Methods, systems, storage media, and program products for assessing depression levels using electroencephalogram (EEG) signals.
By combining hybrid multi-scale convolution and Transformer encoders, the problems of lightweight and high accuracy in EEG signal assessment in existing technologies are solved, realizing a lightweight method for assessing depression levels that is suitable for clinical and wearable devices, and improving the accuracy and interpretability of depression assessment.
Patent Information
- Application Number
- CN202610113071.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-27
- Publication Date
- 2026-05-26
AI Technical Summary
Existing methods for assessing depression levels using EEG signals are highly influenced by subjective factors. Traditional models have large parameter scales and are difficult to deploy in resource-constrained clinical environments, making it difficult to achieve lightweight and highly accurate analysis.
A lightweight network model employing a hybrid multi-scale convolution and self-attention mechanism is used. This model combines a multi-scale convolution feature enhancement module and a Transformer encoder. The hybrid multi-scale convolution feature enhancement module extracts multi-level local features, and a residual fusion layer is constructed at the back end of the Transformer encoder to achieve lightweight and highly accurate depression assessment.
It improves the accuracy and interpretability of depression assessment, has a high degree of lightweight design, and is suitable for clinical EEG monitoring devices, wearable terminals, and edge computing nodes, enabling rapid and accurate assessment of depression levels.
Smart Images
Figure CN122075012A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of depression severity assessment technology, and more specifically, to a method, system, storage medium, and program product for assessing depression severity using electroencephalogram (EEG) signals. Background Technology
[0002] Depression profoundly impacts an individual's quality of life and casts a heavy shadow on society. To better understand and address this prevalent mental illness, researchers are constantly exploring more objective and scientific methods of assessment and diagnosis. Against this backdrop, electroencephalography (EEG) has become an important auxiliary tool.
[0003] Electroencephalography (EEG) is a non-invasive technique that records the brain's electrical activity using electrodes placed on the scalp. Its advantages, including millisecond-level temporal resolution, cost-effectiveness, and ease of operation, make it a highly valuable tool in neuroscience and mental health research. EEG can sensitively capture dynamic changes in brain neurophysiological activity, providing important evidence for revealing the abnormal characteristics of depression at the brain functional level.
[0004] In clinical practice, the assessment of depression levels typically relies on interviews with psychiatrists and patient self-rating scales. However, these methods are highly subjective and may lead to biased diagnostic results. EEG, as an objective neurophysiological measurement tool, can overcome the subjective limitations of traditional scale methods, providing multidimensional brain function information and helping physicians more accurately assess the degree of depression.
[0005] In recent years, electroencephalograms (EEGs) have been considered a promising auxiliary diagnostic tool for depression due to their advantages such as high temporal resolution, non-invasiveness, and low cost. However, depression EEG signals are characterized by low signal-to-noise ratios, subtle feature variations, and significant inter-individual variability, making it difficult for traditional recognition methods based on handcrafted features or single-scale convolutional neural networks to fully extract key patterns. Furthermore, existing deep learning models typically have large parameter scales, making them difficult to deploy in resource-constrained clinical environments and hindering interpretability and real-time analysis.
[0006] On the other hand, multi-scale convolution can extract multi-level local features from different receptive fields, while self-attention mechanisms have the ability to model long-range dependencies and overall semantic information. However, research on effectively integrating the two while maintaining both lightweight design and high accuracy is still relatively limited. Therefore, there is an urgent need for a lightweight network model that can combine multi-scale local feature extraction and global information modeling capabilities, and is suitable for automatic EEG analysis of depression.
[0007] Based on this, a method for identifying the severity of depression that integrates multi-scale convolution, attention mechanisms, and end-to-end classification structures is of great significance. It can effectively improve EEG pattern recognition capabilities and enhance the reliability and objectivity of depression assessment.
[0008] The preceding description is intended to provide general background information and does not necessarily constitute prior art. Summary of the Invention
[0009] The purpose of this invention is to provide a method, system, storage medium, and program product for assessing the degree of depression using electroencephalogram (EEG) signals. This method provides an objective and interpretable neural network decision-making basis for the quantitative auxiliary diagnosis and individualized treatment of depression.
[0010] This invention provides a method for assessing the degree of depression using electroencephalogram (EEG) signals, comprising the following steps:
[0011] S1: For the original multi-channel EEG signal Perform zero-mean and bandpass filtering preprocessing;
[0012] S2: Input the preprocessed signal into the hybrid multi-scale convolutional feature enhancement module to obtain multi-scale local feature maps:
[0013] ;
[0014] S3: By linear projection Flatten into a patch sequence ;
[0015] S4: Will After being added to the learnable position code, it is input into the L-layer Transformer encoder, and the output of the l-th layer is Z1 = LN(MHSA(LN(Z)). 1-1 ) ) +Z 1-1 ) l=1, ...,L;
[0016] S5: Build a residual fusion layer at the back end of the Transformer encoder to obtain the final features:
[0017] F1 = Z1 + UpSample( ) ;
[0018] S6: Map F1 through global average pooling and a fully connected layer to a probability vector of depression severity:
[0019] y = Softmax( W·GAP(F1) + b ) ;
[0020] S7: Output depression classification results based on the principle of maximum probability. .
[0021] Further, in step S1, C is the number of channels and T is the number of sampling points; in step S2, k is the kernel size, d is the dilatation rate, and M is the number of single-branch output channels. s is the step size; in step S3 Where P is the number of patches, P is the patch length, and D is the embedding dimension; in step S4, MHSA(·) is multi-head self-attention, and LN(·) is layer normalization; in step S5, UpSample(·) is linear interpolation and channel alignment operation; in step S6, K is the depression level. , These are learnable parameters.
[0022] Furthermore, the hybrid multi-scale convolutional feature enhancement module in step S2 includes batch normalization (BN) and activation function ReLU (·) after each branch, and the three branch outputs are concatenated in the channel dimension and then fused across channels through 1×1 convolution.
[0023] Further, the calculation formula for MHSA(·) in step S4 is: Attention(Q, K, V) = Softmax( V, where Q, K, These are the query, key, and value matrix, respectively, D. h = h represents the number of attention heads.
[0024] Furthermore, in step S5, UpSample(·) uses one-dimensional linear interpolation to restore the time dimension of F0 to be consistent with Z1, and adjusts the number of channels to D through 1×1 convolution.
[0025] Furthermore, step S6 also includes introducing Dropout (p=0.5) and L2 weight decay λ=1×10 -4 .
[0026] The present invention also provides an electroencephalogram (EEG) signal depression assessment system, which is applied to the above-mentioned EEG signal depression assessment method.
[0027] Furthermore, the EEG signal depression severity assessment system includes a signal acquisition unit, a preprocessing unit, a HyScale-CRF network inference unit, a result output unit, and a model update unit. The signal acquisition unit is used to acquire resting-state and task-state EEG signals according to the international 10-20 system. The preprocessing unit is used to perform bandpass filtering (0.5-45Hz) and independent component removal of electrooculography (EOG). The network structure of the HyScale-CRF network inference unit consists of a hybrid multi-scale convolutional module, a Transformer encoder, a residual fusion layer, and a classification head connected sequentially. The result output unit is used to display or store the depression severity grading results. The model update unit is used to incrementally update the HyScale-CRF parameters through a federated learning mechanism without uploading local data.
[0028] Furthermore, the HyScale-CRF network inference unit is deployed on an edge computing chip, with a model parameter size of <500kB, a single-channel 30s EEG fragment inference time of <50ms, and a power consumption of <100mW.
[0029] The present invention also provides a storage medium storing a computer program / instructions thereon, characterized in that the computer program / instructions, when executed by a processor, implement the steps of the electroencephalogram (EEG) signal depression assessment method of claim 1.
[0030] The present invention also provides a program product, including a computer program / instructions, characterized in that, when the computer program / instructions are executed by a processor, they implement the steps of the electroencephalogram (EEG) signal depression assessment method of claim 1.
[0031] The present invention provides a method for assessing the degree of depression using EEG signals. HyScale-CRF constructs parallel hybrid-scale dilated convolutional branches at the input end to synchronously enhance multi-granularity spatiotemporal receptive field features. A residual cross-layer fusion layer is introduced at the output end of the Transformer encoder to identically superimpose deep global self-attention representations with intermediate local convolutional features, achieving dual backpropagation of gradients and semantics. The overall design is modular and lightweight, significantly reducing the number of parameters compared to traditional baseline models, enabling end-to-end training. Experiments based on resting-state and task-state EEG data demonstrate that HyScale-CRF outperforms existing methods in classification accuracy, recall, and overall performance metrics for mild, moderate, and severe depression, while significantly improving inference speed. It can be directly deployed in clinical EEG monitoring devices, wearable terminals, or edge computing nodes, providing objective and interpretable neural network decision-making support for the quantitative auxiliary diagnosis and individualized treatment of depression. Attached Figure Description
[0032] Figure 1This is a flowchart illustrating the brainwave signal depression assessment method provided in Embodiment 1 of the present invention.
[0033] Figure 2 This is a schematic diagram of the modules of the electroencephalogram (EEG) depression assessment system provided in Embodiment 2 of the present invention.
[0034] The reference numerals and components involved in the accompanying drawings are shown below:
[0035] 1. Signal Acquisition Unit
[0036] 2. Preprocessing unit
[0037] 3. HyScale-CRF network inference unit
[0038] 31. Hybrid Multi-Scale Convolution Module
[0039] 32. Transformer encoder
[0040] 33. Residual Fusion Layer
[0041] 34. Classification Head
[0042] 4. Result Output Unit
[0043] 5. Model Update Unit Detailed Implementation
[0044] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.
[0045] The terms "first," "second," "third," "fourth," etc., used in the specification and claims of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.
[0046] Example 1
[0047] Figure 1 This is a flowchart illustrating the brainwave signal-based depression assessment method provided in Embodiment 1 of the present invention. Please refer to... Figure 1 The method for assessing the degree of depression using electroencephalogram (EEG) signals provided in this embodiment of the invention includes the following steps:
[0048] S1: For the original multi-channel EEG signal Zero-mean and bandpass filtering preprocessing is performed, where C is the number of channels and T is the number of sampling points;
[0049] S2: Input the preprocessed signal into the hybrid multi-scale convolutional feature enhancement module to obtain multi-scale local feature maps:
[0050] ;
[0051] Where k is the kernel size, d is the dilatancy ratio, and M is the number of output channels per branch. s is the step size;
[0052] Specifically, the hybrid multi-scale convolutional feature enhancement module includes batch normalization (BN) and activation function ReLU (·) after each branch, and the outputs of the three branches are concatenated in the channel dimension and then fused across channels through 1×1 convolution;
[0053] S3: By linear projection Flatten into a patch sequence ;in, P is the number of patches, P is the patch length, and D is the embedding dimension;
[0054] S4: Will After being added to the learnable position code, it is input into the L-layer Transformer encoder, and the output of the l-th layer is Z1 = LN(MHSA(LN(Z)). 1-1 ) ) + Z 1-1 ) , l=1,…,L, where MHSA(·) is multi-head self-attention and LN(·) is layer normalization;
[0055] Specifically, the formula for calculating MHSA(·) is: Attention(Q, K, V) = Softmax( V, where Q, K, These are the query, key, and value matrix, respectively, D. h = h represents the number of attention heads;
[0056] S5: Construct a residual fusion layer at the back end of the Transformer encoder to obtain the final feature: F1 = Z1 + UpSample( ) UpSample(·) is a linear interpolation and channel alignment operation.
[0057] Specifically, UpSample(·) uses one-dimensional linear interpolation to restore the time dimension of F0 to be consistent with Z1, and adjusts the number of channels to D through 1×1 convolution;
[0058] S6: Map F1 to a probability vector of depression severity through global average pooling and a fully connected layer: y = Softmax( W·GAP(F1) + b ) Where K is the depression level, , These are learnable parameters;
[0059] Furthermore, step S6 also includes introducing Dropout (p=0.5) and L2 weight decay λ=1×10 -4 ;
[0060] S7: Output depression classification results based on the principle of maximum probability. .
[0061] The present invention provides a method for assessing the degree of depression using EEG signals, employing a hybrid-scale convolutional enhancement mechanism: a hybrid-scale convolutional module is introduced at the input of the Transformer, achieving multi-scale feature extraction through parallel structures of different convolutional kernel sizes (e.g., 1×3, 1×5, 1×7) and dilated convolutions; this module can simultaneously capture local fluctuations and cross-channel dynamic patterns of EEG signals in both the temporal and spatial domains, achieving multi-granularity feature fusion; compared to traditional single-scale convolutional embedding methods, the hybrid-scale convolutional module of the present invention enhances the expressive power of local structures of EEG signals while ensuring model lightweightness, and is more sensitive to the identification of subtle changes in the degree of depression;
[0062] The Transformer-Residual Fusion structure (Residual Fusion after TransformerEncoding) sets a residual fusion layer after the output of the Transformer encoder, which performs identity mapping and superposition on intermediate layer features and deep semantic information to achieve cross-layer feature interaction and semantic alignment. This residual fusion mechanism can improve the gradient fluidity of deep information, strengthen the synergistic effect of global dependencies and local features, and significantly improve the robustness and generalization performance of the model under complex EEG signals.
[0063] Unified Multi-Scale Hierarchical Interaction: This framework integrates the local modeling capabilities of convolutional layers with the global dependency learning mechanism of Transformer. It captures multi-scale features at the front end and performs hierarchical residual interaction at the back end, achieving overall consistent modeling of spatiotemporal features. This design overcomes the problems of insufficient high-frequency EEG feature capture and fragmented feature representation in existing Transformers, enabling the model to more accurately reflect the neural dynamic changes in different brain regions and time periods.
[0064] Lightweight and End-to-End Design: The model adopts a modular design and can be trained end-to-end. The number of parameters is reduced by about 30%-40% compared to the traditional EEG-Conformer, while the inference speed is significantly improved. This lightweight structure facilitates efficient deployment in clinical EEG monitoring equipment, wearable EEG acquisition terminals, and edge computing environments.
[0065] Domain-Specific Optimization (EEG): Targeting the multi-channel, low signal-to-noise ratio, and time-varying characteristics of EEG signals, this invention has made specific optimizations in convolution kernel configuration, normalization layers, and time-series embedding strategies, significantly improving the model's discriminative ability and robustness in the depression severity grading task.
[0066] Example 2
[0067] This embodiment provides a brainwave signal depression assessment system, which is applied to the brainwave signal depression assessment method in Embodiment 1.
[0068] Figure 2 This is a schematic diagram of the modules of the EEG signal depression assessment system provided in Embodiment 2 of the present invention. Please refer to... Figure 2 The EEG signal depression severity assessment system of this embodiment includes a signal acquisition unit 1, a preprocessing unit 2, a HyScale-CRF network inference unit 3, a result output unit 4, and a model update unit 5. The signal acquisition unit 1 is used to acquire resting-state and task-state EEG signals according to the international 10-20 system. The preprocessing unit 2 is used to perform bandpass filtering of 0.5-45Hz and independent component removal of electrooculography. The network structure of the HyScale-CRF network inference unit 3 is composed of a hybrid multi-scale convolutional module 31, a Transformer encoder 32, a residual fusion layer 33, and a classification head 34 connected in sequence. The result output unit 4 is used to display or store the depression severity grading results. The model update unit 5 is used to incrementally update the HyScale-CRN parameters through a federated learning mechanism without uploading local data.
[0069] Furthermore, the HyScale-CRF network inference unit 3 is deployed on an edge computing chip, with a model parameter size of <500kB, a single-channel 30s EEG fragment inference time of <50ms, and a power consumption of <100mW.
[0070] Example 3
[0071] This embodiment provides a storage medium storing a computer program / instruction, which, when executed by a processor, implements the steps of the electroencephalogram (EEG) signal depression assessment method of Embodiment 1.
[0072] Example 4
[0073] This embodiment provides a program product, including a computer program / instructions, which, when executed by a processor, implements the steps of the electroencephalogram (EEG) signal depression assessment method of Embodiment 1.
[0074] As can be seen from the above description, the advantages of this invention are:
[0075] The present invention relates to a method for assessing the degree of depression using EEG signals. HyScale-CRF constructs parallel hybrid-scale dilated convolutional branches at the input end to simultaneously enhance multi-granularity spatiotemporal receptive field features. A residual cross-layer fusion layer is introduced at the output end of the Transformer encoder to identically superimpose deep global self-attention representations with intermediate local convolutional features, achieving dual backpropagation of gradients and semantics. The overall design employs a modular and lightweight approach, significantly reducing the number of parameters compared to traditional baseline models, enabling end-to-end training. Experiments based on resting-state and task-state EEG data demonstrate that HyScale-CRF outperforms existing methods in classification accuracy, recall, and overall performance metrics for mild, moderate, and severe depression, while significantly improving inference speed. It can be directly deployed in clinical EEG monitoring devices, wearable terminals, or edge computing nodes, providing objective and interpretable neural network decision-making support for the quantitative auxiliary diagnosis and individualized treatment of depression.
[0076] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A method for assessing the degree of depression using electroencephalogram (EEG) signals, characterized in that, Includes the following steps: S1: For the original multi-channel EEG signal Perform zero-mean and bandpass filtering preprocessing; S2: Input the preprocessed signal into the hybrid multi-scale convolutional feature enhancement module to obtain multi-scale local feature maps: ; S3: By linear projection Flatten into a patch sequence ; S4: Will After being added to the learnable positional code, it is input into the L-layer Transformer encoder, and the output of the l-th layer is Z1 = LN(MHSA(LN(Z)). 1-1 ) ) + Z 1-1 ) l=1, ...,L; S5: Build a residual fusion layer at the back end of the Transformer encoder to obtain the final features: F1=Z1+UpSample( ) ; S6: Map F1 through global average pooling and a fully connected layer to a probability vector of depression severity: y = Softmax(W·GAP(F1) + b) ; S7: Output depression classification results based on the principle of maximum probability. .
2. The method for assessing the degree of depression using electroencephalogram (EEG) signals according to claim 1, characterized in that, In step S1, C represents the number of channels and T represents the number of sampling points; In step S2, k is the convolution kernel size, d is the dilatation rate, and M is the number of single-branch output channels. s is the step size; In step S3 P is the number of patches, P is the patch length, and D is the embedding dimension; In step S4, MHSA(·) represents multi-head self-attention, and LN(·) represents layer normalization. In step S5, UpSample(·) is a linear interpolation and channel alignment operation; In step S6, K represents the depression level. , These are learnable parameters.
3. The method for assessing the degree of depression using electroencephalogram (EEG) signals according to claim 1, characterized in that, The hybrid multi-scale convolutional feature enhancement module in step S2 includes batch normalization (BN) and activation function ReLU (·) after each branch, and the outputs of the three branches are concatenated in the channel dimension and then fused across channels through 1×1 convolution.
4. The method for assessing the degree of depression using electroencephalogram (EEG) signals according to claim 2, characterized in that, The formula for calculating MHSA(·) in step S4 is: Attention(Q, K, V) = Softmax( V, where Q, K, These are the query, key, and value matrix, respectively, D. h = h represents the number of attention heads.
5. The method for assessing the degree of depression using electroencephalogram (EEG) signals according to claim 2, characterized in that, In step S5, UpSample(·) is interpolated using one-dimensional linear interpolation. The time dimension is restored to be consistent with Z1, and the number of channels is adjusted to D through 1×1 convolution.
6. The method for assessing the degree of depression using electroencephalogram (EEG) signals according to claim 1, characterized in that, Step S6 further includes introducing Dropout (p=0.5) and L2 weight decay λ=1×10 -4 .
7. A brainwave signal-based depression assessment system, characterized in that, The electroencephalogram (EEG) signal depression assessment system is applied to the EEG signal depression assessment method according to any one of claims 1-6.
8. The EEG signal depression severity assessment system according to claim 7, characterized in that, The EEG signal depression assessment system includes a signal acquisition unit (1), a preprocessing unit (2), a HyScale-CRF network inference unit (3), a result output unit (4), and a model update unit (5). The signal acquisition unit (1) is used to acquire resting and task-state EEG signals according to the international 10-20 system; The preprocessing unit (2) is used to perform bandpass filtering of 0.5-45Hz and independent component removal of ocular electrical activity; The network structure of the HyScale-CRF network inference unit (3) consists of a hybrid multi-scale convolutional module (31), a Transformer encoder (32), a residual fusion layer (33), and a classification head (34) connected in sequence. The result output unit (4) is used to display or store the depression severity grading results; The model update unit (5) is used to incrementally update the HyScale-CRN parameters without uploading local data through the federated learning mechanism.
9. The EEG signal depression assessment system according to claim 8, characterized in that, The HyScale-CRF network inference unit (3) is deployed on an edge computing chip, with a model parameter size of <500 kB, a single-channel 30 s EEG fragment inference time of <50 ms, and a power consumption of <100 mW.
10. A storage medium having a computer program / instruction stored thereon, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the electroencephalogram (EEG) depression assessment method of claim 1.
11. A program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the electroencephalogram (EEG) depression assessment method of claim 1.