An auxiliary diagnostic system for attention deficit hyperactivity disorder based on a comprehensive cognitive experimental set and a multi-stream dual-path decision neural network model.

Abstract

This application provides a response testing system for Attention Deficit Hyperactivity Disorder (ADHD). The system comprises multiple cognitive response experiments, all designed and optimized based on cognitive psychology and cognitive neuroscience data, providing a comprehensive assessment of the subject. During the experiments, EEG data is collected from the subject. Based on this data, the frequency bands of interest in the response experiments are processed, and the data transformation and feature extraction are fed into a Multi-Stream Dual-Line Decision Neural Network (MSDD) model. This model selects different learning schemes based on different data types, learning the EEG data features and comparing their similarity to healthy norms, thereby achieving the purpose of assisting diagnosis.

Description

Technical Field

[0001] This application relates to the fields of physiological signal acquisition and medical device technology, specifically to a testing and judgment system for attention deficit hyperactivity disorder (ADHD) and related cognitive impairment. Technical Background

[0002] Attention deficit hyperactivity disorder (ADHD) is a disorder that commonly occurs in adolescents and young children. ADHD patients, compared to their peers, are characterized by significant difficulty concentrating, short attention span, hyperactivity, or impulsivity. Its prevalence is generally reported to be 3%–5%, with a male-to-female ratio of 4:1.

[0003] Due to the heterogeneity of ADHD, its complications, and the global shortage of clinicians to diagnose it, and because hyperactive children are often perceived as healthy and active in kindergartens and daycares, the diagnosis of ADHD is often delayed. The brains of people with ADHD also differ from those of healthy individuals. The prefrontal cortex of people with ADHD has abnormally low levels of dopamine, severely impairing attention, cognitive processes, and working memory. Compared to a normal brain, individuals with ADHD have fewer neuronal connections to the prefrontal cortex.

[0004] Several articles have concluded that the utility of electroencephalography (EEG) is not yet reliable enough for the clinical diagnosis of children or adults. A clearer understanding of the relationship between EEG features and ADHD symptom characteristics may lead to the development of more objective diagnostic tools and help guide personalized treatment strategies.

[0005] Most current cognitive experiments are single-faceted and lack design considerations for all aspects of ADHD or for individualized approaches. Most cognitive experimental paradigms are too broad, making it difficult to accurately identify specific brain regions and related connectivity.

[0006] Current classification models using neural networks are mostly designed for a single data type and do not fully utilize the model's advantages. Although most models can achieve good classification accuracy on existing public datasets, their ability to predict new samples and diseases of different natures still needs improvement. Summary of the Invention

[0007] This application discloses an auxiliary diagnostic system for ADHD, designed to address the current shortage of specialist doctors in ADHD, the often-overlooked nature of ADHD leading to late detection and delayed treatment. The system assists specialist doctors in diagnosis, reducing diagnostic time.

[0008] The operating platform is used to provide an operating environment for interactive testing, monitor physiological indicators in real time, and collect EEG data from the interaction.

[0009] This invention provides an auxiliary diagnostic system for Attention Deficit Hyperactivity Disorder (ADHD) based on a comprehensive multi-task cognitive experiment set and a multi-stream dual-row decision neural network model. The auxiliary diagnostic system includes the following components: a task set based on multiple cognitive task paradigms related to ADHD and associated cognitive impairment; an EEG device for collecting the subject's EEG information; a platform for simple preprocessing of the EEG data; and a model for converting the EEG data and using it as input for feature extraction to determine the evaluation results of each cognitive experiment on the subject.

[0010] The present invention discloses an auxiliary diagnostic system for attention deficit hyperactivity disorder (ADHD) based on a comprehensive multi-task cognitive experimental set and a multi-stream dual-row decision neural network model. The diagnostic system includes the following steps:

[0011] A set of multiple cognitive experimental task paradigms based on attention deficit hyperactivity disorder (ADHD) syndrome;

[0012] Obtain basic information about the subjects and EEG data information on their participation in the entire set of cognitive experiments;

[0013] Based on the acquired EEG data of the subjects, simple preprocessing was performed to remove artifacts and noise, and the target EEG data information was exported in segments according to the Mark markers in the experiment.

[0014] The exported preprocessed EEG data is fed into the corresponding neural network model to determine the difference between the patient's EEG data and the healthy norm for each experiment, thereby classifying the data.

[0015] This process is performed on EEG data from various cognitive task paradigms. The model then makes a comprehensive diagnosis, and if the data exceeds a set threshold, the patient is diagnosed with attention deficit hyperactivity disorder (ADHD).

[0016] Through a collection of cognitive experiments, the different types of EEG signals generated by different experiments are compared and classified with healthy norms in a multi-stream bilinear decision neural network model, and different experiments have corresponding classification results.

[0017] The model performs different data transformations on the input EEG data, using two different neural network models to classify different data types. For time-series EEG data, a Stochastic Bilevel Optimization RNN network model is used, while for brain topography data, a Residual Network network model is used.

[0018] Furthermore, in this embodiment of the invention, each experiment in the cognitive experiment set tests and diagnoses the subjects from different aspects.

[0019] Furthermore, in this embodiment of the invention, after obtaining the subject's EEG data information, the data needs to be filtered, rereferenced, de-ocularized, artifact removed, and segmented for correlation processing to export the preprocessed EEG data.

[0020] Furthermore, in this embodiment of the invention, the system memory contains pre-trained patient data and health norm-related data information.

[0021] Furthermore, in this embodiment of the invention, the final diagnostic result is obtained by comprehensive evaluation of multiple experiments, and the system finally uses the most discriminative original features to be sent to the shallow classifier support vector machine to make the final judgment.

[0022] Furthermore, in this embodiment of the invention, different experiments employ different testing strategies, thereby revealing the subject's problems and enabling corresponding response strategies to be developed based on the system's diagnostic results.

[0023] Furthermore, in this embodiment of the invention, the diagnostic criteria for Attention Deficit Hyperactivity Disorder (ADHD) of the American Psychiatric Association DSM-5, which consists of nine diagnostic criteria, are used to pre-train the final evaluation model if a patient has six or more symptoms.

[0024] Compared with the prior art, the beneficial effects of the embodiments of this application are as follows:

[0025] The cognitive task set selected in this application embodiment is based on multiple cognitive experimental paradigms related to ADHD and its associated cognitive impairment diseases, such as reaction time, attention, control, inhibitory reaction time, and anti-interference ability, thereby obtaining comprehensive EEG signal data reflecting brain network function. For the data obtained from each experiment, after preprocessing, data transformation is performed to generate power spectrum maps, low-resolution electromagnetic tomography, etc., and the processed original EEG data are all input into the MSDD model. Different data types use models with relative advantages over the data. For example, the Recurrent Neural Network (RNN) model has advantages in dealing with time-series data because it can establish contextual dependencies in time series and can handle sequence data of variable length. However, the biggest problem faced by traditional RNNs during training is gradient vanishing and gradient exploding, which makes it impossible to effectively learn long sequence data. If artificial intelligence is applied to medical diagnosis, it must be rigorous, and the model must emphasize stability. Therefore, the models selected in this patent all have the advantage of stability. SBO-RNN borrows the idea of ​​the SGD (Stochastic Gradient Descent) algorithm and replaces the original model calculation formula with ordinary differential equations, replacing it with a two-level optimization problem. This not only enhances the model's stability but also increases its interpretability to some extent. Through the use of residual connections, ResNet can easily scale to networks with dozens or even hundreds of layers, extracting deeper features without suffering performance degradation or gradient vanishing issues. By learning from the data itself, the neural network automatically adjusts weight parameters to extract features, ultimately confirming the subject's relevant information.

[0026] The above description is merely an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above description and other purposes, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0028] Figure 1 This is a schematic diagram of the auxiliary diagnostic system for attention deficit hyperactivity disorder (ADHD) provided in an embodiment of the present invention.

[0029] Figure 2This is a schematic diagram of the auxiliary diagnostic system provided in an embodiment of the present invention.

[0030] Figure 3 This is a schematic diagram of the structure of the multi-stream two-row decision neural network model provided in this embodiment of the invention.

[0031] Figure 4 This is a flowchart of Experiment 1 of the reaction testing system of the present invention.

[0032] Figure 5 This is a flowchart of Experiment 2 of the reaction testing system of the present invention.

[0033] Figure 6 This is a flowchart of Experiment 3 of the reaction testing system of the present invention.

[0034] Figure 7 This is the flowchart of Experiment 4 of the reaction testing system of the present invention.

[0035] Figure 8 This is the flowchart of Experiment 5 of the reaction testing system of the present invention.

[0036] Figure 9 This is a flowchart of Experiment Six of the Reaction Testing System of the present invention.

[0037] Figure 10-1 and 10-2 This is a schematic diagram of the main hardware device of the present invention. Detailed Implementation

[0038] To enable those skilled in the art to better understand the technical solutions of this application, the application will be described in detail below with reference to the accompanying drawings and specific embodiments. The embodiments of this application will be further described in detail with reference to the accompanying drawings and specific examples, but these are not intended to limit the scope of this application.

[0039] In this application, each step is shown in the figure. The arrows shown in the figure are only examples of the execution order. This application involves multiple experiments, and the execution order of the experiments is not limited to the order described in the application. The order can be changed as long as it does not affect the order of the experiments in the execution content.

[0040] All terms used in this experiment (including technical or scientific terms) have the same meaning as understood by one of ordinary skill in the art to which this application pertains, unless otherwise specifically defined. Techniques, methods, and apparatus known to one of ordinary skill in the art may not be discussed in detail, but where appropriate, such techniques, methods, and apparatus should be considered part of the specification.

[0041] The following is combined Figure 1 The ADHD assisted diagnostic system described in this invention includes:

[0042] Operating platform 20 is used to provide an operating environment for interactive testing, monitor physiological indicators in real time, and collect interactive EEG data;

[0043] The cognitive experiment set 30 is used for stimulation experiments that synchronously monitor and collect EEG data corresponding to the experiments, so that the EEG data has analytical significance.

[0044] A multi-stream dual-row decision neural network model 40 is used to process EEG data, extract features, learn classification, and obtain diagnostic results.

[0045] The operating platform used the E-Prime tool to present the cognitive experiments, displaying graphics on a computer screen. Auditory evoked responses were provided by stereo headphones worn by the subjects. The actiCHampPlus amplifier and BrainVision EEG recorder were used to collect EEG data. A 64-channel EEG acquisition system was used, positioned according to the international 10-20 lead standard, with an A / D sampling frequency selectable from 256Hz to 1kHz. The reference electrodes were both earlobes. Two additional channels were used to simultaneously record electrooculography (EOG) signals during the experiment to remove EOG noise. The experiment was conducted in a shielded room. Considering the needs for spatial information acquisition and data processing speed, a 32-lead EEG was used in the audiovisual cognitive event-related EEG experiment. The Analyzer EEG analysis system was used to analyze the EEG data, and PyCharm was used to classify the EEG data using Python.

[0046] The Analyzer EEG analysis system analyzes EEG data, including:

[0047] Filtering involves setting the waveform bandwidth appropriately based on the frequency of the signal to be analyzed, removing unnecessary signals. In the Analyzer's "Transformations," select "Filters." LowCutoff sets the low-cut filter value, where Frequency [Hz] is the frequency and TimeConstant [s] is the time constant. Only one of these parameters needs to be set; they can be converted to each other using F = 1 / (2πT). HighCutoff sets the high-cut filter value. Notch sets the AC filter, determined by the AC voltage at the time of data acquisition. For example, in the US (110V), select 60Hz, while in China (220V), select 50Hz. EnableIndividualChannelFilters allows setting the bandwidth value individually for each channel. This option should not be selected if filtering is applied to all channels.

[0048] Ocular correction (ICA) is used to correct for electrooculography (EOG) components. However, it may also correct for non-EOG components. When performing ICA analysis, it is best to have noise-free data. Therefore, it is recommended to filter the data before using ICA to correct EOG.

[0049] In “Transformations,” select “Ocular Correction ICA.” This generally relies on an algorithm for judgment, and choose semi-automatic mode. Select ValueTriggerAlgorithm, which determines whether it is an electrooculographic signal based on a set standard.

[0050] Artifact Removal: In "Transformations," select "Raw Data Inspection." In the "Inspection Method" section, select "Semiautomatic Inspection." The "IndividualChannelMode" checkbox: For a given electrode point, if a segment exceeds the criteria set, only data from that electrode point will be removed in subsequent steps. If unselected, data from all electrode points within that time period will be removed. In the "Channels" section, all EEG leads except the two Ophthalmic EEG leads should be selected. In the "Criteria" section, set the detection criteria for semi-automatic and automatic analysis. Gradient(x): Gradient change value between two sampling points; Max-Min(x): Maximum absolute value of waveform change; Intervallength: Set the time period within which the maximum allowable waveform change value is within that time period. Amplitude: Waveform change scale value; Set the maximum and minimum allowable voltage values. LowActivity(x): Minimum waveform variation; the minimum allowable variation (large-small) within a set time period. Note: If artifact removal is performed on continuous EEG signals before segmentation, since there is no baseline correction at this time, the Max-Min standard is generally selected instead of the Amplitude standard in the examination criteria. The opposite is true if artifact removal is performed after segmentation. If semi-automatic mode is selected, after the system completes the detection, you can observe the artifacts selected according to the standard. If you believe that some of the "artifacts" selected by the system are not artifacts, you can press Shift + mouse click to remove the marked "artifacts". If some artifacts are not automatically recognized, you can also click on the two endpoints while in Shift mode.

[0051] Segmentation: This involves extracting the EEG signals to be averaged based on the markers. In "Transformations," select "Segmentation" under "Segment Analysis Functions," choose the corresponding marker number, and in the "BasedonTime" property, set the start time of the segment in "Start [ms]" (generally the first 200ms of the marker to prepare for baseline correction) and the end time in "End [ms]." Then, in "Transformations," select "Average" under "Segment Analysis Functions" to perform the average.

[0052] For baseline correction, select “BaselineCorrection” in “Transformations”, then select “SegmentAnalysisFunctions”. “Begin[ms]” is the start time of the calibration reference, and “End[ms]” is the end time of the calibration reference.

[0053] The operational procedure for Experiment 1 of the aforementioned cognitive experiment set is as follows:

[0054] In the traditional antisaccades experiment, after the fixation point, an object is presented on the screen in one of the following positions: up, down, left, or right. The task objective is for the participant to fixate in the opposite direction of the object. Participants must not only suppress the reflexive saccades directed towards the object but also replan their saccade behavior. The traditional antisaccades experiment procedure is as follows:

[0055] Step 1: Present the fixation point for 1 second

[0056] Step 2: Present the interfering target for 1.2 seconds, while simultaneously making an inhibitory response by moving the eyes to the opposite position.

[0057] This experiment has been widely used in areas such as inhibition of responses and attention. In this system, the experiment was improved for different gender groups. Firstly, because the reverse saccade test itself has undergone numerous optimizations due to its "blank effect," the optimized experiment focuses on refining the reverse saccade blank paradigm to achieve better results. The specific experimental procedure is as follows: Figure 3 As shown:

[0058] Step 1: Present the fixation point for 1 second

[0059] Step 2: Present a blank page for 1 second to release attention.

[0060] Step 3: Present the interfering target for 1.2 seconds, while simultaneously making an inhibitory response by moving the eyes to the opposite position.

[0061] The experiment consisted of a practice module and a formal experimental module, with 4 trials and 20 experimental stimuli respectively. Each trial had a 5-minute rest interval, and a total of five trials were conducted. Different colors of interference targets were used for participants of different genders. Males have more rod cells in their retinas, which are more prominent in processing dark colors such as black, white, and brown. Females, on the other hand, have more cone cells in their retinas, which process color and detail and are particularly sensitive to bright colors (red, yellow, and orange). Therefore, we used different approaches for different genders, increasing the difficulty of the inhibitory response and the intensity of the interference to achieve better results.

[0062] The operational procedure for Experiment 2 of the aforementioned cognitive experiment set is as follows:

[0063] In human cognitive psychology, perception is viewed as the organization and interpretation of sensory information. During perceptual processing, there is a tendency to prioritize the processing of objects of optimal size. Among the key factors are the size of the viewing angle and the position of the image; that is, people tend to prioritize processing targets that are within their eye's viewing angle. Based on this, the specific experimental design is as follows: Figure 4 As shown:

[0064] Step 1: Present the fixation point for 1 second.

[0065] Step 2: An audio message will be emitted for 1 second.

[0066] Step 3: Larger letters or numbers will quickly appear in the center of the screen, while smaller letters or numbers will form the composition. (45ms)

[0067] Step 4: Within 1 second, the participants need to judge whether the small-shaped letters or numbers match the audio content. If the letters or numbers match the audio content, they should respond with different buttons. If they do not match the audio content, no button needs to be pressed.

[0068] The experiment consisted of a practice module and a formal experimental module, with 10 trials and 30 experimental stimuli respectively. Each trial had a 5-minute rest interval, and a total of five trials were conducted. This experiment was a multi-task response experiment, requiring participants to not only judge whether an audio signal matched a small-shaped letter or number, but also to determine whether matching numbers or letters required pressing the corresponding key. Utilizing the characteristics of perceptual processing in cognition, large-shaped letters or numbers were designed as distractors to increase the difficulty of the inhibitory response and thus achieve better training results.

[0069] The operational procedure for Experiment 3 of the aforementioned cognitive experiment set is as follows:

[0070] In the Flanker response test, a series of arrows will appear in the center of the screen. Participants need to focus on the central arrow; to look to the left, press the "F" key, and to look to the right, press the "J" key. The traditional Flanker test procedure is as follows:

[0071] Step 1: Present the gaze point for 800ms.

[0072] Step 2: Display a horizontal arrow in the center of the screen for 800ms.

[0073] Step 3: Subjects need to eliminate interference from both sides and complete the arrow direction judgment within 800ms.

[0074] However, traditional two-sided interference experiments lack a crucial element: non-directional interference items. Furthermore, determining the center position in a horizontal string is often slower than in a vertical string, affecting participants' judgment; they tend to locate the center position before making a judgment. Therefore, the horizontal string was changed to a vertical string, the arrows were replaced with graphic arrows, and non-directional interference items ("□", "〇") were added, dividing the experiment into three stimulus types. The specific experimental procedure is as follows: Figure 5 As shown:

[0075] Step 1: Present the gaze point for 800ms.

[0076] Step 2: Display a vertical string of graphics in the center of the screen for 800ms.

[0077] Step 3: Subjects need to eliminate interference from both sides and complete the arrow direction judgment within 800ms.

[0078] The experiment was divided into a practice module and a formal experiment module, with 4 trials and 30 experimental stimuli respectively. The rest interval between each experiment was 5 minutes, and a total of five sets of experiments were conducted.

[0079] The operational procedure for Experiment 4 of the aforementioned cognitive experiment set is as follows:

[0080] In experimental paradigm four, participants are required to respond to stimuli displayed on a screen according to the direction of arrows on the screen. The stimuli appear in the center of the screen, and the arrows appear on either the left or right. A key aspect of the experiment is that sometimes the arrow position and the response button position are the same, and sometimes they are opposite. In this case, participants need to ignore the arrow position and focus only on the arrow direction when pressing the corresponding button. However, due to the conflicting arrow positions, response inhibition is necessary, often leading to incorrect responses. The specific experimental procedure is as follows: Figure 6 As shown:

[0081] Step 1: Present the gaze point for 800ms.

[0082] Step 2: Display the arrow graphic on the screen.

[0083] Step 3: Subjects need to eliminate the influence of positional conflicts and complete the arrow direction judgment and make the corresponding response within 800ms.

[0084] The experiment was divided into a practice module and a formal experiment module, with 4 trials and 22 experimental stimuli respectively. The rest interval between each experiment was 5 minutes, and a total of five sets of experiments were conducted.

[0085] The operational procedure for Experiment 5 of the aforementioned cognitive experiment set is as follows:

[0086] In Paradigm 5, participants were required to react and make a judgment after a prompt appeared on the screen. The prompt might conflict with the final response, yet participants still needed to make the correct response. The specific experimental procedure was as follows: Figure 7 As shown:

[0087] Step 1: Present the fixation point for 1 second

[0088] Step 2: A prompt with an arrow pointing in the direction of the arrow will appear in the center of the screen for 1.2 seconds. The prompt will then appear in the same location as the arrow.

[0089] Step 3: The screen displays a blank screen for 200ms.

[0090] Step 4: A red diamond-shaped square is displayed on the left or right side of the screen, and the subject responds accordingly.

[0091] In cognitive psychology, subjects typically process cues within 200ms before their attention shifts elsewhere; this phenomenon is known as "return inhibition." This is why a 200ms blank page is placed between the cue and the response. The experiment aims to influence patients to focus their attention on whether the cue and response are consistent. The experiment consisted of a practice module with 4 trials and a formal experimental module with 26 stimuli, each with a 5-minute rest interval, for a total of five trials.

[0092] The operational procedure for Experiment Six of the aforementioned cognitive experiment set is as follows:

[0093] Experiment 6 used a variant of the CPT paradigm, a continuous response test now widely used in ADHD diagnosis. Traditional CPT targets the letters "X" and "O," requiring subjects to respond only when the letter "O" appears. The ratio of "X" to "O" is approximately 75% and 25%, respectively. The specific procedure is as follows:

[0094] Step 1: Present the fixation point for 2 seconds

[0095] Step 2: White letters are displayed on a black background in the center of the screen for 200ms. Subjects are required to respond accordingly.

[0096] Step 3: The screen displays a 200ms interval.

[0097] The optimization primarily focuses on modifying the stimuli. Different stimuli elicit different brainwave responses. For younger children, letters are relatively unfamiliar, so the experiment uses graphic stimuli, specifically square and circular shapes. Subjects respond to squares but not to circles. This eliminates interference from prior knowledge, targeting only the stimuli that elicit a response. Furthermore, the ratio of responsive to and non-responsive stimuli is reversed to emphasize the subject's voluntary control over inhibitory responses, achieving better testing results. Specific experimental procedures are as follows... Figure 8 As shown:

[0098] Step 1: Present the fixation point for 2 seconds

[0099] Step 2: A white graphic (square or circle) is displayed on a black background in the center of the screen for 200ms. The subject responds as instructed.

[0100] Step 3: The screen displays a 500ms interval.

[0101] The experiment was divided into a practice module and a formal experiment module, with 12 trials and 25 experimental stimuli respectively. The rest interval between each experiment was 5 minutes, and a total of five sets of experiments were conducted.

[0102] The significance of this experiment lies in its improvement upon the traditional reverse saccade test. The traditional reverse saccade test immediately presents a distraction target after the fixation point, and this distraction target is a single, black color. The improved experiment, due to the "blank space effect," first presents a blank page to allow attention to be released after the fixation point. Then, a brown or red target is presented; brown is for men, and red is for women. Because different genders have different sensitivities to color, the experimental results also reflect that the improved experiment places a stronger test on the subjects' concentration. The EEG, through FFT transformation, reveals more high-frequency brainwaves. This experiment falls under visual conflict, testing the subjects' reaction and control abilities at the visual level. Both the EEG data and theoretical knowledge demonstrate the significant challenge this experiment places on the subjects' reaction and attention.

[0103] The significance of Experiment Two lies in its use of perceptual processing conflict to test the subjects' responsiveness. Because of the varying priorities in perceptual processing (overall and local), the large letters and numbers in the experiment act as a distraction, requiring subjects to maintain focus and eliminate this distraction to complete the experiment successfully, thus increasing its difficulty. Both the experimental results and theoretical knowledge demonstrate the test of the subjects' responsiveness and attention.

[0104] The significance of Experiment 3 lies in the fact that the improved experiment includes an additional scenario compared to the traditional Flanker experiment: non-directional interference. This is achieved by using circular and square shapes for interference. Compared to the letter-based Flanker experiment, this method better explains the neutral situation and is more convincing. Furthermore, presenting stimuli in a vertical string allows for faster location of the center position compared to a horizontal string, eliminating interference from the search for the target. The experimental results show that the improved EEG ERP analysis indicates greater activity in the prefrontal cortex around 200ms, and FFT analysis shows more high-frequency brainwaves than in the traditional Flanker experiment. This experiment involves interference-based conflict, requiring subjects to quickly eliminate interference from both sides to react. Both the experimental results and theoretical knowledge demonstrate that this experiment tests the subjects' reaction ability, attention, and resistance to interference.

[0105] The significance of Experiment 4 lies in its use of positional conflict to test the subjects' reaction abilities. Due to habitual thinking, subjects tend to use their left hand to make judgments when the position is on the left, and their right hand when the position is on the right. Subjects sometimes need to concentrate and overcome this conflict to make the correct response. Both the experimental results and theoretical knowledge demonstrate that this experiment tests the subjects' reaction abilities and attention span.

[0106] The significance of Experiment 5 lies in testing the subjects' responsiveness by using cue information that conflicts with the target stimulus. The cue information is mostly correct, with a small portion being incorrect. Subjects need to maintain focus and not become distracted because most of the information is correct, or release their attention after preparing in advance. Instead, they need to pay attention throughout the process to avoid being misled by incorrect cues. Both the experimental results and theoretical knowledge demonstrate that this experiment tests the subjects' responsiveness and attention.

[0107] The significance of Experiment Six lies in the fact that continuous response stimuli test a person's attention span. This experiment involves testing attention deficits and impulsive emotions in ADHD patients. The ability to respond correctly reflects the subject's level of attention, and the ability to control the response reflects the ability to control impulsive behavior, which can greatly assist the system in diagnosis.

[0108] This cognitive experimental system comprehensively tests the subjects' responsiveness and concentration in terms of visual conflict, spatial location conflict, perceptual information processing conflict, cue information conflict, continuous response, and interference. By testing the subjects' speed of conflict processing, correctness, and reaction speed, the EEG is analyzed, and then a deep learning model is used for feature learning and prediction to distinguish subjects with different characteristics, ultimately achieving the purpose of assisting diagnosis.

[0109] The multi-stream dual-row decision neural network model consists of two different neural network models. Leveraging their respective advantages, they classify various types of data and ultimately make a joint decision. The neural network models are SBO-RNN and ResNet.

[0110] The SBO-RNN described above is a special recurrent neural network structure. Compared with other neural network structures, the main characteristics of SBO-RNN include:

[0111] In traditional optimization problems, we typically face a single objective function and a set of constraints, finding the optimal solution through optimization algorithms. However, bi-level optimization problems are more complex, involving two nested optimization problems. In this case, we need to optimize one optimization problem under the constraints of the other. The core idea is as follows:

[0112]

[0113] Let ht represent the rate of change of the variable ht in the gradient flow over time. To find the minimum value of the objective function, ht must be a local minimum of the preceding and following states. Because the more compact the EEG time series data, the higher the correlation, the local minimum is taken, thus reducing the likelihood of gradient explosion. Furthermore, each subtraction between the initial state h and the hidden state will inevitably produce a difference, so the rate of change ht ≠ 0, thus reducing the likelihood of gradient vanishing.

[0114] The goal of the SBO-RNN model is to model and solve two-level optimization problems with uncertainty. It uses a recurrent neural network structure to model the uncertainty and trains the network to learn the relationships between decision variables and constraints in the optimization problem. Through backpropagation, SBO-RNN automatically adjusts the network parameters using gradient descent to minimize the objective function.

[0115] The ResNet described is a classic convolutional neural network architecture, where the convolutional layer calculation formula is as follows:

[0116]

[0117] Where represents the neuron in the i-th row and j-th column of layer l, represents the weight in the p-th row and q-th column of the k-th convolutional kernel in layer l, represents the (i+p-1)-th and (j+q-1)-th elements in the k-th feature map of layer l-1, bl represents the bias term of the convolutional layer in l, and f(x) represents the activation function. P and Q represent the size of the convolutional kernel, and K represents the number of convolutional kernels.

[0118] Convolutional layers in convolutional networks introduce the concept of local receptive fields, allowing the model to capture local features in data such as images more precisely. Simultaneously, the weight-sharing mechanism in convolutional layers reduces the number of model parameters that need to be trained, accelerating the training and testing process. However, compared to other neural network architectures, ResNet also has the following characteristics:

[0119] ResNet introduces residual connections, allowing some layers to be skipped and information from lower layers to be directly passed to higher layers. This connection method allows the network to directly learn residual information, making it easier to optimize and train deep networks. Residual connections can use identity mappings or 1x1 convolutions for dimensionality adaptation.

[0120] The basic building block in ResNet is the residual block. Each residual block contains multiple convolutional layers and non-linear activation functions, as well as residual connections that bypass these layers. The introduction of residual blocks allows the network to learn residual functions instead of directly learning low-level features, thereby helping to improve the network's expressive power and optimize performance.

[0121] ResNet has a deeper network structure to extract more complex and abstract features. By using residual connections, ResNet can be scaled up to tens or even hundreds of layers without incurring performance issues or gradient vanishing problems.

[0122] By performing random transformations, rotations, and translations on the input data, the training dataset can be effectively expanded, and the model's generalization ability can be improved. Data augmentation techniques have also been widely used in convolutional models.

[0123] Due to the extensive experimental content, the collected EEG datasets are quite large. Therefore, both recurrent neural networks (RNNs) and convolutional neural networks (CNNs) are prone to model instability issues such as vanishing and exploding gradients when processing this data. The application of the two models mentioned above aims to avoid these problems. After the final model predicts the data, it integrates multiple results into simpler category data, which is then used by the shallow model SVM mentioned above for final discrimination. The SVM model has strong interpretability, simple structure, and fast computation speed, which can improve the diagnostic speed of the system to a certain extent.

[0124] In summary, this invention provides an auxiliary diagnostic system for Attention Deficit Hyperactivity Disorder (ADHD) based on a comprehensive multi-task cognitive experimental set and a multi-stream dual-row decision neural network model. This invention acquires EEG signal data for each subject through different cognitive experiments, thereby improving the individual differences in EEG data regarding ADHD and increasing the classification accuracy of the system model. The model learns from different transformed or processed data, extracting features for accurate classification and recognition, thus achieving the purpose of auxiliary diagnosis.

[0125] The above description is illustrative and not restrictive, and is not intended to limit the invention. For example, the processing of EEG data can be diverse, and cognitive experimental sets can be better selected or modified. The sequence numbers of the above embodiments are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0126] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the protection scope of this invention.

Claims

1. An auxiliary diagnostic system for attention deficit hyperactivity disorder (ADHD) based on a comprehensive multi-task cognitive experimental set and a multi-stream dual-row decision neural network model, characterized in that, The diagnostic system includes the following steps: A set of multiple cognitive experimental task paradigms based on attention deficit hyperactivity disorder (ADHD) syndrome; Obtain basic information about the subjects and EEG data information on their participation in the entire set of cognitive experiments; Based on the acquired EEG data of the subjects, simple preprocessing was performed to remove artifacts and noise, and the target EEG data information was exported in segments according to the Mark markers in the experiment. The exported preprocessed EEG data is fed into the corresponding neural network model to determine the difference between the patient's EEG data and the healthy norm for each experiment, thereby classifying the data. This process is performed on EEG data from various cognitive task paradigms. The model then makes a comprehensive diagnosis, and if the data exceeds a set threshold, the patient is diagnosed with attention deficit hyperactivity disorder (ADHD). Through a collection of cognitive experiments, the different types of EEG signals generated by different experiments are compared and classified with healthy norms in a multi-stream bilinear decision neural network model, and different experiments have corresponding classification results. The model performs different data transformations on the input EEG data, using two different neural network models to classify different data types. For time-series EEG data, a Stochastic Bilevel Optimization RNN network model is used, while for brain topography data, a Residual Network network model is used.

2. The auxiliary diagnostic system according to claim 1, characterized in that, Each experiment in the cognitive experiment set tests and diagnoses the subjects from different perspectives.

3. The auxiliary diagnostic system according to claim 1, characterized in that, After obtaining the EEG data from the subjects, the data needs to be filtered, rereferenced, de-oculars removed, artifact removed, and segmented for correlation processing before the preprocessed EEG data is exported.

4. The auxiliary diagnostic system according to claim 1, characterized in that, The system memory contains pre-trained patient data and health norm-related data.

5. The auxiliary diagnostic system according to claim 1, characterized in that, The final diagnosis is derived from a comprehensive evaluation of multiple experiments. The system ultimately uses the most discriminative native features to make the final judgment using a shallow classifier, the support vector machine.

6. The auxiliary diagnostic system according to claim 2, characterized in that, Different experiments employ different testing strategies, which can reveal the problems of the subjects, and the systematic diagnostic results can lead to corresponding coping strategies.

7. The auxiliary diagnostic system according to claim 5, characterized in that, According to the American Psychiatric Association's DSM-5 diagnostic criteria for Attention Deficit Hyperactivity Disorder (ADHD), which consists of nine diagnostic criteria, a patient can be diagnosed with ADHD if they have six or more symptoms. This standard was used as a reference to pre-train the final assessment model.

Images