A depression assessment system based on brain-computer interface technology

Abstract

The present invention relates to the technical field of deep learning, and provides a depression assessment system based on brain-computer interface technology. The system includes a data acquisition module, a signal preprocessing module, a feature extraction module, a depression assessment module, and an optimization assessment module. Among them, the feature extraction module extracts the key features of EEG signals and biological signals. The assessment module combines multi-scale convolution and non-linear transformation technologies to construct a multi-scale assessment network, effectively processes these features and generates assessment results. The optimization assessment module uses a double-inertia approximate parameter update method to optimize the trainable parameters, improving the adaptability and stability of the system. Combining the above technologies, the present invention provides a more accurate, reliable, and real-time depression assessment system, which can effectively solve the deficiencies of traditional assessment systems and significantly improve the objectivity and accuracy of assessment results.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning, and particularly to a depression assessment system based on brain-computer interface technology. Background Art

[0002] With the rapid development of technology, brain-computer interface technology (BCI) has become an important research direction in neuroscience, psychology, and the medical field. Especially in the assessment of neuropsychiatric diseases, electroencephalogram (EEG), as the main bioelectrical signal reflecting brain activity, is widely used in the research of psychological disorders such as depression and anxiety. However, there are still many deficiencies in existing depression assessment systems. First of all, most traditional depression assessment methods rely on means such as questionnaires, clinical interviews, and physiological signal analysis. Although these methods are widely used in clinics, due to their strong subjectivity, long assessment cycles, and lack of real-time nature, they cannot sensitively capture the fluctuations of patients' emotions, resulting in poor accuracy and timeliness of assessment results. Secondly, many existing systems still focus on the analysis of a single signal source, such as only using bio-signals such as EEG signals or heart rate variability (HRV), while ignoring the comprehensive analysis of multi-modal signals. Methods based on a single signal source often fail to comprehensively reflect the complex physiological and psychological states of depression patients, thus affecting the accuracy and reliability of assessment results.

[0003] In addition, when processing these bio-signals, traditional systems often fail to fully consider the spatio-temporal characteristics of the signals, resulting in insufficient ability to capture complex physiological dynamic changes. Existing technologies lack sufficient precision and adaptability in the signal processing process and cannot efficiently improve system performance. Especially when dealing with individual differences of different patients, it is often difficult to achieve accurate prediction and real-time assessment. Summary of the Invention

[0004] In view of the problems of strong subjectivity, long evaluation cycle, lack of real-time performance, etc. existing in traditional depression evaluation systems, the present invention provides a depression evaluation system based on brain-computer interface technology; this system collects multi-modal biological signals and comprehensively analyzes these signals by combining deep learning algorithms, so as to achieve accurate and real-time evaluation of depression; specifically, the present invention first extracts key features from EEG signals and biological signal data, providing rich physiological and psychological state information for subsequent analysis; then, the system combines multi-scale convolution and non-linear transformation technologies to construct a multi-scale evaluation network, which can efficiently process EEG features and biological signal features and generate comprehensive depression evaluation results; finally, through the dual-inertia approximation parameter update method, the trainable parameters in the multi-scale evaluation network are optimized, so as to ensure that the system maintains good adaptability and stability among different environments and individuals; integrating the above technologies, the present invention provides a more accurate and reliable depression evaluation system, effectively solving the deficiencies in traditional evaluation systems and significantly improving the objectivity and real-time performance of evaluation.

[0005] The present invention provides a depression evaluation system based on brain-computer interface technology, and this system includes a data acquisition module, a signal preprocessing module, a feature extraction module, a depression evaluation module, and an optimization evaluation module;

[0006] The data acquisition module has electrodes covering the brain region and connected to the earlobes to reduce interference caused by the movement of the subject, and collects EEG signal data; a biological signal sensor is used to collect heart rate variability data and skin conductance data, and the EEG signal data, heart rate variability data, and skin conductance data are used as the original signal data;

[0007] The signal preprocessing module performs data cleaning and data standardization on the original signal data to generate cleaned signal data; data cleaning includes removing artifacts, band-pass filtering, and ICA independent component analysis; data standardization includes time series alignment and individual baseline adjustment;

[0008] The feature extraction module extracts EEG features and biological signal features of the cleaned signal data;

[0009] The depression evaluation module combines multi-scale convolution and non-linear transformation technologies to construct a multi-scale depression network, processes EEG features and biological signal features through the multi-scale depression network, and generates depression prediction results; risk assessment, personalized monitoring, mental health management, and emotional state trend analysis are carried out according to the depression prediction results;

[0010] The optimization evaluation module optimizes the trainable parameters of the multi-scale depression network through the dual-inertia approximation parameter update method; the trainable parameters include weights and biases.

[0011] Furthermore, the EEG features include power spectrum analysis features and brain region connectivity features; the biosignal features include heart rate variability features, skin conductance response features, and respiratory rate features.

[0012] Furthermore, the process of the depression assessment module generating the depression prediction result specifically includes the following steps:

[0013] Step S1: Min - Max normalize the EEG features and biosignal features to balance different signal scales, then perform temporal alignment and channel expansion to generate a normalized biosignal tensor;

[0014] Step S2: Extract local and global spatio - temporal features of the normalized biosignal tensor through multi - scale convolution to generate multi - scale spatio - temporal features; the multi - scale convolution is divided into dynamic stride convolution, intermediate layer convolution, and dilated convolution;

[0015] Step S3: Construct the Hyper - Sig - V3 activation function through the Bipolar Sigmoid transformation term and the composite non - linear transformation term, and process the multi - scale spatio - temporal features through the Hyper - Sig - V3 activation function to generate a non - linear feature tensor;

[0016] Step S4: Perform pooling processing and batch normalization processing on the non - linear feature tensor to generate a standardized feature tensor;

[0017] Step S5: Process the standardized feature tensor through a fully - connected layer to generate the depression prediction result.

[0018] Furthermore, Step S2 specifically includes the following steps:

[0019] Step S21: Use dynamic stride convolution to adaptively adjust the time window of the normalized biosignal tensor to generate dynamic stride features; and calculate the local change trend of the dynamic stride features in combination with the recursive residual learning strategy to generate short - time recursive time - dependent features;

[0020] Recursive residual learning can perform adaptive memory in the time dimension, enabling the multi - scale depression network to simultaneously consider the change relationship between the current time step and multiple past time steps. By calculating the residual change between adjacent time steps, the time features are recursively updated, enabling the multi - scale depression network to capture the signal fluctuation patterns within a short time, including the power changes in the α and β frequency bands of EEG, the short - time rhythm fluctuations of HRV, and the mutation trend of GSR under stress conditions;

[0021] Step S22: Capture more complex time patterns and local features in the short - time recursive time - dependent features through intermediate layer convolution to extract temporal pattern features;

[0022] Step S23: Based on the temporal pattern features, dilated convolution is used to extract global spatio-temporal features, and by combining short-term recurrent time-dependent features and temporal pattern features, multi-scale spatio-temporal features are generated; the model's perception ability of the long-term dependence pattern and global structural information of EEG signals and biological signals is enhanced.

[0023] Furthermore, the evaluation module is optimized, and the process of optimizing the trainable parameters of the multi-scale depression network specifically includes the following steps:

[0024] Step B1: Calculate the first inertial propulsion term and the second inertial propulsion term of the weights and biases; the first inertial propulsion term is used for parameter update, including the first inertial weight propulsion term and the first inertial bias propulsion term; the second inertial propulsion term is used for gradient estimation, including the second inertial weight propulsion term and the second inertial bias propulsion term;

[0025] Step B2: Construct a weight local quadratic approximation according to the first inertial weight propulsion term and the second inertial weight propulsion term to obtain an optimizable smooth approximation function;

[0026] Step B3: Construct a bias local quadratic approximation according to the first inertial bias propulsion term and the second inertial bias propulsion term to obtain an optimizable bias update function;

[0027] Step B4: Define a regularization term, update the weights according to the optimizable smooth approximation function, and update the biases according to the optimizable bias update function.

[0028] Adopting the above solution, the beneficial effects obtained by the present invention are as follows:

[0029] The present invention provides a depression evaluation system based on brain-computer interface technology. This system realizes the comprehensive acquisition and analysis of multi-modal biological signals. By combining EEG signals and biological signals, the accuracy and real-time performance of depression evaluation are effectively improved; through the feature extraction module analyzing the key features of various biological signals, the system can comprehensively capture the physiological changes and emotional fluctuations of depression patients, thereby providing more accurate and detailed physiological and psychological state information for evaluation; this method overcomes the limitations of traditional single-signal-source evaluation and enhances the comprehensiveness and integrity of the evaluation system;

[0030] In addition, the present invention constructs a multi-scale evaluation network by combining multi-scale convolution and nonlinear transformation technology, effectively improving the processing ability of EEG features and biological signal features; the multi-scale convolution network can extract local and global spatio-temporal features in the signal, so that the evaluation process is not limited to short-term changes but can also capture long-term dynamic changes; this innovation significantly improves the evaluation effect of depression, especially enhancing the adaptability of the system in different individuals and situations; the introduction of nonlinear transformation further enhances the system's ability to identify complex signal patterns, making the evaluation results more accurate;

[0031] Finally, the present invention optimizes the trainable parameters of the multi-scale evaluation network through a dual-inertia approximation parameter update method, thereby improving the stability and reliability of the system among different environments and individuals; this optimization method enables the system to adapt to individual differences, maintain high performance during long-term use, and avoid the biases caused by individual differences in traditional evaluation systems; integrating these technical advantages, the depression evaluation system of the present invention not only improves the objectivity and accuracy of the evaluation, but also enhances its real-time performance and applicability, providing strong support for the precise evaluation of depression. Brief Description of the Drawings

[0032] Figure 1 It is a schematic diagram of the modules of a depression evaluation system based on brain-computer interface technology provided by the present invention;

[0033] Figure 2 It is a comparison diagram of parameter updates provided in Embodiments VI and VII. Detailed Description of the Embodiments

[0034] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0035] Embodiment 1, according to Figure 1 , the present invention provides a depression evaluation system based on brain-computer interface technology, and the system includes a data acquisition module, a signal preprocessing module, a feature extraction module, a depression evaluation module, and an optimization evaluation module;

[0036] The data acquisition module has electrodes covering the brain area and connected to the earlobes to reduce interference caused by the movement of the subject and collect EEG signal data; a biosignal sensor is used to collect heart rate variability data and skin conductance data, and the EEG signal data, heart rate variability data, and skin conductance data are used as the original signal data;

[0037] The signal preprocessing module performs data cleaning and data standardization on the original signal data to generate cleaned signal data; data cleaning includes removing artifacts, band-pass filtering, and ICA independent component analysis; data standardization includes time series alignment and individual baseline adjustment;

[0038] The feature extraction module extracts EEG features and biosignal features of the cleaned signal data;

[0039] The depression assessment module combines multi-scale convolution and non-linear transformation technologies to construct a multi-scale depression network. The multi-scale depression network processes EEG features and biosignal features to generate depression prediction results. Risk assessment, personalized monitoring, mental health management, and emotional state trend analysis are performed based on the depression prediction results.

[0040] The optimization assessment module optimizes the trainable parameters of the multi-scale depression network through a dual-inertia approximation parameter update method. The trainable parameters include weights and biases.

[0041] Example 2: This example is based on Example 1. In this example, the EEG features include power spectrum analysis features and brain region connectivity features; the biosignal features include heart rate variability features, skin conductance response features, and respiratory rate features.

[0042] Example 3: This example is based on Example 2. In this example, the process of the depression assessment module generating depression prediction results specifically includes the following steps:

[0043] Step S1: Min-Max normalize the EEG features and biosignal features to balance different signal scales, then perform temporal alignment and channel expansion to generate a normalized biosignal tensor.

[0044] Step S2: Extract local and global spatio-temporal features of the normalized biosignal tensor through multi-scale convolution to generate multi-scale spatio-temporal features. The multi-scale convolution is divided into dynamic stride convolution, intermediate layer convolution, and dilated convolution.

[0045] Step S3: Construct a Hyper-Sig-V3 activation function through a Bipolar Sigmoid transformation term and a composite non-linear transformation term. Process the multi-scale spatio-temporal features through the Hyper-Sig-V3 activation function to generate a non-linear feature tensor.

[0046] ;

[0047] Among them, represents the multi-scale spatio-temporal features, represents the output value of the Hyper-Sig-V3 activation function, that is, the non-linear feature tensor; , , , and represent adjustable parameters, represents the linear term, represents the quadratic non-linear term, which controls the non-linearity of the Bipolar Sigmoid; represents the Bipolar Sigmoid transformation term, Represents the adjustable parameter for exponential transformation, Represents the exponential transformation term, Represents the exponential growth inhibition term, Represents the logarithmic correction term, Represents the composite non - linear transformation term, Represents the linear trend term;

[0048] Step S4: Perform pooling processing and batch normalization on the non - linear feature tensor to generate a standardized feature tensor;

[0049] Step S5: Process the standardized feature tensor through a fully - connected layer to generate the depression prediction result.

[0050] Example 4: This example is based on Example 2. In this example, the process of the depression assessment module generating the depression prediction result specifically includes the following steps:

[0051] Step E1: Perform Min - Max normalization on the EEG feature and the bio - signal feature to balance different signal scales, then perform time - series alignment and channel expansion to generate a normalized bio - signal tensor;

[0052] Step E2: Extract local and global spatio - temporal features of the normalized bio - signal tensor through multi - scale convolution to generate multi - scale spatio - temporal features;

[0053] Step E3: Process the multi - scale spatio - temporal features through the ReLU activation function to generate a non - linear feature tensor;

[0054] Step E4: Perform pooling processing and batch normalization on the non - linear feature tensor to generate a standardized feature tensor;

[0055] Step E5: Process the standardized feature tensor through a fully - connected layer to generate the depression prediction result.

[0056] Example 5: This example is based on Example 3. In this example, Step S2 specifically includes the following steps:

[0057] Step S21: Use dynamic - step convolution to adaptively adjust the time window of the normalized bio - signal tensor to generate dynamic - step features; and combine the recursive residual learning strategy to calculate the local change trend of the dynamic - step features to generate short - term recursive time - dependent features;

[0058] Dynamic - step calculation formula:

[0059] ;

[0060] Where, Represents the convolution kernel index, Represents the time - step index, Denotes at time step , the dynamic stride corresponding to the -th convolutional kernel; Denotes the base stride, Denotes the stride adjustment coefficient, Denotes the stride adjustment weight, Denotes the bias parameter, Denotes the Sigmoid function, Denotes the current time step of the input signal value, i.e., the normalized biosignal tensor;

[0061] Recursive residual learning can perform adaptive memory in the time dimension, enabling the multi-scale depression network to simultaneously consider the change relationship between the current time step and multiple past time steps. By calculating the residual change between adjacent time steps, the time features are recursively updated, enabling the multi-scale depression network to capture the signal fluctuation patterns within a short time, including the power changes in the α and β frequency bands of EEG, the short-time rhythm fluctuations of HRV, and the mutation trend of GSR under stress conditions;

[0062] Step S22: Capture more complex time patterns and local features in the short-time recursive time-dependent features through intermediate layer convolution to extract temporal pattern features;

[0063] Step S23: Based on the temporal pattern features, extract global spatio-temporal features through dilated convolution, combine the short-time recursive time-dependent features and the temporal pattern features to generate multi-scale spatio-temporal features; enhance the model's perception ability of the long-time dependence patterns and global structural information of EEG signals and biosignals.

[0064] Example Six, according to Figure 2 , this example is based on Example Five. In this example, the optimization evaluation module optimizes the process of the trainable parameters of the multi-scale depression network, specifically including the following steps:

[0065] Step B1: Calculate the first inertial propulsion term and the second inertial propulsion term of the weight and the bias; the first inertial propulsion term is used for parameter update, including the first inertial weight propulsion term and the first inertial bias propulsion term; the second inertial propulsion term is used for gradient estimation, including the second inertial weight propulsion term and the second inertial bias propulsion term;

[0066] Step B2: Construct a weight local quadratic approximation according to the first inertial weight propulsion term and the second inertial weight propulsion term to obtain an optimizable smooth approximation function;

[0067] ;

[0068] Among them, Denotes the number of iteration rounds, Denotes the multi-scale depression network layer index; represents the weight matrix of the layer, represents the regularization coefficient in the approximation function, represents the optimizable smooth approximation function; represents the penalty term function, represents the first inertia weight advancement term, represents the penalty term function at the function value of the first inertia weight advancement term; represents the second inertia weight advancement term, represents the linear approximation offset direction, represents the partial derivative symbol, represents the transpose symbol, represents the first-order Taylor expansion term; represents the quadratic regularization term;

[0069] Step B3: Construct a bias local quadratic approximation based on the first inertia bias advancement term and the second inertia bias advancement term to obtain an optimizable bias update function;

[0070] Step B4: Define the regularization term, update the weights according to the optimizable smooth approximation function, and update the bias according to the optimizable bias update function.

[0071] Example 7, according to Figure 2 , this example is based on Example 5. In this example, the process of the optimization evaluation module for optimizing the trainable parameters of the multi-scale depression network specifically includes: updating the weights and biases of the multi-scale depression network by the gradient descent method.

[0072] The above describes the present invention and its implementation manners. Such description is not restrictive. What is shown in the drawings is only one of the implementation manners of the present invention, and the actual structure is not limited thereto; generally speaking, if those of ordinary skill in the art are inspired by it and design similar structural manners and embodiments without creative efforts without departing from the purpose of the present invention's creation, they shall fall within the protection scope of the present invention.

Claims

1. A depression assessment system based on brain-computer interface technology, comprising a signal preprocessing module, wherein the signal preprocessing module generates clean signal data, characterized in that: The system also includes a feature extraction module, a depression assessment module and an optimization assessment module; Feature extraction module, extracting EEG features and biological signal features of clean signal data; Depression assessment module, which constructs a multi-scale depression network, processes EEG features and biological signal features through the multi-scale depression network, and generates depression prediction results; The optimization evaluation module optimizes the trainable parameters of the multi-scale depression network through the dual-inertia approximate parameter update method; The process of optimizing the evaluation module and optimizing the trainable parameters of the multi-scale depression network specifically includes the following steps: Step B1: Calculate the first inertia weight propulsion term, the first inertia bias propulsion term, the second inertia weight propulsion term and the second inertia bias propulsion term of the weights and biases; Step B2: construct a weighted local quadratic approximation according to the first inertia weight propulsion term and the second inertia weight propulsion term to obtain an optimizable smooth approximation function; ； in, represents the iteration round, Represents the multi-scale depression network layer index; Indicates The weight matrix of the layer, represents the regularization coefficient in the approximate function, It indicates that the smooth approximate function can be optimized; represents the penalty function, represents the first inertia weight propulsion term, Represents the penalty function Function value of the first inertia weight propulsion term; represents the second inertia weight propulsion term, represents the linear approximate offset direction, represents the symbol of partial derivative, represents the transpose symbol, represents the first-order Taylor expansion term; represents the quadratic regularization term; Step B3: constructing a bias local quadratic approximation according to the first inertia bias propulsion term and the second inertia bias propulsion term to obtain an optimizable bias update function; Step B4: Define a regularization term, update the weights according to an optimizable smooth approximation function, and update the bias according to an optimizable bias update function.

2. A depression assessment system based on brain-computer interface technology according to claim 1, characterized in that: EEG features include power spectrum analysis features and brain area connectivity features; biological signal features include heart rate variability features, skin conductance response features and respiratory rate features.

3. The depression assessment system based on brain-computer interface technology according to claim 1, characterized in that: Trainable parameters include weights and biases.

4. The depression assessment system based on brain-computer interface technology according to claim 1, characterized in that: The depression assessment module generates a depression prediction result, which specifically includes the following steps: Step S1: Processing EEG features and biosignal features to generate a normalized biosignal tensor; Step S2: extracting the features of the normalized biological signal tensor through multi-scale convolution to generate multi-scale spatiotemporal features; Step S3: construct a Hyper-Sig-V3 activation function, process multi-scale spatiotemporal features through the Hyper-Sig-V3 activation function, and generate a nonlinear feature tensor; Step S4: Processing the nonlinear feature tensor to generate a standardized feature tensor; Step S5: Process the standardized feature tensor to generate depression prediction results.

5. A depression assessment system based on brain-computer interface technology according to claim 4, characterized in that: In step S3, the Hyper-Sig-V3 activation function is constructed by using a Bipolar Sigmoid transformation term and a composite nonlinear transformation term.

6. The depression assessment system based on brain-computer interface technology according to claim 4, characterized in that: Step S2 specifically includes the following steps: Step S21: Adopting dynamic step-length convolution to perform time window adaptive adjustment on the normalized biological signal tensor to generate dynamic step-length features; combining recursive residual learning strategy to calculate the local change trend of the dynamic step-length features to generate short-term recursive time-dependent features; Step S22: capturing short-term recursive time-dependent features through intermediate layer convolution and extracting temporal pattern features; Step S23: Based on the temporal pattern features, global spatiotemporal features are extracted through dilated convolution to generate multi-scale spatiotemporal features.

Images