A system and method for identifying attention deficit hyperactivity disorder (ADHD)

Abstract

This invention discloses a system and method for identifying Attention Deficit Hyperactivity Disorder (ADHD), belonging to the technical field of ADHD. The system includes: a data processing module for preprocessing and feature extraction of acquired EEG data; a multi-source feature fusion mechanism followed by data augmentation strategies for the extracted raw feature data; a multi-source fusion feedback regulation network model, employing a CNN-GRU parallel modeling structure; a training module for training the model using augmented data, dynamically adjusting key hyperparameters based on performance feedback on the validation set; and a classification module for classifying test samples using the trained multi-source fusion feedback regulation network model and outputting the final identification result. This invention effectively improves the accuracy, robustness, and generalization performance of ADHD EEG identification.

Description

Technical Field

[0001] This invention belongs to the field of attention deficit hyperactivity disorder (ADHD) technology, specifically relating to an ADHD identification system and identification method. Background Technology

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

[0003] In recent years, deep learning (DL) has demonstrated powerful automatic feature extraction and modeling capabilities in EEG decoding tasks. Convolutional neural networks (CNNs) excel at capturing local spatial features, while recurrent neural networks (such as GRU and LSTM) perform exceptionally well in temporal modeling. For example, embedding a Long Short-Term Memory (LSTM) network into a CNN structure for predicting epileptic seizures demonstrates the potential of deep temporal modeling in EEG analysis. Another study proposed an RNN structure based on GRU, achieving performance comparable to LSTM while reducing the number of parameters. Furthermore, a hybrid neural network combining GRU and CNN was constructed, fusing EEG microstate features with RQA dynamic parameters to achieve auditory attention recognition 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 number of subjects in publicly available ADHD-related EEG data is limited, making it a typical small-sample training scenario within the deep learning framework. Even with time-window segmentation to obtain multiple sample segments, it's difficult to fully cover the high-dimensional feature space, easily leading to model overfitting. Second, most existing models primarily use statistical features or raw time-series signals as input, lacking the ability to characterize the dynamic changes in brain function, thus 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 dynamic EEG patterns. While extracting manual statistical features from segmented EEG signals for ADHD classification validated the impact of preprocessing and temporal resolution on classification performance, this method is also limited to static feature description and fails to model the temporal evolution of brain function. Furthermore, clinical EEG data often exhibits non-stationarity and noise perturbations, making it difficult for statically configured training parameters to adapt to performance fluctuations at different stages, lacking effective adaptive adjustment mechanisms.

[0004] The application of microstate analysis in ADHD identification is still in its early stages. Most current research has not yet deeply integrated it with statistical features, nor does it possess the ability to model its non-stationary structure. Despite this, research on microstate analysis in ADHD identification remains relatively limited, facing several technical challenges: First, standard microstate templates exhibit inter-individual variability, and direct use across subjects may affect matching quality and feature stability; second, microstate features are typically modeled in isolation, lacking fusion strategies with other EEG features; and third, traditional analysis methods struggle to embed dynamic microstate sequences into deep learning frameworks for end-to-end modeling. Therefore, effectively incorporating microstate structural information into deep models, integrating its spatiotemporal dynamic features and statistical representation capabilities, and constructing a more expressive multi-source feature system has become an important research direction for intelligent EEG decoding. Summary of the Invention

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

[0006] This invention is achieved through the following technical solution:

[0007] A system for identifying attention deficit hyperactivity disorder (ADHD) includes: 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;

[0008] The extracted raw feature data is first processed using a multi-source feature fusion mechanism to achieve the synergistic expression of static and dynamic EEG patterns. Then, data augmentation strategies are applied, including adding Gaussian noise perturbation, cutout occlusion, and augmenting the dataset construction.

[0009] Multi-source fusion feedback regulation network model: 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.

[0010] Training module: The enhanced data is input 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 to achieve closed-loop adaptive optimization.

[0011] Classification module: The multi-source fusion feedback regulation network model, which has been trained, classifies the test samples and outputs the final recognition result.

[0012] Preferably, the preprocessing of the data processing module is standardized preprocessing;

[0013] Specifically, the EEG statistical features involve extracting the required 23 types of features from the mne-features library to construct a feature set.

[0014] Furthermore, by introducing three types of information theory features—sample entropy, singular value entropy, and singular value Fisher information—we obtain the final set of EEG features.

[0015] Microstate feature extraction includes GFP peak extraction, global microstate template construction and numbering standardization, microstate sequence generation and feature extraction, and global interpretation variance calculation.

[0016] Preferably, the GFP peak extraction specifically involves first calculating the global field power at each time step of the EEG data;

[0017] The construction and unification of the global microstate template involves the following steps: Microstate clustering is performed on the GFP peak frames of each participant, with the number of clusters set to K=4, resulting in four microstate topologies at the individual participant level. Subsequently, the cluster centers of all participants are merged, and global clustering is performed using the ModKMeans algorithm to form a unified global microstate template. To ensure consistency in state numbering, a standard microstate template is introduced as a reference for calculating cluster centers. With reference template Spatial correlation between them;

[0018] The numbers were reordered according to the highest relevance to ensure that all subjects' microstate templates were numbered consistently and uniformly labeled as A, B, C, and D for subsequent identification.

[0019] The microstate sequence generation and feature extraction are specifically performed by using a unified global template to perform microstate prediction on each segment of EEG data, generating a label sequence, and extracting statistical features such as duration, frequency of occurrence, coverage, and state transition probability.

[0020] Specifically, the calculation of global explained variance involves introducing the Global Explained Variance (GEV) index to evaluate the fitting ability of the microstate template to changes in the EEG spatial topology.

[0021] In summary, the micro-state feature output format is as follows: each EEG segment is ultimately converted into a structured feature set in a unified format, including:

[0022] It includes 12-dimensional microstate statistical features, 12-dimensional state transition probability features, four optional GEV features, and auxiliary information.

[0023] Preferably, the addition of Gaussian noise perturbation in the data processing module specifically involves adding Gaussian noise with a mean of zero and a standard deviation of 0.05 to the original EEG feature sequence.

[0024] The Cutout occlusion in the data processing module specifically involves randomly selecting a continuous time segment in each sample sequence and setting the feature values ​​within that interval to zero.

[0025] The augmented dataset construction in the data processing module involves merging the original samples, Gaussian perturbation samples, and occluded samples to generate an augmented training set three times the size, while maintaining consistency with the original labels.

[0026] Preferably, the multi-source feature fusion mechanism of the data processing module specifically involves aligning the extracted micro-state features and EEG statistical features according to the sample dimension and then concatenating them to form a unified input feature matrix.

[0027] Preferably, the CNN-GRU parallel modeling structure in the multi-source fusion feedback regulation network model includes convolutional branches, GRU branches, and feature fusion and classification;

[0028] Convolutional branch: Taking the preprocessed and feature-extracted EEG feature sequence as input, local temporal features are extracted through two separable one-dimensional convolutional layers with different kernel sizes. Batch normalization and activation functions are connected after the convolutional layers, and residual connections are introduced after the second convolutional layer to enhance feature flow and training stability. The convolutional output is transformed into a one-dimensional local feature representation through max pooling and flattening operations, and then regularized through a Dropout layer.

[0029] GRU branch: Using the same input sequence as input, the Gated Recurrent Unit (GRU) is used to extract global temporal dependency information, the hidden state of the last time step is used as the global temporal representation, and regularization is performed through the Dropout layer;

[0030] Feature fusion and classification: The outputs of the two branches are fused into a unified high-dimensional feature representation through a concatenation operation. The fusion result is then transformed by a fully connected layer and input into a Sigmoid activation unit to output a binary classification prediction result.

[0031] Preferably, the performance feedback regulation mechanism in the multi-source fusion feedback regulation network model includes dynamic loss weighting mechanism, hyperparameter resampling and structure reconstruction mechanism, and policy reset and robustness enhancement mechanism;

[0032] Dynamic loss weighting mechanism: Based on the accuracy and loss value of the model on the validation set in each round of training, a performance feedback signal is constructed, and the sample weighting factor in the loss function is dynamically adjusted;

[0033] Hyperparameter resampling and structural 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 local optima.

[0034] Policy reset and robustness enhancement mechanism: If the parameter tuning mechanism does not bring performance improvement in consecutive training rounds, the current weighted parameters are randomly reset and new values ​​are sampled from the specified distribution.

[0035] Preferably, the training module includes, in sequence, a training process of augmented data input, model training, dynamic tuning, closed-loop optimization, and training control mechanism;

[0036] Model training: The training process is based on the constructed multi-source fusion feedback regulation network model. The model is input with augmented datasets, and the multi-source feature fusion mechanism is combined to achieve collaborative modeling of static and dynamic EEG features. Temporal modeling is performed using the CNN-GRU parallel structure.

[0037] Dynamic tuning and closed-loop optimization: A performance feedback adjustment mechanism is introduced during training. Feedback signals are constructed through validation set performance metrics to update the model structure and training path in real time.

[0038] Training control mechanism: An early stopping mechanism and a hyperparameter resampling mechanism are introduced. If the model's performance does not improve in several consecutive training rounds, the hyperparameter reconstruction and weight reset process is triggered.

[0039] Preferably, the classification module includes, in sequence, the classification and output process of sample preprocessing and feature extraction, model inference, classification output, and result generation;

[0040] 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 those in the training stage are extracted, including micro-state features and statistical features, to ensure that the input feature dimensions are consistent with the training model.

[0041] Model inference: The extracted features of the test sample are input into the trained multi-source fusion feedback regulation network model. Feature mapping and temporal 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.

[0042] Classification output: The fused features are processed through a fully connected layer and a Sigmoid activation unit to output a predicted probability value. If the predicted probability is greater than a set threshold θ, it is judged as ADHD category; otherwise, it is judged as control category.

[0043] Results generation: The final output of the model is the class label of the test sample and its corresponding predicted probability, which is used to support ADHD status identification in clinical or auxiliary diagnostic scenarios.

[0044] A method for identifying attention deficit hyperactivity disorder (ADHD), the method using an ADHD identification system as described above, the method comprising the following steps:

[0045] Step 1: Preprocess and extract features from the acquired EEG data;

[0046] Step 2: First, use a multi-source feature fusion mechanism on the raw feature data extracted in Step 1, and then apply data augmentation strategies;

[0047] Step 3: Construct a multi-source fusion feedback regulation network model, which uses a CNN-GRU parallel modeling structure;

[0048] Step 4: Input the enhanced data from Step 2 into the model for training. During the training process, the key hyperparameters are dynamically adjusted by the performance feedback adjustment mechanism, which adjusts the key training parameters based on the performance of the validation set.

[0049] Step 5: Classify the test samples using the trained multi-source fusion feedback regulation network model and output the final recognition result.

[0050] The beneficial effects of this invention are:

[0051] This invention achieves high-precision recognition of ADHD EEG signals by integrating EEG microstate features and multi-source temporal features, and by introducing a performance feedback adjustment mechanism, significantly improving the robustness, generalization ability and interpretability of the model.

[0052] This invention achieves a comprehensive characterization of the dynamic features of EEG signals by integrating EEG microstate features with time-domain, frequency-domain, and nonlinear dynamic features, resulting in classification accuracy and F1 score significantly higher than existing technologies.

[0053] This invention employs a data augmentation strategy of Gaussian noise perturbation and cutout occlusion to expand the training sample space and effectively avoid overfitting under small sample conditions.

[0054] This invention employs a performance feedback-based hyperparameter dynamic optimization mechanism, enabling the model to automatically adjust the Dropout rate, regularization term, and learning rate, thereby improving convergence speed and reducing sensitivity to noise and non-stationary data.

[0055] This invention employs parallel local convolution and global temporal modeling, which can efficiently capture the multi-scale spatiotemporal dependencies of EEG signals and improve the feature expressiveness of the model.

[0056] This invention, through EEG microstate analysis, not only improves recognition performance but also reveals dynamic abnormal patterns in the neural functional networks of ADHD patients, providing interpretable reference for clinical diagnosis. Attached Figure Description

[0057] Figure 1 This is a logical structure diagram of the performance feedback adjustment mechanism of the present invention.

[0058] Figure 2 This is a heatmap of the performance of each model in the embodiments of the present invention on four evaluation indicators.

[0059] Figure 3 This invention provides four micro-state topographic maps, where A represents micro-state A, B represents micro-state B, C represents micro-state C, and D represents micro-state D.

[0060] Figure 4 This is a comparison chart of the average duration of microstates in this invention between groups.

[0061] Figure 5 This is an inter-group comparison chart of the microstate time coverage of the present invention. Detailed Implementation

[0062] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0063] Implementation Method 1

[0064] This embodiment provides a system for identifying attention deficit hyperactivity disorder (ADHD). The system includes a data processing module that preprocesses and extracts features from acquired electroencephalogram (EEG) data. The feature extraction includes multi-source features, including microstate feature extraction and EEG statistical features.

[0065] 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 the DSM-IV criteria, while the control group had no history of mental disorders or neurological diseases. EEG data were collected according to the international 10-20 system, including 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 standardized preprocessing, including bandpass filtering, artifact removal, and independent component analysis.

[0066] The extracted raw feature data is first subjected to a multi-source feature fusion mechanism, and then data augmentation strategies are applied. The data augmentation strategies include adding Gaussian noise perturbation, cutout occlusion, and augmenting the dataset construction.

[0067] 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 respectively.

[0068] A deep classification model integrating data augmentation, temporal modeling, and performance feedback adjustment mechanisms is constructed. The deep classification model includes a data augmentation strategy component, a CNN-GRU fusion feature modeling component, and a performance feedback adjustment mechanism; the logical structure of this mechanism is as follows: Figure 1 As shown;

[0069] Training module: The enhanced data is input 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 key training parameters such as Dropout rate, regularization term and learning rate based on the performance of the validation set, so as to achieve closed-loop adaptive optimization.

[0070] Classification module: The multi-source fusion feedback regulation network model, which has been trained, classifies the test samples and outputs the final recognition result.

[0071] Preferably, the data processing module performs standardized preprocessing, which aims to enhance the ability to identify and correct multiple types of artifacts, and to preserve the original neural signal features as much as possible.

[0072] First, to suppress low-frequency drift and high-frequency interference, the raw EEG signal was bandpass filtered from 0.5 to 45 Hz. This band covers the main brainwave rhythms (delta, theta, alpha, beta, and low-frequency gamma) while filtering out high-frequency artifacts such as power supply interference (50 Hz) and eye movement / electromyography. The filtering was implemented using an FIR filter to ensure stable phase response and preserve timing structure.

[0073] Subsequently, the ArtifactSubspaceReconstruction (ASR) algorithm is 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 anomalous components back to the reference space, thereby achieving a "repair-like" purification of the signal. To improve artifact adaptability, this invention sets a more conservative dual threshold parameter [0.25, 0.75] to enhance the fault tolerance of drastically fluctuating segments.

[0074] Based on ASR processing, an average re-referencing operation 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).

[0075] Subsequently, the ICA algorithm is used to extract potential neural source signals. ICA, as a spatial filter, decomposes the observed signal into a set of statistically independent components. This invention utilizes the `pop_runica` function in EEGLAB to achieve independent component separation and combines it with the ICLabel tool for automatic classification of each component. To enhance artifact removal accuracy, only components identified as "Brain" with a confidence level higher than 0.7 are retained and used to reconstruct the EEG signal to remove non-neural source artifacts such as eye movement, electromyography, electrocardiography, and line noise.

[0076] This preprocessing strategy effectively suppresses multi-class artifact interference while preserving neural information to the maximum extent, providing a reliable input foundation for subsequent micro-state feature extraction and deep classification models (MSRLNet) that combine performance feedback adjustment mechanisms.

[0077] The raw EEG signal without preprocessing contained significant high-frequency artifacts and background noise interference. However, after ASR, rereference and ICA removal, the high-frequency components were significantly reduced and the low-frequency rhythm structure was clearer, demonstrating the effectiveness of this preprocessing scheme in removing artifacts and preserving neural information.

[0078] Feature extraction includes:

[0079] Specifically, the EEG statistical features were extracted using the mne-features library, comprising 26 commonly used EEG features. Then, a random forest feature importance assessment method based on the Gini index was used to filter the features, removing three categories with low contribution and retaining 23 categories to construct a preliminary feature set. These features included standard deviation, peak amplitude, skewness, kurtosis, root mean square value, quantiles, Hearst 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 curves, spectral entropy, band energy, spectral edge frequencies, and energy of wavelet decomposition coefficients.

[0080] Building upon this foundation, and considering the characteristics of ADHD EEG signals in terms of information complexity and non-stationarity—namely, significant dynamic fluctuations and structural changes in the time domain, multi-scale non-stationary features in the frequency domain, and large pattern differences among individuals—to enhance the feature set's ability to characterize such highly time-varying and complex patterns, three information-theoretic 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).

[0081] Ultimately, it comprises 26 categories and 40 dimensions to comprehensively describe the wave nature, spectral distribution, and multi-scale complexity of signals, supporting subsequent model training and discriminant analysis;

[0082] Sample entropy measures the magnitude of the change in self-similarity of a 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 measures the change in local non-uniformity and orderliness of the singular value distribution.

[0083] Preferably, the sample entropy is defined as follows:

[0084]

[0085] in, Indicates dimension as Tolerance is Under the conditions, with the first The number of similar subsequences;

[0086] Singular value entropy is defined as:

[0087]

[0088] in, The first singular value decomposition of the signal represents the... A singular value, To preserve the dimensional threshold of the principal energy;

[0089] The singular value Fisher information expression is:

[0090]

[0091] in, , To prevent the use of a small constant that prevents division by zero (i.e., a minimal constant introduced by preventing the denominator from being zero), which is used to measure the intensity and orderliness of local variations in the probability distribution of singular values, this invention sets... .

[0092] Microstate feature extraction includes GFP peak extraction, global microstate template construction and unified numbering, microstate sequence generation and feature extraction, and global explained variance calculation. Based on the microstate label sequence, feature indicators representing spatiotemporal dynamic behavior of EEG are extracted, including the duration, frequency of occurrence, coverage and inter-state transition probability of each microstate, which are used to characterize the short-term stability and state switching patterns of the brain.

[0093] To characterize the transient stable topological states in electroencephalogram (EEG) signals, this invention employs a microstate analysis method to model signal segments. This method can effectively reveal the short-term information processing of the brain's dynamic network and extract structured features containing multi-dimensional indicators such as persistence, transition, and spatial distribution. The entire microstate feature extraction process includes GFP peak extraction, global microstate template construction and unified numbering, microstate sequence generation and feature extraction, and global interpreted variance calculation.

[0094] Preferably, the GFP peak extraction specifically involves first truncating the EEG data (20–60 seconds), performing passband filtering (0.5–45 Hz), and resetting the average reference; then, based on the Global Field Power (GFP) index, identifying the time points with the strongest EEG activity for subsequent microstate clustering.

[0095] The formula is as follows:

[0096]

[0097] in, Indicates the first Each channel at time electric potential, It is the average electrode potential at that moment. The value represents 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.

[0098] The construction and unification of the global microstate template are as follows: the GFP peak topology map of each participant is clustered into four microstate topologies at the individual participant level using the ModKMeans algorithm (cluster number K=4); then, the cluster centers of all participants are merged and global clustering is performed using the ModKMeans algorithm to form a unified global microstate template; to ensure consistency in state numbering, a standard microstate template is introduced as a reference for calculating cluster centers. With reference template The spatial correlation between them, and the matching criteria are:

[0099]

[0100] in, The first cluster obtained by clustering A central diagram, Using the template image as a reference, the pair with the highest relevance is finally selected; the numbers are then reordered according to the highest relevance to ensure that all subjects' microstate template numbers are consistent, and they are uniformly labeled as A, B, C, and D for subsequent identification.

[0101] The microstate sequence generation and feature extraction process involves using a unified global template to predict the microstates of each segment of EEG data, thereby obtaining a label sequence. Based on the microstate sequence, the following statistical features are extracted:

[0102] 1. Duration (MeanDuration);

[0103]

[0104] in Let k be the number of times the microstate k occurs. The duration of the i-th occurrence;

[0105] 2. Occurrence: The number of times microstate k occurs. ;

[0106] 3. Coverage;

[0107]

[0108] in This represents the total duration occupied by microstate k, where T is the total duration of the entire EEG segment;

[0109] 4. State Transition Probability (TransitionMatrix): Construct a 4×4 state transition matrix, retaining the 12 off-diagonal elements after removing the diagonal (self-transition) as features;

[0110] The Global Explained Variance (GEV) is calculated as follows: To evaluate the ability of a microstate template to fit changes in the spatial topology of EEG, the Global Explained Variance (GEV) index is introduced, defined as follows:

[0111]

[0112] in This indicates that all tags are in a state. A set of time points, Let be the spatial distribution of EEG at time t. For microstate templates, The Pearson correlation coefficient between the two is given.

[0113] Finally, the feature output format is as follows: Each EEG fragment is ultimately converted into a structured feature set in a unified format, containing:

[0114] 12-dimensional microstate statistical features (4 states × duration, frequency, coverage);

[0115] 12-dimensional state transition probability features (removing off-diagonal elements of the transition matrix after self-transition).

[0116] Four optional GEV features; auxiliary information: corresponding sample labels and file index names.

[0117] The aforementioned micro-state features constitute one of the core inputs supporting the training of the ADHD automatic identification model, and serve as an important feature channel for the MSRLNet model.

[0118] Preferably, in the data augmentation strategy section: 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 robustness and generalization ability of the model.

[0119] The CNN-GRU fusion feature modeling part: a multi-branch structure is designed to extract local temporal convolutional features and global temporal dependency information respectively, and the feature fusion is used to achieve a comprehensive characterization of the dynamic pattern of EEG.

[0120] Performance feedback adjustment mechanism: A lightweight adjustment module is designed to dynamically adjust key hyperparameters (such as loss weights, Dropout rate, and learning rate) during training based on the model's performance feedback on the validation set. This mechanism borrows the idea of ​​reward feedback 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.

[0121] The multi-source feature fusion mechanism, CNN-GRU parallel modeling structure, data augmentation strategy, and performance feedback adjustment mechanism work together to constitute the data processing, network model, and training strategy of the MSRLNet model. This improves the accuracy of automatic ADHD identification while maintaining interpretability and stability. The structural design and mathematical modeling process of each module will be described below.

[0122] Data augmentation strategy section: Gaussian noise perturbation and cutout occlusion are used to enhance the diversity of EEG sequences and construct an augmented dataset;

[0123] The CNN-GRU fusion feature modeling part: a multi-branch structure is designed to extract local temporal convolutional features and global temporal dependency information respectively, and the feature fusion is used to achieve a comprehensive characterization of the dynamic pattern of EEG.

[0124] Performance feedback tuning mechanism: Design a lightweight tuning mechanism to dynamically adjust key hyperparameters during training based on the model's performance feedback on the validation set.

[0125] Preferably, the Gaussian noise perturbation specifically involves: Gaussian noise enhancement aims to simulate the unavoidable background electrophysiological perturbations during actual data acquisition, by adding small-amplitude noise to the original features to enhance the model's robustness to minor perturbations; for the original input feature matrix ,in, Indicates the number of samples. Given the time dimension length for each sample, the enhanced feature representation is as follows:

[0126]

[0127] in, Indicates and Gaussian white noise matrices of the same shape, standard deviation Set to 0.05. This strategy is used to simulate weak electrophysiological perturbations during the data acquisition process, improving the model's robustness to input uncertainties;

[0128] Random occlusion enhancement specifically refers to Cutout, a temporal occlusion enhancement strategy designed to improve a model's ability to discriminate when faced with missing local features. Its basic idea is: for each training sample sequence... Randomly select the starting position in its sequence and from position Start, length is The continuous time window is zeroed out to enhance the sample representation. :

[0129]

[0130] in, This indicates the length of the obstructed window, which is set in this invention as follows: This means occluding 10% of the time steps in the sample to enhance the model's discrimination ability in cases of missing local features.

[0131] The augmentation dataset is constructed by concatenating the original, noisy, and occluded samples from the training set, thereby increasing the number of training samples from [previous data type]. Expand to The corresponding labels are also expanded simultaneously; the original sample X and the noisy sample X are combined. With occluded samples The data is concatenated along the sample dimension to form the enhanced training data:

[0132]

[0133] The corresponding tag set has also been expanded accordingly:

[0134]

[0135] With the aforementioned enhancement strategies, despite the limited number of original subjects, the training data was expanded approximately threefold in the feature space, significantly alleviating the training bias caused by insufficient individual samples. While maintaining label consistency, the diversity of the feature space was expanded, thereby mitigating training bias under small sample conditions and improving the model's generalization ability and robustness.

[0136] Preferably, the convolutional branch modeling specifically involves the input of the convolutional branch being an EEG data matrix. ,in, Indicates the time step. Indicates the feature dimension (number of fused features) at each time point;

[0137] This branch first uses two separable one-dimensional convolutional layers (SeparableConv1D) with different kernel sizes to extract short-term and medium-term temporal dependency structures, respectively; the first layer has a kernel size of 3, and the second layer has a kernel size of 5, defined as follows:

[0138]

[0139] in, Indicates the kernel size as Separable one-dimensional convolution operation, BN for batch normalization, and ReLU for activation function;

[0140] To enhance the network's ability to fuse features at different scales while suppressing the degradation problem of deep networks, a residual connection is introduced after the output of the second layer; input After 1×1 convolution compression and Add them together and activate them to obtain the final convolution output:

[0141]

[0142]

[0143] The convolutional features are then temporally compressed using max pooling and flattened into a one-dimensional vector. To reduce the risk of overfitting, regularization is finally performed using a Dropout layer.

[0144]

[0145] The result This represents the local temporal features of the convolutional branches;

[0146] The timing modeling branch specifically involves modeling the long-term dependency characteristics of the signal; the GRU branch directly uses the input... As the sequence input, the hidden state of the last time step is output as the global temporal representation;

[0147] This branch uses a single-layer GRU unit to extract the global temporal dependency of EEG sequences. The core calculation formula is as follows:

[0148] Update Gate: Reset Door:

[0149] Candidate state:

[0150] Final output status:

[0151]

[0152]

[0153]

[0154] in This represents the Hadamard product. The activation function is Sigmoid; the hidden state at the last time step is used as the output of the GRU, and after Dropout regularization, it is denoted as:

[0155]

[0156] Feature fusion and classification output specifically involves combining features from CNN and GRU branches. and The two are fused through feature concatenation operations to construct a unified high-dimensional representation:

[0157]

[0158] The fused features are then fed into a fully connected layer for feature transformation, and the final predicted probability is output through a Sigmoid unit.

[0159]

[0160] in, The prediction result represents the probability that the input sample belongs to the target category (such as ADHD).

[0161] Preferably, to enhance the training stability and generalization ability of the model in complex neural signal classification tasks, this invention introduces a performance feedback adjustment mechanism to dynamically adjust the loss weights and key hyperparameters during training. This mechanism essentially borrows from the reward feedback and policy update concepts in reinforcement learning, but it does not involve the traditional agent-environment interaction, state transition modeling, or policy function learning. Therefore, more accurately, it is a simplified control strategy for training optimization tasks. Specifically, the dynamic loss weighting mechanism involves constructing a performance feedback reward signal R based on the validation set performance after each training round to measure the performance of the current model state. Its calculation method is as follows:

[0162]

[0163] in, This indicates the accuracy rate of the current round of verification. This is the corresponding verification loss;

[0164] The larger the reward value, the closer the model is to high accuracy and low error;

[0165] Based on this reward feedback, the current loss weighting factor Update using the following incremental strategy:

[0166]

[0167] in, To update the step size, used to adjust the weighting strength;

[0168] To ensure training stability, weights Limited to a fixed range:

[0169]

[0170] Weighting factor It is used to construct a weighted loss function, thereby adjusting the sample error in a weighted manner; for binary classification tasks, this invention uses a weighted binary cross-entropy loss function in the following form:

[0171]

[0172] 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.

[0173] The hyperparameter resampling and structural reconstruction mechanism works as follows: In traditional training strategies, hyperparameter settings are usually statically specified before training, which cannot effectively adapt to dynamically evolving error distributions. This study introduces a policy control mechanism to monitor continuous performance stagnation during model training and triggers the model reconstruction process when the following conditions are met:

[0174]

[0175] in, For performance tolerance factor, For a pre-set no-boost threshold, ;

[0176] 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:

[0177]

[0178]

[0179]

[0180] After resampling, the model structure is automatically reinitialized and training starts again with the new configuration, avoiding the original model from getting stuck in a performance bottleneck.

[0181] The strategy reset and robustness mechanism specifically refers to the following: to further enhance the robustness and exploration capabilities of the agent system, if the loss weighting factor w is continuously... If the round weight update does not bring performance improvement, the system will resample the initial weights from the specified uniform distribution and reset them. :

[0182]

[0183] This strategy is essentially a parameter perturbation method inspired by the "policy perturbation" of reinforcement learning. This mechanism is used to escape local optima and improve the diversity of parameter search and the overall robustness of the training process.

[0184] Specifically, the performance of the MSRLNet model in the ADHD identification task is evaluated. Through the display of core indicators, verification of innovative points, ablation experiments and baseline comparison, the effectiveness and superiority of the model are demonstrated layer by layer, so as to verify the effectiveness of the model and the actual role of key modules.

[0185] The workflow of the training module is as follows: (1) Augmented data input:

[0186] The extracted raw feature data is first subjected to a multi-source feature fusion mechanism, and then data augmentation strategies are applied. The data augmentation strategies include adding Gaussian noise perturbation, cutout occlusion, and augmenting the dataset. The dataset is then constructed and concatenated into a training input dataset and a label set, which are used as the training input for the multi-source fusion feedback regulation network model.

[0187] (2) Model training: The training process is based on the constructed multi-source fusion feedback regulation network model. The multi-source feature fusion mechanism is combined with the augmented dataset as the input of the model. The CNN-GRU parallel structure is used for time-series modeling. The multi-source feature fusion mechanism realizes the synergy of static and dynamic EEG features.

[0188] (3) Dynamic adjustment of parameters and closed-loop optimization: A performance feedback adjustment mechanism is introduced during the training process. Feedback signals are 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 update the model structure and training path in real time accordingly.

[0189] (4) Training control mechanism: In order to improve training stability and convergence efficiency, the process further introduces an early stopping mechanism and a hyperparameter resampling mechanism. If the model has no performance improvement in several consecutive training rounds, the hyperparameter reconstruction and weight reset process is triggered to achieve robust optimization.

[0190] The workflow of the classification module is as follows: (1) Preprocessing and feature extraction of the test sample: The EEG data sample to be tested is preprocessed in a standardized manner, and multi-source features consistent with those in the training stage are extracted, including micro-state features and statistical features, to ensure that the input feature dimension is consistent with the training model.

[0191] (2) Model reasoning: The extracted features of the test sample are input into the trained multi-source fusion feedback regulation network model. The feature mapping and temporal 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.

[0192] (3) Classification output: The fused features are output as predicted probability values ​​through a fully connected layer and a Sigmoid activation unit. If the predicted probability is greater than the set threshold θ (e.g., θ=0.5), it is judged as ADHD category; otherwise, it is judged as control category.

[0193] (4) Result generation: The model finally 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 diagnostic scenarios.

[0194] The automatic ADHD identification method can be deployed and used in any of the following EEG recognition systems:

[0195] (1) Computer-based software system: The method is embedded in the software program on the computer. It performs batch recognition and analysis by calling EEG data files and outputs ADHD recognition results. It is suitable for scientific research and data post-processing tasks.

[0196] (2) Analysis system integrated into EEG acquisition device: The method is integrated into a hardware system with EEG acquisition function to realize integrated acquisition and recognition processing, which is suitable for screening purposes in routine medical or educational scenarios;

[0197] (3) Portable EEG recognition terminal: The method is deployed in a portable device with basic computing capabilities, such as a handheld analyzer, tablet or laptop, for ADHD recognition assistance in a small area.

[0198] Experimental design and reliability description:

[0199] All experiments in this study were performed 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 appropriate initial learning rate and batch size, and a maximum training epoch of 500. To avoid overfitting, an early stopping mechanism was introduced during training, automatically terminating training when the validation set loss did not significantly decrease over several consecutive epochs. Simultaneously, a dynamic learning rate adjustment strategy was employed, gradually reducing the learning rate as training performance approached saturation to achieve more refined model tuning.

[0200] Given that data in ADHD EEG recognition tasks typically exhibits a structure characterized by "a limited number of subjects and a large number of sample segments per subject," this invention introduces a time-series-based feature-level data augmentation strategy during the model training phase. This strategy includes Gaussian noise perturbation and random occlusion mechanisms to enhance the model's generalization ability and robustness. The former introduces minute perturbations into the original features to improve the model's robustness to minor noise interference; the latter strengthens the model's adaptability to local information loss by randomly occluding parts of the time window information. The augmented samples maintain consistency with the original samples in terms of feature dimensions and are concatenated, resulting in a three-fold expansion of the training sample space. This strategy effectively alleviates training bias caused by the scarcity of individual samples, providing more abundant and diverse optimization support for deep models, and further enhancing the learning ability and generalization performance of ADHD EEG differential patterns.

[0201] To further enhance the stability and adaptability of the model during training, this invention also introduces a performance feedback adjustment mechanism. This mechanism constructs a reward signal based on the difference between the accuracy and loss value on the validation set in each round, dynamically updating key training parameters, including the Dropout rate, L2 regularization coefficient, and learning rate. Once the model performance is detected to have stagnated for several consecutive rounds of training, the system will automatically trigger hyperparameter resampling and structural reconstruction, thereby effectively avoiding the problem of getting trapped in local optima and improving the model's training robustness and generalization ability.

[0202] For data preprocessing, EEG statistical features and microstate features were first fused at the sample level and normalized using Z-score standardization. Then, a 5-fold subject-wise cross-validation scheme was used, where all 121 subjects were randomly divided into five non-overlapping subsets. In each iteration, one subset was used as the test set, and the remaining four subsets were used for training and validation (with 10% further divided into a validation set to enable Early-Stopping). This strategy ensured that data from the same subject did not appear simultaneously in both the training and testing phases, and after five rounds, the mean ± standard deviation of each indicator was taken as the final performance.

[0203] After model training is complete, the system will automatically save the model weights and hyperparameter combinations that correspond to the optimal performance on the validation set, and output the accuracy change curve, loss trend, confusion matrix, and various classification evaluation metrics during the training process, facilitating subsequent visualization and experimental analysis. The above training process, from data construction and model design to optimization strategies and evaluation mechanisms, achieves systematic integration, ensuring the effectiveness and reliability of the proposed MSRLNet model in ADHD EEG recognition tasks.

[0204] Evaluation metrics: To comprehensively evaluate the classification performance of the proposed model in the ADHD EEG recognition task, this invention uses several commonly used metrics, including accuracy, Cohen's Kappa coefficient (κ), F1 score, precision, recall, and root mean square error (RMSE), and combines confusion matrix to visualize the model's discriminative ability.

[0205] Accuracy is used to measure the overall correctness of the model's classification, and the calculation formula is as follows:

[0206]

[0207] in, , These represent the number of correctly classified positive and negative samples, respectively. , These represent the number of positive and negative samples that were misclassified, respectively. The closer the accuracy value is to 1, the stronger the overall classification ability of the model.

[0208] To further evaluate the consistency between the model classification results and the actual labels, this invention introduces the Cohen's Kappa coefficient. As a supplementary evaluation indicator, it is defined as follows:

[0209]

[0210] in, This represents the actual probability of consistency between the model's prediction and the true label. Kappa represents the expected consistency probability under the random classification assumption. The closer the Kappa value is to 1, the higher the consistency between the model's predictions and the actual labels, and the stronger the robustness.

[0211] In addition to overall accuracy and consistency metrics, this invention also evaluates the model's ability to identify positive samples and its error distribution. Precision measures how many samples predicted as positive by the model are actually positive, while recall indicates how many samples the model correctly identifies as actually positive. The F1 score, as the harmonic mean of precision and recall, more comprehensively reflects the model's classification ability under class imbalance conditions, and its calculation formula is as follows:

[0212]

[0213]

[0214]

[0215] Furthermore, this invention introduces the root mean square error (RMSE) to measure the overall deviation between the predicted result and the true label, calculated as follows:

[0216]

[0217] in, and They represent the first The smaller the RMSE (Real Label and Model Prediction) of a sample, the closer the model prediction is to the true result.

[0218] To further analyze the classification results from a visual perspective, this invention generates a confusion matrix, which visually displays the model's classification accuracy and misclassification between the ADHD group and the control group. This facilitates the identification of the distribution characteristics of false positive and false negative samples and helps to understand the model's discrimination mechanism.

[0219] In summary, this invention 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.

[0220] To comprehensively evaluate the effectiveness and robustness of MSRLNet in the automatic EEG recognition task for ADHD, this section first presents the model's core performance metrics under five-fold cross-validation (Table 1), and then compares the results with representative models from existing studies on the same dataset (Table 2).

[0221] Table 1 shows the performance of MSRLNet in five-fold cross-validation.

[0222]

[0223] As can be seen, MSRLNet achieved high scores of over 98% on all six evaluation metrics, with a K value close to 1, indicating that the prediction results have extremely high consistency with the true labels. The error (RMSE) was controlled within 0.11, which fully demonstrates that the model still maintains stable performance in scenarios with sparse samples and class imbalance.

[0224] Table 2 Horizontal Comparison with Existing Models

[0225]

[0226] To more intuitively compare the performance of each model on key performance indicators, Figure 2 Furthermore, heatmaps were used to illustrate the specific values ​​and relative strengths of different models across four metrics: Accuracy, F1 score, Precision, and Recall. It is clear that MSRLNe occupies the highest color intensity zone in all four core metrics, demonstrating its comprehensive leading performance; Att-CNN and GPC are close to or slightly inferior to this model in some metrics, while traditional machine learning methods show significant gaps in F1 and Precision.

[0227] Compared to EEG-MHCNet, which is based on a multi-head convolutional attention mechanism, as proposed by Sanchis J, MSRLNet improves accuracy by 22.81%, demonstrating the strong ability of multi-source fusion to capture weak spatiotemporal patterns. Compared to the latest Att-CNN, it still improves accuracy by 0.02% and F1 score by 0.68%, indicating that adaptive tuning of micro-state features can still bring considerable gains at the top level. Compared to traditional machine learning methods such as GPC, CNN+LR, and LASSO+SVM, MSRLNet's overall advantage remains between 1.3% and 5.5%, and it significantly narrows the precision-recall gap, demonstrating the adaptability of deep temporal modeling to complex non-stationary signals.

[0228] Based on a combination of quantitative metrics and cross-sectional comparisons, MSRLNet achieved the highest accuracy and F1 scores to date in the automatic ADHD EEG recognition task, and outperformed or matched the best deep learning models in Kappa, Precision, and Recall metrics. Its microstate fusion structure, data augmentation strategy, and adaptive adjustment mechanism collectively contributed to its performance advantages, while also providing an interpretable physiological basis for abnormal brain dynamics. Future research will further examine the model's generalization ability and clinical applicability through multimodal data and cross-center validation.

[0229] Ablation and Baseline Comparison

[0230] To deeply analyze the specific contributions of each functional module in the MSRLNet model to the performance of automatic ADHD identification, this invention designed a series of ablation experiments. These experiments specifically isolated and evaluated key components of the model structure—the convolutional branch (CNN), the recurrent branch (GRU), the micro-state feature fusion module, the performance feedback adjustment mechanism (RL_Agent), and data augmentation strategies (including Gaussian noise and random occlusion)—examining their roles in the complete model architecture and the performance fluctuations caused by their absence. The experimental results are shown in Table 3.

[0231] Table 3 Ablation Experiment Results: Impact Analysis of Each Module on MSRLNet Performance

[0232]

[0233] Overall, each module makes an irreplaceable positive contribution to the model's performance. Particularly noteworthy is the precipitous drop in performance after removing the CNN branch. Accuracy plummeted to 77.20%, the Kappa coefficient dropped sharply to 0.533, and the F1 score was only 80.62%, indicating that the CNN structure plays a crucial role in capturing local spatial patterns of EEG signals. In contrast, removing the GRU branch had a relatively mild impact, maintaining the model accuracy at 97.27% and achieving a Kappa of 0.955. This suggests that the GRU branch primarily provides support in modeling temporal dependencies and capturing global contextual information. Together, these two modules form the backbone of MSRLNet's processing of dynamic EEG features.

[0234] Regarding feature fusion, the elimination of microstate feature modules also led to a significant degradation in model performance, with accuracy dropping to 88.70% and Kappa decreasing to 0.866. Combining the performance of Precision (87.25%) and Recall (92.68%), it is evident that the model's boundary in distinguishing ADHD from control samples became blurred, indicating that microstate features possess unique discriminative value in revealing the spatiotemporal transient patterns of neural activity and are significantly complementary to statistical features.

[0235] Furthermore, 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.

[0236] It is worth mentioning that data augmentation strategies also have a positive impact on the model's generalization ability. After removing data augmentation, the model accuracy dropped to 94.82%, and the Kappa was 0.895, reflecting that Gaussian noise perturbation and the Cutout operation provide effective support in suppressing overfitting and simulating signal fluctuation scenarios, which is especially suitable for small-sample EEG learning tasks.

[0237] Overall, the complete MSRLNet model performed best 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 only 0.104. These results fully validate the rationality of the module design and the effectiveness of the collaborative mechanism. Specifically, the CNN-GRU dual-branch structure and micro-state feature fusion jointly provide the dominant support for classification performance, while data augmentation and adaptive adjustment mechanisms offer strong guarantees for model generalization ability and convergence efficiency, forming the foundation for MSRLNet's efficient and robust EEG recognition framework.

[0238] In the aforementioned ablation experiments, the introduction of microstate features was verified 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 the results from two aspects: classification performance comparison and feature statistics.

[0239] The complete model, after incorporating microstate features, exhibits superior classification accuracy and misclassification control. Specifically, the recognition accuracy for the control group (negative class) reached 98.4%, and for the ADHD group (positive class) it reached 98.7%. Compared to the model without microstate features, its accuracy decreased significantly, with the control group achieving only 84.0% and the ADHD group achieving 92.7%. The model showed a significant increase in misclassifications under the condition of missing microstate features, indicating that microstate parameters play a crucial role in the discriminative model.

[0240] like Figure 3 As shown, microstate analysis is an important method for modeling the spatiotemporal features of EEG. The four types of microstates, A, B, C, and D, which are divided based on the instantaneous spatial patterns of whole brain electroencephalogram (EEG) signals, have been extensively studied and are considered to be closely related to specific functional neural network activities.

[0241] Microstate A: Usually associated with bilateral temporal lobes and language-related network activity, 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 shift; Microstate C: Is thought to be associated with the posterior default mode network (pDMN), participating in introspection, integrative processing, and cognitive regulation; Microstate D: Associated with the midline frontal lobe (ACC) and prefrontal control network, commonly seen in executive function, impulse inhibition, and situational regulation tasks.

[0242] Therefore, the differences in the expression of different microstates between ADHD and normal subjects may reflect fundamental differences in the neural basis of their brain functional connectivity networks in areas such as activation, attention maintenance, impulse control, and integration regulation. The following section will delve deeper into this structural shift based on statistical analysis of microstate parameters.

[0243] Microstate features not only demonstrate strong discriminative power at the model performance level, but also reveal a systematic shift in the dynamic EEG patterns of ADHD at the physiological mechanism level. Table 4 summarizes the differences in typical microstate transition pathways between the two groups. ADHD patients tend to rapidly transition from stable state A to active state B, and frequently cycle between D and B, exhibiting a sustained high-activation pattern, while the probability of transitioning to the integrative state C is significantly reduced. This phenomenon reflects that their neural activity is biased towards "activation-excitation," lacking effective regulation and convergence mechanisms.

[0244] Table 4. Differences in microstate transition pathways between ADHD and control groups.

[0245]

[0246] In addition, Tables 5 and 6 show the inter-group differences in mean duration and coverage of microstates.

[0247] Table 5. Differences in the average duration of microstates

[0248]

[0249] Table 6. Differences in Microstate Time Coverage

[0250]

[0251] In addition, to more intuitively demonstrate the differences between ADHD and normal subjects in various microstate characteristics, Figure 4 and Figure 5 Box plots comparing the average duration and time coverage of four types of microstates were drawn.

[0252] In summary, ADHD participants exhibited a systematic shift in microstate dynamics, primarily including more frequent entry into hyperactivation states (B) and reduced transitions to integrative states (C). These shifts not only align with the clinical manifestations of attention deficit and increased impulsivity but also validate the physiological value and discriminative potential of microstate features in identifying abnormal EEG dynamics in ADHD.

[0253] This invention focuses on the performance of the MSRLNet model in ADHD EEG recognition tasks, conducting a systematic empirical study and multi-dimensional analysis. Through five-fold cross-validation on publicly available ADHD datasets, MSRLNet achieved excellent results on multiple core metrics such as accuracy, F1 score, and Kappa coefficient. Its overall performance is stable and the error is controllable, demonstrating strong adaptability and generalization ability to small sample sizes and non-stationary EEG signals.

[0254] In the comparative analysis of models, MSRLNet achieved performance improvements compared to existing representative deep learning models (such as Att-CNN and EEGMCHNet) and traditional machine learning methods (such as GPC and SVM). In particular, it maintained a leading position in metrics that measure model stability and discriminative power, such as F1 and Precision, which verified the effectiveness of the multi-source fusion architecture and deep temporal modeling strategy.

[0255] Ablation experiments further revealed the functional contributions of each module in the model: the dual-branch structure composed of CNN and GRU is the key support for modeling spatiotemporal features; micro-state features significantly improve the model's discriminative ability; data augmentation strategies effectively mitigate the risk of overfitting; and the performance feedback adjustment mechanism provides a robust training and optimization path for the model. The synergistic cooperation among these modules jointly constructs a systematic guarantee for the model's performance.

[0256] Meanwhile, the introduction of microstate features not only showed significant gains in model classification performance, but also revealed the dynamic imbalance in the activation, integration and regulation states of ADHD patients at the neural mechanism level, demonstrating high physiological interpretability and providing a theoretical basis and practical support for the subsequent construction of interpretable EEG recognition models.

[0257] In addition, the performance feedback adjustment mechanism constructs a closed-loop training optimization process through performance feedback, which improves training efficiency and enhances the model's ability to dynamically focus on key features. It is one of the core strategies to promote accurate and efficient model learning.

[0258] In summary, the MSRLNet model proposed in this invention achieves synergistic optimization of performance, robustness, and interpretability in the ADHD EEG recognition task, validating the scientific rigor and advancement of the model architecture and design strategy. Future research will focus on multi-center data, cross-modal fusion, and clinical application to further promote the model's practicality and generalization.

[0259] Implementation Method 2

[0260] This embodiment provides a method for identifying attention deficit hyperactivity disorder (ADHD). The identification method uses an ADHD identification system as described in Embodiment 1, and the identification method includes:

[0261] 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.

[0262] Step 2: First, use a multi-source feature fusion mechanism on the raw feature data extracted in Step 1, and then apply data augmentation strategies. Data augmentation strategies include adding Gaussian noise perturbation, cutout occlusion, and augmenting the dataset construction.

[0263] 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.

[0264] Step 4: Input the enhanced data from Step 2 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 to achieve closed-loop adaptive optimization.

[0265] Step 5: Classify the test samples using the trained multi-source fusion feedback regulation network model and output the final recognition result.

[0266] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope 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 for preprocessing and feature extraction of the acquired EEG data; First, a multi-source feature fusion mechanism is used on the extracted raw feature data, and then data augmentation strategies are applied. Multi-source fusion feedback regulation network model: Construct a multi-source fusion feedback regulation network model, which adopts a CNN-GRU parallel modeling structure; Training module: The enhanced data is input into the model for training. During the training process, the key hyperparameters are dynamically adjusted by the performance feedback adjustment mechanism, which adjusts the key training parameters based on the performance of the validation set. Classification module: The multi-source fusion feedback regulation network model, after being trained, classifies the test samples and outputs the final recognition result; The data processing module performs standardized preprocessing. Specifically, the EEG statistical features involve extracting the required 23 types of features from the mne-features library to construct a feature set. Furthermore, by introducing three types of information theory features—sample entropy, singular value entropy, and singular value Fisher information—we obtain the final set of EEG features. Microstate feature extraction includes GFP peak extraction, global microstate template construction and numbering unification, microstate sequence generation and feature extraction, and global interpreted variance calculation; Specifically, GFP peak extraction involves first calculating the global field power at each moment of the EEG data; The construction and numbering of the global microstate template are as follows: microstate clustering is performed on the GFP peak frames of each subject, and the number of clusters is set to K=4, resulting in 4 microstate topology maps at the individual 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; to achieve consistency in state numbering, a standard microstate template is introduced as a reference, and the spatial correlation between the cluster centers and the reference template is calculated. The numbers were reordered according to the highest relevance to ensure that all subjects' microstate templates were numbered consistently and uniformly labeled as A, B, C, and D for subsequent identification. The microstate sequence generation and feature extraction are specifically performed by using a unified global template to perform microstate prediction on each segment of EEG data, generating a label sequence, and extracting statistical features such as duration, frequency of occurrence, coverage, and state transition probability. Specifically, the calculation of global explained variance involves introducing the Global Explained Variance (GEV) index to evaluate the fitting ability of the microstate template to changes in the EEG spatial topology. In summary, the micro-state 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, 4 optional GEV features, and auxiliary information.

2. The attention deficit hyperactivity disorder (ADHD) identification system according to claim 1, characterized in that, The addition of Gaussian noise perturbation in the data processing module specifically involves adding Gaussian noise with a mean of zero and a standard deviation of 0.05 to the original EEG feature sequence. The Cutout occlusion in the data processing module specifically involves randomly selecting a continuous time segment in each sample sequence and setting the feature values ​​within the randomly selected continuous time segment interval to zero. The augmented dataset construction in the data processing module involves merging the original samples, Gaussian perturbation samples, and occluded samples to generate an augmented training set three times the size, while maintaining consistency with the original labels.

3. The attention deficit hyperactivity disorder (ADHD) identification system according to claim 2, characterized in that, The multi-source feature fusion mechanism of the data processing module specifically involves aligning the extracted micro-state features and EEG statistical features according to the sample dimensions and then concatenating them to form a unified input feature matrix.

4. The attention deficit hyperactivity disorder (ADHD) 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 convolutional branches, GRU branches, and feature fusion and classification. Convolutional branch: Taking the preprocessed and feature-extracted EEG feature sequence as input, local temporal features are extracted through two separable one-dimensional convolutional layers with different kernel sizes. Batch normalization and activation functions are connected after the convolutional layers, and residual connections are introduced after the second convolutional layer to enhance feature flow and training stability. The convolutional output is transformed into a one-dimensional local feature representation through max pooling and flattening operations, and then regularized through a Dropout layer. GRU branch: Using the same input sequence as input, the Gated Recurrent Unit (GRU) is used to extract global temporal dependency information, the hidden state of the last time step is used as the global temporal representation, and regularization is performed 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 concatenation operation. The fusion result is then transformed by a fully connected layer and input into a Sigmoid activation unit to output a binary classification prediction result.

5. The attention deficit hyperactivity disorder (ADHD) identification system according to claim 4, characterized in that, The performance feedback regulation mechanism in the multi-source fusion feedback regulation network model includes dynamic loss weighting mechanism, hyperparameter resampling and structure reconstruction mechanism, and policy reset and robustness enhancement mechanism; Dynamic loss weighting mechanism: Based on the accuracy and loss value of the model on the validation set in each round of training, a performance feedback signal is constructed, and the sample weighting factor in the loss function is dynamically adjusted; Hyperparameter resampling and structural 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 local optima. Policy reset and robustness enhancement mechanism: If the parameter tuning mechanism does not bring performance improvement in consecutive training rounds, the current weighted parameters are randomly reset and new values ​​are sampled from the specified distribution.

6. The attention deficit hyperactivity disorder (ADHD) identification system according to claim 1, characterized in that, The training module includes, in sequence, the training process of augmented 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 augmented datasets, and the multi-source feature fusion mechanism is combined to achieve collaborative modeling of static and dynamic EEG features. Temporal modeling is performed using the CNN-GRU parallel structure. Dynamic tuning and closed-loop optimization: A performance feedback adjustment mechanism is introduced during training. Feedback signals are constructed through validation set performance metrics to update the model structure and training path in real time. Training control mechanism: An early stopping mechanism and a hyperparameter resampling mechanism are introduced. If the model's performance does not improve in several consecutive training rounds, the hyperparameter reconstruction and weight reset process is triggered.

7. The attention deficit hyperactivity disorder (ADHD) identification system according to claim 1, characterized in that, The classification module includes, in sequence, the classification and output process of sample preprocessing and feature extraction, model inference, classification output, and result generation. 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 those in the training stage are extracted, including micro-state features and statistical features, to ensure that the input feature dimensions are consistent with the training model. Model inference: The extracted features of the test sample are input into the trained multi-source fusion feedback regulation network model. Feature mapping and temporal 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. Classification output: The fused features are processed through a fully connected layer and a Sigmoid activation unit to output a predicted probability value. If the predicted probability is greater than a set threshold θ, it is judged as ADHD category; otherwise, it is judged as control category. Results generation: The final output of the model is the class label of the test sample and its corresponding predicted probability, which is used to support ADHD status identification in clinical or auxiliary diagnostic scenarios.

8. A method for identifying attention deficit hyperactivity disorder, characterized in that, The identification method uses an attention deficit hyperactivity disorder (ADHD) identification system as described in any one of claims 1-7, and the identification method includes the following steps: Step 1: Preprocess and extract features from the acquired EEG data; Step 2: First, use a multi-source feature fusion mechanism on the raw feature data extracted in Step 1, and then apply data augmentation strategies; Step 3: Construct a multi-source fusion feedback regulation network model, which uses a CNN-GRU parallel modeling structure; Step 4: Input the enhanced data from Step 2 into the model for training. During the training process, the key hyperparameters are dynamically adjusted by the performance feedback adjustment mechanism, which adjusts the key training parameters based on the performance of the validation set. Step 5: Classify the test samples using the trained multi-source fusion feedback regulation network model and output the final recognition result.

Images