain a binarized brain functional connectivity map. For example, the thresholding process uses a "sparseness thresholding method," with the following specific rules: the sparsity of the preset brain functional connectivity matrix is ​​0.1~0.3 (i.e., retaining 10%~30% of the strongest connections in the matrix), the correlation coefficient threshold corresponding to this sparsity is calculated, and elements in the matrix with absolute values ​​below the threshold are set to 0, while elements with absolute values ​​above or equal to the threshold are set to 1. The sparsity value needs to be pre-calibrated based on the subject's resting-state fMRI data: for subjects with high brain network connectivity density, the sparsity is 0.1~0.2; for subjects with low connectivity density, the sparsity is 0.2~0.3, ensuring that the connections retained after thresholding are statistically significant (P<0.05, verified by permutation test). S14, calculate the inverse average of the shortest path lengths between all node pairs in the brain functional connectivity graph to obtain the global brain network efficiency value at the current moment.

[0024] As described in steps S11-S14 above, the preset time window length depends on a trade-off between temporal resolution and computational stability. For data segments from all brain regions within this time window, the Pearson correlation coefficient between any two different brain region time series is calculated. The correlation coefficient calculation results of all brain region pairs are arranged into a symmetric square matrix, which yields the dynamic functional connectivity matrix at the current moment. The element values ​​of this matrix reflect the degree of linear synchronization of neural activity in two brain regions on low-frequency oscillations within the window time, effectively capturing the covariation patterns of blood oxygenation signal waveforms between brain regions. Thresholding is applied to this dynamic functional connectivity matrix. By using a preset threshold rule, elements in the matrix with absolute values ​​below the threshold are set to zero, while elements with values ​​above or equal to the threshold are set to 1. After this binarization operation, the matrix is ​​transformed into an unweighted, undirected simple graph adjacency matrix, defining the topological connectivity of the brain functional network within the current time window. By filtering out weak connections that may be caused by noise, strong connections that are statistically significant or physiologically important are retained, thus revealing the backbone structure of the brain network. Based on the graph represented by the binarized adjacency matrix, the shortest path algorithm in graph theory is used to calculate the shortest path length between all pairs of nodes in the graph. For each pair of connected nodes, its transmission efficiency is calculated as the reciprocal of the shortest path length; for disconnected pairs of nodes, its efficiency is counted as zero. Finally, the sum of the efficiency values ​​of all pairs of nodes is divided by the total number of pairs of nodes to obtain the global brain network efficiency value at the current time.

[0025] This invention, through the aforementioned steps, ensures that the global brain network efficiency value, a core state variable of the system, is a reliable signal with high signal-to-noise ratio, clear physiological interpretation, and computational stability. The combination of Pearson correlation and thresholding generates a connection pattern estimate that is relatively robust to instantaneous fluctuations, avoiding misjudgments of the overall network structure due to drastic fluctuations at individual time points. Furthermore, the global efficiency calculation based on shortest paths is extremely sensitive to subtle changes in the network's topology. For example, the opening or closing of a critical connection can significantly alter the shortest path length between many node pairs, causing significant fluctuations in the global efficiency value. This high sensitivity means that any meaningful functional reorganization of the network state can be sensitively captured and quantified by this indicator. Therefore, through the standardized transformation process from the raw signal to the efficiency value described above, the time series input to the prediction model is guaranteed to more accurately reflect the intrinsic dynamics of the brain's functional network, providing high-quality and predictable signals.

[0026] Formula for calculating the efficiency value of a global brain network:

[0027] In the formula, express The global brain network efficiency value at time (corresponding to the current time). This represents the total number of nodes in the brain region. This represents a binarized brain functional connectivity map after thresholding. Represents the nodes in the graph With nodes The shortest path length (i.e., number of edges) between two nodes is calculated using a shortest path algorithm in graph theory (such as the Floyd-Warshall algorithm). If there is no path between two nodes, then... =∞, its reciprocal is treated as 0.

[0028] This formula abstracts the network as an information transmission channel, calculating the average transmission efficiency between all node pairs. Its core lies in using the reciprocal of the shortest path length to define the efficiency between node pairs. This simulates the assumption that information is transmitted along the optimal path. Calculating the average efficiency of all node pairs ensures that this indicator reflects the global properties of the network, rather than local characteristics. This makes it sensitive to changes in the overall network topology while exhibiting robustness to minor fluctuations or noise in local connections. Therefore, it provides a state observation variable with ideal mathematical properties for solving signal delay problems. The formula generates... The value has a well-defined boundary (between 0 and 1) and is a continuous variable, which provides good numerical stability for subsequent prediction models. More importantly, this index has a non-linear and significant response to changes in network topology, especially the connectivity of critical connections. The loss of a critical connection may lead to a sharp increase in the shortest path length between multiple node pairs, resulting in a significant decrease in the global efficiency value. This sensitivity means... It can act as a high-gain state sensor, reflecting in a timely and significant manner the meaningful structural reorganization of the brain's functional network caused by internal activity or external stimuli. In the closed-loop process of addressing delay issues, a stable, sensitive, and interpretable state variable is crucial. The stability of this formula ensures the smoothness of the time series, which is beneficial for trend prediction in the model; its sensitivity guarantees that any potential network effects triggered by stimuli can be effectively captured, becoming the learning target of the predictive model; and its explicit graph theory interpretation links the abstract brain state with the specific network information transmission capacity, making improving or reducing efficiency a clear and reasonable regulatory objective.

[0029] In one embodiment of the present invention, the step of obtaining the time series of historical global brain network efficiency values ​​and current and historical stimulus parameters, and using a nonlinear prediction model to predict the global brain network efficiency value at the next sampling time, includes: S21, Obtain the efficiency time series vector consisting of the global brain network efficiency values ​​at M historical moments. T(where T represents transpose), and obtain the stimulation history sequence consisting of stimulation parameters at the current time and L historical time points, where L is determined according to the hemodynamic delay time. The stimulation parameters at each time point include the stimulation frequency and stimulation current amplitude (L is set to be slightly larger than the number of sampling points corresponding to the average hemodynamic delay time (e.g., 6-8 seconds)). For example, the length M of the efficiency time series vector is 4 to 8, corresponding to the global brain network efficiency values ​​at 4 to 8 historical moments. The basis for this value is: when M is less than 4, the model cannot capture the short-term evolution trend of brain network efficiency; when M is greater than 8, redundant historical information will be introduced, increasing the computational complexity of the model and reducing real-time performance. The specific value can be adjusted according to the prediction accuracy requirements: if higher prediction accuracy is required, M is 6 to 8; if real-time performance is to be prioritized, M is 4 to 6. The value of M must match the sampling frequency of functional magnetic resonance imaging (fMRI) (for example, when the sampling frequency is 1Hz, M=6 corresponds to 6 seconds of historical data). S22, The stimulus history sequence is input into a temporal convolutional network to extract the stimulus history feature vector; S23, the efficiency time series vector is concatenated with the stimulus history feature vector to obtain a fused feature vector. T ; S24, the fused feature vector is input into a multilayer perceptron, the multilayer perceptron containing at least one hidden layer, and the output is... Predicted global brain network efficiency at time step .

[0030] As described in steps S21-S24 above, the implementation of the nonlinear prediction model includes a cascaded and specialized data processing pipeline. First, the model receives two input sequences: an efficiency time series vector consisting of global brain network efficiency values ​​at M historical moments. T(Where T represents transpose), and obtain the stimulation history sequence consisting of stimulation parameters from the current time and L historical time points, where L is determined based on the hemodynamic delay time. The stimulation parameters at each time point include the stimulation frequency and the stimulation current amplitude (L is set to be slightly larger than the number of sampling points corresponding to the average hemodynamic delay time (e.g., 6-8 seconds). The value of L is determined by dividing the average hemodynamic delay time by the system sampling interval to ensure coverage of the complete historical window where the stimulation may have an effect). Then, the stimulation history sequence is fed into a temporal convolutional network module, which consists of at least one causal convolutional layer. This layer automatically extracts from the original stimulation parameter sequence... A low-dimensional stimulus history feature vector containing temporal patterns is generated. Then, the efficiency time-series vector representing the brain's state history is concatenated with the stimulus history feature vector representing the control history pattern, fusing them into a comprehensive feature vector. This concatenation operation is performed along the feature dimension, aiming to jointly represent state and control information in the same vector space. Finally, the fused feature vector is fed into a multilayer perceptron, which includes an input layer, at least one hidden layer with a nonlinear activation function, and an output layer. Neurons in the hidden layer perform complex nonlinear transformations on the input features, and the output layer ultimately produces a scalar value for the next sampling time. Predicted global brain network efficiency at time step The weight parameters of the entire model are optimized during the offline phase using supervised training with historically collected stimulus-efficiency paired data sequences, aiming to minimize the prediction error. It should be noted that the stimulus-efficiency paired data sequence refers to a set of structured data with strict temporal correspondence used for supervised training of the nonlinear prediction model. This sequence consists of two parts: first, the stimulus history sequence, which refers to the transcranial alternating current stimulation parameters actually applied to the subject at multiple consecutive historical sampling moments, including at least the stimulation frequency and stimulation current amplitude at each moment; second, the efficiency response sequence, which refers to the global brain network efficiency value calculated in real time at subsequent corresponding moments after the stimulation is applied, due to changes in neural activity and hemodynamic delay. Each pair of paired data clearly establishes a causal relationship between a specific stimulus history input and its resulting network efficiency output observed at a specific future moment. The entire training dataset consists of a large number of such "input-output" paired samples extracted from different time segments, where the model input is a combination of a fixed-length, continuous historical stimulus sequence and a historical efficiency sequence, and the model's training target (i.e., label) is the true efficiency value at the moment immediately following that historical window. By optimizing model parameters in this data-driven manner, the aim is to enable the model to learn autonomously from the data and internalize the complex dynamic mapping relationship between stimulus intervention and delayed network state response, thereby gaining the ability to accurately predict the future efficiency of the brain network. This is the core premise and engineering foundation for the realization of the entire closed-loop advanced regulation scheme.

[0031] This invention, through a nonlinear prediction model with a specific architecture, can create a data-driven system capable of end-to-end learning the complex dynamic mapping between stimulus intervention and delayed network state responses. The components of the model architecture are as follows: the temporal convolutional network is specifically designed to handle temporal stimulus sequences; its causal convolutional properties are a rigid requirement for real-time online operation, while its dilated convolutional structure enables it to effectively capture long-range temporal dependencies, which is crucial for modeling hemodynamic delays of several seconds. Concatenating stimulus features with historical efficiency features is based on a key assumption: the future network state is simultaneously determined by its own historical evolutionary inertia and the legacy effects of a series of past stimulus interventions; therefore, these two types of information must be fused at the decision-making level. The multilayer perceptron, as a general function approximator, uses its multilayer nonlinear transformation structure to give the model the ability to learn arbitrarily complex nonlinear relationships, used to fit unknown and potentially very complex functional relationships between fused features and future efficiency values.

[0032] This invention intrinsically and adaptively models the temporal dynamics of transcranial alternating current stimulation (TCD) effects through a nonlinear prediction model, thereby generating reliable predictions of future brain network states. When processing stimulus history, the trainable convolutional kernel weights of the temporal convolutional network automatically adjust their attention to stimulus signals at different historical moments during training. This means the model can autonomously learn from the data that, for example, stimuli applied in the most recent 2-4 seconds may have the greatest impact on the current prediction, while earlier stimuli may have a diminishing influence. This is essentially a data-driven, non-parametric estimation of the stimulus effect's time window and contribution distribution, replacing the imprecise models in traditional methods that require manually preset fixed delay parameters. By combining the learned stimulus temporal effects with the network's self-evolutionary history, the multilayer perceptron ultimately determines the most likely efficiency level the network will reach in the next moment. The entire model, as a whole, produces a forward-looking and accurate state predictor. The global brain network efficiency prediction provides information about the future, enabling the control system to advance the decision point from the present (when the actual state is unknown) to the past. Based on the prediction of the future, it can formulate control commands that should be issued now, thereby contributing to solving the delay problem.

[0033] In one embodiment of the present invention, the temporal convolutional network extracts temporal features from the stimulus history sequence through causal convolution operations, wherein the causal convolution operations satisfy: Its computational output depends only on the stimulus parameters at the current and historical moments; The temporal receptive field length of the convolutional kernel used for feature extraction does not exceed the length of the stimulus history sequence; The weight and bias parameters in the causal convolution operation are obtained by training the nonlinear prediction model with the goal of minimizing the network efficiency prediction error (where minimizing the network efficiency prediction error refers to the loss function that needs to be minimized during the model training phase, with the goal of making the model's prediction value as close as possible to the actual measured value).

[0034] As described in the above scheme, specifically, let the stimulus history sequence be the... The stimulus parameter vector at each historical moment is (Including the stimulation frequency and current amplitude at that moment), where For each output channel The causal convolutional layer in the 1st... Output characteristics at time step The formula is obtained by weighted summation of the input sequence:

[0035] in, Indicates the causal convolutional layer at the 1st... Time, Number Feature values ​​on each output channel; This represents the kernel size, a preset hyperparameter that determines the range of influence of historical stimuli considered by the model. This ensures that calculations use only data from the current and historical moments. Show the first The convolution kernel corresponding to the nth output channel is at the nth... The weight parameters at each position are optimized during the model training phase using the backpropagation algorithm and are used to extract specific temporal patterns in the stimulus history. Indicates the first The bias parameters of each output channel are trainable parameters of the model, which are learned during training. Indicates the first The stimulus parameter vector at time; Finally, the output characteristics of all output channels The stimulus history feature vector is formed by combining these features.

[0036] As described above, this invention further enhances the targeted solution to the hemodynamic delay problem. Its inventiveness lies in: explicitly limiting the convolution operation to only use the current and past timeframes through the formula for causal convolution layer operations. The input strictly prevents the leakage of future information, which ensures the physical feasibility and causal rationality of the predictive model in real-time closed-loop control; and stimulates historical sequences. Through trainable weights The weighted combination essentially allows the model to automatically learn the contribution of stimuli at different historical moments to the prediction of the current network state (i.e., the temporal course of the stimulus effect). This enables the model to intrinsically characterize the delay, accumulation, and decay process of the stimulus effect from its application to its manifestation in the blood oxygen level-dependent signal, rather than relying on a fixed linear assumption. This specific operation is closely coupled with the sequence length (determined by the delay), the feature fusion step, and the subsequent MLP predictor in the aforementioned steps, together forming an organic overall prediction scheme that is deeply designed to solve the problem of anticipatory regulation.

[0037] In one embodiment of the present invention, the preset individualized target brain network efficiency is determined through the following steps: S31, acquire functional magnetic resonance imaging data of subjects in a resting state to calculate their baseline brain network efficiency; S32, Obtain the subject's clinical diagnosis or cognitive task performance to determine the expected direction and amplitude coefficient of network efficiency modulation; S33, obtain the individualized target brain network efficiency based on the baseline brain network efficiency and the network efficiency adjustment direction and amplitude coefficient.

[0038] Formula for calculating the efficiency of individualized target brain networks:

[0039] in, Indicates the efficiency of individualized target brain networks. Indicates baseline brain network efficiency. The coefficient representing the direction and magnitude of network efficiency adjustment is a percentage coefficient preset based on the intervention target.

[0040] Formula for calculating prediction efficiency deviation:

[0041] in, This indicates prediction efficiency bias. Indicates the efficiency of individualized target brain networks. This represents the predicted value for global brain network efficiency.

[0042] As described in steps S31-S33 above, it can quantify the inherent differences in brain network properties among individuals. Different individuals have significantly different resting-state network integration efficiency due to differences in brain structure, age, pathological state, and even cognitive style. Directly using a uniform and absolute efficiency value as the regulatory target for all subjects ignores this biological heterogeneity and may lead to insufficient stimulation for some individuals and excessive stimulation for others.

[0043] After obtaining individualized baseline values, a clear direction and magnitude of regulation are determined. This is achieved by introducing an adjustable percentage coefficient, which is closely linked to specific clinical or research intervention goals. For example, for cognitive enhancement interventions aimed at improving working memory, It can be set to a positive value; however, for interventions aimed at reducing hypersynchronization of the epileptic network... This can then be set to a negative value. The specific value can be determined based on previous research data, the clinical judgment of physicians, or by associating it with expected improvement in behavioral performance on a specific cognitive task. For example, if the goal is set to improve network efficiency to 115% of the baseline level, then... =0.15. Finally, the efficiency of the individualized target brain network is obtained through the formula for calculating the efficiency of the individualized target brain network.

[0044] In one embodiment of the present invention, the step of querying a pre-stored stimulation parameter-network efficiency response surface for the target brain region based on the prediction efficiency deviation, and mapping to obtain a set of advanced stimulation parameters includes: S41, obtain the stimulation parameter-network efficiency response surface pre-constructed for the target brain region; the stimulation parameter-network efficiency response surface takes the stimulation frequency and stimulation current amplitude as input and the induced network efficiency change as output; S42, using the predicted efficiency deviation as the desired output, find the cost function on the stimulus parameter-network efficiency response surface. Minimum stimulation frequency and stimulation current amplitude; In the cost function, Represents the cost function, This represents the stimulus parameter-network efficiency response surface on the stimulus parameter The corresponding predicted efficiency change is obtained by pre-fitting the response surface through prior experiments (applying stimuli with different parameters and recording efficiency changes) or simulations based on physiological models. Indicates the frequency of stimulation. Indicates the amplitude of the stimulation current. This indicates a bias in prediction efficiency; S43, the found stimulation frequency and stimulation current amplitude are used as the advanced stimulation parameters.

[0045] For example, the advanced stimulation parameters need to meet the following safety range constraints: stimulation frequency The value range is 1~20Hz (covering physiologically relevant frequency bands such as theta wave, alpha wave, and beta wave), and the amplitude of the stimulation current is... The value range is 0.5~2.0mA (to meet the clinical safety standards for transcranial alternating current stimulation and avoid scalp burns or nerve damage due to excessive current). If the stimulation parameters obtained by optimization through response surface methodology exceed the above range, the parameters will be automatically adjusted to the nearest safe boundary value.

[0046] As described in steps S41-S43 above, the stimulus parameter-network efficiency response surface is a mathematical function or lookup table that takes two-dimensional stimulus parameters as input and one-dimensional efficiency change as output. In practice, the construction of this stimulus parameter-network efficiency response surface requires a brief individualized calibration phase. During this phase, under the premise of ensuring safety, multiple sets of exploratory stimuli with different frequencies and current combinations, short durations and moderate intensities are applied to the subject, and the actual change in network efficiency before and after each stimulus is recorded simultaneously. Using these sparse measured data points, the system can construct a continuous or discrete response surface model covering the target parameter range through interpolation or fitting algorithms, thereby obtaining an empirical mathematical description of the stimulus-response characteristics of the subject's brain.

[0047] After obtaining the prediction efficiency bias and the stimulus parameter-network efficiency response surface, the system executes an optimization search process. The goal is to find a set of stimulus parameters within the domains of stimulus frequency and stimulus current amplitude such that the change in prediction efficiency corresponding to these stimulus parameters on the stimulus parameter-network efficiency response surface is as close as possible to the desired prediction efficiency bias. Specifically, this is achieved by minimizing the cost function. To achieve this, the system can use gradient descent, grid search, or more efficient optimization algorithms to solve this optimization problem, and finally output the stimulation frequency and stimulation current amplitude that minimize the cost function, which are the calculated advanced stimulation parameters.

[0048] Because the brain's response to electrical stimulation is non-linear and varies from person to person, constructing an individualized stimulation parameter-network efficiency response surface allows the system to possess a unique stimulus-response characteristic map of the subject's brain. By using the predicted efficiency deviation as a query value and searching for the optimal parameters on this characteristic map in reverse, a model predictive control strategy is achieved. Through these technical features, a reliable conversion from state prediction to precise control commands is realized, ensuring that the advanced stimulation parameters selected by the system are not based on guesswork or group averages, but rather on the optimal intervention plan most likely to produce the desired efficiency change, inferred from the subject's own historical response data. This makes the advanced stimulation commands not only forward-looking in time, but also highly individualized and precisely adapted in intensity and pattern.

[0049] In one embodiment of the present invention, after the step of generating a transcranial alternating current stimulation signal based on the pre-stimulation parameters and applying it to the target brain region, the method further includes: S61, continue to acquire blood oxygen level-dependent signals in real time and calculate the actual global brain network efficiency value generated; S62, compare the actual global brain network efficiency value with the predicted global brain network efficiency value, and calculate the prediction error; For example, the preset threshold is an adaptive dynamic threshold, which is calculated as follows: In the first 10 to 20 sampling periods after the closed-loop control is started, the initial data of the prediction error is collected, the mean μ and standard deviation σ of the initial error are calculated, and the preset threshold is set to μ + 2σ; In the subsequent control process, the mean μ and standard deviation σ are updated every 5 sampling periods, and the threshold size is dynamically adjusted. The physical meaning of this threshold is: when the prediction error exceeds "mean + 2σ", it indicates that the model prediction deviation is significantly deviating from the normal fluctuation range, and parameter fine-tuning or response surface refitting needs to be started to ensure the control accuracy. For example, if the initial error mean μ = 0.02 and standard deviation σ = 0.01, then the preset threshold is 0.04. When the prediction error is greater than 0.04, the update mechanism is triggered. S63, if the prediction error exceeds a preset threshold, then at least one of the following is triggered: The parameters of the nonlinear prediction model are fine-tuned online. Local refitting of the stimulus parameter-network efficiency response surface.

[0050] For example, the specific process of online fine-tuning is as follows: 1) Use the fusion feature vector of the latest sampling period and the actual global brain network efficiency value as new training samples; 2) Set the learning rate to 1e-5~1e-4 (far lower than the learning rate in the offline training stage) to avoid drastic fluctuations in parameters; 3) Use the stochastic gradient descent algorithm to update the hidden layer weights and bias parameters of the multilayer perceptron 1~3 times; 4) After updating, verify the prediction error of the model. If the error drops below the preset threshold, stop fine-tuning; if it is still higher than the threshold, 1~2 more iterations can be added, but the total number of fine-tuning iterations shall not exceed 5 times to prevent the model from overfitting.

[0051] For example, the specific method for local refitting is as follows: 1) Extract the applied stimulus parameters and the corresponding actual network efficiency changes, and label this data pair as high-weight samples (weight coefficient of 2.0, higher than the sample weight of 1.0 in the initial calibration stage); 2) Determine the local parameter interval where the stimulus parameter is located: the frequency interval is [ -1Hz, +1Hz], current amplitude range is [ -0.2mA, +0.2mA]; 3) Call the Gaussian process regression algorithm to refit only the response surface within the local interval, keeping the surface structure outside the interval unchanged; 4) Calculate the prediction error within the local interval after fitting. If the error is less than 0.01, the fitting is valid; otherwise, expand the local interval to [ -2Hz, +2Hz], [ -0.4mA, [+0.4mA], re-execute the fitting process; As described in steps S61-S63 above, in specific implementation, after each application of prediction-based advance stimulation, the system initiates a delayed verification cycle. At the corresponding moment of this cycle, the system synchronously collects the actual blood oxygen level-dependent signal and calculates the actual observed value of the global brain network efficiency at that moment. After obtaining this actual value, the system immediately compares it with the predicted value at the same time previously, and calculates the prediction error. This prediction error is defined as the absolute or relative deviation between the model output and the physical observation, and is a key indicator for quantifying the prediction accuracy of the model under the current operating environment. Through the above steps, the performance monitoring of the model itself can be treated as an endogenous variable of the closed-loop system, focusing not only on whether the goal is achieved, but also on whether the prediction is accurate. This is because, under the advance regulation paradigm, the generation of control commands depends entirely on the output of the prediction model, and the model accuracy directly determines the efficiency and safety of the entire system. A system that only performs feedforward actions without verifying its own prediction reliability has uncontrollable long-term operational risks. When the calculated prediction error exceeds a preset, dynamically adjustable adaptive threshold, the system determines that the current model performance has deviated significantly and then triggers an internal update procedure. This threshold is usually set based on the statistical distribution of historical errors, for example, set to twice the standard deviation of the recent error mean, so as to achieve sensitive detection of abnormal fluctuations.

[0052] The system's update strategy is as follows: If the error primarily stems from short-term changes in state dynamics, online fine-tuning of the nonlinear prediction model is initiated. The system uses the newly generated input-output data pair as a new training sample, updating the model parameters with one or several gradients at an extremely low learning rate. This process is incremental learning, designed to allow for minor drifts in the model parameters to reflect the latest dynamic trends while avoiding catastrophic forgetting. If error analysis suggests a potential fundamental change in the stimulus-response relationship, the system initiates local refitting of the individualized response surface. The system incorporates the applied stimulus parameters and their actual changes in network efficiency as new high-weight data points into the original dataset used to construct the response surface, and calls a surface fitting algorithm to rapidly update the model in the affected region.

[0053] like Figure 2 As shown, the present invention also provides a closed-loop controlled transcranial alternating current stimulation system, comprising: The acquisition and calculation module is used to acquire the blood oxygen level dependent signal of the target brain region in real time, and calculate the global brain network efficiency value at the current moment based on the blood oxygen level dependent signal. The efficiency prediction module is used to obtain the time series of historical global brain network efficiency values ​​and current and historical stimulus parameters, so as to predict the global brain network efficiency value at the next sampling time through a nonlinear prediction model. The comparison calculation module is used to compare the predicted global brain network efficiency with the preset individualized target brain network efficiency and calculate the prediction efficiency deviation. The parameter mapping module is used to query the pre-stored stimulation parameter-network efficiency response surface for the target brain region based on the prediction efficiency deviation, and map to obtain a set of advanced stimulation parameters. The advanced stimulation parameters are used to be issued before the neural activity corresponding to the next sampling time occurs. The synchronization control module is used to generate a transcranial alternating current stimulation signal based on the advanced stimulation parameters and apply it to the target brain region.

[0054] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of a closed-loop controlled transcranial alternating current stimulation method.

[0055] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of a closed-loop controlled transcranial alternating current stimulation method.

[0056] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0057] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A closed-loop controlled transcranial alternating current stimulation method, characterized in that, include: Real-time acquisition of blood oxygen level-dependent signals of the target brain region, and real-time calculation of the global brain network efficiency value at the current moment based on the blood oxygen level-dependent signals; The time series of historical global brain network efficiency values, along with current and historical stimulus parameters, are obtained to predict the global brain network efficiency at the next sampling time using a nonlinear prediction model. Specifically, this includes: Efficiency time series vector consisting of global brain network efficiency values ​​at historical moments is obtained, as well as stimulation history sequence consisting of stimulation parameters at the current moment and historical moments, wherein the number of historical moments corresponding to the stimulation parameters is determined according to hemodynamic delay time, and the stimulation parameters include stimulation frequency and stimulation current amplitude. The stimulus history sequence is input into a temporal convolutional network to extract stimulus history feature vectors; The efficiency time series vector is concatenated with the stimulus history feature vector to obtain a fused feature vector; The fused feature vector is input into a multilayer perceptron, which contains at least one hidden layer, and outputs a prediction of the global brain network efficiency at the next time step. The predicted global brain network efficiency is compared with the preset individualized target brain network efficiency, and the predicted efficiency deviation is calculated. Based on the predicted efficiency deviation, the pre-stored stimulation parameter-network efficiency response surface for the target brain region is queried, and a set of advanced stimulation parameters is obtained by mapping. The advanced stimulation parameters are used to issue before the neural activity corresponding to the next sampling time occurs. A transcranial alternating current stimulation signal is generated based on the aforementioned pre-stimulation parameters and applied to the target brain region.

2. The closed-loop controlled transcranial alternating current stimulation method according to claim 1, characterized in that, The steps of acquiring the blood oxygen level-dependent signal of the target brain region in real time and calculating the global brain network efficiency value at the current moment based on the blood oxygen level-dependent signal include: Based on a preset time window, extract the time series of multiple brain regions in the blood oxygen level dependent signal at the current and nearby times; Based on the time series of each brain region, the Pearson correlation coefficient between every two brain regions is calculated to construct the dynamic functional connectivity matrix at the current moment; The dynamic functional connectivity matrix is ​​thresholded to obtain a binarized brain functional connectivity map. The global brain network efficiency value at the current moment is obtained by calculating the inverse average of the shortest path lengths between all node pairs in the brain functional connectivity graph.

3. The closed-loop controlled transcranial alternating current stimulation method according to claim 1, characterized in that, The temporal convolutional network extracts temporal features from the stimulus history sequence through causal convolution operations, wherein the causal convolution operations satisfy: Its computational output depends only on the stimulus parameters at the current and historical moments; The temporal receptive field length of the convolutional kernel used for feature extraction does not exceed the length of the stimulus history sequence; The weight and bias parameters in the causal convolution operation are obtained by training the nonlinear prediction model with the goal of minimizing the network efficiency prediction error.

4. The closed-loop controlled transcranial alternating current stimulation method according to claim 1, characterized in that, The preset individualized target brain network efficiency is determined through the following steps: Acquire functional magnetic resonance imaging data of subjects in a resting state to calculate their baseline brain network efficiency; Obtain the subjects' clinical diagnosis or cognitive task performance to determine the expected direction and magnitude coefficient of network efficiency modulation; Individualized target brain network efficiency is obtained based on the baseline brain network efficiency and the network efficiency adjustment direction and amplitude coefficients.

5. The closed-loop controlled transcranial alternating current stimulation method according to claim 1, characterized in that, The step of querying the pre-stored stimulation parameter-network efficiency response surface for the target brain region based on the prediction efficiency deviation, and mapping it to obtain a set of advanced stimulation parameters includes: Obtain a pre-constructed stimulation parameter-network efficiency response surface for the target brain region, wherein the stimulation parameter-network efficiency response surface takes stimulation frequency and stimulation current amplitude as input and the induced network efficiency change as output. Using the predicted efficiency deviation as the desired output, the stimulation frequency and stimulation current amplitude that minimize the cost function are sought on the stimulation parameter-network efficiency response surface. The found stimulation frequency and stimulation current amplitude are used as the advanced stimulation parameters.

6. The closed-loop controlled transcranial alternating current stimulation method according to claim 1, characterized in that, After the step of generating a transcranial alternating current stimulation signal based on the pre-stimulation parameters and applying it to the target brain region, the method further includes: Continue to acquire blood oxygen level-dependent signals in real time and calculate the actual global brain network efficiency value generated; The actual global brain network efficiency value is compared with the predicted global brain network efficiency value, and the prediction error is calculated. If the prediction error exceeds a preset threshold, at least one of the following will be triggered: The parameters of the nonlinear prediction model are fine-tuned online. Local refitting of the stimulus parameter-network efficiency response surface.

7. A closed-loop controlled transcranial alternating current stimulation system, characterized in that, include: The acquisition and calculation module is used to acquire the blood oxygen level dependent signal of the target brain region in real time, and calculate the global brain network efficiency value at the current moment based on the blood oxygen level dependent signal. An efficiency prediction module is used to acquire time series of historical global brain network efficiency values ​​and current and historical stimulation parameters to predict the global brain network efficiency at the next sampling time using a nonlinear prediction model. Specifically, this includes: acquiring an efficiency time series vector composed of global brain network efficiency values ​​from historical times; and acquiring a stimulation history sequence composed of stimulation parameters from the current and historical times, wherein the number of historical times corresponding to the stimulation parameters is determined based on hemodynamic delay time, and the stimulation parameters include stimulation frequency and stimulation current amplitude; inputting the stimulation history sequence into a temporal convolutional network to extract a stimulation history feature vector; concatenating the efficiency time series vector with the stimulation history feature vector to obtain a fused feature vector; and inputting the fused feature vector into a multilayer perceptron, which contains at least one hidden layer, to output a predicted global brain network efficiency value for the next time step. The comparison calculation module is used to compare the predicted global brain network efficiency with the preset individualized target brain network efficiency and calculate the prediction efficiency deviation. The parameter mapping module is used to query the pre-stored stimulation parameter-network efficiency response surface for the target brain region based on the prediction efficiency deviation, and map to obtain a set of advanced stimulation parameters. The advanced stimulation parameters are used to be issued before the neural activity corresponding to the next sampling time occurs. The synchronization control module is used to generate a transcranial alternating current stimulation signal based on the advanced stimulation parameters and apply it to the target brain region.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images