Personalized attention rehabilitation training system for adhd children

Abstract

This invention relates to the field of medical signal processing technology, specifically to a personalized attention rehabilitation training system for children with ADHD. The system includes: a data acquisition module that acquires EEG signals and interactive behavior data of the target child during an interference-free benchmark testing phase; a benchmark determination module that generates a benchmark feature probability distribution matrix using a graph network algorithm based on the EEG power ratio characteristics of the EEG signals and the interactive response variability characteristics of the interactive behavior data; a noise reduction module that acquires a real-time grid jump frequency matrix during the target child's rehabilitation training phase and inputs it into the graph network algorithm engine to obtain the real-time feature probability distribution matrix; and an intervention control module that determines the feature distribution difference divergence index between the real-time feature probability distribution matrix and the benchmark feature probability distribution matrix, and then issues corresponding instructions. This invention achieves adaptive closed-loop regulation based on individual differences, ensuring a stable and reliable rehabilitation intervention for the attention state of children with ADHD.

Description

Technical Field

[0001] This invention relates to the field of medical signal processing technology, and more specifically to a personalized attention rehabilitation training system for children with ADHD. Background Technology

[0002] In neurofeedback rehabilitation training for attention deficit hyperactivity disorder (ADHD), the system typically uses portable EEG devices to collect the EEG signals of the children being tested, and combines this with key press response features extracted from a smart terminal. Based on these physiological and behavioral characteristics, the system assesses the children's attentional state in real time, and then dynamically adjusts the task difficulty of the rehabilitation software.

[0003] In practical applications, children with ADHD inevitably exhibit brief hyperactive behaviors such as instantaneous teeth clenching or head turning. These hyperactive behaviors trigger high-frequency electromyographic interference signals, causing abnormal sudden jumps in the feature values ​​extracted by the system within a very short period of time. Existing processing methods for such transient abnormal data suffer from two main engineering flaws: if a fixed numerical threshold is used for truncation and removal, the system frequently discards original sampling frames containing valid information, leading to data input queue breaks and consequently causing high-frequency misjudgments and oscillating jumps in system control commands; if a sliding time window mean filtering is used for smoothing, the system needs to rely on the accumulation of data over a relatively long period, which introduces computational delays, causing the system's intervention actions to lag behind the actual mental state of the children, resulting in poor regulatory effects of rehabilitation training. Summary of the Invention

[0004] To address the problem that existing methods cannot achieve stable and reliable closed-loop regulation of rehabilitation training based on the individual physiological differences of children with ADHD, the present invention aims to provide a personalized attention rehabilitation training system for children with ADHD. The specific technical solution adopted is as follows: This invention provides a personalized attention rehabilitation training system for children with ADHD, the system comprising: The data acquisition module is used to acquire EEG signals and interactive behavior data of the target children during an interference-free benchmark testing phase; The benchmark determination module is used to construct a benchmark grid jump frequency matrix based on the EEG power ratio characteristics presented by EEG signals and the interaction response variability characteristics presented by interactive behavior data; and to iteratively generate a benchmark feature probability distribution matrix based on the benchmark grid jump frequency matrix using a graph network algorithm. The noise reduction module is used to obtain the real-time grid jump frequency matrix during the rehabilitation training phase of the target child; the real-time grid jump frequency matrix is ​​input into the graph network algorithm engine to obtain the real-time feature probability distribution matrix after filtering out high-frequency transient interference; The intervention control module is used to determine the feature distribution difference divergence index of the real-time feature probability distribution matrix relative to the benchmark feature probability distribution matrix; and to issue corresponding instructions based on the feature distribution difference divergence index.

[0005] Preferably, before constructing the baseline grid jump frequency matrix, the following steps are also included: The absolute coordinate boundaries of the two-dimensional coordinate system are determined based on the maximum value of the EEG power ratio presented in the historical data of the target children and the maximum value of the interaction response variability feature presented in the interaction behavior data.

[0006] Preferably, the construction of the baseline grid jump frequency matrix based on the EEG power ratio characteristics presented by EEG signals and the interaction response variability characteristics presented by interaction behavior data includes: During the benchmark testing phase, instantaneous EEG power ratio features and interactive response variability features are extracted at fixed time steps, and two-dimensional feature coordinate points are determined based on the instantaneous EEG power ratio features and interactive response variability features. The two-dimensional feature coordinate points at each time step are mapped to a two-dimensional feature grid to obtain a reference grid node jump sequence composed of continuous grid block indices; Based on the grid block index pairs of adjacent time steps in the reference grid node jump sequence, a reference grid jump frequency matrix is ​​constructed to record the number of directed jumps.

[0007] Preferably, the step of iteratively generating the baseline feature probability distribution matrix based on the baseline grid jump frequency matrix using a graph network algorithm includes: The reference grid jump frequency matrix is ​​input into the graph network algorithm engine, the damping coefficient is set, and the steady-state probability value of each grid block is output through iterative calculation of the graph network algorithm. The steady-state probability values ​​of each grid block are normalized to obtain the baseline characteristic probability distribution matrix.

[0008] Preferably, after generating the baseline feature probability distribution matrix, the method further includes: During the benchmark testing phase, local data is continuously extracted through a sliding window of preset length to generate multiple continuous local feature probability distribution matrices. Calculate the local KL divergence value of each of the local feature probability distribution matrices relative to the baseline feature probability distribution matrix, and all the local KL divergence values ​​constitute the baseline divergence sequence; Based on the mean and standard deviation of all data in the benchmark divergence sequence, the tolerance threshold for individual differences is determined.

[0009] Preferably, obtaining the real-time grid jump frequency matrix includes: During the rehabilitation training phase, a real-time analysis window of preset length is used to slide with a fixed step size; the real-time two-dimensional feature coordinate points of each time step within each real-time analysis window are extracted. The real-time two-dimensional feature coordinate points are mapped to the two-dimensional feature grid. If the coordinate points exceed the absolute coordinate boundary, they are forcibly constrained to the boundary value, and a real-time grid node jump sequence is generated. Based on the grid block index pairs of adjacent time steps in the real-time grid node jump sequence, a real-time grid jump frequency matrix is ​​constructed.

[0010] Preferably, the step of inputting the real-time grid jump frequency matrix into the graph network algorithm engine to obtain the real-time feature probability distribution matrix after filtering out high-frequency transient interference includes: The real-time grid jump frequency matrix is ​​input into the graph network algorithm engine, and the damping coefficient is set. Through iterative calculation of the graph network algorithm, the steady-state probability value of each grid block is output. The steady-state probability values ​​of each grid block are normalized to generate the real-time feature probability distribution matrix.

[0011] Preferably, determining the feature distribution difference divergence index of the real-time feature probability distribution matrix relative to the benchmark feature probability distribution matrix includes: The real-time feature probability distribution matrix is ​​smoothed by an exponentially weighted moving average to obtain a smoothed real-time feature probability distribution matrix. Calculate the divergence value of the smoothed real-time feature probability distribution matrix relative to the baseline feature probability distribution matrix, and use the divergence value as the feature distribution difference divergence index.

[0012] Preferably, issuing corresponding instructions based on the feature distribution difference divergence index includes: When the number of times the real-time acquired feature distribution difference divergence index exceeds the individual difference tolerance threshold continuously reaches a preset number, the target child is determined to have entered the cognitive state continuous over-limit stage. When the target child enters a stage of sustained cognitive overload and is not within the instruction cooling-off period, issue an instruction to reduce the difficulty of the task.

[0013] Preferably, the acquisition of the EEG power ratio feature includes: converting the EEG signal into a frequency domain signal using a fast Fourier transform algorithm; and using the frequency domain energy ratio of the Theta wave band to the Beta wave band in the frequency domain signal as the EEG power ratio feature. The acquisition of the interaction response variability feature includes: for any time step, acquiring the response time corresponding to the most recent preset first number of historical valid key presses at any time step, and using the standard deviation of all the response times as the interaction response variability feature for any time step.

[0014] The present invention has at least the following beneficial effects: This invention constructs a baseline grid jump frequency matrix based on EEG signals and interactive behavior data of target children during an interference-free benchmark testing phase, and generates a baseline feature probability distribution matrix using a graph network algorithm. This achieves accurate modeling of the individual physiological and behavioral characteristics of the target children. During the rehabilitation training phase, the system utilizes the real-time grid jump frequency matrix and graph network algorithm engine to effectively filter out high-frequency transient interference, avoiding data breaks or misjudgments caused by sudden hyperactivity, thus improving the stability and real-time performance of signal processing. By calculating the feature distribution difference divergence index between the real-time feature probability distribution matrix and the baseline matrix, the system can accurately assess changes in the target children's attention state and issue corresponding intervention instructions accordingly, achieving adaptive closed-loop control based on individual differences. This system overcomes the response delay and control oscillation problems introduced by fixed threshold elimination or mean filtering in traditional methods, improving the control accuracy and robustness of the rehabilitation training system, and ensuring stable, reliable, and personalized rehabilitation intervention for the attention state of children with ADHD. Attached Figure Description

[0015] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a structural block diagram of a personalized attention rehabilitation training system for children with ADHD, provided in an embodiment of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following detailed description of a personalized attention rehabilitation training system for children with ADHD, proposed according to the present invention, is provided in conjunction with the accompanying drawings and preferred embodiments.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] The following description, in conjunction with the accompanying drawings, details a specific scheme for a personalized attention rehabilitation training system for children with ADHD provided by this invention.

[0020] Example of a personalized attention rehabilitation training system for children with ADHD: This embodiment proposes a personalized attention rehabilitation training system for children with ADHD, such as... Figure 1 As shown, a personalized attention rehabilitation training system for children with ADHD in this embodiment includes a data acquisition module, a benchmark determination module, a noise reduction module, and an intervention control module.

[0021] The data acquisition module is used to acquire EEG signals and interactive behavior data of the target children during an interference-free benchmark test phase. The benchmark determination module is used to construct a benchmark grid jump frequency matrix based on the EEG power ratio characteristics presented by EEG signals and the interaction response variability characteristics presented by interactive behavior data; and to iteratively generate a benchmark feature probability distribution matrix based on the benchmark grid jump frequency matrix using a graph network algorithm. The noise reduction module is used to obtain the real-time grid jump frequency matrix during the rehabilitation training phase of the target child; the real-time grid jump frequency matrix is ​​input into the graph network algorithm engine to obtain the real-time feature probability distribution matrix after filtering out high-frequency transient interference; The intervention control module is used to determine the feature distribution difference divergence index of the real-time feature probability distribution matrix relative to the benchmark feature probability distribution matrix; and to issue corresponding instructions based on the feature distribution difference divergence index.

[0022] The data acquisition module is used to acquire EEG signals and interactive behavior data of the target children during an interference-free benchmark testing phase.

[0023] First, in the interference-free benchmark testing phase, the data acquisition module of the personalized attention rehabilitation training system for children with ADHD acquires the target child's EEG signals and interactive behavior data. By continuously advancing the timeline with a fixed time step of 1 second, the EEG signals at each time step can be captured.

[0024] A benchmark determination module is used to construct a benchmark grid jump frequency matrix based on the EEG power ratio characteristics presented by EEG signals and the interaction response variability characteristics presented by interactive behavior data; based on the benchmark grid jump frequency matrix, a benchmark feature probability distribution matrix is ​​iteratively generated using a graph network algorithm.

[0025] Given the distinct individual differences in the daily physiological baseline and keystroke speed limits of different test children, a fixed, universal coordinate system cannot fully encompass the actual range of data generated. Therefore, it is necessary to first read the historical extreme values ​​of the target children and pre-define a unified two-dimensional feature grid with absolute boundaries to provide an equivalent spatial comparison scale for subsequent baseline and real-time data.

[0026] Specifically, firstly, the system retrieves the target child's historical rehabilitation records stored in the system. These records include EEG signals and interactive behavior data. The EEG signals in the historical records are divided into multiple time steps. The EEG signals from each time step are extracted, and the time-domain EEG signals of each time step are converted to frequency-domain signals using a Fast Fourier Transform (FFT) algorithm. After conversion, the frequency-domain energy ratio of the Theta wave band to the Beta wave band in the frequency-domain signal is calculated, and this ratio is used as the EEG power ratio feature. The maximum value of the EEG power ratio feature for all time steps in the historical rehabilitation records is obtained. Simultaneously, the reaction time corresponding to the most recent preset number of valid key presses in the historical rehabilitation data is obtained. The standard deviation of all reaction times is used as the interactive response variability feature for each time step, and the maximum value of the interactive response variability feature for all time steps in the historical rehabilitation records is obtained. In this embodiment, the preset number of first presses is 5. In specific applications, the implementer can set this number according to specific circumstances.

[0027] After obtaining the maximum values ​​of the EEG power ratio characteristics and the maximum values ​​of the interaction response variability characteristics for all time steps in the historical rehabilitation records, these two maximum values ​​are multiplied by a preset magnification factor, respectively, and the results are used as the absolute coordinate boundaries of the horizontal and vertical axes of the two-dimensional coordinate system. In this embodiment, the preset magnification factor is 1.2; in specific applications, the implementer can set it according to the specific circumstances.

[0028] After setting the absolute coordinate boundaries, the step size of the horizontal and vertical axes is dynamically calculated and adjusted according to the determined absolute coordinate boundary range to ensure that the divided two-dimensional feature grid is fixed at a size of 4×4 or 5×5, thereby generating a two-dimensional feature grid composed of independent grid blocks. Specifically, since subsequent modules require a fixed-length sliding time window, this embodiment sets the real-time analysis window length to include 30 sample points to statistically analyze the jump connectivity probability between grids. If the number of grid blocks is too large, projecting 30 sample points onto a large number of blocks will result in most blocks being unvisited, leading to a sparse jump frequency matrix generated subsequently. A sparse matrix causes the steady-state probability output by the graph network algorithm to be dominated by the random constants at the algorithm's underlying layer, thus losing its topological evaluation significance that reflects the true data distribution.

[0029] To prevent the transition matrix from becoming extremely sparse, a grid granularity constraint rule is implemented here. The total number of blocks in the resulting 2D feature grid (i.e., grid size M×K) must be less than or equal to the total number of samples contained in the subsequent real-time analysis window. For example, with a total of 30 samples in the analysis window, the system adjusts the grid's horizontal and vertical axis segmentation step size to create a grid size of 4×4 (16 blocks) or 5×5 (25 blocks). After completing the segmentation operation and satisfying the quantity constraint, the entirety of these M×K blocks is stored as a global static measurement reference system.

[0030] The benchmark determination module will then construct a benchmark grid jump frequency matrix based on the EEG power ratio characteristics of EEG signals and the interaction response variability characteristics of interactive behavior data presented in the benchmark testing phase without additional interference.

[0031] Specifically, in the benchmark testing phase without additional interference, at time step t, the EEG signals from the past second are acquired. After capturing the EEG signals, a Fast Fourier Transform (FFT) algorithm is used to convert them into frequency domain signals. The frequency domain energy ratio of the Theta wave band to the Beta wave band is calculated, and this ratio is used as the EEG power ratio feature for that time step. The larger this value, the deeper the target child's cerebral cortex is in a state of low arousal (inattention). Simultaneously, the reaction time corresponding to the most recent preset number of valid key presses at time step t is acquired. The standard deviation of all reaction times obtained at this time is used as the interaction response variability feature for time step t. The larger this value, the more discrete the target child's behavioral control is, and the more varied the target child's reaction speed. Through these methods, the EEG power ratio feature and interaction response variability feature for each time step in the benchmark testing phase can be obtained.

[0032] Next, based on the EEG power ratio characteristics and interaction response variability characteristics at each time step of the benchmark test, two-dimensional feature coordinate points are determined; the numerical value of these coordinate points is then compared with the absolute coordinate boundaries of the two-dimensional feature grid. Through this comparison, the system determines the specific independent block in which the coordinate point falls and obtains the unique discrete grid index corresponding to that independent block. ,in, As time goes by... As the process continues, the system continuously acquires a series of discrete grid indices and connects them in one-dimensional order according to their chronological occurrence, ultimately generating a reference grid node transition sequence that records the evolution path of the reference state space. .

[0033] Since this embodiment will subsequently employ a graph network iterative algorithm, it requires inputting an edge set data structure that expresses the topological connectivity between nodes. Therefore, the system needs to convert the jump sequence of the baseline grid nodes into a two-dimensional frequency occurrence adjacency matrix.

[0034] Specifically, based on the total number of blocks in the two-dimensional feature grid ( ), construct a memory array with all initial values ​​of 0 and a size of After constructing the air-to-ground transition matrix, an iterative accumulation operation based on the time step order is initiated. First, the baseline determination module obtains the baseline grid node jump sequence. Corresponding time step Time starting grid index and time step End point grid index After obtaining the two adjacent grid indices, the reference determination module locates the row coordinates in the air-to-ground transition matrix as follows: Column coordinates are The directed edge element position is located; after locating this element position, the value at that position is incremented by 1, which indicates that the feature state has changed from the block. Pointing to block The directed transfer. Then, the benchmark determination module acquires the time steps. Time starting grid index and corresponding time steps End point grid index After obtaining the coordinates, the reference determination module locates the coordinate position in the transition matrix. The directed edge element is selected, and the value at that element's position is incremented by 1. This iterative loop continues until the jump sequence of the baseline grid nodes is traversed. After traversing all adjacent index pairs up to the last time step, the baseline determination module will finally complete the cumulative operation and output the transition matrix, defining it as the baseline grid jump frequency matrix containing the statistical sum of all directed jump paths. This matrix records the interconnection and repetitive jumping patterns of the target child's features between different spatial grid blocks in an undisturbed state.

[0035] Furthermore, the benchmark determination module will output a steady-state probability distribution target model after filtering out occasional interference features, based on the benchmark grid jump frequency matrix and utilizing the node connectivity evaluation mechanism of the graph network algorithm.

[0036] Considering that during interference-free benchmark testing, children may occasionally experience brief, unexpected physiological responses such as slight coughing or involuntary head turning. Due to the objective differences in the response delays of different physiological signals, such brief disturbances can cause sudden changes in transient EEG characteristics while long-term interactive characteristics fail to respond. Therefore, the feature coordinates induced by these disturbances appear as temporal jump trajectories lacking symmetrical return flow in the two-dimensional feature grid; for example, there are only directed edges jumping into edge blocks, but no reverse edges that immediately jump back from the edge block to the core block. To suppress their influence without discarding the original sampling frames containing these jump values, the benchmark grid jump frequency matrix is ​​used as a set of directed graph edges expressing the feature state transition relationships. This set is input into the PageRank graph network algorithm engine, which sets a damping coefficient and outputs the steady-state probability value of each grid block through iterative calculation of the graph network algorithm. In this embodiment, the damping coefficient is 0.85. In specific applications, the implementer can set it according to the specific situation. The purpose of this setting is to force the remaining... The random jump probability is evenly distributed across all blocks in the two-dimensional feature grid. Since some edge blocks are only visited once, dead-end nodes with zero out-degree are formed in the graph theory model. Without a full grid probability distribution as a fallback, the algorithm engine will decay the steady-state probability of these nodes to absolute zero after iterative calculations (i.e., ...). This would directly lead to fatal mathematical errors in subsequent divergence calculations, such as the logarithm being an argument or the denominator being zero. By uniformly distributing the remaining probabilities to all blocks, the system completely eliminates dead zones where the absolute value of the final output matrix is ​​0.

[0037] After setting the damping coefficient, the PageRank algorithm engine uses the baseline grid jump frequency matrix. The algorithm engine uses the recorded out-degree and in-degree distributions of nodes to initiate a Markov chain-style steady-state probability iterative distribution calculation. During the iterative calculation, the algorithm engine continuously distributes the probability score of each grid block according to the proportion of its jumps to other blocks. For asymmetric edge grid nodes that lack backflow characteristics (very small in-degree) or form isolated unidirectional jumps (zero out-degree), the algorithm engine naturally reduces their accumulated probability after multiple iterations. When the difference in probability distributions calculated in two adjacent iterations is less than a preset convergence constant, the baseline determination module determines that the network steady state has been established and stops the iterative calculation. In this embodiment, the preset convergence constant is... In specific applications, implementers can configure it according to the specific circumstances.

[0038] After the iterative calculation is completed, the algorithm engine outputs a value with the same size as the two-dimensional feature mesh and a size of [missing value]. The two-dimensional floating-point matrix is ​​output. After outputting this matrix, the benchmark determination module performs normalization processing on the steady-state probability values ​​corresponding to all blocks within the matrix. The data normalization method is existing technology and will not be elaborated further here. The sum of the normalized results of the steady-state probability values ​​corresponding to all blocks equals... Define this normalized matrix as the baseline feature probability distribution matrix. The data is then latched and used as the final distribution pattern for determining the cognitive dispersion of the tested children. There are many methods for data normalization, which will not be elaborated upon here.

[0039] The baseline feature probability distribution matrix obtained in this embodiment aims to output an individual difference tolerance threshold specific to the target child by measuring the fluctuation range of the inherent feature distribution of the tested children under optimal baseline conditions, replacing the inflexible global fixed constant judgment. This embodiment will extract the extreme values ​​of the fluctuation of the KL divergence corresponding to the local area in the benchmark testing phase as the reference upper limit.

[0040] Specifically, all baseline grid node jump sequences recorded during the benchmarking phase are reloaded. In these baseline sequences, a baseline sliding analysis window with the same length as the subsequent formal training phase is enabled. This analysis window can contain 30 sample points with a step size of 1 second.

[0041] As the baseline sliding analysis window continuously slides across the benchmark testing timeline, the processing logic for the benchmark grid jump frequency matrix and benchmark feature probability distribution matrix from the above steps is used to extract the local jump frequency matrix for each sliding window within a brief 30-second period. This matrix is ​​then input into the PageRank algorithm engine for probability iteration and normalization with a damped coefficient, and a preset minimum positive number is added. To prevent division by zero, a series of time-continuous local feature probability distribution matrices are generated. .

[0042] Furthermore, using the formula for calculating relative entropy, the entropy value of each local feature probability distribution matrix relative to the baseline feature probability distribution matrix is ​​calculated and used as the local KL divergence value. At this point, all the calculated local KL divergence values ​​constitute the baseline divergence sequence.

[0043] After obtaining the baseline divergence sequence, the mean and standard deviation of all data in the baseline divergence sequence are calculated. Based on the mean and standard deviation of all data in the baseline divergence sequence, an individual difference tolerance threshold is determined. Specifically, the sum of the mean and standard deviation is set as the upper limit of the characteristic distribution fluctuation of the target child under the condition of no severe cognitive decline or hyperactivity interference. After extracting this fluctuation upper limit, the system multiplies the sum by a preset tolerance amplification factor, and the product is used as the individual difference tolerance threshold for the target child. The preset tolerance amplification factor is a constant set by the system to reserve intervention margin. In this embodiment, the preset tolerance amplification factor is 1.2. In specific applications, the implementer can set it according to the specific situation.

[0044] The noise reduction module is used to obtain the real-time grid jump frequency matrix during the rehabilitation training phase of the target child; the real-time grid jump frequency matrix is ​​input into the graph network algorithm engine to obtain the real-time feature probability distribution matrix after filtering out high-frequency transient interference.

[0045] During the formal rehabilitation training phase for the target children, real-time continuous EEG signals and interactive behavior data are continuously received, and a two-dimensional probability matrix representing the current macroscopic state with high-frequency interference filtered out is output at a fixed frequency.

[0046] Specifically, a real-time analysis window of preset length is used to slide with a fixed step size; real-time two-dimensional feature coordinate points of each time step within each real-time analysis window are extracted. In this embodiment, the preset length is 30 seconds, corresponding to 30 independent data sampling nodes, and the fixed step size is 1 second, that is, a window advancement operation is performed every 1 second, that is, the oldest set of sampled data is discarded and the latest set of sampled data is included. In specific applications, the implementer can set it according to the specific situation. At each time step as the real-time analysis window continues to advance, the system acquires the real-time frontal EEG signal of the subject child in the past 1 second. After acquiring the real-time EEG signal, the real-time EEG signal is processed using the Fast Fourier Transform algorithm, and the corresponding EEG power ratio feature and interaction response variability feature are calculated. The EEG power ratio feature and interaction response variability feature are used as the horizontal and vertical coordinates, respectively, thus obtaining the real-time two-dimensional feature coordinate points. The method for acquiring the EEG power ratio feature and interaction response variability feature at this time refers to the method for acquiring the EEG power ratio feature and interaction response variability feature at each time step in the benchmark test stage, which will not be elaborated further here.

[0047] After calculating the characteristics of the brainwave power ratio Simultaneously, the system backtracks the real-time interactive behavior data of the test software. During the backtracking of the real-time interaction logs, the system is not limited by whether any keystrokes occurred within the current second, and extracts data from the historical record up to a point in time from the current time. The system calculates the reaction time for the five most recent valid key presses. If, during the extraction of reaction time for valid key presses, the system fails to extract five valid key presses from the past 60 seconds of historical data, it determines that the child's current interaction is in a state of depletion. Since the system cannot calculate the time standard deviation, it directly assigns the current interaction response variability feature to the maximum value of the interaction response variability feature across all time steps in the historical rehabilitation file, multiplied by a preset penalty coefficient. If the system successfully extracts five valid key presses, it calculates the standard deviation of the reaction time for these five historical valid key presses and uses this standard deviation as the interaction response variability feature. In this embodiment, the preset penalty coefficient is 1.15. In specific applications, the implementer can set it according to the specific circumstances.

[0048] Acquiring EEG power ratio characteristics With interaction response variability characteristics Afterwards, These two one-dimensional values ​​are concatenated using coordinates. After the concatenation operation is completed, the system generates the current time step. Two-dimensional feature coordinates points .

[0049] Due to the extracted real-time two-dimensional feature coordinate points As these are unbounded floating-point numbers, the real-time features generated by the target child during rehabilitation training may exceed the historical maximum values ​​recorded during the benchmark testing phase. Directly projecting coordinates exceeding the boundaries onto the 2D feature grid will prevent the system from extracting valid grid indices, leading to program errors. Therefore, the system needs to implement a mandatory boundary truncation mapping rule at this point.

[0050] Specifically, the system extracts 30 sequentially generated two-dimensional feature coordinate points contained within the current real-time analysis window. Extract these 30 two-dimensional feature coordinate points. Then, the system retrieves the stored fixed two-dimensional feature mesh (size: After retrieving the two-dimensional feature mesh, these 30 two-dimensional feature coordinate points are compared one by one with the absolute coordinate boundaries (maximum / minimum values ​​of the horizontal and vertical axes) of the two-dimensional feature mesh in chronological order of their generation. During the comparison, if the two-dimensional feature coordinate points... any dimension value ( or If the value exceeds the absolute coordinate boundary of the two-dimensional feature mesh, the system triggers a truncation protection mechanism, forcibly constraining the out-of-bounds value to the maximum or minimum value of the corresponding boundary. For example, if the current... If the value exceeds the maximum value boundary set on the horizontal axis of the grid, the system will... The value is forcibly assigned to the boundary value equal to the maximum value of the horizontal axis. This boundary truncation protection operation ensures that all coordinate points can be successfully mapped to valid blocks within the 2D feature mesh. After determining the location of each valid block, the system obtains the discrete mesh index corresponding to that block. ,in, Retrieve all 30 discrete grid indices within the current sliding window. Then, these discrete grid indices are arranged in one-dimensional order according to time. After the arrangement is completed, the system generates a real-time grid node jump sequence representing the feature space evolution path within the current 30 seconds. .

[0051] Considering that existing technologies typically employ time-window mean filtering when processing high-frequency electromyographic jumps caused by instantaneous hyperactivity in children with ADHD, this introduces significant feedback response lag. This embodiment utilizes a graph network iterative algorithm to naturally attenuate the probability of isolated edge grids without bidirectional backflow, thereby outputting a stable real-time probability distribution matrix without discarding any original sampled data points or introducing additional mean calculation delay.

[0052] Specifically, the real-time grid node jump sequence This is converted into a matrix structure that records the jump direction and number of jumps. The system constructs a matrix of size in memory. A void transition matrix with all initial elements set to 0 is constructed. After constructing the void transition matrix, the system initiates an iterative accumulation operation based on the time step sequence, first extracting the real-time grid node jump sequence. Mid-time step Time starting grid index and time step End point grid index Locate the row coordinate in the air-to-air transition matrix. Column coordinates are The system extracts the positions of directed edge elements and increments the value at that position by 1. This process is repeated sequentially up to the time step. The last endpoint grid index. After completing the accumulation operation of all consecutive index pairs, the system outputs a real-time grid jump frequency matrix containing statistics on the number of directed jumps within the current real-time analysis window. .

[0053] In the real-time grid jump frequency matrix After construction, the system inputs it into the PageRank algorithm engine. Similar to the baseline determination module, the system sets a damping coefficient of 0.85 for the algorithm engine, forcing the remaining... The walk probability is evenly distributed across all blocks in the two-dimensional feature grid, thus completely eliminating the absolute zero region within the real-time probability matrix. After setting this damping coefficient, the algorithm engine initiates a probabilistic iterative solution operation based on the real-time Markov chain steady-state model. During the iterative solution operation, the algorithm engine uses the real-time grid jump frequency matrix... Record the out-degree and in-degree distributions and dynamically adjust the probability allocation of each grid block.

[0054] Since the cognitive fatigue of the children in the test is manifested as a continuous and repetitive evolution between adjacent feature states (with symmetrical in-degree and out-degree), while hyperactivity artifacts such as sudden teeth clenching or unexpected head turning are manifested as isolated state jumps lacking temporal backflow (unidirectional pointing to the edge, with a very small or zero out-degree), the algorithm engine naturally identifies and greatly attenuates the steady-state probability scores of these isolated edge blocks after iteration.

[0055] Once the iterative computation reaches the convergence constant criterion, the algorithm engine stops iterating. After stopping iteration, the system extracts the size of the algorithm engine's output, which is also [size missing]. The system generates a two-dimensional floating-point steady-state probability matrix. To compress this matrix into a compliant discrete probability distribution, the system performs a global normalization process on the two-dimensional floating-point matrix. During the division step of the normalization operation, the system uniformly adds the aforementioned preset minimum positive number to the denominator. After performing the normalization operation, the sum of the probabilities of the system generating all blocks equals This is a two-dimensional matrix. After generating this matrix, the system formally defines it as the real-time feature probability distribution matrix. Once defined, the system will display the real-time feature probability distribution matrix after filtering out high-frequency transient interference. The data is output to the next stage as a benchmark for measuring the current macroscopic cognitive state. As the real-time analysis window slides forward in 1-second increments, the system continues to process data and outputs the latest real-time feature probability distribution matrix to the next stage at an update frequency of 1 Hz.

[0056] An intervention control module is used to determine the feature distribution difference divergence index of the real-time feature probability distribution matrix relative to the baseline feature probability distribution matrix; and to issue corresponding instructions based on the feature distribution difference divergence index.

[0057] As training time increases, the decline in cognitive attention of the tested children is objectively manifested as a gradual deviation of the overall energy center of the real-time feature probability distribution matrix from the baseline feature probability distribution matrix. The system needs to compress these two two-dimensional probability matrices, each of size M×K, into a scalar value that can intuitively reflect the degree of state deviation.

[0058] Due to the baseline characteristic probability distribution matrix It originates from the full data sample throughout the entire benchmark testing phase, and the real-time feature probability distribution matrix The data is derived from local samples within a 30-second window during the target child's rehabilitation training phase. The distribution pattern of local samples is naturally sharper than that of the full sample. To eliminate the inherent statistical fluctuations caused by the extreme asymmetry in sample size, the system first performs temporal smoothing on the real-time feature probability distribution matrix before performing divergence calculation.

[0059] Specifically, the exponentially weighted moving average (EMA) formula is used to smooth the real-time feature probability distribution matrix. The smoothing result from the previous time step is used to weight and correct the matrix at the current time step, resulting in the smoothed real-time feature probability distribution matrix. The specific formula for this smoothing process is as follows: in, This represents the element value of the i-th block in the smoothed real-time feature probability distribution matrix. Indices representing discrete grid blocks in a two-dimensional feature grid. This represents the probability value of the i-th block in the smoothed matrix calculated in the previous time step. This represents the probability value of the i-th block in the real-time feature probability distribution matrix at the current time step. This represents the preset timing smoothing coefficient.

[0060] In this embodiment, the preset time-series smoothing coefficient is set to 0.1. In specific applications, the implementer can set it according to the specific circumstances. This smoothing operation effectively filters out local extreme value oscillations caused by short time windows and outputs a smoothed real-time matrix.

[0061] Using the formula for relative entropy, the entropy value of the smoothed real-time feature probability distribution matrix relative to the baseline feature probability distribution matrix is ​​calculated, and this value is used as the divergence value. The divergence value is then used as the feature distribution difference divergence index. The formula for calculating the divergence value can be expressed as: in, This represents the divergence index of the feature distribution calculated at the current time step. This represents the element value of the i-th block in the baseline feature probability distribution matrix. This represents the total number of blocks in a two-dimensional feature grid. This indicates a preset minimum positive number.

[0062] This value represents the relative probability of the i-th block in the current real-time state to the probability of the i-th block in the baseline ideal state. A preset minimum positive number is introduced to prevent the denominator of the formula from being 0.

[0063] The feature distribution divergence index is an asymmetric, absolutely non-negative metric. When the child's state is perfectly aligned with the baseline, the value approaches 0. The larger the value, the more drastic the divergence of the child's feature distribution from their ideal cognitive baseline at that moment, indicating a more severe loss of attention.

[0064] Since an increase in the divergence index in a single instance may stem from a brief distraction by ambient light or sound in the child, directly triggering a difficulty reduction based on a single out-of-bounds action would lead to frequent fluctuations in the task parameters of the training software. Furthermore, after the system issues a difficulty adjustment command, the target child's physiological and cognitive state reacts, resulting in an objective physiological lag of several seconds to tens of seconds reflected in the EEG signal. If the system does not implement protective logic, during these tens of seconds before the index returns to normal, the system will continuously issue downgrade commands, causing the task difficulty of the rehabilitation software to drop to zero in a very short time, thus causing system control deadlock and ineffective training. To address these engineering problems, the system incorporates a dual protection mechanism during the judgment and comparison process: a time-cumulative out-of-bounds judgment and an intervention command cooldown period.

[0065] Specifically, when the number of times the real-time acquired feature distribution difference divergence index exceeds the individual difference tolerance threshold reaches a preset number, it indicates that the target child has exceeded the extreme boundary of its own allowed feature drift, and the target child is judged to have entered the cognitive state continuous over-limit stage; in this embodiment, the preset number is 3 times, and in specific applications, the implementer can set it according to the specific situation.

[0066] Once the target child enters a state of sustained cognitive overload and is not within the instruction cooldown period, the system sends a task difficulty reduction instruction to the rehabilitation training software performing the interactive operation. Upon receiving this instruction, the rehabilitation training software executes specific difficulty adjustment actions within the software interface, such as reducing the movement speed of the visual tracking target or reducing the number of distractions displayed on the screen simultaneously.

[0067] At the same moment the downgrade command is successfully issued, the system immediately initiates a command cooling-off window of a preset duration. In this embodiment, the preset duration is 15 seconds, but the implementer can set it according to specific circumstances in actual applications. During the 15-second command cooling-off window, the system continues to receive EEG features, update the grid jump frequency matrix, and calculate the real-time feature distribution difference divergence index normally. However, the system forcibly suspends (blocks) all intervention judgment logic, no longer issues any new task difficulty adjustment commands, and forcibly locks the out-of-bounds cumulative counter to a zero state. When the 15-second command cooling-off window ends, the system releases the suspension of the judgment logic and resumes normal out-of-bounds cumulative counter and threshold comparison operations.

[0068] Through the aforementioned complete closed-loop design, the cognitive load of the software training task on the test children was smoothly and controllably reduced. As the cognitive load decreased, the target children's attention gradually recovered, and their output real-time feature probability distribution matrix subsequently converged back towards the target model's baseline feature probability distribution matrix. Ultimately, the system monitored that the feature distribution difference divergence index gradually fell back below the individual difference tolerance threshold, thus completing a personalized neurofeedback rehabilitation closed loop based entirely on the individual physiological baselines of the test children, while ensuring that the system did not crash and the control did not deadlock.

[0069] This embodiment constructs a baseline grid jump frequency matrix based on the EEG signals and interactive behavior data of the target child during an interference-free benchmark test phase, and generates a baseline feature probability distribution matrix using a graph network algorithm, achieving accurate modeling of the individual physiological and behavioral characteristics of the target child. During the rehabilitation training phase, the system utilizes the real-time grid jump frequency matrix and graph network algorithm engine to effectively filter out high-frequency transient interference, avoiding data breaks or misjudgments caused by sudden hyperactivity, thus improving the stability and real-time performance of signal processing. By calculating the feature distribution difference divergence index between the real-time feature probability distribution matrix and the baseline matrix, the system can accurately assess changes in the target child's attention state and issue corresponding intervention instructions accordingly, achieving adaptive closed-loop control based on individual differences. This system overcomes the response delay and control oscillation problems introduced by fixed threshold elimination or mean filtering in traditional methods, improving the control accuracy and robustness of the rehabilitation training system, and ensuring stable, reliable, and personalized rehabilitation intervention for the attention state of children with ADHD.

[0070] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A personalized attention rehabilitation training system for children with ADHD, characterized in that, The system includes: The data acquisition module is used to acquire EEG signals and interactive behavior data of the target children during an interference-free benchmark testing phase; The benchmark determination module is used to construct a benchmark grid jump frequency matrix based on the EEG power ratio characteristics presented by EEG signals and the interaction response variability characteristics presented by interactive behavior data; and to iteratively generate a benchmark feature probability distribution matrix based on the benchmark grid jump frequency matrix using a graph network algorithm. The noise reduction module is used to obtain the real-time grid jump frequency matrix during the rehabilitation training phase of the target child; the real-time grid jump frequency matrix is ​​input into the graph network algorithm engine to obtain the real-time feature probability distribution matrix after filtering out high-frequency transient interference; The intervention control module is used to determine the feature distribution difference divergence index of the real-time feature probability distribution matrix relative to the benchmark feature probability distribution matrix; and to issue corresponding instructions based on the feature distribution difference divergence index.

2. The personalized attention rehabilitation training system for children with ADHD according to claim 1, characterized in that, Before constructing the baseline grid jump frequency matrix, the following is also included: The absolute coordinate boundaries of the two-dimensional coordinate system are determined based on the maximum value of the EEG power ratio presented in the historical data of the target children and the maximum value of the interaction response variability feature presented in the interaction behavior data.

3. The personalized attention rehabilitation training system for children with ADHD according to claim 2, characterized in that, The baseline grid jump frequency matrix is ​​constructed based on the EEG power ratio characteristics presented by EEG signals and the interaction response variability characteristics presented by interactive behavior data, including: During the benchmark testing phase, instantaneous EEG power ratio features and interactive response variability features are extracted at fixed time steps, and two-dimensional feature coordinate points are determined based on the instantaneous EEG power ratio features and interactive response variability features. The two-dimensional feature coordinate points at each time step are mapped to a two-dimensional feature grid to obtain a reference grid node jump sequence composed of continuous grid block indices; Based on the grid block index pairs of adjacent time steps in the reference grid node jump sequence, a reference grid jump frequency matrix is ​​constructed to record the number of directed jumps.

4. The personalized attention rehabilitation training system for children with ADHD according to claim 3, characterized in that, The step of iteratively generating a baseline feature probability distribution matrix based on the baseline grid jump frequency matrix using a graph network algorithm includes: The reference grid jump frequency matrix is ​​input into the graph network algorithm engine, the damping coefficient is set, and the steady-state probability value of each grid block is output through iterative calculation of the graph network algorithm. The steady-state probability values ​​of each grid block are normalized to obtain the baseline characteristic probability distribution matrix.

5. The personalized attention rehabilitation training system for children with ADHD according to claim 4, characterized in that, After generating the baseline feature probability distribution matrix, the following steps are also included: During the benchmark testing phase, local data is continuously extracted through a sliding window of preset length to generate multiple continuous local feature probability distribution matrices. Calculate the local KL divergence value of each of the local feature probability distribution matrices relative to the baseline feature probability distribution matrix, and all the local KL divergence values ​​constitute the baseline divergence sequence; Based on the mean and standard deviation of all data in the benchmark divergence sequence, the tolerance threshold for individual differences is determined.

6. The personalized attention rehabilitation training system for children with ADHD according to claim 2, characterized in that, The process of obtaining the real-time grid jump frequency matrix includes: During the rehabilitation training phase, a real-time analysis window of preset length is used to slide with a fixed step size; the real-time two-dimensional feature coordinate points of each time step within each real-time analysis window are extracted. The real-time two-dimensional feature coordinate points are mapped to the two-dimensional feature grid. If the coordinate points exceed the absolute coordinate boundary, they are forcibly constrained to the boundary value, and a real-time grid node jump sequence is generated. Based on the grid block index pairs of adjacent time steps in the real-time grid node jump sequence, a real-time grid jump frequency matrix is ​​constructed.

7. The personalized attention rehabilitation training system for children with ADHD according to claim 1, characterized in that, The step of inputting the real-time grid jump frequency matrix into the graph network algorithm engine to obtain the real-time feature probability distribution matrix after filtering out high-frequency transient interference includes: The real-time grid jump frequency matrix is ​​input into the graph network algorithm engine, and the damping coefficient is set. Through iterative calculation of the graph network algorithm, the steady-state probability value of each grid block is output. The steady-state probability values ​​of each grid block are normalized to generate the real-time feature probability distribution matrix.

8. The personalized attention rehabilitation training system for children with ADHD according to claim 1, characterized in that, The step of determining the feature distribution difference divergence index of the real-time feature probability distribution matrix relative to the benchmark feature probability distribution matrix includes: The real-time feature probability distribution matrix is ​​smoothed by an exponentially weighted moving average to obtain a smoothed real-time feature probability distribution matrix. Calculate the divergence value of the smoothed real-time feature probability distribution matrix relative to the baseline feature probability distribution matrix, and use the divergence value as the feature distribution difference divergence index.

9. The personalized attention rehabilitation training system for children with ADHD according to claim 5, characterized in that, The issuance of corresponding instructions based on the feature distribution difference divergence index includes: When the number of times the real-time acquired feature distribution difference divergence index exceeds the individual difference tolerance threshold continuously reaches a preset number, the target child is determined to have entered the cognitive state continuous over-limit stage. When the target child enters a stage of sustained cognitive overload and is not within the instruction cooling-off period, issue an instruction to reduce the difficulty of the task.

10. The personalized attention rehabilitation training system for children with ADHD according to claim 2, characterized in that, The acquisition of the EEG power ratio feature includes: converting the EEG signal into a frequency domain signal using a fast Fourier transform algorithm; and using the frequency domain energy ratio of the Theta wave band to the Beta wave band in the frequency domain signal as the EEG power ratio feature. The acquisition of the interaction response variability feature includes: for any time step, acquiring the response time corresponding to the most recent preset first number of historical valid key presses at any time step, and using the standard deviation of all the response times as the interaction response variability feature for any time step.

Images