Multi-flow graph neural network-based hyperactivity disorder intelligent decision support method and system

Abstract

The application provides a hyperactivity disorder auxiliary diagnosis method based on a multi-flow graph neural network and space-time constraints, comprising: acquiring motion sensor data of a user in a human-computer interaction scene, and extracting time domain statistical features of the motion sensor data; acquiring a space-time constraint graph of the human-computer interaction scene based on the time domain statistical features; acquiring integral graph vector representations of the space-time constraint graph; and fusing all the integral graph vector representations to obtain a classification result of the human-computer interaction behavior of the user. The application also provides a hyperactivity disorder auxiliary diagnosis system based on a multi-flow graph neural network and space-time constraints, and a data processing system for realizing hyperactivity disorder auxiliary diagnosis based on a multi-flow graph neural network.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a multi-flow graph neural network-based intelligent decision support method and system for ADHD. Background Technology

[0002] Attention deficit hyperactivity disorder (ADHD) is a common childhood mental and behavioral disorder with a prevalence rate as high as 5%-10%. Its main symptoms include inattention, hyperactivity, and impulsivity, and these symptoms may not gradually disappear with age. Therefore, researching diagnostic methods for ADHD and providing early intervention is of great significance for controlling the condition.

[0003] Existing clinical diagnostic methods for ADHD mainly include rating scales, neurophysiological examination techniques, cognitive neurophysiological examination techniques, and brain imaging and functional brain imaging techniques. Rating scales, lacking quantitative descriptions, are easily influenced by physician subjectivity; neurophysiological examination techniques have low accuracy due to the presence of comorbidities; cognitive neurophysiological examination techniques and brain imaging and functional brain imaging techniques are easily affected by patient anxiety during the testing process, and the latter are expensive and can cause harm to the body. To address the shortcomings of these diagnostic methods, researchers have recently designed a series of ADHD-assisted diagnostic methods based on wearable and interactive technologies. These methods assess ADHD symptoms through interactive scenario-based games, collecting the subject's motion data during the game using wearable devices, achieving a more objective and non-invasive quantitative assessment through natural sensory interaction.

[0004] During scene interaction tasks performed by wearable devices, there are interactive relationships between the tester and the task nodes, and spatial relationships between the task nodes. However, traditional convolutional neural networks are limited to data structured in an ordered grid in Euclidean space, making it difficult to capture such complex neighborhood information. In contrast, graph neural networks can be used to analyze the characteristics of unstructured data, giving them a natural advantage in modeling data related to interactive scenes. However, most existing graph neural network models applied to wearable motion sensor data are based on the human skeletal structure and still cannot capture the temporal constraints reflecting the dynamic interaction process or the spatial constraints reflecting the static physical scene arrangement in scene interaction tasks.

[0005] Therefore, there is an urgent need to design a graph neural network method based on spatiotemporal constraints in interactive scenarios, which can mine unstructured information contained in interactive scenarios through graph representation learning. Summary of the Invention

[0006] To address the aforementioned problems, this invention utilizes a multi-flow graph neural network based on spatiotemporal constraints to achieve intelligent decision support (assisted diagnosis) for ADHD. This intelligent decision support method for ADHD based on a multi-flow graph neural network includes: acquiring motion sensor data from a user in a human-computer interaction scenario; extracting the temporal statistical features of the motion sensor data; acquiring a spatiotemporal constraint graph of the human-computer interaction scenario based on the temporal statistical features; acquiring the integral graph vector representation of the spatiotemporal constraint graph; and fusing all the integral graph vector representations to obtain the user's attention classification result for human-computer interaction behavior (ADHD intelligent decision support result).

[0007] The ADHD intelligent decision support method based on multi-flow graph neural network of the present invention includes a first time constraint graph and a second time constraint graph, wherein the time-domain statistical features include a first time constraint graph, a second time constraint graph, a first spatial constraint graph, and a second spatial constraint graph. The first time-domain statistical features of the motion sensor data are extracted manually, and the first time constraint graph and the first spatial constraint graph of the human-computer interaction scene are obtained based on the first time-domain statistical features. A deep residual network is used as an automatic feature extractor to extract the second time-domain statistical features of the motion sensor data, and the second time constraint graph and the second spatial constraint graph of the human-computer interaction scene are obtained based on the second time-domain statistical features.

[0008] The ADHD intelligent decision support method based on multi-flow graph neural network described in this invention uses ResNet18 network as the automatic feature extractor.

[0009] The ADHD intelligent decision support method based on multiflow graph neural network of the present invention obtains the whole graph vector representation H1 of the first temporal constraint graph, the whole graph vector representation H2 of the second temporal constraint graph, the whole graph vector representation H3 of the first spatial constraint graph, and the whole graph vector representation H4 of the second spatial constraint graph through a vector extraction network; wherein, the vector extraction network includes a graph convolutional neural network, a regularization operation layer, a Dropout operation layer, and a global average pooling layer.

[0010] The ADHD intelligent decision support method based on multi-flow graph neural networks described in this invention uses the following convolution calculation formula for the graph convolutional neural network:

[0011]

[0012] Let be the adjacency matrix of this spatiotemporal constraint graph. Let X be the degree matrix of the spatiotemporal constraint graph, and let X be the feature matrix composed of the time-domain statistical features. W is the parameter matrix. H is the feature matrix output by the convolutional neural network of this graph. N is the number of digital nodes in this human-computer interaction scenario, F is the feature dimension of X, and F′ is the feature dimension of H.

[0013] The multi-flow graph neural network-assisted diagnosis method for ADHD described in this invention obtains the classification result Y = softmax(ReLU(H5W5+b5)) by linear transformation and summation operations, and by passing it through a fully connected classification layer, where H5 = (H1W1+b1) + (H2W2+b2) + (H3W3+b3) + (H4W4+b4), and W1, b1, W2, b2, W3, b3, W4, b4, W5, and b5 are the fusion parameters of the fully connected classification layer.

[0014] This invention also proposes an intelligent decision support system for ADHD based on a multiflow graph neural network, comprising: a feature extraction module for acquiring motion sensor data of a user in a human-computer interaction scenario, and extracting the temporal statistical features of the motion sensor data; a spatiotemporal constraint graph construction module for acquiring the spatiotemporal constraint graph of the human-computer interaction scenario based on the temporal statistical features; a graph convolution module for acquiring the whole graph vector representation of the spatiotemporal constraint graph; and a fusion and classification module for fusing all the whole graph vector representations to obtain the attention classification result of the user's human-computer interaction behavior.

[0015] The ADHD intelligent decision support system based on a multi-flow graph neural network of the present invention includes a time-domain statistical feature comprising a first time constraint graph and a second time-domain statistical feature, and a spatiotemporal constraint graph comprising a first time constraint graph, a second time constraint graph, a first spatial constraint graph, and a second spatial constraint graph. The feature extraction module includes: a manual feature extraction module for manually extracting the first time-domain statistical feature of the motion sensor data; and an automatic feature extraction module for extracting the second time-domain statistical feature of the motion sensor data using a deep residual network as an automatic feature extractor. The spatiotemporal constraint graph construction module is based on the first... The first temporal constraint map and the first spatial constraint map of the human-computer interaction scenario are obtained based on temporal statistical features, and the second temporal constraint map and the second spatial constraint map of the human-computer interaction scenario are obtained based on the second temporal statistical features. The graph convolution module obtains the whole graph vector representation H1 of the first temporal constraint map, the whole graph vector representation H2 of the second temporal constraint map, the whole graph vector representation H3 of the first spatial constraint map, and the whole graph vector representation H4 of the second spatial constraint map through a vector extraction network. The vector extraction network includes a graph convolutional neural network, a regularization operation layer, a Dropout operation layer, and a global average pooling layer.

[0016] The present invention also proposes a computer-readable storage medium storing computer-executable instructions, characterized in that, when the computer-executable instructions are executed, ADHD intelligent decision support based on multi-flow graph neural networks as described above is implemented.

[0017] The present invention also proposes a data processing device, including the computer-readable storage medium as described above, wherein when the processor of the data processing device retrieves and executes the computer-executable instructions in the computer-readable storage medium, the data processing device implements ADHD intelligent decision support based on a multi-flow graph neural network. Attached Figure Description

[0018] Figure 1 This is the architecture diagram of the interactive scenario-driven multi-flow graph neural network model of the present invention.

[0019] Figure 2 This is a schematic diagram of motion sensor wearing and Schulte grid scene testing.

[0020] Figure 3 This is a diagram illustrating the time constraint composition method for children with ADHD.

[0021] Figure 4 This is a schematic diagram of the spatial constraint composition method of the Schulte Grid.

[0022] Figure 5 This is a schematic diagram of the graph convolution module structure of the multi-flow graph neural network model of the present invention.

[0023] Figure 6 This is a schematic diagram of the data processing device of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0025] To address the shortcomings of the aforementioned methods, this invention proposes a spatiotemporally constrained multi-flow graph neural network-assisted ADHD diagnostic model, which uses graph neural networks to mine unstructured information contained in interactive scenarios. Specifically, the model comprehensively captures relevant information of the interactive scenario by fusing a temporal constraint graph based on the dynamic human-computer interaction process and a spatial constraint graph based on the static physical scene arrangement relationship. Furthermore, it fully mines the time-frequency statistical features and temporal variation features of motion sensors by fusing manual features based on empirical knowledge and automatic features based on deep learning feature extractors.

[0026] The multi-flow graph neural network ADHD-assisted diagnostic model of this invention adopts a graph construction method based on spatiotemporal constraints, constructing a temporal constraint graph reflecting the dynamic human-computer interaction process and a spatial constraint graph reflecting the static physical scene arrangement. The fusion of these spatiotemporal constraint graphs comprehensively captures relevant information about the interaction scene. Furthermore, it integrates human experience and knowledge with deep learning intelligent processing for feature extraction. This invention introduces manual features based on human experience and knowledge, along with automatic features based on a deep learning feature extractor, fusing human experience and knowledge with deep learning intelligent processing to comprehensively mine multidimensional information from motion sensor data.

[0027] Figure 1 The architecture of a spatiotemporally constrained multi-flow graph neural network-based ADHD-assisted diagnostic model is demonstrated. The model consists of four main modules: feature extraction module, spatiotemporally constrained graph construction module, graph convolution module, and fusion and classification module.

[0028] The functions and components of each module are as follows:

[0029] Feature Extraction Module: This module consists of two sub-modules: a manual feature extraction module and an automatic feature extraction module. Manual feature extraction primarily relies on specific professional knowledge to extract the time-frequency statistical features of motion sensor data throughout the task. While this method is fast and efficient and does not introduce additional model parameters, it loses some of the temporal variation characteristics of the motion sensor data and deeper information that is difficult for human experience to capture. Therefore, this module introduces an automatic feature extractor based on deep learning to supplement the manual feature extraction.

[0030] Spatiotemporal constraint graph construction module: The construction of the temporal constraint graph is based on the dynamic human-computer interaction process, while the construction of the spatial constraint graph is based on the static physical scene arrangement relationship. The fusion of spatiotemporal constraint graphs is used to fully explore the unstructured information contained in the interaction scene.

[0031] Graph Convolution Module: This module learns the graph representation of the constructed scene graph based on a graph convolutional neural network, and finally reads out the corresponding whole graph representation vector.

[0032] Fusion and Classification Module: This module first fuses the four different whole graph representation vectors obtained by the graph convolution module. After the fused vector is passed through a fully connected layer, the final ADHD intelligent decision support classification result is obtained, including: normal children (TD), children with attention deficit hyperactivity disorder (ADHD-I), or children with mixed ADHD (ADHD-C).

[0033] This invention's model achieves ADHD diagnosis by modeling motion sensor data and interactive scene-related information of the test subject during scenario testing. This section uses the Schulte Grid interactive scene as an example to introduce the specific implementation method of this invention. The testing rules for this scene are described below: Figure 2 As shown in (a), during the scenario test, each child wore six wearable sensors (left and right wrists, left and right ankles, neck and waist) to collect motion data such as triaxial acceleration and angular velocity during the task. Figure 2 (b) shows the test scenario for the Schulte Grid scenario. The scenario presents a randomly distributed number grid to the test child on a large touchscreen computer. The test child is required to touch the position of each number on the screen in ascending order with their finger. The scenario task is set with three difficulty levels: 2×2, 3×3 and 4×4. The test child needs to complete all the tests in order.

[0034] 1. Problem Definition

[0035] For each child in the test, the number nodes in the Schulte Grid scene are used as nodes in a graph. The relevant input data is represented by a graph G = (V, E, A, X), where V is the set of vertices in the graph (|V| = 29, with three difficulty levels: 2×2, 3×3, and 4×4), and E is the set of edges in the graph. It is an adjacency matrix of a graph, where each node contains a d-dimensional eigenvector. This eigenvector extracts motion sensor data from the previous node to the current node. This is the feature vector matrix of the graph nodes. After transforming the interactive scenario into a graph representation, the ADHD auxiliary diagnosis problem is transformed into a graph-level classification task in a graph neural network. The input is the graph representation information of the child being tested, and the output is the category of the child being tested (normal attention, attention deficit hyperactivity disorder, mixed ADHD), which serves as the intelligent decision support result for ADHD. This support result is presented to the clinician as a basis for the clinician to determine whether the child being tested has attention deficit hyperactivity disorder.

[0036] 2. Feature Extraction Module

[0037] 2.1 Manual Node Feature Extraction Based on Time-Frequency Statistical Features of Sensor Data

[0038] Each sensor generates an acceleration vector a = {a1, a2, ... a} within the test period t. i ,...a t},in It is the composite value of the three-axis acceleration at time i. Based on the generated acceleration vector a, this part extracts a total of 16 features, including 7 time-domain statistical features (mean, standard deviation, mode, maximum value, minimum value, range, number of points greater than the mean) and 9 frequency-domain statistical features (DC component, amplitude mean, amplitude standard deviation, amplitude skewness, amplitude kurtosis, shape mean, shape standard deviation, shape skewness, shape kurtosis).

[0039] Time-domain statistical features refer to statistical features related to time-series changes. n represents the size of the time window used for feature extraction, and a... i Let represent the composite acceleration of the i-th data in the window. The calculation method of the extracted time-domain statistical features is shown in Table 1.

[0040]

[0041]

[0042] Table 1

[0043] Frequency domain statistical features are mainly calculated and discovered using fast Fourier transform. N represents the number of time windows, and C(i) represents the frequency amplitude value of the i-th window. The calculation method of the extracted frequency domain statistical features is shown in Table 2.

[0044]

[0045] Table 2

[0046] 2.2 Automatic Node Feature Extraction Based on Temporal Variation Characteristics of Sensor Data

[0047] Manual feature extraction mainly relies on specific professional knowledge to extract the time-frequency statistical features of motion sensor data throughout the testing process. To a certain extent, this process loses the temporal variation features of the motion sensor data. In order to capture this deep temporal variation information, this module uses ResNet18 as a feature extractor to perform temporal one-dimensional convolution on the motion data to achieve automatic node feature extraction.

[0048] 3. Spatiotemporal constraint graph construction module

[0049] In order to fully capture the spatiotemporal constraint information in the interactive scene, this invention introduces two mapping methods: temporal constraint mapping based on the scene interaction click sequence and spatial constraint mapping based on the inherent arrangement relationship of digital nodes in the scene.

[0050] 3.1 Time-Constrained Graph Construction

[0051] The time-constrained graphing method is primarily based on the dynamic human-computer interaction process during testing, that is, constructing the graph according to the time sequence of the tester's correct clicks on digital nodes. Specifically, for two nodes u and v, a connection is established between u and v only if the time interval between u and v in the click sequence is less than 'a', where 'a' is a set hyperparameter. For example... Figure 3 The experiment presented a time constraint for a child with ADHD. The connections between nodes 1 to 6 were relatively tight, while the connections after node 7 were missing. This indicates that the child had high attention and responsiveness during the first 6 nodes, but lost attention after node 7, resulting in a delay in the clicking interaction. This reflects the child's poor performance in sustained attention.

[0052] 3.2 Spatial Constraints in Layout

[0053] Spatial constraint graph construction is primarily based on the static physical scene layout, that is, graph construction is performed according to the inherent arrangement of digital nodes in the scene. Specifically, a complete graph is first constructed, then the Floyd-Warshall algorithm is executed to calculate the shortest path between each pair of nodes, and finally a threshold hyperparameter β is set. If the shortest path distance between two points is less than β, a connection is established. The specific algorithm flow is shown in Table 3. Figure 4 The spatial constraint diagram of the 4×4 Schulte grid is shown when β=1.

[0054]

[0055]

[0056] Table 3

[0057] 4. Graph Convolution Module

[0058] Since medical datasets typically have a small number of samples, to avoid overfitting due to a mismatch between the number of parameters and the sample size, graph convolution is implemented based on a traditional Graph Convolutional Network (GCN) with regularization and Dropout operations added to the module. Furthermore, to stabilize the learning process, batch normalization is added after graph convolution. The final module composition of the graph convolutional layer is as follows: Figure 5 As shown, the convolution calculation formula for GCN is:

[0059]

[0060] in, Let be the adjacency matrix of the scene graph. Let be the degree matrix of the scene graph. It is a feature matrix (initially composed of features obtained from motion sensor data after feature extraction by a feature extraction module). It is a parameter matrix. This is the feature matrix after convolution, where N is the number of nodes, and F and F′ are the input and output feature dimensions, respectively. GAP (Global Average Pooling) is a global average pooling layer used to read out the final whole image vector representation.

[0061] 5. Integration and Classification Module

[0062] Let H1, H2, H3, and H4 be the complete graph representation vectors obtained after graph convolution layers for the temporal and spatial constraint maps corresponding to manual and automatic features, respectively. To reduce the introduction of parameters, the fusion of multiple data streams is achieved by simple addition after linear transformation, followed by a fully connected classification layer to obtain the final result. The calculation formula is as follows:

[0063] H5=(H1W1+b1)+(H2W2+b2)+(H3W3+b3)+(H4W4+b4)

[0064] Y = softmax(ReLU(H5W5+b5))

[0065] Where W1, b1, W2, b2, W3, b3, W4, b4, W5, b5 are the parameters of the corresponding fully connected classification layer.

[0066] 6. Experimental Verification

[0067] To further verify the effectiveness of the proposed spatiotemporally constrained multi-flow graph neural network ADHD auxiliary diagnostic model, the inventors conducted experiments on data collected from a Schulte grid scene. The dataset included 83 normal children (TD), 50 children with attention deficit hyperactivity disorder (ADHD-I), and 19 children with mixed ADHD (ADHD-C). The ratio of the training set to the test set was 3:2. The baseline models selected for the comparative experiments included Support Vector Machine (SVM), Random Forest (RF), ResNet18, Vgg11, LSTM, and skeleton-based graph convolutional neural network (S-GCN). The results of the comparative experiments are shown in Table 4, and the comparison of network module parameter quantities is shown in Table 5. To verify the effectiveness of each module in the proposed model, the inventors also conducted ablation experiments, the results of which are shown in Table 6.

[0068]

[0069] Table 4

[0070] Network module Parameters ResNet18 3861123 Vgg11 34559427 LSTM 11641858 GCNLayer 33536

[0071] Table 5

[0072] Module Ablation test results Automatic features + time constraints 67.88％ Automatic features + spatial constraints 68.27％ Automatic features + spatiotemporal constraints 71.54％ Manual features + spatial constraints 82.12％ Manual features + time constraints 82.69％ Manual features + spatiotemporal constraints 83.08％ Manual features + automatic features + spatiotemporal constraints 85.19％

[0073] Table 6

[0074] The results in Tables 4 and 5 show that the model proposed in this invention achieved the highest accuracy in the comparative experiment with fewer parameters, demonstrating the effectiveness of the model proposed in this invention. The results of the ablation experiment in Table 6 verify the effectiveness of each module in the model proposed in this invention.

[0075] Figure 6 This is a schematic diagram of the data processing apparatus of the present invention. Figure 6 As shown, embodiments of the present invention also provide a computer-readable storage medium and a data processing apparatus. The computer-readable storage medium of the present invention stores computer-executable instructions. When these computer-executable instructions are executed by the processor of the data processing apparatus, the aforementioned ADHD auxiliary diagnostic method based on spatiotemporal constraints and multi-flow graph neural networks is implemented. Those skilled in the art will understand that all or part of the steps in the above method can be implemented by a program instructing related hardware (e.g., processor, FPGA, ASIC, etc.), and the program can be stored in a readable storage medium, such as a read-only memory, a disk, or an optical disk. All or part of the steps in the above embodiments can also be implemented using one or more integrated circuits. Accordingly, each module in the above embodiments can be implemented in hardware, for example, by implementing its corresponding function through an integrated circuit, or it can be implemented in the form of a software functional module, for example, by a processor executing a program / instruction stored in memory to implement its corresponding function. Embodiments of the present invention are not limited to any particular combination of hardware and software.

[0076] Compared with existing technologies, the multi-flow graph neural network model based on spatiotemporal constraints proposed in this invention can solve the problem that existing algorithm models cannot capture spatiotemporal constraint information in human-computer interaction scenarios. In addition, the fusion of manual and automatic features can achieve more comprehensive feature extraction of motion sensor data, providing decision support for ADHD clinical diagnosis. This not only improves the current ADHD clinical diagnosis rate but also provides theoretical basis and experimental guidance for subsequent research.

[0077] The above embodiments are only used to illustrate the present invention and are not intended to limit the present invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all equivalent technical solutions also fall within the scope of the present invention, and the patent protection scope of the present invention should be defined by the claims.

Claims

1. A method for intelligent decision support in ADHD based on multi-flow graph neural networks, characterized in that, include: Acquire motion sensor data from users in human-computer interaction scenarios and extract the temporal statistical features of the motion sensor data; Spatiotemporal constraint graphs of human-computer interaction scenarios are obtained based on temporal statistical features; Obtain the full-scale vector representation of the spatiotemporal constraint graph; By fusing all whole-image vector representations, we obtain the attention classification results of the user's human-computer interaction behavior, which serve as the intelligent decision support results for the user's ADHD. The time-domain statistical features include first time-domain statistical features and second time-domain statistical features, and the spatiotemporal constraint graph includes first time constraint graph, second time constraint graph, first spatial constraint graph, and second spatial constraint graph, wherein: The first temporal statistical features of the motion sensor data are extracted manually, and the first temporal constraint map and the first spatial constraint map of the human-computer interaction scene are obtained based on the first temporal statistical features. Using a deep residual network as an automatic feature extractor, the second temporal statistical features of motion sensor data are extracted, and the second temporal constraint map and the second spatial constraint map of the human-computer interaction scene are obtained based on the second temporal statistical features. The vector extraction network is used to obtain the whole graph vector representations H1 of the first temporal constraint graph, H2 of the second temporal constraint graph, H3 of the first spatial constraint graph, and H4 of the second spatial constraint graph. The vector extraction network includes a graph convolutional neural network, a regularization operation layer, a Dropout operation layer, and a global average pooling layer.

2. The ADHD intelligent decision support method based on multi-flow graph neural networks as described in claim 1, characterized in that, The ResNet18 network was used as the automatic feature extractor.

3. The ADHD intelligent decision support method based on multi-flow graph neural networks as described in claim 1, characterized in that, The formula for calculating convolution in a graph convolutional neural network is: in, The adjacency matrix of the spatiotemporal constraint graph. The degree matrix of the spatiotemporal constraint graph. The feature matrix is ​​composed of time-domain statistical features. , For the parameter matrix, , The feature matrix output by the graph convolutional neural network. , The number of digital nodes in human-computer interaction scenarios for Feature dimensions, for The feature dimensions.

4. The ADHD intelligent decision support method based on multi-flow graph neural network as described in claim 3, characterized in that, The classification result is obtained through linear transformation and summation operations, followed by a fully connected classification layer. ; in, , , , , , These are the fusion parameters for the fully connected classification layer.

5. A smart decision support system for ADHD based on multiflow graph neural networks, characterized in that, include: The feature extraction module is used to acquire motion sensor data of users in human-computer interaction scenarios, so as to extract the temporal statistical features of the motion sensor data; The spatiotemporal constraint graph construction module is used to obtain spatiotemporal constraint graphs of human-computer interaction scenarios based on temporal statistical features; The graph convolution module is used to obtain the full graph vector representation of the spatiotemporal constraint graph; The fusion and classification module is used to fuse all whole image vector representations to obtain the attention classification results of user human-computer interaction behavior; The time-domain statistical features include the first time-domain statistical features and the second time-domain statistical features; the spatiotemporal constraint graph includes the first time constraint graph, the second time constraint graph, the first spatial constraint graph, and the second spatial constraint graph. The feature extraction module includes: The manual feature extraction module is used to extract the first temporal statistical features of motion sensor data through manual feature extraction. An automatic feature extraction module is used to extract the second temporal statistical features of motion sensor data using a deep residual network as an automatic feature extractor. The spatiotemporal constraint graph construction module obtains the first temporal constraint graph and the first spatial constraint graph of the human-computer interaction scenario based on the first temporal domain statistical characteristics, and obtains the second temporal constraint graph and the second spatial constraint graph of the human-computer interaction scenario based on the second temporal domain statistical characteristics. The graph convolution module obtains the complete graph vector representations H1 of the first temporal constraint graph, H2 of the second temporal constraint graph, H3 of the first spatial constraint graph, and H4 of the second spatial constraint graph through a vector extraction network. The vector extraction network includes a graph convolutional neural network, a regularization operation layer, a Dropout operation layer, and a global average pooling layer.

6. A computer-readable storage medium storing computer-executable instructions, characterized in that, When the computer-executable instructions are executed, the ADHD intelligent decision support method based on a multi-flow graph neural network as described in any one of claims 1 to 4 is implemented.

7. A data processing apparatus comprising the computer-readable storage medium as described in claim 6, wherein when the processor of the data processing apparatus retrieves and executes computer-executable instructions in the computer-readable storage medium, the data processing apparatus implements ADHD intelligent decision support based on a multi-flow graph neural network.

Images