Identification system and identification method for attention deficit hyperactivity disorder

Abstract

The invention discloses an attention deficit hyperactivity disorder recognition system and recognition method, and belongs to the technical field of attention deficit hyperactivity disorder. The data processing module is used for carrying out preprocessing and feature extraction on the acquired electroencephalogram data; a multi-source feature fusion mechanism is firstly used for the extracted original feature data, and then a data enhancement strategy is applied; a multi-source fusion feedback regulation network model is constructed, wherein the model is of a CNN-GRU parallel modeling structure; the training module is used for inputting the enhanced data into a model for training, key hyper-parameters are dynamically adjusted by a performance feedback adjusting mechanism in the training process, and the performance feedback adjusting mechanism is used for dynamically adjusting key training parameters according to the performance of the verification set; and the classification module is used for classifying to-be-detected samples through the trained multi-source fusion feedback regulation network model and outputting a final recognition result. The ADHD electroencephalogram recognition method effectively improves the accuracy, robustness and generalization performance of ADHD electroencephalogram recognition.

Description

Technical Field

[0001] The present invention belongs to the technical field of attention deficit hyperactivity disorder, and in particular relates to an attention deficit hyperactivity disorder identification system and an identification method. Background Art

[0002] Attention Deficit Hyperactivity Disorder (ADHD) is a common neurodevelopmental disorder. In recent years, electroencephalography (EEG) has become an important tool for objectively assisting the diagnosis of ADHD due to its high temporal resolution, noninvasiveness, and low cost. Traditional EEG analysis methods rely on time-domain, frequency-domain, or nonlinear statistical features, such as mean, power spectral density, entropy, and Hjorth parameter, to capture the statistical variations of EEG signals across different dimensions. However, these features are often extracted statically and lack in-depth modeling of the temporal structure and spatial topological variations of EEG signals. This makes it difficult to fully reflect the dynamic abnormalities of brain function in children with ADHD, limiting further improvement in classification model performance.

[0003] In recent years, deep learning (DL) has demonstrated powerful capabilities for automatic feature extraction and modeling in EEG decoding tasks. Convolutional neural networks (CNNs) excel at capturing local spatial features, while recurrent neural networks (such as GRUs and LSTMs) excel at temporal modeling. For example, a long short-term memory (LSTM) network was embedded within a CNN structure for epileptic seizure prediction, demonstrating the potential of deep temporal modeling in EEG analysis. Another study proposed a GRU-based RNN architecture, which achieved comparable performance to LSTM while reducing the number of parameters. Furthermore, a hybrid neural network combining GRU and CNN was constructed. By fusing EEG microstate features with RQA dynamic parameters, auditory attention recognition was achieved in the absence of speech stimulation, validating the value of microstates in high-temporal-resolution EEG modeling. Despite this, most existing deep learning models still face several challenges. First, the limited number of subjects in publicly available EEG data related to ADHD represents a typical small-sample training scenario within deep learning frameworks. Even with multiple segmented samples obtained through time windowing, it is difficult to fully cover the high-dimensional feature space, which can easily lead to model overfitting. Second, most existing models still primarily rely on statistical features or raw time series signals as input, lacking the ability to capture the dynamic structure of brain function, limiting their performance in modeling complex EEG rhythms. For example, most MI-EEG classification models directly input raw EEG data into deep networks. While this eliminates the need for manual feature design, it struggles to capture more complex EEG dynamic patterns. Manual statistical feature extraction from segmented EEG signals for ADHD classification validated the impact of preprocessing and temporal resolution on classification performance. However, this approach is similarly limited to static feature descriptions and fails to model the temporal evolution of brain activity. Furthermore, clinical EEG data is often characterized by nonstationarity and noise perturbations. Static training parameters are difficult to adapt to performance fluctuations across different phases and lack effective adaptive adjustment mechanisms.

[0004] The application of microstate analysis in ADHD identification is still in its infancy. Most current studies have not yet deeply integrated it with statistical features, and lack the ability to model its non-stationary structure. Despite this, research on microstate analysis in ADHD identification is still relatively limited, and there are multiple technical challenges: First, standard microstate templates vary between individuals, and direct use across subjects may affect matching quality and feature stability; second, microstate features are usually modeled in isolation, lacking strategies for integration with other EEG features; and third, traditional analysis methods make it difficult to embed microstate dynamic sequences into deep learning frameworks for end-to-end modeling. Therefore, how to effectively introduce microstate structural information into deep models, integrate their spatiotemporal dynamic features with statistical representation capabilities, and construct a more expressive multi-source feature system has become an important research direction for current EEG intelligent decoding. Summary of the Invention

[0005] The present invention provides an attention deficit hyperactivity disorder (ADHD) recognition system and method, aiming to improve the modeling ability and generalization performance of the model for ADHD EEG dynamic abnormal patterns through the synergistic effect of a multi-source feature fusion mechanism, a CNN-GRU parallel modeling structure, a data enhancement strategy, and a performance feedback adjustment mechanism.

[0006] The present invention is achieved through the following technical solutions: A recognition system for attention deficit hyperactivity disorder, comprising: a data processing module for preprocessing and feature extraction of acquired EEG data, wherein the feature extraction includes multi-source features including microstate feature extraction and EEG statistical features; The extracted raw feature data is first fused using a multi-source feature fusion mechanism to achieve the coordinated expression of static and dynamic EEG patterns. Data enhancement strategies are then applied, including adding Gaussian noise perturbations, cutout occlusion, and enhancing dataset construction. Multi-source fusion feedback regulation network model: Construct a multi-source fusion feedback regulation network model with a CNN-GRU parallel modeling structure to capture local and global multi-scale temporal dependencies. Training module: The enhanced data is input into the model for training. During the training process, the performance feedback adjustment mechanism dynamically adjusts key hyperparameters based on the performance of the validation set, achieving closed-loop adaptive optimization. Classification module: Classify the test samples through the trained multi-source fusion feedback regulation network model and output the final recognition results.

[0007] Preferably, the preprocessing of the data processing module is standardized preprocessing; Specifically, EEG statistical features are extracted from the mne-features library to construct a feature set of 23 required features; Then, we further introduce three types of information theory features: sample entropy, singular value entropy and singular value Fisher information, and finally obtain the constructed EEG feature set; Microstate feature extraction includes GFP peak extraction, global microstate template construction and numbering unification, microstate sequence generation and feature extraction, and global explained variance calculation.

[0008] Preferably, GFP peak extraction is specifically as follows: first, the global field power at each moment of the EEG data is calculated; The global microstate template construction and numbering unification are as follows: microstate clustering is performed on the GFP peak frame of each subject, and the number of clusters is set to K=4, obtaining four microstate topological maps at the subject level; then the cluster centers of all subjects are merged and global clustering is performed using the ModKMeans algorithm to form a unified global microstate template; in order to achieve consistency in state numbering, a standard microstate template is introduced as a reference to calculate the cluster centers With reference template The spatial correlation between Reorder the numbers by the largest correlation to ensure that all microstate templates of the subjects are numbered consistently and are uniformly labeled A, B, C, and D for subsequent identification. Microstate sequence generation and feature extraction specifically involves using a unified global template to perform microstate prediction on each EEG data segment, generating a label sequence, and extracting statistical features such as duration, frequency, coverage, and state transition probability. The global explained variance calculation is as follows: In order to evaluate the fitting ability of the microstate template to the EEG spatial topological changes, the global explained variance GEV indicator is introduced; In summary, the microstate feature output format is as follows: each EEG segment is ultimately converted into a structured feature set in a unified format, including: 12-dimensional microstate statistical features, 12-dimensional state transition probability features, optional 4 GEV features, and auxiliary information.

[0009] Preferably, the data processing module adds Gaussian noise perturbation by adding Gaussian noise with a mean of zero and a standard deviation of 0.05 to the original EEG feature sequence; The Cutout occlusion of the data processing module is specifically to randomly select a continuous time segment in each sample sequence and set the feature value in this interval to zero; The enhanced dataset of the data processing module is constructed by merging the original samples, Gaussian perturbation samples and occlusion samples to generate an enhanced training set of three times the size, while keeping the original labels consistent.

[0010] Preferably, the multi-source feature fusion mechanism of the data processing module is to align the extracted microstate features and EEG statistical features according to the sample dimension and then splice them to form a unified input feature matrix.

[0011] Preferably, the CNN-GRU parallel modeling structure in the multi-source fusion feedback regulation network model includes a convolution branch, a GRU branch, and feature fusion and classification; Convolutional branch: Taking the EEG feature sequence after preprocessing and feature extraction as input, two separable one-dimensional convolutional layers with different convolution kernel sizes are used to extract local temporal features. Batch normalization and activation functions are connected after the convolutional layer, and a residual connection is introduced after the second convolutional layer to enhance feature flow and training stability. The convolution output is max-pooled and flattened to form a one-dimensional local feature representation, which is then regularized by a dropout layer. GRU branch: takes the same input sequence as input, uses the gated recurrent unit (GRU) to extract global temporal dependency information, uses the hidden state of the last time step as the global temporal representation, and performs regularization through the Dropout layer; Feature fusion and classification: The outputs of the two branches are fused into a unified high-dimensional feature representation through a splicing operation. The fusion result is transformed by the fully connected layer and input into the Sigmoid activation unit to output the binary classification prediction result.

[0012] Preferably, the performance feedback regulation mechanism in the multi-source fusion feedback regulation network model includes a dynamic loss weighting mechanism, a hyperparameter resampling and structure reconstruction mechanism, and a strategy reset and robustness enhancement mechanism; Dynamic loss weighting mechanism: Based on the accuracy and loss value of each round of model training on the validation set, a performance feedback signal is built to dynamically adjust the sample weighting factor in the loss function; Hyperparameter resampling and structure reconstruction mechanism: When the model performance does not improve significantly within several consecutive training rounds, the model reconstruction process is automatically triggered. Key training parameters are resampled from the predefined hyperparameter search space and the model structure is reinitialized accordingly to escape the local optimum. Strategy reset and robustness enhancement mechanism: If the parameter adjustment mechanism does not bring performance improvement within consecutive training rounds, the current weighted parameters are randomly reset and new values ​​are sampled from the specified distribution.

[0013] Preferably, the training module includes, in sequence, a training process of enhanced data input, model training, dynamic tuning, closed-loop optimization, and training control mechanism; Model training: The training process is based on the constructed multi-source fusion feedback regulation network model. The model is input with an enhanced dataset and combined with a multi-source feature fusion mechanism to achieve collaborative modeling of static and dynamic EEG features. The CNN-GRU parallel structure is used for time series modeling. Dynamic adjustment and closed-loop optimization: A performance feedback adjustment mechanism is introduced during the training process. The feedback signal is built based on the performance indicators of the validation set, and the model structure and training path are updated in real time. Training control mechanism: Introducing the early stopping mechanism and hyperparameter resampling mechanism. If the model's performance does not improve after several consecutive rounds of training, the hyperparameter reconstruction and weight reset process will be triggered.

[0014] Preferably, the classification module includes the classification and output process of preprocessing and feature extraction of the sample to be tested, model reasoning, classification output and result generation in sequence; Preprocessing and feature extraction of test samples: The EEG data samples to be tested are standardized and preprocessed according to the processing method of the data processing module, and multi-source features consistent with the training stage are extracted, including microstate features and statistical features, to ensure that the input feature dimensions are consistent with the training model; Model inference: The extracted features of the sample to be tested are input into the trained multi-source fusion feedback regulation network model. The model uses the multi-source feature fusion mechanism and CNN-GRU parallel modeling structure to perform feature mapping and time series modeling, and outputs the fused high-dimensional representation vector. Classification output: The fused features are output through the fully connected layer and the Sigmoid activation unit to output the predicted probability value. If the predicted probability is greater than the set threshold θ, it is judged as the ADHD category; otherwise, it is judged as the control category; Result generation: The model ultimately outputs the category labels of the test samples and their corresponding predicted probabilities, which are used to support ADHD status identification in clinical or auxiliary diagnosis scenarios.

[0015] A method for identifying attention deficit hyperactivity disorder (ADHD), using the above-mentioned ADHD identification system, comprising the following steps: Step 1: Preprocess and extract features of the acquired EEG data; Step 2: First use the multi-source feature fusion mechanism on the original feature data extracted in step 1, and then apply the data enhancement strategy; Step 3: Construct a multi-source fusion feedback regulation network model, which is a CNN-GRU parallel modeling structure; Step 4: The data enhanced in step 2 is fed into the model for training. During the training process, the key hyperparameters are dynamically adjusted by the performance feedback adjustment mechanism. The performance feedback adjustment mechanism dynamically adjusts the key training parameters based on the performance of the validation set. Step 5: Classify the test samples through the trained multi-source fusion feedback regulation network model and output the final recognition results.

[0016] The beneficial effects of the present invention are: The present invention achieves high-precision recognition of ADHD EEG signals by fusing EEG microstate features with multi-source time series features and introducing a performance feedback adjustment mechanism, significantly improving the robustness, generalization ability and interpretability of the model.

[0017] The present invention achieves a comprehensive characterization of the dynamic characteristics of EEG signals by integrating EEG microstate characteristics with time domain, frequency domain and nonlinear dynamic characteristics, making the classification accuracy and F1 value significantly higher than the existing technology.

[0018] The present invention adopts the data enhancement strategy of Gaussian noise perturbation and Cutout occlusion to expand the training sample space and effectively avoid the overfitting problem under small sample conditions.

[0019] The present invention adopts a dynamic hyperparameter optimization mechanism based on performance feedback, which enables the model to automatically adjust the dropout rate, regularization term and learning rate, improve the convergence speed and reduce the sensitivity to noise and non-stationary data.

[0020] The present invention adopts local convolution and global temporal series modeling in parallel, which can efficiently capture the multi-scale spatiotemporal dependencies of EEG signals and improve the feature expression of the model.

[0021] Through EEG microstate analysis, the present invention not only improves recognition performance, but also reveals the dynamic abnormal patterns of ADHD patients in the neural functional network, providing an interpretable reference basis for clinical diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a logical structure diagram of the performance feedback adjustment mechanism of the present invention.

[0023] Figure 2 It is a heat map of the performance of each model in the embodiment of the present invention on the four evaluation indicators.

[0024] Figure 3 These are four topographic maps of microstates of the present invention, where A is microstate A, B is microstate B, C is microstate C, and D is microstate D.

[0025] Figure 4 This is a comparison chart between groups of the average duration of the microstate of the present invention.

[0026] Figure 5 It is a comparison chart between groups of the microstate time coverage of the present invention. DETAILED DESCRIPTION

[0027] The following is a clear and complete description of the technical solutions in the embodiments of this application in conjunction with the drawings in the specification of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0028] Implementation Method 1 This embodiment provides an attention deficit hyperactivity disorder recognition system, which includes a data processing module: preprocessing and feature extraction of acquired EEG data, wherein the feature extraction includes multi-source features including microstate feature extraction and EEG statistical features; The experimental data for this study were obtained from the publicly available EEG dataset "EEGdataforADHD / Controlchildren." This dataset includes 121 children aged 7-12 years (61 with ADHD and 60 healthy controls). All subjects were both male and female. The ADHD group was diagnosed by a professional psychiatrist according to DSM-IV criteria, while the control group had no history of psychiatric or neurological disorders. EEG acquisition was based on the international 10-20 system, consisting of 19 leads (Fz, Cz, Pz, C3, T3, C4, T4, Fp1, Fp2, F3, F4, F7, F8, P3, P4, T5, T6, O1, O2), with a sampling frequency of 128 Hz and earlobe reference. All EEG data underwent unified preprocessing, including bandpass filtering, artifact removal, and independent component analysis.

[0029] The extracted original feature data is first subjected to a multi-source feature fusion mechanism, followed by a data enhancement strategy, which includes adding Gaussian noise perturbation, cutout occlusion, and enhancing the dataset construction. Multi-source fusion feedback regulation network model: Construct a multi-source fusion feedback regulation network model, which is a CNN-GRU parallel modeling structure, used to capture local and global multi-scale temporal dependencies respectively.

[0030] Construct a deep classification model that integrates data enhancement, time series modeling, and performance feedback adjustment mechanism. The deep classification model includes a data enhancement strategy part, a CNN-GRU fusion feature modeling part, and a performance feedback adjustment mechanism. The logical structure of the mechanism is as follows: Figure 1 As shown; Training module: The enhanced data is input into the model for training. During the training process, the performance feedback adjustment mechanism dynamically adjusts key hyperparameters. The performance feedback adjustment mechanism dynamically adjusts key training parameters such as the dropout rate, regularization term, and learning rate based on the performance of the validation set to achieve closed-loop adaptive optimization. Classification module: Classify the test samples through the trained multi-source fusion feedback regulation network model and output the final recognition results.

[0031] Preferably, the preprocessing of the data processing module is standardized preprocessing, which aims to enhance the ability to identify and correct multiple types of artifacts and preserve the original neural signal characteristics as much as possible; First, to suppress low-frequency drift and high-frequency interference, the raw EEG signal is bandpass filtered from 0.5 to 45 Hz. This frequency band covers the main EEG rhythms (delta, theta, alpha, beta, and low-frequency gamma) while also removing high-frequency artifacts such as power supply interference (50 Hz) and eye movements and myoelectricity. Filtering is performed using an FIR filter to ensure a stable phase response and preserve the temporal structure.

[0032] The Artifact Subspace Reconstruction (ASR) algorithm was then introduced to automatically identify and reconstruct high-amplitude transient artifacts. Unlike traditional artifact removal methods, ASR constructs a clean subspace through principal component analysis (PCA) and projects abnormal components back into the reference space, achieving a "restorative" signal purification. To improve artifact resilience, the present invention uses a more conservative dual threshold parameter of [0.25, 0.75] to enhance tolerance to high-frequency fluctuations.

[0033] On the basis of ASR processing, average re-referencing is further applied to reduce the influence of the reference electrode on the overall potential distribution, providing a more stable spatial filtering basis for subsequent independent component analysis (ICA).

[0034] The ICA algorithm is then used to extract potential neural source signals. As a spatial filter, ICA decomposes the observed signal into a set of statistically independent components. This paper utilizes the pop_runica function in EEGLAB to achieve independent component separation and combines it with the ICLabel tool to automatically classify the components. To enhance artifact removal accuracy, only components classified as "brain" are retained. Components with a confidence level greater than 0.7 are used to reconstruct the EEG signal to eliminate non-neural artifacts such as eye movements, electromyography, electrocardiography, and line noise.

[0035] This preprocessing strategy effectively suppresses multi-class artifact interference while retaining neural information to the maximum extent, providing a reliable input basis for subsequent microstate feature extraction and deep classification model (MSRLNet) combined with performance feedback regulation mechanism.

[0036] The original EEG signals without preprocessing contained significant high-frequency artifacts and background noise interference. However, after ASR, re-referencing and ICA removal, the high-frequency components were significantly reduced and the low-frequency rhythm structure became clearer, proving the effectiveness of this preprocessing scheme in removing artifacts and retaining neural information.

[0037] Feature extraction includes: Specifically, the EEG statistical features were extracted using the mne-features library to extract 26 common EEG features. Then, the random forest feature importance evaluation method based on the Gini index was used to screen and remove three features with low contribution. 23 features were retained to construct a preliminary feature set. The features included standard deviation, peak amplitude, skewness, kurtosis, root mean square value, quantile, Hurst exponent, approximate entropy, decorrelation time, power spectrum, Hjorth mobility (frequency domain), Hjorth complexity (frequency domain), Hjorth mobility (time domain), Hjorth complexity (time domain), Higuchi fractal dimension, Katz fractal dimension, number of zero crossings, line length, linear regression of log-log frequency curve, spectral entropy, band energy, spectral edge frequency, and energy of wavelet decomposition coefficients. On this basis, combined with the performance of ADHD EEG signals in terms of information complexity and non-stationarity - that is, significant dynamic fluctuations and structural changes in the time domain, multi-scale non-stationary characteristics in the frequency domain, and large pattern differences between different individuals - in order to enhance the feature set's ability to characterize such high time-variability and complex patterns, three types of information theory features from the mne-features library are further introduced: sample entropy (used to measure the irregularity and complexity of time series), singular value entropy (used to measure the uniformity of information distribution after signal matrix decomposition), and singular value Fisher information (used to characterize the sensitivity of signal structure in the frequency and time domains).

[0038] Finally, a total of 26 categories and 40 dimensions were included to comprehensively describe the volatility, spectral distribution and multi-scale complexity of the signal, supporting subsequent model training and discriminant analysis; Sample entropy is a measure of the change in the self-similarity of the signal under adjacent embedding dimensions; singular value entropy is an entropy measure based on the singular value distribution of the signal reconstruction matrix; singular value Fisher information is a measure of the local unevenness and orderliness of the singular value distribution; Preferably, the sample entropy is defined as follows:

[0039] in, Represents the dimension , the tolerance is Under the conditions, The number of similar subsequences; The singular value entropy is defined as:

[0040] in, The first singular values, The dimensionality threshold for retaining the main energy; The expression of the singular value Fisher information is:

[0041] in, , In order to prevent the small constant from being divided by zero, that is, to prevent the extremely small constant introduced by the denominator being zero, the present invention is set to be .

[0042] Microstate feature extraction includes GFP peak extraction, global microstate template construction and numbering unification, microstate sequence generation and feature extraction, and global explained variance calculation. Based on the microstate label sequence, characteristic indicators representing the spatiotemporal dynamic behavior of EEG are extracted, including the duration, occurrence frequency, coverage and inter-state transition probability of each microstate, which are used to characterize the short-term stability and state switching rules of the brain.

[0043] To characterize transient stable topological states in EEG signals, this paper uses microstate analysis to model signal segments. This method effectively reveals the short-term information processing of dynamic brain networks and extracts structured features encompassing multidimensional indicators such as persistence, transferability, and spatial distribution. The entire microstate feature extraction process includes GFP peak extraction, global microstate template construction and numbering, microstate sequence generation and feature extraction, and global explained variance calculation.

[0044] Preferably, GFP peak extraction is specifically performed by first performing time-limited truncation (20–60 seconds), passband filtering (0.5–45 Hz), and average reference resetting on the EEG data; then, based on the global field power (GFP) indicator, the time point with the strongest EEG activity is identified for subsequent microstate clustering; The formula is as follows:

[0045] in, Indicates the Channels at time The electric potential, is the average electrode potential at that moment, is the number of channels; the larger the GFP value, the greater the difference in spatial potential distribution at that moment, and the more likely it is to correspond to a typical microstate.

[0046] The global microstate template is constructed and numbered uniformly as follows: the GFP peak topology map of each subject is clustered by the ModKMeans algorithm (number of clusters K = 4) to obtain four microstate topology maps at the subject level; then the cluster centers of all subjects are merged and globally clustered by the ModKMeans algorithm to form a unified global microstate template; in order to achieve consistency in state numbering, the standard microstate template is introduced as a reference to calculate the cluster centers With reference template The spatial correlation between them, the matching criteria is:

[0047] in, Indicates the clustering result A central diagram, For the template map, the maximum correlation pairing is finally selected; the numbers are reordered according to the maximum correlation to ensure that the numbers of all subject microstate templates are consistent and uniformly marked as A, B, C, and D for subsequent identification.

[0048] Microstate sequence generation and feature extraction specifically involves using a unified global template to predict the microstates of each EEG data segment to obtain a label sequence. Based on the microstate sequence, the following statistical features are extracted: 1. Duration (MeanDuration);

[0049] in is the number of occurrences of microstate k, is the duration of the i-th occurrence; 2. Occurrence: Count the number of occurrences of microstate k ; 3. Coverage

[0050] in represents the total duration of microstate k, and T is the total duration of the entire EEG segment; 4. State transition probability (Transition Matrix) constructs a 4×4 state transition matrix, retaining the 12 off-diagonal elements after removing the diagonal (self-transitions) as features; The global explained variance (GEV) is calculated as follows: In order to evaluate the ability of the microstate template to fit the changes in EEG spatial topology, the global explained variance (GEV) indicator is introduced and defined as follows:

[0051] in Indicates that all tags are in state A collection of time points, is the EEG spatial distribution at time t, is the microstate template, is the Pearson correlation coefficient between the two; Finally, the feature output format is: Each EEG segment is finally converted into a structured feature set in a unified format containing: 12-dimensional microstate statistical features (4 states × duration, frequency, coverage); 12-dimensional state transition probability features (off-diagonal elements of the transition matrix after removing self-transitions); Optional 4 GEV features; auxiliary information: corresponding sample labels and file index names The above microstate features constitute one of the core inputs supporting the training of the ADHD automatic recognition model and serve as an important feature channel of the MSRLNet model.

[0052] Optimally, in the data enhancement strategy, Gaussian noise perturbation and cutout occlusion are used to enhance the diversity of EEG sequences, effectively expanding the training set sample space and improving the model robustness and generalization ability; CNN-GRU fusion feature modeling: Design a multi-branch structure to extract local temporal convolution features and global temporal dependency information respectively, and achieve a comprehensive characterization of EEG dynamic patterns through feature fusion; Performance feedback adjustment mechanism: A lightweight adjustment module is designed to dynamically adjust key hyperparameters during training (such as loss weight, dropout rate, and learning rate) based on the model's performance feedback on the validation set. This mechanism draws on the reward-feedback concept in reinforcement learning, but differs from the agent-environment interaction and policy learning process in traditional reinforcement learning. In essence, it is a simplified feedback control strategy for training optimization tasks.

[0053] The MSRLNet model's data processing, network model, and training strategy are synergistically composed of a multi-source feature fusion mechanism, a CNN-GRU parallel modeling architecture, a data augmentation strategy, and a performance feedback mechanism. This improves the accuracy of automatic ADHD recognition while also ensuring interpretability and stability. The following describes the structural design and mathematical modeling process for each module.

[0054] Data enhancement strategy: Gaussian noise perturbation and cutout occlusion are used to enhance the diversity of EEG sequences and construct an enhanced dataset; CNN-GRU fusion feature modeling: Design a multi-branch structure to extract local temporal convolution features and global temporal dependency information respectively, and achieve a comprehensive characterization of EEG dynamic patterns through feature fusion; Performance feedback adjustment mechanism: Design lightweight adjustments to dynamically adjust key hyperparameters during training based on the model's performance feedback on the validation set.

[0055] Preferably, the Gaussian noise perturbation is specifically as follows: Gaussian noise enhancement is intended to simulate the inevitable background electrophysiological perturbation in the actual acquisition process, and to enhance the robustness of the model to small perturbations by adding slight noise to the original features; for the original input feature matrix ,in, represents the number of samples, For the time dimension length of each sample, the enhanced feature is expressed as:

[0056] in, Represents Gaussian white noise matrix of the same shape, standard deviation Set to 0.05. This strategy is used to simulate weak electrophysiological perturbations during the acquisition process and improve the robustness of the model to input uncertainties; Random occlusion enhancement is specifically: Cutout is a temporal occlusion enhancement strategy that aims to improve the model's ability to distinguish when local features are missing. The basic idea is: for each training sample sequence , randomly selecting a starting position in its sequence , and from the position Start, length The continuous time window is set to zero, and the enhanced sample is expressed as :

[0057] in, Indicates the occlusion window length, which is set to , that is, occluding 10% of the time steps in the sample to enhance the model's discrimination ability in the absence of local features.

[0058] The specific steps of constructing the enhanced dataset are as follows: original, noisy and occluded samples are spliced ​​in the training set, so that the number of training samples is increased from Expand to , the corresponding labels are also expanded synchronously; the original sample X, the noise sample With occlusion samples Splicing in the sample dimension to form enhanced training data:

[0059] The corresponding tag set is also expanded synchronously to:

[0060] With the help of the above enhancement strategy, although the original number of subjects is limited, the training data is expanded by about three times in the feature space, significantly alleviating the training bias problem caused by insufficient individual samples. While maintaining label consistency, the diversity expansion of the feature space is achieved, thereby alleviating the training bias under small sample conditions and improving the generalization ability and robustness of the model.

[0061] Preferably, the convolution branch modeling is specifically as follows: the input of the convolution branch is the EEG data matrix ,in, represents the time step, Represents the feature dimension (number of fused features) at each time point; This branch first uses two separable one-dimensional convolutional layers (SeparableConv1D) with different convolution kernel sizes to extract short-term and medium-term temporal dependency structures respectively. The convolution kernel size of the first layer is 3 and that of the second layer is 5, which are defined as follows:

[0062] in, Indicates that the convolution kernel size is Separable one-dimensional convolution operation, BN is batch normalization, and ReLU is the activation function; In order to improve the network's ability to integrate features of different scales and suppress the degradation problem of deep networks, a residual connection is introduced after the second layer output; After 1×1 convolution compression and Add and activate to get the final convolution output:

[0063]

[0064] The convolutional features are then compressed in time using the MaxPooling operation and flattened into a one-dimensional vector. To reduce the risk of overfitting, they are finally regularized using the Dropout layer:

[0065] The result That is the local time feature representation of the convolution branch; The timing modeling branch is specifically, in order to model the long-term dependency characteristics of the signal, the GRU branch directly takes the input As sequence input, the hidden state of the last time step is output as the global temporal representation; This branch uses a single-layer GRU unit to extract the global temporal dependency of the EEG sequence. The core calculation formula is as follows: Update Gate: Reset the gate:

[0066] Candidate status:

[0067] Final output status:

[0068]

[0069]

[0070]

[0071] in represents the Hadamard product, is the Sigmoid activation function; the hidden state of the last time step is used as the GRU output and is recorded as follows after Dropout regularization:

[0072] Feature fusion and classification output are specifically to combine the features from CNN and GRU branches and The two are fused through feature concatenation to construct a unified high-dimensional representation:

[0073] The fused features are then input to the fully connected layer for feature conversion, and the final prediction probability is output through the Sigmoid unit:

[0074] in, For the prediction result, it indicates the probability that the input sample is the target category (such as ADHD).

[0075] Preferably, in order to enhance the training stability and generalization ability of the model in complex neural signal classification tasks, the present invention introduces a performance feedback regulation mechanism to achieve dynamic adjustment of loss weights and key hyperparameters during the training process. This mechanism essentially draws on the reward feedback and strategy update ideas in reinforcement learning, but does not involve the interaction between the intelligent agent and the environment, state transition modeling or policy function learning in the traditional sense. Therefore, it is more accurately said that it is a simplified control strategy for training optimization tasks. The dynamic loss weighting mechanism is specifically that after each round of training, the system constructs a performance feedback reward signal R based on the performance of the validation set to measure the performance of the current model state, which is calculated as follows:

[0076] in, Indicates the verification accuracy of the current round, is the corresponding verification loss; The larger the reward value, the closer the model is to high accuracy and low error; Based on this reward feedback, the current loss weighting factor Updates are performed using the following incremental strategy:

[0077] in, is the update step size, used to adjust the weighted intensity; To ensure training stability, weights Restricted to a fixed range:

[0078] Weighting Factor It is used to construct a weighted loss function to perform weighted adjustment on the sample error. For the binary classification task, the present invention uses the following form of weighted binary cross entropy loss function:

[0079] This design makes the model highly sensitive to performance fluctuations, thereby enhancing the suppression of generalization errors in the early stages of training and enabling stable updates during the convergence phase; Specifically, in traditional training strategies, hyperparameter settings are usually statically specified before training, which cannot effectively adapt to the dynamically evolving error distribution. This study introduces a policy control mechanism to monitor continuous performance stagnation during model training and trigger the model reconstruction process when the following conditions are met:

[0080] in, is the performance tolerance factor, The no-lift threshold is set in advance. ; After the above conditions are met, the system randomly samples a new set of key hyperparameters from the predefined search space for replacement. The search space is as follows:

[0081]

[0082]

[0083] After resampling, the model structure is automatically reinitialized and training is restarted with the new configuration, preventing the original model from falling into performance bottlenecks; The specific strategy reset and robustness mechanism is that, in order to further improve the robustness and exploration ability of the agent system, if the loss weighting factor w is continuously When the round weight update does not bring performance improvement, the system will resample the initial weights from the specified uniform distribution and reset :

[0084] This strategy is essentially a parameter perturbation method inspired by the "strategy perturbation" of reinforcement learning. This mechanism is used to escape from the local optimal solution, improve the diversity of parameter search and the overall robustness of the training process.

[0085] Specifically evaluate the performance of the MSRLNet model in the ADHD recognition task. Through core indicator display, innovation verification, ablation experiments and baseline comparison, the effectiveness and superiority of the model are demonstrated layer by layer to verify the effectiveness of the model and the actual role of key modules.

[0086] The workflow of the training module is: (1) Enhanced data input: The extracted original feature data is first used using a multi-source feature fusion mechanism, and then a data enhancement strategy is applied. The data enhancement strategy includes adding Gaussian noise perturbation, cutout occlusion and enhancing the dataset to construct and splice it into a training input dataset and label set, which is then used as the training input of the multi-source fusion feedback regulation network model.

[0087] (2) Model training: The training process is based on the constructed multi-source fusion feedback regulation network model, which uses the multi-source feature fusion mechanism combined with the enhanced data set as the input of the model, and the CNN-GRU parallel structure as the time series modeling. Among them, the multi-source feature fusion mechanism realizes the coordination of static and dynamic EEG features.

[0088] (3) Dynamic adjustment of closed-loop optimization: A performance feedback adjustment mechanism is introduced during the training process. The feedback signal is constructed through the performance indicators of the validation set to dynamically adjust key training parameters such as the Dropout rate, L2 regularization term and learning rate, and the model structure and training path are updated in real time accordingly.

[0089] (4) Training control mechanism: In order to improve training stability and convergence efficiency, the process further introduces the early stopping mechanism and hyperparameter resampling mechanism. If the model verification performance does not improve in several consecutive rounds of training, the hyperparameter reconstruction and weight reset process will be triggered to achieve robustness optimization.

[0090] The workflow of the classification module is: (1) Preprocessing and feature extraction of test samples: The EEG data samples to be tested are standardized and preprocessed, and multi-source features consistent with the training stage are extracted, including microstate features and statistical features, to ensure that the input feature dimension is consistent with the training model.

[0091] (2) Model reasoning: The extracted features of the sample to be tested are input into the trained multi-source fusion feedback regulation network model. Feature mapping and time series modeling are performed through the multi-source feature fusion mechanism and CNN-GRU parallel modeling structure in the model, and the fused high-dimensional representation vector is output.

[0092] (3) Classification output: The fused features are output through the fully connected layer and the Sigmoid activation unit to predict the probability value. If the predicted probability is greater than the set threshold θ (for example, θ = 0.5), it is judged as the ADHD category; otherwise, it is judged as the control category.

[0093] (4) Result generation: The model ultimately outputs the category label of the sample to be tested and its corresponding predicted probability, which is used to support ADHD status identification in clinical or auxiliary diagnosis scenarios.

[0094] The ADHD automatic recognition method can be deployed and used in any of the following EEG recognition systems: (1) Computer-based software system: The method is embedded in a computer-side software program, which performs batch recognition and analysis by calling EEG data files and outputs ADHD recognition results, which is suitable for scientific research and data post-processing tasks; (2) Analysis system integrated into EEG acquisition equipment: The method is integrated into a hardware system with EEG acquisition function to achieve integrated processing of acquisition and recognition, which is suitable for screening purposes in routine medical or educational scenarios; (3) Portable EEG recognition terminal: The method is deployed in portable devices equipped with basic computing capabilities, such as handheld analyzers, tablets or laptops, to assist in ADHD recognition in a small range.

[0095] Experimental design and reliability description: All experiments in this study were conducted on a high-performance workstation equipped with an NVIDIA GeForce RTX 2080 Ti GPU. Model building and training were implemented using the TensorFlow 2.x framework to ensure efficient computational performance and experimental reproducibility. During model training, the Adaptive Moment Estimation (Adam) optimizer was used, with a moderate initial learning rate and batch size, and a maximum number of 500 training epochs. To prevent overfitting, an early stopping mechanism was introduced during training. Training was automatically terminated when the validation set loss did not significantly decrease over several consecutive epochs. Furthermore, a dynamic learning rate adjustment strategy was employed to gradually reduce the learning rate as training performance approached saturation, enabling more refined model tuning.

[0096] Given that data in ADHD EEG recognition tasks typically exhibit the structural characteristics of a limited number of subjects and a large number of sample segments for each subject, this paper introduces a time-series-based feature-level data augmentation strategy during model training, including Gaussian noise perturbation and random masking mechanisms, to improve the model's generalization and robustness. The former enhances the model's robustness to minor noise interference by introducing slight perturbations to the original features; the latter strengthens the model's adaptability to local information loss by randomly masking information in some time windows. The augmented samples maintain consistency with the original samples in feature dimensions and are concatenated, resulting in a three-fold expansion of the training sample space. This strategy effectively mitigates the training bias caused by the scarcity of individual samples, provides more robust and diverse optimization support for deep models, and further improves the learning and generalization performance of ADHD EEG differential patterns.

[0097] To further enhance the model's stability and adaptability during training, the present invention also introduces a performance feedback mechanism. This mechanism constructs a reward signal based on the difference between the validation set accuracy and loss value in each round, dynamically updating key training parameters, including the dropout rate, L2 regularization coefficient, and learning rate. If the system detects that the model's performance has stagnated over several consecutive rounds of training, it automatically triggers hyperparameter resampling and structural reconstruction, effectively avoiding the problem of being trapped in a local optimal solution and improving the model's training robustness and generalization ability.

[0098] In terms of data preprocessing, EEG statistical features and microstate features were first fused at the sample level and normalized using Z-score standardization. Subsequently, using a subject-independent 5-fold cross-validation scheme (5-fold subject-wise CV), all 121 subjects were randomly divided into five non-overlapping subsets. One fold was used as the test set, and the remaining four folds were used for training and validation (10% of the folds were further divided into a validation set to enable early-stopping). This strategy ensured that data from the same subject did not appear in both the training and testing phases. After five rotations, the mean ± standard deviation of each metric was taken as the final performance.

[0099] After model training is complete, the system automatically saves the model weights and hyperparameter combinations that achieved the best performance on the validation set. It also outputs the accuracy curve, loss trend, confusion matrix, and various classification evaluation metrics during training, facilitating subsequent visualization and reproducible experimental analysis. This systematic integration of the training process, from data construction and model design to optimization strategies and evaluation mechanisms, ensures the effectiveness and reliability of the proposed MSRLNet model in ADHD EEG recognition tasks.

[0100] Evaluation indicators: To comprehensively evaluate the classification performance of the proposed model in the ADHD EEG recognition task, this paper uses multiple common indicators including accuracy, Cohen's Kappa coefficient (κ), F1 score, precision, recall, and root mean square error (RMSE), and combines the confusion matrix to visualize the model's discriminative ability.

[0101] Among them, the accuracy is used to measure the correctness of the overall classification of the model, and the calculation formula is as follows:

[0102] in, 、 Represents the number of correctly classified positive and negative samples, 、 The closer the accuracy value is to 1, the stronger the overall classification ability of the model is.

[0103] In order to further evaluate the consistency between the model classification results and the actual labels, the present invention introduces Cohen's Kappa coefficient As a supplementary evaluation indicator, it is defined as:

[0104] in, represents the actual probability of agreement between the model prediction result and the true label, It represents the expected probability of agreement under the random classification assumption. The closer the Kappa value is to 1, the higher the consistency between the model prediction and the actual label, and the stronger the robustness.

[0105] In addition to overall accuracy and consistency metrics, this paper also evaluates the model's ability to identify positive samples and the distribution of errors. Precision measures how many of the samples predicted by the model as positive are actually positive, while recall indicates how many of the samples the model correctly identifies as positive. The F1 score, the harmonic mean of precision and recall, more comprehensively reflects the model's classification ability in the case of class imbalance. Its calculation formula is as follows:

[0106]

[0107]

[0108] In addition, the present invention also introduces the root mean square error (RMSE) to measure the overall deviation between the predicted result and the true label. The calculation formula is:

[0109] in, and Respectively represent The smaller the RMSE is, the closer the model prediction is to the true result.

[0110] In order to further analyze the classification results from a visual perspective, the present invention draws a confusion matrix (Confusion Matrix) to intuitively display the classification accuracy and misclassification of the model between the ADHD group and the control group, thereby facilitating the identification of the distribution characteristics of false positive and false negative samples and assisting in understanding the model's discrimination mechanism.

[0111] In summary, this paper systematically verifies the classification ability and practical application value of the proposed model MSRLNet in the ADHD EEG recognition task through multi-dimensional performance evaluation indicators and result visualization analysis.

[0112] To comprehensively evaluate the effectiveness and robustness of MSRLNet in the ADHD EEG automatic recognition task, this section first presents the core performance of the model under five-fold cross-validation (Table 1) and compares the results with representative models previously studied on the same dataset (Table 2). Table 1 Performance of MSRLNet in five-fold cross validation

[0113] As can be seen, MSRLNet achieved high scores of over 98% on all six evaluation indicators, with a K value close to 1, indicating that the prediction results are highly consistent with the true labels. The error (RMSE) is controlled within 0.11, which fully demonstrates that the model still maintains stable performance in scenarios with sparse samples and class imbalance.

[0114] Table 2 Horizontal comparison with existing models

[0115] In order to more intuitively compare the performance of each model on key performance indicators, Figure 2 Heat maps further illustrate the specific values ​​and relative strengths of different models across four metrics: Accuracy, F1 score, Precision, and Recall. It's clear that MSRLNe achieves the highest color intensity across all four core metrics, demonstrating its overall leading performance. Att-CNN and GPC are close to or slightly behind this model in some metrics, while traditional machine learning methods lag significantly behind in F1 and Precision.

[0116] Compared with EEG-MHCNet based on the multi-head convolutional attention mechanism of SanchisJ et al., MSRLNet has improved by 22.81% in Accuracy, showing the strong capture ability of multi-source fusion for weak spatiotemporal patterns; compared with the latest Att-CNN, Accuracy is still improved by 0.02% and F1 is improved by 0.68%, indicating that adaptive tuning of micro-state features can still bring considerable gains at the top level; compared with traditional machine learning methods such as GPC, CNN+LR, LASSO+SVM, the overall advantage of MSRLNet remains between 1.3-5.5%, and the Precision-Recall gap is significantly narrowed, reflecting the adaptability of deep time series modeling to complex non-stationary signals.

[0117] Based on a comprehensive analysis of quantitative metrics and horizontal comparisons, MSRLNet achieved the highest Accuracy and F1 scores on the ADHD EEG automatic recognition task, and was ahead of or on par with the best deep learning models in metrics such as Kappa, Precision, and Recall. Its microstate fusion architecture, data augmentation strategy, and adaptive regulation mechanism contribute to its performance advantages and provide an interpretable physiological basis for abnormal brain dynamics. Subsequent research will further examine the model's generalization and clinical applicability through multimodal data and cross-center validation.

[0118] Ablation and Baseline Comparison To deeply analyze the specific contributions of each functional module in the MSRLNet model to the automatic ADHD recognition performance, we designed a series of ablation experiments. We specifically removed and evaluated key components of the model architecture: the convolutional branch (CNN), the recurrent branch (GRU), the microstate feature fusion module, the performance feedback adjustment mechanism (RL_Agent), and the data augmentation strategy (including Gaussian noise and random occlusion). We examined their role in the complete model architecture and the performance fluctuations caused by their absence. The experimental results are shown in Table 3.

[0119] Table 3 Ablation experiment results: Analysis of the impact of each module on MSRLNet performance

[0120] Overall, each module makes an irreplaceable positive contribution to model performance. Particularly noteworthy is that removing the CNN branch leads to a precipitous decline in model performance, with accuracy plummeting to 77.20%, the Kappa coefficient plummeting to 0.533, and the F1 value dropping to only 80.62%. This demonstrates that the CNN structure plays a central role in capturing local spatial patterns in EEG signals. In contrast, removing the GRU branch has a relatively mild impact, maintaining model accuracy at 97.27% and a Kappa of 0.955. This suggests that it primarily supports modeling temporal dependencies and capturing global contextual information. These two components work together to form the backbone of MSRLNet for processing dynamic EEG features.

[0121] In terms of feature fusion, eliminating the microstate feature module also led to a significant degradation in model performance, with accuracy dropping to 88.70% and Kappa dropping to 0.866. Combined with the Precision (87.25%) and Recall (92.68%) performance, the model's ability to distinguish between ADHD and control samples became blurred, demonstrating that microstate features possess unique discriminative value in revealing spatiotemporal transient patterns of neural activity, and are significantly complementary to statistical features.

[0122] In addition, after removing the performance feedback adjustment mechanism, the model accuracy dropped to 96.56%. Although the absolute value remained at a high level, the stability and convergence speed of the training process decreased significantly, indicating that this module is of key significance in optimizing hyperparameter configuration and enhancing the robustness of the training process.

[0123] It's worth noting that the data augmentation strategy also has a positive impact on the model's generalization ability. After removing data augmentation, the model's accuracy dropped to 94.82%, with a Kappa of 0.895. This demonstrates that Gaussian noise perturbation and cutout operations effectively support overfitting and simulate signal fluctuations, making them particularly suitable for small-sample EEG learning tasks.

[0124] Overall, the complete MSRLNet model achieved optimal performance across all evaluation metrics, achieving an accuracy of 98.90%, a Kappa coefficient of 0.979, an F1 score of 98.98%, and the lowest root mean square error (RMSE) of 0.104. These results fully validate the rationality of the design of each module and the effectiveness of the collaborative mechanism. The CNN-GRU dual-branch architecture and microstate feature fusion provide the primary support for classification performance, while data augmentation and adaptive adjustment mechanisms provide strong guarantees for model generalization and convergence efficiency, forming the foundation of MSRLNet's efficient and robust EEG recognition framework.

[0125] In the aforementioned ablation experiments, the introduction of microstate features was shown to significantly improve the classification performance of the ADHD EEG automatic recognition model. To further clarify the actual contribution of microstate features in neural representation and classification tasks, this section analyzes them from two perspectives: classification performance comparison and feature statistics.

[0126] The full model, after incorporating microstate features, demonstrated improved classification accuracy and error reduction. Specifically, the recognition accuracy for the control group (negative class) reached 98.4%, and for the ADHD group (positive class) reached 98.7%. Compared to the model without microstate features, the accuracy dropped significantly, reaching only 84.0% for the control group and 92.7% for the ADHD group. The model experienced a significant increase in error rates when microstate features were missing, demonstrating the critical role of microstate parameters in the discriminative model.

[0127] like Figure 3 As shown in Figure 3, microstate analysis is an important way to model the spatiotemporal characteristics of EEG. The four types of microstates A, B, C, and D, which are divided based on the instantaneous spatial patterns of whole-brain electrical signals, have been widely studied and are considered to be closely related to specific functional neural network activities.

[0128] Microstate A: Usually associated with bilateral temporal lobe and language-related network activities, representing a stable or initial cognitive state; Microstate B: Mainly corresponds to the right frontoparietal network or visual attention area, often associated with processes such as increased alertness, sensory mobilization and attention shifting; Microstate C: Believed to be related to the posterior default mode network (pDMN), involved in introspection, integrative processing and cognitive regulation; Microstate D: Associated with the midline area of ​​the frontal lobe (ACC) and the prefrontal control network, commonly seen in executive function, impulse inhibition and situational regulation tasks.

[0129] Therefore, the differences in the expression of different microstates between ADHD and normal subjects may reflect differences in the neural basis of their brain functional connectivity networks in terms of activation, attention maintenance, impulse control, and integrated regulation. Next, we will further reveal this structural shift based on statistical analysis of microstate parameters.

[0130] Microstate features not only demonstrate strong discriminative power in terms of model performance but also reveal systematic shifts in ADHD EEG dynamic patterns at the physiological mechanism level. Table 4 summarizes the differences between the two groups in typical microstate transition pathways. ADHD patients tend to transition rapidly from stable state A to activated state B, frequently cycling between D and B, exhibiting a pattern of sustained high activation, while the probability of transitioning to integrated state C is significantly reduced. This phenomenon reflects that their neural activity tends to be "activation-excitement" oriented and lacks effective regulatory and convergence mechanisms.

[0131] Table 4 Difference analysis of microstate transition pathways between ADHD and control groups

[0132] In addition, Tables 5 and 6 show the inter-group differences in the mean duration (MeanDuration) and time coverage (Coverage) of microstates.

[0133] Table 5 Differences in average duration of microstates

[0134] Table 6 Differences in microstate time coverage

[0135] At the same time, in order to more intuitively show the differences between ADHD and normal subjects in the characteristics of each microstate, Figure 4 and Figure 5 A box plot comparing the average duration and time coverage of the four types of microstates is drawn.

[0136] In summary, ADHD subjects exhibited systematic shifts in microstate dynamics, primarily including more frequent entry into the hyperactive state (B) and reduced transitions to the integrated state (C). These shifts not only align with the clinical features of attentional dysregulation and increased impulsivity, but also validate the physiological value and discriminative potential of microstate characteristics in identifying abnormal EEG dynamics in ADHD.

[0137] This paper conducts a systematic empirical study and multi-dimensional analysis of the MSRLNet model's performance in ADHD EEG recognition tasks. Through five-fold cross-validation on a publicly available ADHD dataset, MSRLNet achieves excellent results across multiple core metrics, including accuracy, F1 score, and Kappa coefficient. Its overall performance is stable and error-controllable, demonstrating strong adaptability and generalization capabilities to small sample sizes and non-stationary EEG signals.

[0138] In the model comparison analysis, MSRLNet achieved performance improvements compared with existing representative deep learning models (such as Att-CNN, EEGMHCNet) and traditional machine learning methods (such as GPC, SVM, etc.), especially maintaining its leading position in indicators such as F1 and Precision that measure model stability and discriminability, verifying the effectiveness of the multi-source fusion architecture and deep time series modeling strategy.

[0139] Ablation experiments further reveal the functional contributions of each model module: the dual-branch structure composed of CNN and GRU is the key support for modeling spatiotemporal features, microstate features significantly enhance the model's discriminative capabilities, data augmentation strategies effectively mitigate the risk of overfitting, and the performance feedback adjustment mechanism provides a robust training optimization path for the model. The synergistic cooperation between these modules provides a systematic guarantee for model performance.

[0140] At the same time, the introduction of microstate features not only shows significant gains in model classification performance, but also reveals the dynamic imbalance of ADHD patients in activation, integration and regulation states at the neural mechanism level, reflecting a high degree of physiological interpretability, and providing a theoretical basis and practical support for the subsequent construction of an interpretable EEG recognition model.

[0141] In addition, the performance feedback adjustment mechanism builds a closed-loop training optimization process through performance feedback, which improves training efficiency while enhancing the model's dynamic focus on key features. It is one of the core strategies to promote accurate and efficient model learning.

[0142] In summary, the proposed MSRLNet model achieves a coordinated optimization of performance, robustness, and interpretability in the ADHD EEG recognition task, validating the scientific and advanced nature of the model architecture and design strategy. Further research will be conducted in areas such as multi-center data, cross-modal integration, and clinical application to promote the practical application and generalization of the model.

[0143] Implementation Method 2 This embodiment provides a method for identifying attention deficit hyperactivity disorder (ADHD). The method uses the ADHD identification system described in embodiment 1. The method includes: Step 1: Preprocess and extract features from the acquired EEG data. Feature extraction includes multi-source features such as microstate feature extraction and EEG statistical features. Step 2: The original feature data extracted in step 1 is first subjected to a multi-source feature fusion mechanism, followed by a data enhancement strategy, which includes adding Gaussian noise perturbation, cutout occlusion, and enhancing the dataset construction. Step 3: Construct a multi-source fusion feedback regulation network model. The model is a CNN-GRU parallel modeling structure, which is used to capture local and global multi-scale temporal dependencies respectively. Step 4: The data enhanced in step 2 is fed into the model for training. During the training process, the performance feedback adjustment mechanism dynamically adjusts key hyperparameters based on the performance of the validation set, achieving closed-loop adaptive optimization. Step 5: Classify the test samples through the trained multi-source fusion feedback regulation network model and output the final recognition results.

[0144] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention, and should all be included in the scope of protection of the present invention.

Claims

1. A system for identifying attention deficit hyperactivity disorder, characterized in that: The recognition system includes: a data processing module: preprocessing and feature extraction of acquired EEG data; The extracted original feature data is first subjected to a multi-source feature fusion mechanism, and then a data enhancement strategy is applied; Multi-source fusion feedback regulation network model: Construct a multi-source fusion feedback regulation network model, which is a CNN-GRU parallel modeling structure; Training module: The augmented data is fed into the model for training. During the training process, key hyperparameters are dynamically adjusted by the performance feedback adjustment mechanism. The performance feedback adjustment mechanism dynamically adjusts key training parameters based on the performance of the validation set. Classification module: Classify the test samples through the trained multi-source fusion feedback regulation network model and output the final recognition results.

2. The attention deficit hyperactivity disorder recognition system according to claim 1, characterized in that: The preprocessing of the data processing module is standardized preprocessing; Specifically, EEG statistical features are extracted from the mne-features library to construct a feature set of 23 required features; Then, we further introduce three types of information theory features: sample entropy, singular value entropy and singular value Fisher information, and finally obtain the constructed EEG feature set; Microstate feature extraction includes GFP peak extraction, global microstate template construction and numbering unification, microstate sequence generation and feature extraction, and global explained variance calculation.

3. The attention deficit hyperactivity disorder recognition system according to claim 2, characterized in that: Specifically, GFP peak extraction involves first calculating the global field power at each moment of the EEG data; The global microstate template construction and numbering unification are as follows: microstate clustering is performed on the GFP peak frame of each subject, the number of clusters is set to K = 4, and four microstate topological maps at the subject level are obtained; The cluster centers of all subjects were then merged and globally clustered using the ModKMeans algorithm to form a unified global microstate template. To achieve consistency in state numbering, a standard microstate template was introduced as a reference, and the spatial correlation between the cluster centers and the reference template was calculated. Reorder the numbers by the largest correlation to ensure that all microstate templates of the subjects are numbered consistently and are uniformly labeled A, B, C, and D for subsequent identification. Microstate sequence generation and feature extraction specifically involves using a unified global template to perform microstate prediction on each EEG data segment, generating a label sequence, and extracting statistical features such as duration, frequency, coverage, and state transition probability. The global explained variance calculation is as follows: In order to evaluate the fitting ability of the microstate template to the EEG spatial topological changes, the global explained variance GEV indicator is introduced; In summary, the microstate feature output format is as follows: each EEG segment is ultimately converted into a structured feature set in a unified format, including: 12-dimensional microstate statistical features, 12-dimensional state transition probability features, optional 4 GEV features and auxiliary information.

4. The attention deficit hyperactivity disorder identification system according to claim 1, characterized in that: The Gaussian noise perturbation added in the data processing module is specifically to add Gaussian noise with a mean of zero and a standard deviation of 0.05 to the original EEG feature sequence; The Cutout occlusion of the data processing module is specifically to randomly select a continuous time segment in each sample sequence and set the feature value within the randomly selected continuous time segment to zero; The enhanced dataset of the data processing module is constructed by merging the original samples, Gaussian perturbation samples and occlusion samples to generate an enhanced training set of three times the size, while keeping the original labels consistent.

5. The attention deficit hyperactivity disorder identification system according to claim 4, characterized in that: The multi-source feature fusion mechanism of the data processing module is to align the extracted microstate features and EEG statistical features according to the sample dimension and then splice them to form a unified input feature matrix.

6. The attention deficit hyperactivity disorder identification system according to claim 1, characterized in that: The CNN-GRU parallel modeling structure in the multi-source fusion feedback regulation network model includes convolution branches, GRU branches, and feature fusion and classification; Convolutional branch: Taking the EEG feature sequence after preprocessing and feature extraction as input, two separable one-dimensional convolutional layers with different convolution kernel sizes are used to extract local temporal features. Batch normalization and activation functions are connected after the convolutional layer, and a residual connection is introduced after the second convolutional layer to enhance feature flow and training stability. The convolution output is max-pooled and flattened to form a one-dimensional local feature representation, which is then regularized by a dropout layer. GRU branch: takes the same input sequence as input, uses the gated recurrent unit (GRU) to extract global temporal dependency information, uses the hidden state of the last time step as the global temporal representation, and performs regularization through the Dropout layer; Feature fusion and classification: The outputs of the two branches are fused into a unified high-dimensional feature representation through a splicing operation. The fusion result is transformed by the fully connected layer and input into the Sigmoid activation unit to output the binary classification prediction result.

7. The attention deficit hyperactivity disorder identification system according to claim 6, characterized in that: The performance feedback regulation mechanism in the multi-source fusion feedback regulation network model includes a dynamic loss weighting mechanism, a hyperparameter resampling and structure reconstruction mechanism, and a strategy reset and robustness enhancement mechanism; Dynamic loss weighting mechanism: Based on the accuracy and loss value of each round of model training on the validation set, a performance feedback signal is built to dynamically adjust the sample weighting factor in the loss function; Hyperparameter resampling and structure reconstruction mechanism: When the model performance does not improve significantly within several consecutive training rounds, the model reconstruction process is automatically triggered. Key training parameters are resampled from the predefined hyperparameter search space and the model structure is reinitialized accordingly to escape the local optimum. Strategy reset and robustness enhancement mechanism: If the parameter adjustment mechanism does not bring performance improvement within consecutive training rounds, the current weighted parameters are randomly reset and new values ​​are sampled from the specified distribution.

8. The attention deficit hyperactivity disorder identification system according to claim 1, characterized in that: The training module includes the training process of enhanced data input, model training, dynamic adjustment and closed-loop optimization, and training control mechanism; Model training: The training process is based on the constructed multi-source fusion feedback regulation network model. The model is input with an enhanced dataset and combined with a multi-source feature fusion mechanism to achieve collaborative modeling of static and dynamic EEG features. The CNN-GRU parallel structure is used for time series modeling. Dynamic adjustment and closed-loop optimization: A performance feedback adjustment mechanism is introduced during the training process. The feedback signal is built based on the performance indicators of the validation set, and the model structure and training path are updated in real time. Training control mechanism: Introducing the early stopping mechanism and hyperparameter resampling mechanism. If the model's performance does not improve after several consecutive rounds of training, the hyperparameter reconstruction and weight reset process will be triggered.

9. The attention deficit hyperactivity disorder identification system according to claim 1, characterized in that: The classification module includes the classification and output process of sample preprocessing and feature extraction, model inference, classification output and result generation in sequence; Preprocessing and feature extraction of test samples: The EEG data samples to be tested are standardized and preprocessed according to the processing method of the data processing module, and multi-source features consistent with the training stage are extracted, including microstate features and statistical features, to ensure that the input feature dimensions are consistent with the training model; Model inference: The extracted features of the sample to be tested are input into the trained multi-source fusion feedback regulation network model. The model uses the multi-source feature fusion mechanism and CNN-GRU parallel modeling structure to perform feature mapping and time series modeling, and outputs the fused high-dimensional representation vector. Classification output: The fused features are output through the fully connected layer and the Sigmoid activation unit to output the predicted probability value. If the predicted probability is greater than the set threshold θ, it is judged as the ADHD category; otherwise, it is judged as the control category; Result generation: The model ultimately outputs the category labels of the test samples and their corresponding predicted probabilities, which are used to support ADHD status identification in clinical or auxiliary diagnosis scenarios.

10. A method for identifying attention deficit hyperactivity disorder, characterized in that: The identification method uses an attention deficit hyperactivity disorder identification system according to any one of claims 1 to 9, and the identification method comprises the following steps: Step 1: Preprocess and extract features of the acquired EEG data; Step 2: First use the multi-source feature fusion mechanism on the original feature data extracted in step 1, and then apply the data enhancement strategy; Step 3: Construct a multi-source fusion feedback regulation network model, which is a CNN-GRU parallel modeling structure; Step 4: The data enhanced in step 2 is fed into the model for training. During the training process, the key hyperparameters are dynamically adjusted by the performance feedback adjustment mechanism. The performance feedback adjustment mechanism dynamically adjusts the key training parameters based on the performance of the validation set. Step 5: Classify the test samples through the trained multi-source fusion feedback regulation network model and output the final recognition results.

Images