Children magnetoencephalogram personalized stimulation normal form optimization method based on reinforcement learning

Abstract

The invention provides a children magnetoencephalogram personalized stimulation normal form optimization method based on reinforcement learning, and the method comprises the steps: S1, continuously collecting multi-channel brain magnetic signals of a to-be-detected child through an optical pump magnetometer array, and synchronously collecting behavior data and physiological data; s2, brain magnetic signal features, behavior features and physiological features are extracted and fused to construct a real-time state feature vector; s3, inputting the real-time state feature vector into a reinforcement learning model to obtain a predicted Q value of a stimulation adjustment action, and dynamically adjusting parameters of a magnetoencephalogram stimulation normal form of the to-be-tested child based on the predicted Q value; and S4, acquiring new multichannel brain magnetic signals, behavior data and physiological data of the to-be-tested child based on the parameter-adjusted brain magnetic map stimulation normal form, and calculating to obtain a comprehensive reward value to update the reinforcement learning model. The method has the advantages that the magnetoencephalogram stimulation normal form can be intelligently and dynamically adjusted, and the multi-channel magnetoencephalogram signal quality of the child to be tested is improved.

Description

Technical Field

[0001] This invention relates to the technical field of biomedical engineering, and more specifically, to a method for optimizing personalized stimulation paradigms based on reinforcement learning in children's magnetoencephalography (MEG). Background Technology

[0002] Magnetoencephalography (MEG) is a non-invasive neurophysiological technique that reflects brain function by measuring the weak magnetic fields generated by the activity of neurons in the brain. Compared to electroencephalography (EEG), MEG is not affected by the conductivity of the skull and scalp, has higher spatial resolution, and maintains millisecond-level temporal resolution, giving it unique advantages in brain function research and clinical diagnosis, especially in the localization of epileptic foci.

[0003] Epilepsy is a common neurological disorder in childhood, with a high incidence rate. Some children suffer from drug-resistant epilepsy, requiring surgical removal of the epileptogenic focus to control the condition. Precise localization of the epileptogenic focus is crucial for the success of epilepsy surgery. Magnetoencephalography (MEG) in children, due to its high spatiotemporal resolution, can effectively capture the origin of epileptiform discharges and is widely used in preoperative evaluation for epilepsy.

[0004] In recent years, with the development of optically pumped magnetometers (OPM), not only have their sensitivity and portability achieved great breakthroughs, but their cost and maintenance costs are also lower. This makes up for the shortcomings of traditional magnetoencephalography (MEG) devices based on superconducting quantum interference devices, such as high operating costs and the long distance between the probe and the head of the subject. Its wearable feature also allows children to make a certain degree of head movement during MEG scanning, improving children's compliance and comfort.

[0005] While OPM-MEG offers new opportunities for the diagnosis of childhood epilepsy, several challenges remain. These challenges include fully utilizing OPM-MEG's advantages to address data quality issues arising from individual differences during magnetoencephalography (MEG) data acquisition, such as short attention spans, easy distractibility, and poor cooperation. Furthermore, dynamic optimization of the stimulation paradigm to maximize effective signals and minimize artifacts is also crucial. Existing MEG stimulation paradigms are mostly fixed and preset, failing to adapt to significant individual differences in children's cognitive abilities, attention spans, emotional states, and brain development stages. This results in poor data quality, low experimental efficiency, and an inability to achieve truly personalized assessments. Therefore, a method that can intelligently and dynamically adjust MEG stimulation paradigms is urgently needed to improve the quality of children's MEG signals, thereby enhancing the accuracy and efficiency of epileptic focus localization. Summary of the Invention

[0006] The technical problem this invention aims to solve is how to intelligently and dynamically adjust the magnetoencephalography (MEG) stimulation paradigm to improve the quality of multi-channel MEG signals in children under testing. To overcome the shortcomings of the existing technology (or related technology), this invention provides a method for optimizing personalized stimulation paradigms for children's MEG based on reinforcement learning.

[0007] This invention provides a method for optimizing personalized stimulation paradigms based on reinforcement learning for children's magnetoencephalography (MEG). The method involves having the child wear an optically pumped magnetometer array and observing alternating images of positive, neutral, and negative emotions on a screen. After each emotion image is displayed, the child presses a corresponding button to categorize it, thus performing an emotion image recognition task. The method includes the following steps: Step S1: During the process of the child performing the emotion picture recognition task, the optical pump magnetometer array continuously collects the multi-channel brain magnetic signals of the child and simultaneously collects the child's behavioral and physiological data. Step S2: Extract the magnetic brain signal features, behavioral features, and physiological features from the multi-channel magnetic brain signal, the behavioral data, and the physiological data, respectively, and fuse them to construct a real-time state feature vector; Step S3: Input the real-time state feature vector into the pre-built reinforcement learning model to obtain the predicted Q value of the stimulus adjustment action, and dynamically adjust the parameters of the magnetoencephalogram stimulation paradigm of the child to be tested based on the predicted Q value of the stimulus adjustment action. Step S4: Based on the adjusted magnetoencephalography stimulation paradigm, new multi-channel magnetoencephalography signals, behavioral data, and physiological data of the child to be tested are collected, and a comprehensive reward value is calculated to update the reinforcement learning model. Then, the process returns to step S2.

[0008] Compared with existing technologies, the personalized stimulation paradigm optimization method based on reinforcement learning for children's magnetoencephalography (MEG) has the following advantages: In this invention, multi-channel magnetoencephalography (MEG) signals, behavioral data, and physiological data are acquired in step S1; feature extraction and real-time state feature vectors are constructed in step S2; parameters of the MEG stimulation paradigm are adjusted in step S3; and data re-acquisition and reinforcement learning model updates are performed in step S4. Throughout the process, the Q-value prediction of the real-time state feature vector composed of MEG signal features, behavioral features, and physiological features is analyzed based on the reinforcement learning model. By adjusting the MEG stimulation paradigm in real time and in a personalized manner, artifacts caused by non-cooperation or distraction of the child being tested are reduced, thereby obtaining higher quality new multi-channel MEG signals. Finally, the reinforcement learning model is iteratively updated by combining new behavioral and physiological data to achieve higher prediction Q-value accuracy. The quality of multi-channel MEG signals is improved through continuous iterative optimization and adjustment of the MEG stimulation paradigm.

[0009] In one possible implementation, the optically pumped magnetometer array includes multiple extremely weak magnetic measurement sensors arranged in an array based on the spin-free exchange relaxation effect. In step S1, each of the extremely weak magnetic measurement sensors is controlled to continuously acquire the multi-channel magnetoencephalogram (MEG) signals of the child to be tested.

[0010] Compared with existing technologies, the above-mentioned technical solution can utilize the characteristics of extremely high magnetic field sensitivity without spin exchange relaxation effect, and is suitable for weak magnetic field detection of multi-channel brain magnetic signals of children to be tested, which can improve the accuracy of multi-channel brain magnetic signal acquisition.

[0011] In one possible implementation, in step S2, the multi-channel magnetoencephalography (MEG) signal is subjected to 0.1-100Hz bandpass filtering, independent component analysis for artifact removal, and head motion compensation to obtain the MEG signal-to-noise ratio and the number of artifacts as MEG signal features. The task accuracy, root mean square value of head three-dimensional displacement, emotional positivity score, and heart rate variability coefficient of the emotion picture cognition task are extracted from the behavioral data and the physiological data as the behavioral features and the physiological features.

[0012] Compared with existing technologies, the above-mentioned technical solution can effectively remove noise and artifacts through independent component analysis artifact removal and motion compensation technology, significantly improving the purity of the EEG signal features; at the same time, it integrates signal quality (EMG signal-to-noise ratio, number of artifacts), behavioral performance (task accuracy), and physiological state (emotional positivity score, heart rate variability coefficient) as a multi-dimensional state representation, comprehensively reflecting the real-time state of the child being tested.

[0013] In one possible implementation, in step S3, a deep Q-network is used as the reinforcement learning model. The deep Q-network includes an input layer, multiple hidden layers and an output layer connected in sequence. The input layer receives the real-time state feature vector, and the output layer outputs the predicted Q value of each stimulus adjustment action in the discrete action space.

[0014] Compared with existing technologies, the above technical solution can introduce a deep Q-network as a reinforcement learning model, which is suitable for complex multimodal feature inputs and can effectively learn the state-action mapping relationship. By predicting the Q value, the optimal stimulus is selected to adjust the action, thereby achieving dynamic adaptive optimization of the magnetoencephalogram stimulation paradigm.

[0015] In one possible implementation, each hidden layer of the deep Q-network uses a modified linear unit as the activation function, the network is trained using the Adam optimizer with a learning rate of 0.001 and a discount factor γ = 0.95, and an experience replay mechanism is used with an experience replay pool size of 10,000, a batch size of 64, and a target network update frequency of 100 steps.

[0016] Compared with existing technologies, the above-mentioned technical solution can improve training efficiency by reasonably setting hyperparameters such as learning rate, discount factor, and batch size.

[0017] In one possible implementation, the parameters of the magnetoencephalogram stimulation paradigm in step S3 include stimulation intensity, stimulation duration, stimulation presentation interval, task difficulty, and rest interval. The stimulation adjustment actions in the discrete action space include increasing stimulation intensity, decreasing stimulation intensity, shortening stimulation duration, extending stimulation duration, shortening stimulation presentation interval, extending stimulation presentation interval, increasing rest interval, shortening rest interval, pausing experiment, and ending experiment.

[0018] Compared with existing technologies, the above-mentioned technical solution clarifies the parameters of the magnetoencephalography (MEG) stimulation paradigm and the specific content of the stimulation adjustment actions, enabling flexible adjustment of stimulation attributes to adapt to different experimental scenarios.

[0019] In one possible implementation, the multi-channel magnetoencephalogram (MEG) signal includes the MEG signal-to-noise ratio and the number of artifacts, the behavioral data is the task accuracy, the physiological data is the emotional positivity score, and in step S4, the comprehensive reward value is obtained using the following formula: ; in, This represents the overall reward value; Indicates the first weighting coefficient; This represents a reward value that is positively correlated with the signal-to-noise ratio of the brain magnetic signal; This represents the second weighting coefficient; This represents a penalty value that is positively correlated with the number of artifacts. This represents the third weighting coefficient; This represents a reward value that is positively correlated with the accuracy of the task. This represents the fourth weighting coefficient; The reward value represents the positive correlation with the emotional positivity score.

[0020] Compared with existing technologies, the above technical solution can effectively guide the reinforcement learning model to learn in the direction of improving data quality and children's experience, balance multiple optimization objectives such as signal quality, task performance, and children's cooperation, and optimize the focus for different experimental objectives by adjusting four weight coefficients.

[0021] In one possible implementation, the step of updating the reinforcement learning model in step S4 includes: Step A1: Calculate the comprehensive reward value for the current time step, and integrate the complete interaction information of the current time step into an experience tuple. And store it in the experience replay pool, where, Indicates the execution of stimulus adjustment actions The previously mentioned real-time state feature vector, This refers to the stimulus adjustment action selected and executed by the reinforcement learning model. Indicates the execution of stimulus adjustment actions The new comprehensive reward value, Indicates the execution of stimulus adjustment actions Then, the new real-time state feature vector of the child to be tested; Step A2: When the amount of data in the experience replay pool reaches a preset threshold, a small batch of experience tuples is randomly extracted from the experience replay pool for the reinforcement learning model to learn. The loss function is defined by calculating the difference between the target Q value and the predicted Q value. Then, the Adam optimizer is used to calculate the gradient of the loss function with respect to the intrinsic parameters of the reinforcement learning model through the backpropagation algorithm, and the weights of the reinforcement learning model are updated along the gradient descent direction so that the predicted Q value approaches the target Q value.

[0022] Compared with existing technologies, the above technical solution describes the update mechanism of reinforcement learning models, uses an experience replay mechanism to break data correlation, improve sample utilization and training stability, and forms a closed-loop optimization of "state-action-reward-update" to achieve continuous self-improvement. Attached Figure Description

[0023] Figure 1 This is a flowchart of the steps of the present invention; Figure 2 This is a flowchart of the optimized magnetoencephalography stimulation paradigm of the present invention; Figure 3 This is a schematic diagram illustrating the interaction between the reinforcement learning model of the present invention and the environment. Detailed Implementation

[0024] First, those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention. Those skilled in the art can make adjustments as needed to adapt to specific application scenarios.

[0025] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0026] See Figure 1 This invention discloses a method for optimizing personalized stimulation paradigms based on reinforcement learning for children's magnetoencephalography (MEG). The method involves having the child wear an optically pumped magnetometer array and observing alternating images of positive, neutral, and negative emotions on a screen. After each emotion image is displayed, the child presses a corresponding button to categorize it, thus performing an emotion image recognition task. The method includes the following steps: Step S1: During the process of the child performing the emotion picture recognition task, the multi-channel magnetoencephalogram (MEG) signal of the child is continuously collected using an optically pumped magnetometer array, and the behavioral and physiological data of the child are collected simultaneously. Step S2: Extract the brain magnetic signal features, behavioral features, and physiological features from the multi-channel brain magnetic signals, behavioral data, and physiological data, respectively, and fuse them to construct a real-time state feature vector; Step S3: Input the real-time state feature vector into the pre-built reinforcement learning model to obtain the predicted Q value of the stimulus adjustment action, and dynamically adjust the parameters of the magnetoencephalogram stimulation paradigm of the child under test based on the predicted Q value of the stimulus adjustment action. Step S4: Based on the parameter-adjusted magnetoencephalography stimulation paradigm, new multi-channel magnetoencephalography signals, behavioral data, and physiological data of the children to be tested are collected, and the comprehensive reward value is calculated to update the reinforcement learning model. Then, the process returns to step S2.

[0027] This embodiment details how to apply the method of the present invention to dynamically optimize personalized stimulation paradigms based on the collection of OPM-MEG data of children under test when performing emotional picture recognition tasks; In this embodiment, when the children participate in the emotion image recognition task, the task requires them to view three categories of emotion images—positive, neutral, and negative—alternating on the screen, and to press the corresponding button to classify each emotion image after it is displayed. See [link to relevant documentation]. Figure 2 and Figure 3 The specific implementation process is as follows: Experimental setup and data acquisition Multi-channel magnetoencephalography (MEG) signal acquisition: An optically pumped magnetometer (OPM) array, consisting of multiple extremely weak magnetic measurement sensors based on spin-free exchange relaxation, was used to continuously acquire multi-channel MEG signals from children during the task. Simultaneous collection of behavioral and physiological data: Simultaneous collection of task response data (key accuracy and reaction time), head movement data, facial expression data, and physiological data of the children under test; Head motion data was acquired using a triaxial accelerometer with a sampling frequency of 200Hz and a measurement range of ±10° rotation angle and ±5cm displacement; facial expression data was acquired using a camera and the emotional state was calculated using the MTCNN+FER-2013 model. Construction of real-time state feature vectors The original multi-channel magnetoencephalogram (MEG) signals were subjected to a 0.1-100Hz fourth-order Butterworth bandpass filter to remove low-frequency drift and high-frequency noise. The FastICA algorithm was used for independent component analysis to automatically identify and remove physiological artifacts such as eye movements, blinks, and heartbeats. At the same time, the signal spatial projection (SSP) algorithm was used in combination with head motion sensor data to correct motion artifacts. The study extracts features from the magnetic field signals (MFS) and behavioral / physiological features. Specifically, the MFS features include the signal-to-noise ratio (SNR) and the number of artifacts in the MFS signal, which are derived from the event-related magnetic field induced by the image stimulus. Behavioral / physiological features are used to calculate the real-time accuracy and reaction time of the emotional image recognition task. A facial expression recognition algorithm is used to analyze the child's emotional state in real time and quantify it into an emotional positivity score to assess the impact of stimuli on the child's emotions. The root mean square (RMS) value of the three-dimensional head displacement is calculated to reflect the intensity of head movement; a smaller RMS value indicates higher cooperation and head stability in the child. The heart rate variability (HRV) coefficient is calculated to reflect the balance of the autonomic nervous system, which is related to the child's stress and emotional state. These features are standardized and fused into a real-time state feature vector (State) as input to the reinforcement learning model. Action selection and execution based on DQN In this embodiment, a Deep Q-Network (DQN) is used as the reinforcement learning model, and its specific network structure and training parameter settings are as follows: This deep Q-network consists of one input layer, three hidden layers, and one output layer. The dimension of the input layer matches the dimension of the state feature vector. The hidden layers consist of three fully connected layers with 512, 256, and 128 neurons, respectively, all using the ReLU activation function to increase the non-linear expressive power of the deep Q-network. The number of neurons in the output layer is the same as the number of stimulus-adjustment actions in the discrete action space, and it outputs the predicted Q-value corresponding to each stimulus-adjustment action. The deep Q-network is trained using the Adam optimizer with a learning rate of 0.001, β1=0.9, and β2=0.999. To break data correlation and stabilize the training process, the deep Q-network adopts an experience replay mechanism with a replay pool size of 10,000. Each time, 64 samples (batch size) are randomly selected from the replay pool for training and updating, and the update is performed every 100 steps. The input to this reinforcement learning model is the real-time state feature vector constructed in step S2. The discrete action space is designed as a series of discrete adjustment operations on the magnetoencephalogram (MEG) stimulation paradigm, mainly including: (1) Extend / shorten the display time of the next emotion image; (2) Extend / shorten the interval between the presentation of emotional images; (3) Add a short break; (4) Adjust the emotion type of the next emotional picture (for example, when the child is showing a negative emotion, prioritize showing a positive or neutral emotional picture). The reinforcement learning model outputs the predicted Q value of each possible stimulus adjustment action based on the current real-time state feature vector, and selects the stimulus adjustment action with the highest predicted Q value to adjust the emotional picture stimulus parameters in the next round. Reward Calculation and Model Update After the reinforcement learning model performs a stimulus adjustment action, a comprehensive reward value is calculated based on newly acquired multi-channel magnetoencephalogram (MEG) signals, behavioral data, and physiological data. This comprehensive reward value is used to evaluate the quality of the stimulus adjustment action performed. The formula for the comprehensive reward value is as follows: ; in, This represents the overall reward value; Indicates the first weighting coefficient; The reward value represents the positive correlation between the signal-to-noise ratio of the brain magnetic signal; This represents the second weighting coefficient; This represents a penalty value that is positively correlated with the number of artifacts. This represents the third weighting coefficient; This represents a reward value that is positively correlated with task accuracy. This represents the fourth weighting coefficient; This represents a reward value that is positively correlated with the emotional positivity score; Specifically, SNR_reward = (SNR_current-10) / 10, with the baseline SNR set to 10dB; Artifact_penalty = Artifact_rate / 0.2, with the baseline artifact rate set to 20%; Task_Completion_Rate = 0.7 ×(N_correct / N_total) + 0.3 × (1 - T_response / 2), where N_correct is the number of times the corresponding correct button was pressed, N_total is the total number of button presses, and T_response is the reaction time in seconds; Compliance_reward represents the child's cooperation score. After calculating the comprehensive reward value Reward for the current time step, the complete interaction information for the current time step is integrated into an experience tuple and stored in the experience replay pool. Indicates the execution of stimulus adjustment actions The previous real-time state feature vector, This represents the stimulus adjustment action selected and executed by the reinforcement learning model. Indicates the execution of stimulus adjustment actions The new overall reward value, Indicates the execution of stimulus adjustment actions Then, the new real-time state feature vector of the child to be tested; Once the amount of data in the experience replay pool reaches a certain size, the reinforcement learning model will randomly select a mini-batch of experience tuples from the experience replay pool for learning. The formula for calculating the target Q-value is: ; The predicted Q-value is the real-time state feature vector. Perform stimulus adjustment actions The Q value at that time, i.e. The loss function is defined by calculating the difference between the target Q-value and the predicted Q-value. ; Subsequently, the reinforcement learning model employs the Adam optimizer, calculates the gradient of the loss function relative to the parameters of the main network within the reinforcement learning model through the backpropagation algorithm, and updates the weights of the main network along the direction of gradient descent, thereby gradually bringing the predicted Q-value closer to the target Q-value.

[0028] To further increase the stability of training, the parameters of the main network are not updated at every step. Instead, the parameters of the main network are completely copied to the target network every fixed number of steps. This delayed update mechanism can effectively prevent oscillations and divergences during training.

[0029] This embodiment illustrates how the method proposed in this invention can be applied to children with epilepsy to personalize and optimize the photic stimulation paradigm during magnetoencephalography (MEG) examination to induce epileptiform discharges. This scenario corresponds to the application of this invention in clinical diagnosis, especially in the localization of epileptic lesions. In the specific implementation process, the first two steps are the same. In the third step, the input of the reinforcement learning model is the real-time state feature vector constructed in step S2. The discrete action space is designed as a series of discrete adjustment operations on the magnetoencephalogram stimulation paradigm, aiming to dynamically adapt to the real-time state of children with epilepsy, improve the induction efficiency of epileptiform discharges and data quality, mainly including: (1) Increase / decrease the brightness level of the flashing light stimulation; (2) Increase / decrease the frequency of flashing light stimulation; (3) Extend / shorten the duration of a single flash stimulus; (4) Insert a non-stimulating rest interval; The reinforcement learning model outputs the predicted Q value of each possible stimulus adjustment action based on the current real-time state feature vector, and selects the stimulus adjustment action with the highest predicted Q value to adjust the image stimulus parameters in the next round. After the reinforcement learning model performs a stimulus adjustment action, a comprehensive reward value is calculated based on newly acquired multi-channel magnetoencephalogram (MEG) signals, behavioral data, and physiological data. This comprehensive reward value is used to evaluate the quality of the stimulus adjustment action performed. The formula for the comprehensive reward value is as follows: ; Since the flashing stimulus paradigm lacks behavioral feedback, the formula for the comprehensive reward value does not include task completion-related components. Here, SNR_reward = (SNR_current - 10) / 10, with a baseline SNR set to 10 dB; Artifact_penalty = Artifact_rate / 0.2, with a baseline artifact rate set to 20%; Compliance_reward represents the child's cooperation score. After calculating the comprehensive reward value (Reward) for the current time step, the complete interaction information for that time step is integrated into an experience tuple. The data is stored in the experience replay pool. When the amount of data in the experience replay pool reaches a certain scale, a small batch of experience tuples is randomly selected from the pool for learning. The subsequent steps are the same as in Example 1.

[0030] In the description of this invention, the references to "one embodiment," "some embodiments," "in this embodiment," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0031] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for optimizing personalized stimulation paradigms based on reinforcement learning in children's magnetoencephalography (MEG), characterized in that, The children to be tested are first fitted with an optically pumped magnetometer array, and then observe alternating images of positive, neutral, and negative emotions on a screen. After each emotion image is displayed, the children press the corresponding button to classify it, thus performing an emotion image recognition task. The personalized stimulation paradigm optimization method for children's magnetoencephalography includes the following steps: Step S1: During the process of the child performing the emotion picture recognition task, the optical pump magnetometer array continuously collects the multi-channel brain magnetic signals of the child and simultaneously collects the child's behavioral and physiological data. Step S2: Extract the magnetic brain signal features, behavioral features, and physiological features from the multi-channel magnetic brain signal, the behavioral data, and the physiological data, respectively, and fuse them to construct a real-time state feature vector; Step S3: Input the real-time state feature vector into the pre-built reinforcement learning model to obtain the predicted Q value of the stimulus adjustment action, and dynamically adjust the parameters of the magnetoencephalogram stimulation paradigm of the child to be tested based on the predicted Q value of the stimulus adjustment action. Step S4: Based on the adjusted magnetoencephalography stimulation paradigm, new multi-channel magnetoencephalography signals, behavioral data, and physiological data of the child to be tested are collected, and a comprehensive reward value is calculated to update the reinforcement learning model. Then, the process returns to step S2.

2. The method for optimizing personalized stimulation paradigms of children's magnetoencephalography based on reinforcement learning according to claim 1, characterized in that, The optically pumped magnetometer array includes multiple extremely weak magnetic measurement sensors arranged in an array based on the spin-free exchange relaxation effect. In step S1, each of the extremely weak magnetic measurement sensors is controlled to continuously acquire the multi-channel magnetoencephalogram (MEG) signals of the child to be tested.

3. The method for optimizing personalized stimulation paradigms of children's magnetoencephalography based on reinforcement learning according to claim 1, characterized in that, In step S2, the multi-channel magnetoencephalography (MEG) signal is subjected to 0.1-100Hz bandpass filtering, independent component analysis for artifact removal, and head motion compensation to obtain the MEG signal-to-noise ratio and the number of artifacts as MEG signal features. The task accuracy, root mean square value of head three-dimensional displacement, emotional positivity score, and heart rate variability coefficient of the emotion picture cognition task are extracted from the behavioral data and the physiological data as the behavioral features and the physiological features.

4. The method for optimizing personalized stimulation paradigms of children's magnetoencephalography based on reinforcement learning according to claim 1, characterized in that, In step S3, a deep Q-network is used as the reinforcement learning model. The deep Q-network includes an input layer, multiple hidden layers and an output layer connected in sequence. The input layer receives the real-time state feature vector, and the output layer outputs the predicted Q value of each stimulus adjustment action in the discrete action space.

5. The method for optimizing personalized stimulation paradigms of children's magnetoencephalography based on reinforcement learning according to claim 4, characterized in that, The hidden layers of the deep Q-network use modified linear units as activation functions. The network is trained using the Adam optimizer with a learning rate of 0.001, a discount factor γ of 0.95, and an experience replay mechanism. The experience replay pool size is 10,000, the batch size is 64, and the target network update frequency is 100 steps.

6. The method for optimizing personalized stimulation paradigms of children's magnetoencephalography based on reinforcement learning according to claim 4, characterized in that, The parameters of the magnetoencephalogram stimulation paradigm in step S3 include stimulation intensity, stimulation duration, stimulation presentation interval, task difficulty, and rest interval. The stimulation adjustment actions in the discrete action space include increasing stimulation intensity, decreasing stimulation intensity, shortening stimulation duration, extending stimulation duration, shortening stimulation presentation interval, extending stimulation presentation interval, increasing rest interval, shortening rest interval, pausing experiment, and ending experiment.

7. The method for optimizing personalized stimulation paradigms of children's magnetoencephalography based on reinforcement learning according to claim 1, characterized in that, The multi-channel magnetoencephalogram (MEG) signal includes the MEG signal-to-noise ratio and the number of artifacts; the behavioral data is the task accuracy; the physiological data is the emotional positivity score; and in step S4, the comprehensive reward value is obtained using the following formula: ； in, This represents the overall reward value; Indicates the first weighting coefficient; This represents a reward value that is positively correlated with the signal-to-noise ratio of the brain magnetic signal; This represents the second weighting coefficient; This represents a penalty value that is positively correlated with the number of artifacts. Indicates the third weighting coefficient; This represents a reward value that is positively correlated with the accuracy of the task. This represents the fourth weighting coefficient; The reward value represents the positive correlation with the emotional positivity score.

8. The method for optimizing personalized stimulation paradigms of children's magnetoencephalography based on reinforcement learning according to claim 1, characterized in that, The step of updating the reinforcement learning model in step S4 includes: Step A1: Calculate the comprehensive reward value for the current time step, and integrate the complete interaction information of the current time step into an experience tuple. And store it in the experience replay pool, where, Indicates the execution of stimulus adjustment actions The previously mentioned real-time state feature vector, This refers to the stimulus adjustment action selected and executed by the reinforcement learning model. Indicates the execution of stimulus adjustment actions The new comprehensive reward value, Indicates the execution of stimulus adjustment actions Then, the new real-time state feature vector of the child to be tested; Step A2: When the amount of data in the experience replay pool reaches a preset threshold, a small batch of experience tuples is randomly extracted from the experience replay pool for the reinforcement learning model to learn. The loss function is defined by calculating the difference between the target Q value and the predicted Q value. Then, the Adam optimizer is used to calculate the gradient of the loss function with respect to the intrinsic parameters of the reinforcement learning model through the backpropagation algorithm, and the weights of the reinforcement learning model are updated along the gradient descent direction so that the predicted Q value approaches the target Q value.

Images