Depression risk detection method based on adaptive multi-scale neighborhood perception fused with space-time diagram convolution

Abstract

The invention relates to a depression risk detection method based on self-adaptive multi-scale neighborhood perception fusion space-time diagram convolution, and aims to solve the problems that traditional depression diagnosis is subjective and existing electroencephalogram detection space-time feature modeling is insufficient. The method comprises the following steps: acquiring electroencephalogram signals of a testee, and extracting frequency spectrum features through a frequency domain feature extraction module; in combination with an adjacent matrix (independent of an electrode physical distance) which is adaptively learned and normalized in training, embedded features are generated through a graph convolutional neural network; and then capturing long-time-history dependence through a dynamic feature extraction module, realizing feature multi-level fusion by means of a multi-cascade multi-scale convolution module, finally integrating local and global features through an adaptive feature fusion module, and inputting a prediction layer to output a detection result. According to the method, electroencephalogram signal time-space domain joint modeling is achieved, brain region function connection dynamic and multi-scale neural features are effectively captured, information redundancy and key feature loss are avoided, classification accuracy and generalization ability are remarkably improved, depression-related key brain region and function network features can be revealed, and the method has high interpretability and potential clinical application value.

Description

Technical Field

[0001] This invention belongs to the field of electroencephalogram (EEG) signal analysis and artificial intelligence technology, and in particular relates to a method for detecting depression risk based on adaptive multi-scale neighborhood perception fusion spatiotemporal graph convolution. Background Technology

[0002] Depression is a prevalent and serious mental disorder. Its persistent low mood and impaired cognitive function significantly impact patients' physical and mental health and quality of life. Traditional diagnosis of depression relies primarily on patient self-reports and clinicians' experience, which is easily influenced by subjective factors and memory bias, lacking objective physiological support. Electroencephalography (EEG), as a non-invasive neurophysiological detection method, can record changes in brain activity with millisecond-level temporal resolution, providing quantifiable neurophysiological indicators for the objective identification of depression. Research on related detection methods has significant clinical application value.

[0003] Early methods for identifying depression were mostly based on the statistical or frequency domain features of EEG signals, combined with traditional machine learning classifiers to complete the discrimination. With the development of deep learning, convolutional neural networks (CNNs) have been used to automatically extract the temporal features of EEG signals, and two-dimensional CNN structures have further attempted to capture both spatial and temporal features simultaneously. Graph neural networks (GNNs) characterize the functional connectivity of brain regions by treating EEG channels as graph nodes and inter-brain region correlations as edges, achieving more physiologically meaningful modeling. Some studies have also introduced multi-scale temporal convolution, attention mechanisms, or dynamic convolution techniques to try to enhance the ability to capture features of different brain regions and time segments.

[0004] However, existing technologies still have significant limitations: traditional graph convolutional networks mostly construct adjacency matrices based on the physical distance of electrodes, which are not entirely consistent with the actual functional connections of brain regions and are difficult to capture key neural activity patterns; global pooling operations are prone to losing fine-grained local information; most methods rely on fixed time windows or static weighting strategies, which cannot effectively capture the complex changes in transient microstates and continuous rhythms in EEG signals; the spatiotemporal feature fusion methods are insufficient, which can easily lead to information redundancy, cross-domain inconsistencies or loss of key features, limiting the model's comprehensive utilization of multi-level neural features, resulting in insufficient detection accuracy and robustness. There is an urgent need for a depression risk detection scheme that can achieve dynamic modeling and efficient fusion of spatiotemporal features. Summary of the Invention

[0005] To address the problems existing in the background art, this invention provides a method for detecting depression risk based on adaptive multi-scale neighborhood-aware fusion spatiotemporal graph convolution, comprising:

[0006] S1: Acquire the subject's EEG signal data, and use the frequency domain feature extraction module to extract features from the subject's EEG signal to obtain the subject's spectral features;

[0007] S2: Based on the subject's spectral characteristics and the adaptively learned adjacency matrix, a graph convolutional neural network is used to aggregate features and generate the subject's embedded features;

[0008] S3: Input the subject's embedded features and the adaptively learned adjacency matrix into the dynamic feature extraction module to extract the dynamic features of the subject;

[0009] S4: Input the subject's embedded features into multiple cascaded multi-scale convolutional modules and fuse them with the subject's dynamic features at multiple levels to obtain the subject's primary fused features;

[0010] S5: Based on the adjacency matrix of adaptive learning, the subject's initial fusion features are adaptively fused through the adaptive feature fusion module to obtain the subject's final fusion features;

[0011] S6: Input the subject's final fusion features into the prediction layer for prediction to obtain the subject's depression risk detection results.

[0012] Another aspect of the present invention provides a depression risk detection system based on adaptive multi-scale neighborhood-aware fusion spatiotemporal graph convolution, the system comprising a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute the depression risk detection method based on adaptive multi-scale neighborhood-aware fusion spatiotemporal graph convolution.

[0013] Another aspect of the present invention provides a computer storage medium storing a remote monitoring program, which, when executed by a processor, implements the aforementioned method for detecting depression risk based on adaptive multi-scale neighborhood perception fusion spatiotemporal graph convolution.

[0014] The present invention has at least the following beneficial effects

[0015] This invention accurately captures high-level temporal features of EEG signals through a frequency domain feature extraction module. It leverages an adaptively learned adjacency matrix to align with the actual functional connections of brain regions rather than physical distances. Combined with a dynamic feature extraction module and a multi-scale convolution module, it flexibly captures transient microstates, continuous rhythms, and spatiotemporal dependencies of EEG signals at different scales. Furthermore, an adaptive feature fusion module efficiently integrates local fine-grained features with global brain network dynamics. This approach avoids information redundancy, cross-domain inconsistencies, and loss of key features, while also enhancing the model's ability to represent complex neural patterns. Ultimately, it achieves higher classification accuracy and generalization ability on public datasets, revealing key brain regions and functional network features related to depression. It possesses strong interpretability and potential clinical application value, providing an objective and accurate risk detection solution for depression. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0017] Figure 2 This is a schematic diagram of the overall structure of the model of the present invention;

[0018] Figure 3 This is a schematic diagram of the adaptive feature fusion module of the present invention;

[0019] Figure 4 This is a schematic diagram comparing the results of this invention with other existing models on the MODMA dataset;

[0020] Figure 5 This diagram illustrates the comparison of results from the present invention on the PRED + CT dataset with other existing models. Detailed Implementation

[0021] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0022] Please see Figure 1 and Figure 2 On the one hand, it provides a method for detecting depression risk based on adaptive multi-scale neighborhood-aware fusion spatiotemporal graph convolution, including:

[0023] S1: Acquire the subject's EEG signal data, and use the frequency domain feature extraction module to extract features from the subject's EEG signal to obtain the subject's spectral features;

[0024] In this embodiment, the experimental environment used was Python 3.7.16, PyTorch 1.12.1, and CUDA 12.0. All experiments in this invention were conducted on the MODMA and PRED+CT resting-state depression datasets. The MODMA dataset contains resting-state EEG records from 24 patients with major depressive disorder (MDD) and 29 healthy controls (HC). All subjects underwent psychiatric evaluation and completed the PHQ-9 and GAD-7 questionnaires before EEG acquisition. Data were sampled using a 128-channel EEG system at a frequency of 250 Hz. After screening, 3600 MDD samples and 4350 HC samples were obtained. The PRED+CT dataset, based on the international 10–20 electrode system, contains 64-channel EEG data involving 55 MDD patients and 63 healthy controls. Participants were diagnosed by psychiatrists based on BDI scores (HC < 7, MDD > 7) and clinical interviews. Each subject received 2 minutes of resting-state EEG data at a sampling rate of 500 Hz. After screening, a total of 8250 MDD samples and 9450 HC samples were obtained.

[0025] Preferably, the frequency domain feature extraction module includes a one-dimensional convolutional layer, a batch normalization layer, and a Mish activation function cascaded in sequence.

[0026] In this embodiment, the one-dimensional convolutional layer performs sliding convolution operations on the temporal dimension of the original EEG signal, focusing on capturing neural activity features corresponding to different frequency bands, filtering redundant temporal noise, and achieving a preliminary mapping from the original signal to frequency domain features. The batch normalization layer standardizes the features output by the convolutional layer, adjusting the feature distribution to a state where the mean is close to 0 and the variance is close to 1, avoiding training instability caused by excessive differences in feature values, and accelerating the model convergence speed. The Mish activation function introduces a nonlinear transformation, which, compared with traditional activation functions, can better preserve gradient information, avoid gradient saturation, enhance the model's ability to express complex frequency domain features, and further optimize feature discrimination. Finally, the extracted spectral features not only retain the spatial distribution information of different EEG channels, but also condense the key neural activity patterns in the frequency domain dimension.

[0027] S2: Based on the subject's spectral characteristics and the adaptively learned adjacency matrix, a graph convolutional neural network is used to aggregate features and generate the subject's embedded features;

[0028] Preferably, the adaptive learning adjacency matrix includes: before training begins, the adjacency matrix is ​​randomly initialized using a Xavier normal distribution; during training, the parameters of the adjacency matrix are adaptively learned through backpropagation; and after training is completed, the adjacency matrix is ​​symmetrically normalized to obtain the adaptive learning adjacency matrix.

[0029] Preferably, the step of using a graph convolutional neural network to aggregate features and generate the subject's embedded features includes:

[0030]

[0031] in, Indicates the first Weight parameters of layer graph convolution; The adjacency matrix represents adaptive learning; Indicates the first Hidden layer features of layer graph convolution; This represents the Mish activation function; Indicates the first Hidden layer features of layer graph convolution; when When it is 1, Indicates the spectral characteristics of the subjects; when for hour, This represents the embedding characteristics of the subjects.

[0032] In this embodiment, before training begins, the adjacency matrix A is randomly initialized using a Xavier normal distribution to ensure a reasonable initial parameter distribution and avoid gradient anomalies in the early stages of training. During training, the parameters of the adjacency matrix A are optimized synchronously with other model parameters through backpropagation, dynamically adapting to the real-world patterns of brain region functional connectivity and overcoming the limitations of traditional methods that rely on physical distance to construct the adjacency matrix. After training, to stabilize the feature propagation process and avoid gradient vanishing or exploding, the adjacency matrix A is symmetrically normalized to obtain an adaptively learned adjacency matrix. , is represented as:

[0033]

[0034] in, Let A be the degree matrix of the adjacency matrix A.

[0035] Graph convolutional neural networks (GNNs) aggregate spectral features on a spatial structure defined by an adaptive adjacency matrix through multi-layered stacked convolution operations. This process gradually uncovers functional connectivity features between brain regions. These features integrate frequency domain information from EEG signals with spatial functional connectivity information of brain regions, resulting in stronger discriminative power. By using an adaptive adjacency matrix to closely match the actual functional connectivity of brain regions, the aggregated features become more physiologically meaningful. This overcomes the limitation of traditional graph convolution relying on physical distance. Through the progressive aggregation of multi-layered graph convolutions, scattered channel features are transformed into structured spatial connectivity features, preserving local information from each brain region while integrating global connectivity between brain regions, laying the foundation for deep fusion of spatiotemporal features.

[0036] S3: Input the subject's embedded features and the adaptively learned adjacency matrix into the dynamic feature extraction module to extract the dynamic features of the subject;

[0037] Preferably, the dynamic characteristics of the subject include:

[0038]

[0039]

[0040]

[0041]

[0042] in, Indicates the dynamic characteristics of the subjects; Represents the learnable weight parameter matrix; The adjacency matrix represents adaptive learning; This represents depthwise separable convolution; Indicates batch normalization; This represents the Mish activation function; , and Indicates intermediate processing features; This represents the embedding characteristics of the subjects.

[0043] In this embodiment, based on the subject's embedded features and an adaptively learned adjacency matrix, a dynamic feature extraction module captures the long-term temporal dependencies and dynamic change patterns of EEG signals, outputting dynamic features with both spatiotemporal dynamic characteristics. The dynamic feature extraction module achieves accurate extraction of dynamic features of EEG signals through a four-step process: projection transformation, feature propagation, channel fusion, and dynamic enhancement. This relies heavily on the synergistic effect of a series of learnable parameters and specific network layers. The embedded feature projection transformation adjusts the dimension and expression of the embedded features through linear transformation, making them more suitable for subsequent spatial propagation operations based on the adjacency matrix, while simultaneously introducing learnable parameters to enhance the adaptive capability of the module. The adaptive adjacency matrix is ​​then utilized. This method guides the spatial propagation of projected features to uncover dynamic functional connections between brain regions. Unlike traditional feature propagation based on physical distance, this step leverages the real functional connectivity of brain regions obtained through training. It allows each EEG channel's features to absorb relevant information from neighboring channels, thereby capturing the dynamic dependency patterns of brain activity in the spatial dimension, especially the coordinated changes between brain regions during short-term fluctuations. To avoid losing original information during feature propagation, a channel splicing strategy is used to fuse propagated and projected features, while preserving the associated features after spatial propagation. and original projection features This approach achieves complementarity between related and original information, providing a rich information foundation for subsequent dynamic feature enhancement. The fused features are processed using deep separable convolution, batch normalization, and nonlinear activation to ultimately generate dynamic features. The deep separable convolution employs a high dilation rate design, reducing model parameter computation while expanding the receptive field, effectively capturing long-term temporal dependencies in EEG signals and overcoming the limitations of traditional convolution in modeling the temporal dimension. Batch normalization standardizes the convolution output features, stabilizing feature distribution and avoiding training instability caused by feature value fluctuations, thus accelerating model convergence. The Mish activation function better preserves gradient information, alleviates gradient saturation, and enhances the model's ability to express complex dynamic features. Therefore, by combining deep separable convolution with an adaptive adjacency matrix, both short-term fluctuations and long-term dependencies in EEG signals are captured simultaneously, accurately depicting the dynamic changes in brain activity and compensating for the shortcomings of existing methods in temporal modeling. Furthermore, the fusion of spatial propagation features and original projection features avoids information loss while strengthening spatiotemporal correlation, making the features more discriminative.

[0044] S4: Input the subject's embedded features into multiple cascaded multi-scale convolutional modules and fuse them with the subject's dynamic features at multiple levels to obtain the subject's primary fused features;

[0045] Preferably, the primary fusion features of the subjects include: defining the input features of each multi-scale convolutional module as follows: ,when hour, Represents the embedding features of the subjects. Indicates the number of multi-scale convolutional modules, then the th... The output features of the multi-scale convolutional modules are represented as follows: ;when hour, , This represents the dynamic characteristics of the subjects; therefore, the primary fusion characteristics of the subjects. .

[0046] Preferably, the multi-scale convolution module includes: multiple convolutional branches of different scales, each branch including a convolutional layer, a batch normalization (BN) layer, and... The activation function sums and averages the features processed by all convolutional branches to obtain the output features of the multi-scale convolutional module.

[0047] In this embodiment, each multi-scale convolutional module is the core unit for feature refinement. It captures features at different time scales through parallel multi-scale convolutional branches, and then fuses them to output more discriminative features. Each multi-scale convolutional module contains multiple convolutional branches of different scales (the patent preferably has three branches with kernel sizes of 3, 5, and 7, corresponding to short, medium, and long time scales). Each branch consists of a convolutional layer, a BN layer, and a Mish activation function cascaded sequentially. The workflow of each branch is as follows: the convolutional layer performs sliding convolution on the input features using kernels of different sizes. Different kernels correspond to different receptive fields; small kernels capture short-term local fluctuations, while large kernels capture long-term trend features, achieving parallel extraction of multi-scale temporal information. The BN layer standardizes the output features of the convolution, stabilizing the feature distribution and avoiding issues caused by multi-scale convolutions. The training fluctuations caused by differences in degree features are addressed; the Mish activation function introduces a nonlinear transformation to enhance the module's ability to express complex multi-scale features and alleviate gradient saturation problems; the features processed by all convolutional branches are summed and averaged; the features after branch fusion are normalized in time length by an adaptive average pooling layer to generate fixed-dimensional module output features; multi-level fusion gradually integrates multi-scale refined features and dynamic features through module cascading and dynamic feature accumulation, and captures short, medium, and long-term features in parallel through multi-scale convolutional modules, making up for the limitations of single-scale modeling in traditional methods and comprehensively covering the temporal dimension information of EEG signals; a progressive fusion strategy is adopted to gradually couple dynamic features with multi-scale refined features, avoiding information conflicts or redundancy caused by one-time fusion, and improving feature consistency and discriminability.

[0048] Please see Figure 3 S5: Based on the adjacency matrix of adaptive learning, the primary fusion features of the subjects are adaptively fused through the adaptive feature fusion module to obtain the final fusion features of the subjects;

[0049] Preferably, the final fusion characteristics of the subject include:

[0050]

[0051]

[0052]

[0053]

[0054]

[0055]

[0056] in, This indicates the final fusion characteristics of the subjects; Adjacency matrix representing adaptive learning middle Arrive A matrix composed of rows; Represents the central matrix; Indicates the preset parameters; Indicates local features; Represents global features; Indicates the activation function; Indicates the batch normalization layer; This represents one-dimensional convolution; Indicates transpose; This indicates the primary fusion characteristics of the subjects; As an intermediate feature, This represents channel-wise convolution. Indicates adaptive average pooling. This indicates that the pieces are joined along the feature dimension; express of Characteristics of rows; This indicates rounding down to the nearest integer.

[0057] In this embodiment, the primary fused features generated in the steps are processed by the adaptive feature fusion module. Adaptive integration of local and global features eliminates information redundancy and cross-domain inconsistencies, outputting a final fused feature that combines fine-grained details with global correlations. The local feature branch captures fine-grained spatiotemporal details of each EEG channel through feature transposition, one-dimensional convolution, batch normalization, and Mish activation. The global feature branch constructs a central matrix based on an adaptive adjacency matrix, guiding the initial fused features to perform global spatial correlation aggregation, and then optimizes through feature concatenation and deep convolution to obtain globally correlated features. Finally, the element-average fusion strategy balances the weights of local and global features, outputting a final fused feature that combines fine-grained details with global functional correlations, effectively eliminating information redundancy and cross-domain inconsistencies, and providing optimal feature support for subsequent accurate predictions.

[0058] S6: Input the subject's final fusion features into the prediction layer for prediction to obtain the subject's depression risk detection results.

[0059] Preferably, the depression risk test results of the subjects include:

[0060]

[0061] in, This indicates the results of the depression risk test for the subjects; Indicates a fully connected layer; This indicates the final fusion characteristics of the subjects.

[0062] Please refer to Figure 2 , Figure 2 This is a model diagram provided in the example of this application, showing the overall framework of our invention's spatiotemporal graph convolution model based on multi-scale dynamic fusion. The invention mainly consists of four parts: an adaptive graph convolution topology module, a dynamic feature extraction module, a multi-scale convolution module, and an adaptive feature fusion module. Specifically, the adaptive graph convolution topology module dynamically learns the spatial dependencies between EEG electrodes through a trainable graph structure, capturing functional connectivity patterns that evolve over time; the dynamic feature extraction module uses graph convolution to model the dynamic relationships between EEG channels, thereby extracting spatial features, enabling the model to capture topological structures and spatial dependencies related to brain activity; the multi-scale convolution module simultaneously captures short-term fluctuations and long-term trends of EEG signals through deformable convolutions of different receptive fields; and the adaptive feature fusion module fuses global and local EEG features through a flexible dual-branch mechanism, such as... Figure 3 As shown, combining large-scale brain network patterns with subtle local activities ensures that feature learning for MDD is both comprehensive and discriminative. The loss calculation is as follows:

[0063]

[0064] This invention uses cross-entropy as the loss function for model prediction. and They represent the first Each sample in category The true label and the predicted probability; Indicates the number of samples; Indicates the number of categories; This represents the loss function.

[0065] Please see Figure 4 ,pass Figure 4 It can be observed that this invention achieves high performance on the MODMA dataset in key metrics such as accuracy, precision, and F1 score. It demonstrates a significant advantage over the strongest baseline model, EEGNet. Furthermore, MCDGCN substantially outperforms CNN1DNet and SSGCNnet. These improvements are statistically significant (fDR-corrected t-test, p < 0.05), indicating that MCDGCN effectively captures the complex spatiotemporal dynamics of EEG signals, thus constructing a more robust depression detection model.

[0066] Please see Figure 5 ,pass Figure 5It can be observed that this invention comprehensively outperforms all comparable models on the PRED+CT dataset. MCDGCN surpasses the second-best performing model, EEGVit, in all metrics. Compared to DGCNN, TSception, and SSGCNnet, the improvement of MCDGCN is statistically significant (p<0.05, t-test corrected for FDR). These results establish the superiority of MCDGCN (this invention) for MDD (depression) detection on this dataset.

[0067] Another aspect of the present invention provides a depression risk detection system based on adaptive multi-scale neighborhood-aware fusion spatiotemporal graph convolution, the system comprising a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute the depression risk detection method based on adaptive multi-scale neighborhood-aware fusion spatiotemporal graph convolution.

[0068] Another aspect of the present invention provides a computer storage medium storing a remote monitoring program, which, when executed by a processor, implements the aforementioned method for detecting depression risk based on adaptive multi-scale neighborhood perception fusion spatiotemporal graph convolution.

[0069] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0070] In summary, this invention accurately captures high-level temporal features of EEG signals through a frequency domain feature extraction module. It utilizes an adaptively learned adjacency matrix to align with the actual functional connections of brain regions rather than physical distances. Combined with a dynamic feature extraction module and a multi-scale convolution module, it flexibly captures transient microstates, continuous rhythms, and spatiotemporal dependencies at different scales in EEG signals. Furthermore, an adaptive feature fusion module efficiently integrates local fine-grained features with global brain network dynamics. This approach avoids information redundancy, cross-domain inconsistencies, and the loss of key features, while also enhancing the model's ability to represent complex neural patterns. Ultimately, it achieves higher classification accuracy and generalization ability on public datasets, revealing key brain regions and functional network features related to depression. It possesses strong interpretability and potential clinical application value, providing an objective and accurate risk detection solution for depression.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for detecting depression risk based on adaptive multi-scale neighborhood perception fusion spatiotemporal graph convolution, characterized in that, include: S1: Acquire the subject's EEG signal data, and use the frequency domain feature extraction module to extract features from the subject's EEG signal to obtain the subject's spectral features; S2: Based on the subject's spectral characteristics and the adaptively learned adjacency matrix, a graph convolutional neural network is used to aggregate features and generate the subject's embedded features; S3: Input the subject's embedded features and the adaptively learned adjacency matrix into the dynamic feature extraction module to extract the dynamic features of the subject; S4: Input the subject's embedded features into multiple cascaded multi-scale convolutional modules and fuse them with the subject's dynamic features at multiple levels to obtain the subject's primary fused features; S5: Based on the adjacency matrix of adaptive learning, the subject's initial fusion features are adaptively fused through the adaptive feature fusion module to obtain the subject's final fusion features; S6: Input the subject's final fusion features into the prediction layer for prediction to obtain the subject's depression risk detection results.

2. The method for detecting depression risk based on adaptive multi-scale neighborhood perception fusion spatiotemporal graph convolution according to claim 1, characterized in that, The adaptive learning adjacency matrix includes: before training begins, the adjacency matrix is ​​randomly initialized using a Xavier normal distribution; during training, the parameters of the adaptive learning adjacency matrix are adaptively learned through backpropagation; and after training is completed, the adjacency matrix is ​​symmetrically normalized to obtain the adaptive learning adjacency matrix.

3. The method for detecting depression risk based on adaptive multi-scale neighborhood perception fusion spatiotemporal graph convolution according to claim 1, characterized in that, The frequency domain feature extraction module includes a cascaded one-dimensional convolutional layer, a batch normalization layer, and a Mish activation function.

4. The method for detecting depression risk based on adaptive multi-scale neighborhood perception fusion spatiotemporal graph convolution according to claim 1, characterized in that, The method of using graph convolutional neural networks to aggregate features and generate embedded features of subjects includes: in, Indicates the first Weight parameters of layer graph convolution; The adjacency matrix represents adaptive learning; Indicates the first Hidden layer features of layer graph convolution; This represents the Mish activation function; Indicates the first Hidden layer features of layer graph convolution; when When it is 1, Indicates the spectral characteristics of the subjects; when for hour, This represents the embedding characteristics of the subjects.

5. The method for detecting depression risk based on adaptive multi-scale neighborhood perception fusion spatiotemporal graph convolution according to claim 1, characterized in that, The dynamic characteristics of the subjects include: in, Indicates the dynamic characteristics of the subjects; Represents the learnable weight parameter matrix; The adjacency matrix represents adaptive learning; This represents depthwise separable convolution; Indicates batch normalization; This represents the Mish activation function; , and Indicates intermediate processing features; This represents the embedding characteristics of the subjects.

6. The method for detecting depression risk based on adaptive multi-scale neighborhood perception fusion spatiotemporal graph convolution according to claim 1, characterized in that, The primary fusion features of the subjects include: defining the input features of each multi-scale convolutional module as follows: ,when hour, This represents the embedding features of the subjects. Indicates the number of multi-scale convolutional modules, then the th... The output features of the multi-scale convolutional modules are represented as follows: ;when hour, , This represents the dynamic characteristics of the subjects; therefore, the primary fusion characteristics of the subjects. .

7. A method for detecting depression risk based on adaptive multi-scale neighborhood perception fusion spatiotemporal graph convolution according to claim 1 or 6, characterized in that, The multi-scale convolution module includes: multiple convolutional branches of different scales, each branch including a convolutional layer, a batch normalization (BN) layer, and... The activation function sums and averages the features processed by all convolutional branches to obtain the output features of the multi-scale convolutional module.

8. The method for detecting depression risk based on adaptive multi-scale neighborhood perception fusion spatiotemporal graph convolution according to claim 1, characterized in that, The final fusion characteristics of the subjects included: in, This indicates the final fusion characteristics of the subjects; Adjacency matrix representing adaptive learning middle Arrive A matrix composed of rows; Represents the central matrix; Indicates the preset parameters; Indicates local features; Represents global features; Indicates the activation function; Indicates the batch normalization layer; This represents one-dimensional convolution; Indicates transpose; This indicates the primary fusion characteristics of the subjects; As an intermediate feature, This represents channel-wise convolution. Indicates adaptive average pooling. This indicates that the pieces are joined along the feature dimension; express of Characteristics of rows; This indicates rounding down to the nearest integer.

9. A depression risk detection system based on adaptive multi-scale neighborhood perception fusion spatiotemporal graph convolution, characterized in that, The system includes a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute the depression risk detection method based on adaptive multi-scale neighborhood perception fusion spatiotemporal graph convolution as described in any one of claims 1 to 8.

10. A computer storage medium, characterized in that, The computer storage medium stores a remote monitoring program, which, when executed by the processor, implements a depression risk detection method based on adaptive multi-scale neighborhood perception fusion spatiotemporal graph convolution as described in any one of claims 1 to 8.

Images