Depression level assessment system based on body partition perception graph convolution network

Abstract

The application discloses a depression degree evaluation system based on body partition perception graph convolution network, wherein the system comprises the following steps: acquiring a gait depth video of an object to be evaluated, extracting a corresponding original three-dimensional skeleton sequence, preprocessing the original three-dimensional skeleton sequence to obtain a standardized skeleton sequence; performing partition mapping on the standardized skeleton sequence according to upper, middle and lower three human body partitions to obtain upper partition features, middle partition features and lower partition features; inputting each partition feature into a trained depression degree evaluation network to obtain a depression degree evaluation result of the object to be evaluated; wherein the trained depression degree evaluation network performs partition position coding, intra-partition self-attention and cross-partition fusion on each partition feature to obtain refined partition fusion features; inputting the partition fusion features into a graph convolution backbone network to extract global spatio-temporal features; performing pooling and classification on the global spatio-temporal features to output a depression severity evaluation result.

Description

Technical Field

[0001] This application relates to the field of intelligent medical assessment technology, and in particular to a depression assessment system based on a body partitioning perceptual graph convolutional network. Background Technology

[0002] Adolescent depression is characterized by its insidious onset, strong heterogeneity in clinical manifestations, and significant fluctuations in course. Current severity assessments mainly rely on scales, interviews, and physician experience. While scales such as the HAMD-17 can provide standardized scoring criteria, the assessment process is still limited by the experience of the rater, the subjective reports of the subjects, and the dispersed outpatient follow-up scenarios, making it difficult to meet the needs for continuous, objective, and low-interference monitoring.

[0003] Research has attempted to use multimodal signals such as EEG, facial expressions, speech, and eye movements to analyze automated methods for depression identification. However, such methods usually have high requirements for data collection equipment, environmental control, and user cooperation. In practical scenarios, they may also be constrained by factors such as privacy, deployment costs, and long-term compliance.

[0004] Gait, as a naturally occurring behavioral signal, has advantages such as being non-contact, repeatable, and passively acquireable. By acquiring three-dimensional skeleton sequences using a depth camera, quantitative analysis of human spatiotemporal movement patterns can be achieved at a relatively low cost, thus providing a potential objective technical approach for assessing the severity of depression in adolescents.

[0005] Existing gait-based methods for assessing the severity of depression can be broadly categorized into two types: one relies on artificially designed kinematic features such as gait speed, stride length, and joint angles, while the other relies on sequence models or graph convolutional models to learn representations directly from skeleton sequences. The former over-relies on predefined indicators and struggles to adequately characterize the high-order nonlinear dynamics generated by multi-joint coupling; the latter, while possessing end-to-end learning capabilities, still lacks explicit modeling of coarse-grained semantic partitioning of the body.

[0006] Furthermore, clinical and behavioral studies have found that depression-related gait changes often include slower walking speed, reduced arm swing, drooping posture, and decreased head movement. These phenomena suggest that depression-related gait abnormalities are not evenly distributed throughout the body, but are more likely to appear heterogeneously in different parts of the upper, middle, and lower body.

[0007] However, existing methods mostly focus on topological relationship modeling or receptive field expansion, paying insufficient attention to body partition semantics and dynamic selection of key joints. This makes it difficult to stably characterize subtle gait abnormalities related to the severity of adolescent depression, which is very detrimental to the assessment of the severity of adolescent depression. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this application provides a depression severity assessment system based on a body partitioning-aware convolutional network. Addressing the limitations of current methods for assessing the severity of depression in adolescents, this application conducts research based on 3D skeletal gait data of 119 subjects acquired using a depth camera, including 39 healthy subjects, 41 subjects with mild depression, and 39 subjects with moderate to severe depression, for a task assessing the severity of depression in adolescents. Based on this, an upper, middle, and lower body partitioning-aware modeling module and a partitioning feature refinement module are designed. By mapping the human skeleton into three body partitions (upper, middle, and lower) and refining and enhancing key joint features under partitioning constraints, the system improves the ability to recognize subtle depression-related gait cues, achieving a more objective and accurate automatic assessment of the severity of depression in adolescents.

[0009] On the one hand, a depression assessment system based on a body partitioning-aware graph convolutional network is provided, including: The data acquisition and preprocessing module is configured to: acquire gait depth video of the object to be evaluated, extract the corresponding original three-dimensional skeleton sequence, preprocess the original three-dimensional skeleton sequence to obtain a standardized skeleton sequence; The partitioning mapping module is configured to: perform partitioning mapping on the standardized skeleton sequence according to three human body partitions: upper, middle, and lower, to obtain upper partition features, middle partition features, and lower partition features; The depression severity assessment module is configured to: input the features of each partition into a trained depression severity assessment network to obtain the depression severity assessment result of the subject to be assessed; wherein, the trained depression severity assessment network performs partition position encoding, intra-partition self-attention, and cross-partition fusion on the features of each partition to obtain refined partition fusion features; input the partition fusion features into a graph convolutional backbone network to extract global spatiotemporal features; perform pooling and classification on the global spatiotemporal features to output the depression severity assessment result.

[0010] The above technical solution has the following advantages or beneficial effects: (1) This application introduces semantic priors of body parts into the skeleton modeling process by explicitly dividing the human body into three parts: upper, middle and lower. This makes it more suitable for describing the heterogeneous distribution of depression-related gait abnormalities in different body parts.

[0011] (2) By setting up partition location coding, self-attention within partition and cross-partition fusion mechanism, this application can highlight key joints and key moments under partition constraints, reduce the impact of irrelevant motion disturbances on the evaluation results, and improve the ability to express subtle depressive gait cues.

[0012] (3) This application is based on the analysis of the three-dimensional skeleton sequence obtained by the depth camera. It does not require wearable devices and complex contact sensors, and has the advantages of being non-contact, low-intrusion and easy to deploy. Attached Figure Description

[0013] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0014] Figure 1 This is a schematic diagram of the adolescent depressive gait data collection paradigm provided in Embodiment 1 of this application; Figure 2 This is the overall flowchart of the convolutional network based on upper, middle and lower body partition perception maps provided in Embodiment 1 of this application; Figure 3 This is a structural diagram of the partition location encoding module provided in Embodiment 1 of this application; Figure 4 This is a structural diagram of the self-attention module within a partition provided in Embodiment 1 of this application; Figure 5 This is a visualization of the confusion matrix results of the proposed method in the adolescent depression severity assessment task provided in Embodiment 1 of this application. Detailed Implementation

[0015] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0016] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of this application. The terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0017] Example 1 like Figure 2 As shown, this embodiment provides a depression severity assessment system based on a body partitioning-aware graph convolutional network, including: The data acquisition and preprocessing module is configured to: acquire gait depth video of the object to be evaluated, extract the corresponding original three-dimensional skeleton sequence, preprocess the original three-dimensional skeleton sequence to obtain a standardized skeleton sequence; The partitioning mapping module is configured to: perform partitioning mapping on the standardized skeleton sequence according to three human body partitions: upper, middle, and lower, to obtain upper partition features, middle partition features, and lower partition features; The depression severity assessment module is configured to: input the features of each partition into a trained depression severity assessment network to obtain the depression severity assessment result of the subject to be assessed; wherein, the trained depression severity assessment network performs partition position encoding, intra-partition self-attention, and cross-partition fusion on the features of each partition to obtain refined partition fusion features; input the partition fusion features into a graph convolutional backbone network to extract global spatiotemporal features; perform pooling and classification on the global spatiotemporal features to output the depression severity assessment result.

[0018] Furthermore, the partition location encoding is used to generate partition-level temporal location descriptions and inject them back into the joint features of each partition; the intra-partition self-attention is used to adaptively reweight the joint features within each human body partition; and the cross-partition fusion is used to fuse the refined features of multiple partitions in the channel dimension and map them back to a unified feature space.

[0019] Furthermore, the acquisition of gait depth video of the subject to be evaluated and the extraction of the corresponding original 3D skeleton sequence are performed simultaneously using three Azure Kinect DK depth cameras. The depth video resolution is set to 320×288 pixels, the sampling frame rate is set to 30 Hz, and the 3D skeleton sequence is extracted using the Azure Kinect SDK. During the acquisition process, the subject completes a back-and-forth walking task along a preset walking route to obtain a gait skeleton sequence containing multiple complete gait cycles.

[0020] Further, the preprocessing of the original 3D skeleton sequence to obtain a standardized skeleton sequence includes: performing forward-looking segment preservation, camera tilt correction, root node translation normalization, smoothing filtering, and fixed-length resampling or padding on the gait 3D skeleton sequence to obtain a standardized skeleton sequence, specifically including: Original skeleton sequence of each sample Represented as:

[0021] in, , Indicates the number of channels in the three-dimensional coordinate system. Indicates the number of frames in the original sequence. Indicates the number of joints.

[0022] Because the Azure Kinect sensor is installed at an angle relative to the ground, the walking direction is not perfectly aligned with the Z-axis of the original Kinect. Let the... The Z and Y coordinates of the SpineBase joint at the base of the spine (center of the hip) in the frame are respectively... and The angle between the camera's optical axis and the ground is denoted as . Then, the least squares fitting method is used to estimate the tilt angle. :

[0023] in, This indicates the number of frames in the sequence. (The result is...) Then, the three-dimensional coordinates of each joint in each frame are... Around Perform a rotational transformation on the axis to make the corrected Align the axis with the actual walking direction:

[0024] After angle correction, the SpineBase joint is used as the root node to remove the global translation effect. That is, in each frame, all joint coordinates are subtracted from the corresponding frame's SpineBase coordinates, so that the skeleton sequence has translation invariance in the camera coordinate system.

[0025] To reduce noise and smooth the skeleton trajectory, a Gaussian filter was applied to each gait segment. A sliding window of length 5 was used, and the filter was applied to each joint separately. , and The trajectory sequences in the three coordinate directions are filtered independently.

[0026] Since gait segments may still vary in duration, to ensure consistent input size for subsequent processing and network training, each skeleton sequence is resampled to a fixed length of 50 frames, thus obtaining a standardized skeleton sequence. :

[0027] in, Indicates batch size. This indicates the number of coordinate channels, preferably the number of three-dimensional coordinate channels; This represents the length of the unified sequence after resampling or padding, preferably 50; Indicates the number of joints.

[0028] The camera tilt angle relative to the ground is estimated based on the coordinate changes of the root node in the vertical and walking directions. The three-dimensional coordinates of each joint are then rotated around the horizontal axis to complete the tilt angle correction. After tilt angle correction, translation normalization is performed using the SpineBase joint as the root node. The coordinate sequences of each joint are then smoothed and resampled or padded to obtain a standardized skeleton sequence with consistent duration, invariant translation, and suppressed noise.

[0029] Furthermore, the step of partitioning the standardized skeleton sequence according to three human body partitions—upper, middle, and lower—to obtain upper partition features, middle partition features, and lower partition features specifically includes: First, perform partition mapping on the human skeleton, assuming the complete set of joints is... It is divided into three non-overlapping partitions: upper partition, middle partition, and lower partition.

[0030] in, , and These correspond to three human body sections: upper, middle, and lower. The upper section corresponds to the joints related to the head and neck, the middle section corresponds to the joints related to the trunk and upper limbs, and the lower section corresponds to the joints related to the pelvis and lower limbs. For each partition Define the corresponding binary selection mask. Then the partition features Represented as:

[0031] in, Indicates in Element-wise multiplication after broadcasting in a dimension. The mask satisfies... .

[0032] This partition mapping allows for the explicit introduction of semantic priors for body parts and reduces cross-partition interference between different body parts.

[0033] The three human body regions, namely upper, middle, and lower, include: an upper region for representing head and neck movements, a middle region for representing trunk and upper limb movements, and a lower region for representing pelvic and lower limb movements. The original joint set is mapped by setting corresponding binary masks for each region to obtain upper region features, middle region features, and lower region features.

[0034] Furthermore, the step of inputting the features of each partition into the trained depression assessment network to obtain the depression assessment result of the subject to be assessed specifically includes: Construct a training set, which consists of the features of each partition of the known depression severity classification results; The training set is input into the depression assessment network to train the network. Training is stopped when the network's loss function value no longer decreases or the number of iterations exceeds a set number, resulting in the trained depression assessment network.

[0035] Furthermore, such as Figure 3 As shown, the process of performing partition location encoding, intra-partition self-attention, and cross-partition fusion on the features of each partition yields refined partition fusion features. Specifically, partition location encoding includes: For any partition Features are converged along the joint dimensions within the partition to obtain a partition-level frame description. And further generate partition location codes. :

[0036]

[0037] in, Represents the first normalized backbone sequence. Frame, First The features of a joint are represented by the colon ":", which indicates all elements of that dimension. , Indicates the first The set of joint indexes contained in each partition. Indicates the first The number of joints contained in each partition. This represents a feedforward neural network; Then, the encoding Broadcast along the joint dimension and back-inject residuals into all joint features of the current partition:

[0038] in, Indicates the first In the partition, the th Frame, First The enhanced joint features are obtained by adding the original input features to each joint after partitioning and position encoding. The symbol ":" indicates that all feature channels are selected.

[0039] Through the above operations, without compromising the original spatiotemporal resolution, partition-level semantic cues are injected into the subsequent feature refinement process.

[0040] The partition location encoding is used to aggregate the joint features of each partition in each frame to generate a partition-level temporal description, and generate the corresponding partition location encoding through lightweight mapping. The partition location encoding is then injected back into the joint features of the corresponding partition to inject partition identity information and body position semantics.

[0041] Furthermore, such as Figure 4 As shown, the process of performing partition location encoding, intra-partition self-attention, and cross-partition fusion on the features of each partition yields refined partition fusion features. Specifically, intra-partition self-attention includes: Model the joint correlations within each partition, for the first... Frames will partition The internal joints are treated as tokens, and... Transpose ; Intra-partition self-attention is calculated as follows:

[0042] in, , and For learnable projection matrices, To scale the dimensions, This represents the token matrix after the joints in the frame input, representing the body partition; This represents the partition token obtained after attention reweighting; For each frame, According to the zones The corresponding set of original joint indices The feature vectors are then backfilled into the corresponding positions in the complete skeleton layout to obtain the partitioned feature vectors. Specifically, in the first Frame, set of joint indices Features at corresponding positions for: Among them, Transpose( ) indicates that Transpose the joint token dimension and feature channel dimension; This allows for the enhancement of key joint features within a partition while maintaining the alignment of the complete skeleton topology.

[0043] The self-attention within each partition independently calculates the inter-joint correlation frame by frame within each human body partition, in order to highlight key joints related to the severity of depression and suppress irrelevant perturbations under partition constraints. Furthermore, the partition tokens reweighted by self-attention are written back to their original joint positions to maintain the alignment of each partition branch with the complete skeleton layout.

[0044] Furthermore, the partition location encoding, intra-partition self-attention, and cross-partition fusion are performed on the features of each partition to obtain refined partition fusion features. Specifically, cross-partition fusion includes: The three partition features are concatenated along the channel dimension and then mapped back to the original channel dimension to obtain the partition fusion feature. :

[0045] in, and These represent the trainable weights and trainable biases of a linear projection, respectively; the symbol " "Indicates a linear mapping along the channel dimension; These represent the feature tensors obtained after the upper, middle, and lower partitions are refined and backfilled by self-attention within the partitions, respectively. Presentation layer normalization processing; The resulting enhanced features While preserving the results of the partition semantic modeling, it also maintains the feature layout consistent with the original skeleton topology.

[0046] The cross-partition fusion is used to concatenate the refined features of multiple partitions in the channel dimension, and restore them to a unified feature space through linear mapping, normalization and nonlinear activation to obtain partition fusion features. The partition fusion features are input into the CTR-GCN graph convolutional backbone network, and output three severity assessment results of health, mild depression and moderate to severe depression through global average pooling and fully connected classification layers.

[0047] Furthermore, the step of inputting the partitioned fusion features into the graph convolutional backbone network to extract global spatiotemporal features specifically includes: The obtained partition fusion features A graph convolutional backbone network is used for global spatiotemporal modeling. Preferably, the graph convolutional backbone network is a CTR-GCN backbone network.

[0048] Furthermore, the process of pooling and classifying the global spatiotemporal features to output an assessment result of depression severity specifically includes: The high-level features extracted by the backbone network are used to perform three-class classification prediction through a global average pooling layer, a fully connected layer, and a Softmax classification layer, outputting the severity assessment results of depression in three categories: healthy, mild depression, and moderate to severe depression.

[0049] The proposed method was validated using a dataset of adolescent depressive gait. The data was collected in the following manner: To collect gait data from adolescent subjects under controlled conditions, this method employs a standardized gait data collection paradigm, conducting data collection in a hospital corridor environment, such as... Figure 1 As shown in the figure. During the data acquisition process, three Azure Kinect DK depth cameras were used simultaneously to acquire gait depth videos of the subjects' round-trip walking. Subjects completed the round-trip walking task along a preset walking route at a comfortable walking pace of their own choosing to obtain gait segments containing rich temporal contextual information. Considering that continuous walking may affect the subjects' condition and further interfere with the stability of gait performance, rest periods were set between each acquisition round to reduce the impact of fatigue on data quality. Medical personnel were present on-site during the acquisition process to ensure the safety of the subjects and the smooth progress of the acquisition process. In addition, a professional physician assessed the severity of the subjects' gait using the HAMD-17 scale, and the assessment results were used as clinical labels for subsequent modeling and analysis.

[0050] To systematically evaluate the model's generalization ability across individual scenarios, a subject-independent 5-fold cross-validation strategy was adopted: training and test sets were constructed based on the subjects, ensuring that data from the same subject did not appear in both the training and testing phases simultaneously. To avoid potential data leakage and ensure the objectivity of the evaluation, all processes related to model selection were strictly limited to the training fold, and test fold data remained completely isolated within each fold, used only for the final performance evaluation.

[0051] At the level of index aggregation, this study selected accuracy, macro-precision, macro-recall, and macro-F1 score as the main evaluation indicators to comprehensively evaluate the recognition performance of the proposed method in assessing the severity of depression in adolescents from different perspectives.

[0052] To ensure fairness in the comparison, all experiments were conducted under uniform settings. All experiments were performed with a fixed random seed (seed = 42). All models were implemented in PyTorch, using the Adam optimizer, with an initial learning rate of 0.001, a batch size of 8, cross-entropy loss as the supervision signal, and a maximum training epoch of 100. All comparison models were trained and tested under the same experimental settings to ensure the validity and reproducibility of the results comparison.

[0053] Under independent 5-fold cross-validation by the subjects, the proposed method achieved 75.63% accuracy, 75.41% macro-mean precision, 75.84% macro-mean recall, and 75.52% macro-mean F1 score on a sample-level adolescent depression severity assessment task, indicating that the method has relatively stable overall recognition ability across subject testing scenarios. Furthermore, Figure 5 The confusion matrix results for the proposed method are presented. The confusion matrix shows a clear diagonal clustering trend, indicating that the model can correctly classify most samples. The proposed method can effectively distinguish between the healthy group and the depressed group, and maintains good recognition ability for moderate to severe depression samples. However, some confusion still exists between mild depression and adjacent categories, which is consistent with the inherent challenge of assessing the severity of depression in adolescents.

[0054] To evaluate the impact of partition granularity design on model performance in the proposed upper, middle, and lower body partition perception modeling modules, this application conducted partition ablation experiments using the same data preprocessing, training settings, and evaluation protocol as the main experiment. Therefore, partition ablation primarily focuses on the impact of different body partition granularities on model performance. This application designed three partition configurations: (1) Single partition: The whole skeleton is modeled as a whole partition; (2) Dual partitioning: The upper body partition and the middle body partition are merged into one partition, and the lower body partition is modeled as another partition; (3) Three-partition: The upper body partition, middle body partition and lower body partition are modeled as three independent partitions.

[0055] Table 1 reports the results of the partitioning ablation experiment. With increasing partitioning granularity, model performance continuously improved. Compared to the single-partition configuration, the dual-partition configuration increased accuracy from 73.11% to 74.79% and macro-average F1 score from 72.92% to 74.55%, indicating that introducing coarse-grained body partitioning can improve the recognition performance of the adolescent depression severity stratification task. Furthermore, the proposed three-partition strategy achieved the best results across all metrics, with accuracy, macro-average precision, macro-average recall, and macro-average F1 score reaching 75.63%, 75.41%, 75.84%, and 75.52%, respectively. Compared to the dual-partition configuration, the three-partition scheme further improved performance, demonstrating that explicitly modeling the upper, middle, and lower body partitions is more effective than coarse-grained partitioning in capturing depression-related gait patterns.

[0056] Table 1

[0057] To evaluate the contributions of key components in the proposed network, this application conducted component ablation experiments using the same data preprocessing, training settings, and evaluation protocol as the main experiment. Therefore, component ablation primarily focuses on the impact of the upper, middle, and lower body partition perceptual modeling (UML) module, the partition location encoding (PPE) module, and the intra-partition self-attention (IPSA) module on model performance. This application designs five ablation configurations: (1) CTR-GCN: Only the CTR-GCN backbone network is used as the baseline model; (2) CTR-GCN + UML: Based on the CTR-GCN backbone network, an upper, middle and lower body partition perception modeling module (UML) is introduced to evaluate the role of explicit partition priors; (3) CTR-GCN + UML + PPE: Based on the upper, middle and lower body partition perception modeling module (UML), a partition location coding module (PPE) is further added to evaluate the impact of partition identity coding on feature discriminability; (4) CTR-GCN + UML + IPSA: Based on the upper, middle and lower body partition perception modeling module (UML), a partition self-attention module (IPSA) is further added to evaluate the effect of partition adaptive feature refinement; (5) UMLPA-GCN: A complete model that includes upper, middle and lower body partition perception modeling module (UML), partition location encoding module (PPE), partition self-attention module (IPSA) and CTR-GCN backbone network.

[0058] Table 2 reports the results of the component ablation experiments. The ablation results show that each key component has a positive effect on the final performance improvement. Compared with the CTR-GCN baseline, the model accuracy improved from 72.27% to 73.11% after introducing the UML modeling module, indicating that this module can improve the modeling effect of adolescent depression gait by explicitly introducing upper, middle, and lower body partitions as priors. After further adding PPE on the basis of the UML modeling module, the accuracy improved to 73.95%, indicating that partition identity encoding can enhance the partition-level discrimination ability in the feature space. In contrast, after adding IPSA on the basis of UML partitioning, the model accuracy further improved to 74.79%, indicating that adaptive feature refinement within each partition is more conducive to capturing adolescent depression-related gait patterns. When all components work together, the complete model UMLPA-GCN achieves the best results across all metrics, with accuracy, macro-average precision, macro-average recall, and macro-average F1 score reaching 75.63%, 75.41%, 75.84%, and 75.52%, respectively. These results demonstrate that the UML modeling module, PPE, and IPSA have complementary effects, collectively contributing to the overall performance improvement of the model.

[0059] Table 2

[0060] To comprehensively evaluate the effectiveness of the proposed method, this application selected representative baseline methods for comparison, including ST-GCN, 2s-AGCN, MS-G3D, HD-GCN, InfoGCN, and CTR-GCN. All comparison methods were trained and tested under the same subject-independent five-fold cross-validation protocol to ensure fairness. Evaluation metrics included accuracy, macro-precision, macro-recall, and macro-F1 score, with the average of the five-fold results used as the final performance.

[0061] Table 3 reports the comparison results between the baseline methods and the proposed UMLPA-GCN. The results show that the proposed UMLPA-GCN achieves the best performance across all evaluation metrics. Specifically, compared to the strongest baseline CTR-GCN, UMLPA-GCN improves accuracy from 72.27% to 75.63% and macro-average F1 score from 72.25% to 75.52%, representing improvements of 3.36 and 3.27 percentage points, respectively. Furthermore, the proposed method also achieves the highest results in macro-average precision and macro-average recall, indicating more balanced predictive performance across categories. These results demonstrate that by explicitly introducing upper, middle, and lower body partitioning priors and combining them with a partitioning feature refinement mechanism, heterogeneous motion features between different body parts can be captured more effectively, thereby improving the recognition performance of the adolescent depression severity stratification task.

[0062] Table 3

[0063] This application obtains the gait 3D skeleton sequence of the subject to be evaluated, and performs forward vision segment preservation, tilt correction, root node translation normalization, smoothing filtering, and fixed-length resampling or padding on the gait 3D skeleton sequence to obtain a standardized skeleton sequence. The standardized skeleton sequence is then mapped to three body regions: upper, middle, and lower, to obtain upper region features, middle region features, and lower region features. Each region feature is input into a region feature refinement module, and refined region fusion features are obtained through region position encoding, intra-regional self-attention, and cross-regional fusion. These region fusion features are then input into a graph convolutional backbone network to extract global spatiotemporal features, and the results are output after pooling and classification layers to assess the severity of adolescent depression. This application, by explicitly introducing upper, middle, and lower body region priors and combining them with a region feature refinement mechanism, can more effectively capture the heterogeneous distribution of depression-related gait abnormalities in different body parts, improving the objectivity, accuracy, and interpretability of the automatic assessment of adolescent depression severity.

[0064] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A depression severity assessment system based on a body partitioning-aware convolutional network, characterized by: include: The data acquisition and preprocessing module is configured to: acquire gait depth video of the object to be evaluated, extract the corresponding original three-dimensional skeleton sequence, preprocess the original three-dimensional skeleton sequence to obtain a standardized skeleton sequence; The partitioning mapping module is configured to: perform partitioning mapping on the standardized skeleton sequence according to three human body partitions: upper, middle, and lower, to obtain upper partition features, middle partition features, and lower partition features; The depression severity assessment module is configured to: input the features of each partition into a trained depression severity assessment network to obtain the depression severity assessment result of the subject to be assessed; wherein, the trained depression severity assessment network performs partition position encoding, intra-partition self-attention, and cross-partition fusion on the features of each partition to obtain refined partition fusion features; input the partition fusion features into a graph convolutional backbone network to extract global spatiotemporal features; perform pooling and classification on the global spatiotemporal features to output the depression severity assessment result.

2. The depression severity assessment system based on a body partitioning-aware graph convolutional network as described in claim 1, characterized in that, The partition location encoding is used to generate partition-level temporal location descriptions and inject them back into the joint features of each partition. The intra-partition self-attention is used to adaptively reweight the joint features within each human body partition. The cross-partition fusion is used to fuse the refined features of multiple partitions in the channel dimension and map them back to a unified feature space.

3. The depression severity assessment system based on a body partitioning-aware graph convolutional network as described in claim 1, characterized in that, The preprocessing of the original 3D skeleton sequence to obtain a standardized skeleton sequence includes: preserving the forward-looking segments, correcting the camera tilt angle, normalizing the root node translation, smoothing filtering, and performing fixed-length resampling or padding on the gait 3D skeleton sequence to obtain a standardized skeleton sequence, specifically including: Original skeleton sequence of each sample Represented as: ;in, , Indicates the number of channels in the three-dimensional coordinate system. Indicates the number of frames in the original sequence. Indicate the number of joints; let the number of joints be... The Z and Y coordinates of the SpineBase joint at the base of the spine in the frame are respectively and The angle between the camera's optical axis and the ground is denoted as . Then, the least squares fitting method is used to estimate the tilt angle. : in, This indicates the number of frames in the sequence. Then, the three-dimensional coordinates of each joint in each frame are... Around Perform a rotational transformation on the axis to make the corrected Align the axis with the actual walking direction: After angle correction, the global translation effect is removed by using the SpineBase joint as the root node. In each frame, the SpineBase coordinates of the corresponding frame are subtracted from all joint coordinates to make the skeleton sequence translation invariant in the camera coordinate system. A sliding window of length 5 is used, and the translation of each joint is adjusted accordingly. , and The trajectory sequences in the three coordinate directions are filtered independently; each skeleton sequence is resampled to a fixed length of 50 frames to obtain a standardized skeleton sequence. : ;in, Indicates batch size. This indicates the number of coordinate channels, preferably the number of three-dimensional coordinate channels; Indicates the length of the uniform sequence after resampling or padding. Indicates the number of joints.

4. The depression severity assessment system based on a body partitioning-aware graph convolutional network as described in claim 1, characterized in that, The standardized skeleton sequence is partitioned and mapped according to three human body regions: upper, middle, and lower, to obtain upper region features, middle region features, and lower region features. Specifically, this includes: First, perform partition mapping on the human skeleton, assuming the complete set of joints is... It is divided into three non-overlapping partitions: upper partition, middle partition, and lower partition. in, , and These correspond to three human body sections: upper, middle, and lower. The upper section corresponds to the joints related to the head and neck, the middle section corresponds to the joints related to the trunk and upper limbs, and the lower section corresponds to the joints related to the pelvis and lower limbs. For each partition Define the corresponding binary selection mask. Then the partition features Represented as: in, Indicates in Element-wise multiplication after broadcasting in the dimension; the mask satisfies .

5. The depression severity assessment system based on a body partitioning-aware graph convolutional network as described in claim 1, characterized in that, The process of performing partition location encoding, intra-partition self-attention, and cross-partition fusion on the features of each partition yields refined partition fusion features. Specifically, partition location encoding includes: For any partition Features are converged along the joint dimensions within the partition to obtain a partition-level frame description. And further generate partition location codes. : ； ； in, Represents the first normalized backbone sequence. Frame, First The features of each joint, where the symbol ":" represents all elements of that dimension. , Indicates the first The set of joint indexes contained in each partition. Indicates the first The number of joints contained in each partition. This represents a feedforward neural network; Then, the encoding Broadcast along the joint dimension and back-inject residuals into all joint features of the current partition: ； in, Indicates the first In the partition, the th Frame, First The enhanced joint features are obtained by adding the original input features to each joint after partitioning and position encoding, where the symbol ":" indicates that all feature channels are selected.

6. The depression severity assessment system based on a body partitioning-aware graph convolutional network as described in claim 1, characterized in that, The process of performing partition location encoding, intra-partition self-attention, and cross-partition fusion on the features of each partition yields refined partition fusion features. Specifically, intra-partition self-attention includes: Model the joint correlations within each partition, for the first... Frames will partition The internal joints are treated as tokens, and... Transpose ; Intra-partition self-attention is calculated as follows: in, , and For learnable projection matrices, To scale the dimensions, This represents the token matrix after the joints in the frame input, representing the body partition; This represents the partition token obtained after attention reweighting; For each frame, According to the zones The corresponding set of original joint indices The feature vectors are then backfilled into the corresponding positions in the complete skeleton layout to obtain the partitioned feature vectors. Specifically, in the first Frame, set of joint indices Features at corresponding positions for: Among them, Transpose( ) indicates that The joint token dimension and feature channel dimension are transposed.

7. The depression severity assessment system based on a body partitioning-aware graph convolutional network as described in claim 1, characterized in that, The process of performing partition location encoding, intra-partition self-attention, and cross-partition fusion on the features of each partition yields refined partition fusion features. Specifically, cross-partition fusion includes: The three partition features are concatenated along the channel dimension and then mapped back to the original channel dimension to obtain the partition fusion feature. : in, and These represent the trainable weights and trainable biases of a linear projection, respectively; the symbol " "Indicates a linear mapping along the channel dimension; These represent the feature tensors obtained after the upper, middle, and lower partitions are refined and backfilled by self-attention within the partitions, respectively. Presentation layer normalization processing.

8. The depression severity assessment system based on a body partitioning-aware graph convolutional network as described in claim 1, characterized in that, The partition fusion features are input into the graph convolutional backbone network to extract global spatiotemporal features, specifically including: processing the obtained partition fusion features. A graph convolutional backbone network is used for global spatiotemporal modeling; the graph convolutional backbone network is a CTR-GCN backbone network.

9. The depression severity assessment system based on a body partitioning-aware graph convolutional network as described in claim 1, characterized in that, The process of pooling and classifying the global spatiotemporal features to output a depression severity assessment result specifically includes: high-level features extracted by the backbone network are used to perform three-class classification prediction through a global average pooling layer, a fully connected layer, and a Softmax classification layer, outputting three categories of depression severity assessment results: healthy, mild depression, and moderate to severe depression.

10. The depression severity assessment system based on a body partitioning-aware graph convolutional network as described in claim 1, characterized in that, Gait depth videos of the subjects to be evaluated were acquired and the corresponding original 3D skeleton sequences were extracted. This was done by simultaneously acquiring data using three Azure Kinect DK depth cameras. The depth video resolution was set to 320×288 pixels and the sampling frame rate was set to 30 Hz. The 3D skeleton sequences were extracted using the Azure Kinect SDK. During the acquisition process, the subjects completed a back-and-forth walking task along a preset walking route to obtain gait skeleton sequences containing multiple complete gait cycles.

Images