Neural network based electroencephalography data mental health assessment system

Abstract

This invention relates to the fields of electroencephalogram (EEG) signal processing, artificial intelligence, and mental health assessment technology, specifically to a neural network-based EEG data mental health assessment system. The system includes an EEG data input interface, a basic processing subsystem, a core assessment subsystem, and a support subsystem. It receives EEG data packets and forms a timestamped EEG time series, generating system status information and window-level data integrity information containing window incompleteness information and time interval discrete information. Based on the system downsampling information and integrity information, it determines the window status, switches between the main assessment path and a backup assessment path, performs neural network assessments on the EEG time series or dimensionality-reduced EEG features, and outputs mental health assessment results in conjunction with hardware status characteristics. This invention can maintain the stability and continuity of assessment output even under hardware-constrained operating conditions.

Description

Technical Field

[0001] This invention relates to the fields of electroencephalogram (EEG) signal processing, artificial intelligence, and mental health assessment, specifically to a mental health assessment system based on neural network-based EEG data. Background Technology

[0002] Existing mental health assessment programs can use EEG acquisition devices to obtain EEG signals and analyze the acquired EEG data based on neural network models to output mental health indicators such as focus and fatigue. In related technologies, in order to improve the accuracy of evaluation results, continuous and complete EEG time series are usually directly input into neural networks for inference; or, when EEG data is lost, the time series is discontinuous, or the device's operating performance is degraded, the missing data is first interpolated, filled, or simply dimensionality reduced, and then the processed data is input into the evaluation model. However, in scenarios with limited computing power, such as edge lightweight devices, the devices are prone to issues like processor frequency reduction, temperature rise, increased scheduling latency, and data reception queue backlog during operation. This can lead to incomplete EEG data packet reception, discrete sample time intervals, or even momentary communication interruptions. In such cases, the former approach cannot guarantee that the neural network can continuously and stably receive continuous EEG time series that meet the input requirements, while the latter approach not only increases the consumption of additional computing resources but may also introduce features that are inconsistent with the real signals. Therefore, EEG data-based mental health assessment methods in related technologies struggle to balance assessment accuracy, output stability, and continuous operation under limited hardware conditions. Summary of the Invention

[0003] The purpose of this invention is to provide a mental health assessment system based on EEG data from a neural network, which avoids the need for interpolation to fill data when the device's computing power is limited or EEG data is missing, thus avoiding additional computing resource consumption and the introduction of non-realistic features. Furthermore, it can balance assessment accuracy, output stability, and the system's continuous operation capability under limited hardware conditions.

[0004] The objective of this invention can be achieved through the following technical solutions: A neural network-based EEG data-based mental health assessment system includes an EEG data input interface, a basic processing subsystem, a core assessment subsystem, and a support subsystem; the EEG data input interface is used to communicate with EEG acquisition devices and receive EEG data packets. The basic processing subsystem is used to form a timestamped EEG time series based on EEG data packets, generate system status information and window-level data integrity information, which includes window incompleteness information and time interval discrete information. The core assessment subsystem includes a path selection module and a dual-path neural network module. The dual-path neural network module includes a main assessment path and a backup assessment path. The path selection module is used to switch between the main assessment path and the backup assessment path based on system status information and window-level data integrity information. The main assessment path is used to perform neural network assessment on EEG time series, and the backup assessment path is used to perform neural network assessment on dimensionality-reduced EEG features and output mental health assessment results. The support subsystem is used to generate hardware status characteristics based on the original hardware status information, and the hardware status characteristics are input to the core evaluation subsystem.

[0005] Preferably, the core assessment subsystem includes a path selection module, a dual-path neural network module, and a mental health assessment output module; the path selection module is used to receive system status information and window-level data integrity information and generate path selection results; The dual-path neural network module includes a main evaluation path, a backup evaluation path, and a state information combination layer. The main evaluation path is used to process EEG time series, the backup evaluation path is used to process dimensionality-reduced EEG features, and the state information combination layer is used to receive and combine hardware state features and dimensionality-reduced EEG features. The mental health assessment output module is used to receive the assessment results output by the dual-path neural network module and output at least one mental health indicator.

[0006] Preferably, the system status information includes system frequency reduction information; the path selection module is used to determine the window status based on the system frequency reduction information and the window incomplete information and time interval discrete information in the window-level data integrity information, and output the main evaluation path selection result or the backup evaluation path selection result based on the window status; The window status includes stable status, warning status, and high-risk status; when the window status is stable or warning status, the main assessment path is selected. When the window status is high-risk, select the backup assessment path and input the dimensionality-reduced EEG features and hardware status features into the backup assessment path.

[0007] Preferably, the basic processing subsystem includes an EEG data receiving buffer module, a hardware status detection module, a window status determination module, and a data dimensionality reduction module; The EEG data receiving buffer module is used to receive EEG data packets, form EEG time series with timestamps, and provide information on the number of samples within the window, adjacent time intervals, and queue backlog. The hardware status detection module is used to read the real-time frequency, base frequency, temperature, scheduling delay information and queue backlog information of the central processing unit, and determine the central processing unit frequency decay information based on the real-time frequency and base frequency of the central processing unit; The window state determination module is used to generate system state information, window-level data integrity information and window state based on the output of the EEG data receiving buffer module and the hardware state detection module, and output them to the path selection module in the core evaluation subsystem. The data dimensionality reduction module is used to perform timestamp-based dimensionality reduction processing on EEG time series when the path selection result corresponds to the alternative evaluation path, generating fixed-dimensional dimensionality reduction features.

[0008] Preferably, the window state determination module uses frequency attenuation information, temperature information, scheduling delay information and queue backlog information from the hardware state detection module to form system state information, wherein the system frequency reduction information includes at least frequency attenuation information; The window state determination module also uses the number of samples within the window and the adjacent time intervals from the EEG data receiving buffer module to determine the window incomplete information and the time interval discrete information. The window state determination module determines the window state based on system frequency reduction information, window incompleteness information, and time interval discreteness information.

[0009] Preferably, the data dimensionality reduction module divides the current evaluation window into multiple sub-segments and collects the EEG samples in each sub-segment according to the timestamp; The data dimensionality reduction module extracts dimensionality reduction features that characterize the EEG signals and temporal coverage of each sub-segment, and combines window incomplete information and time interval discrete information to generate fixed-dimensional dimensionality reduction features; the fixed-dimensional dimensionality reduction features are input to the backup evaluation path.

[0010] Preferably, the support subsystem includes a hardware state feature generation module and a parameter calibration module; the hardware state feature generation module is used to align frequency attenuation information, temperature information, scheduling delay information and queue backlog information within the same evaluation window, and generate system state information and fixed-dimensional hardware state features accordingly. The parameter calibration module is used to update the window state determination parameters, path switching parameters, and output smoothing parameters based on the hardware status, path selection results, output fluctuations, and inference latency in the historical window, and writes the updated results back to the window state determination module, path selection module, and mental health assessment output module.

[0011] Preferably, the core assessment subsystem also includes a state information combination layer, which receives hardware state features and dimensionality-reduced EEG features, and combines and encodes the dimensionality-reduced EEG features based on the hardware state features to generate combined state features; the backup assessment path outputs mental health assessment results based on the combined state features.

[0012] Preferably, the mental health assessment output module is used to smooth the assessment results of the current assessment window with the assessment results of the previous assessment window; The mental health assessment output module determines a smoothing update coefficient based on the system down-frequency information and window incompleteness information of the current assessment window. The smoothing update coefficient is correlated with the system down-frequency information and window incompleteness information. The mental health assessment output module outputs at least one mental health indicator.

[0013] The beneficial effects of this invention are: 1. This invention sets up an EEG data input interface, a basic processing subsystem, a core evaluation subsystem, and a support subsystem. The basic processing subsystem generates a timestamped EEG time series based on the EEG data packets, generates system status information, and window-level data integrity information containing window incompleteness information and time interval discrete information. This enables the system to reflect the device's operating status and data reception quality simultaneously while continuously acquiring EEG signals, thus making the system no longer dependent on an ideal and complete continuous EEG time series for evaluation. 2. This invention divides the current evaluation window into multiple sub-segments according to timestamps under high-risk conditions by a data dimensionality reduction module, and extracts signal energy features and time coverage features based on the arriving samples in each sub-segment. Then, it combines window incomplete information and time interval discrete information to generate fixed-dimensional dimensionality reduction features. This can preserve the true state of missing data without interpolation or padding, thereby avoiding additional computational resource consumption and feature distortion introduced by non-true data completion. 3. This invention uses a hardware state feature generation module in the support subsystem to align frequency attenuation information, temperature information, scheduling delay information, and queue backlog information within the same evaluation window and generate fixed-dimensional hardware state features. Then, the state information combination layer splices or combines the hardware state features with the dimensionality-reduced EEG features using attention weighting. This enables the backup evaluation path to distinguish between feature changes caused by device frequency reduction and packet loss and changes in the human psychological state itself during inference, thereby reducing the risk of misjudgment and improving the benchmark consistency of evaluation results under constrained working conditions. Attached Figure Description

[0014] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a module architecture diagram of the EEG data-based mental health assessment system based on neural networks in this application embodiment. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Please see Figure 1 A neural network-based EEG data-based mental health assessment system, comprising an EEG data input interface, a basic processing subsystem, a core assessment subsystem, and a support subsystem; The EEG data input interface is used to communicate with EEG acquisition equipment and receive EEG data packets; The basic processing subsystem is used to form a timestamped EEG time series based on EEG data packets, generate system status information and window-level data integrity information, which includes window incompleteness information and time interval discrete information. The core assessment subsystem includes a path selection module and a dual-path neural network module. The dual-path neural network module includes a primary assessment path and a backup assessment path. The path selection module switches between the primary and backup assessment paths based on system status information and window-level data integrity information. The primary assessment path performs neural network assessment on EEG time series, while the backup assessment path performs neural network assessment on dimensionality-reduced EEG features and outputs mental health assessment results. The support subsystem generates hardware status features based on the original hardware status information, and the hardware status features are input to the core assessment subsystem. The core assessment subsystem includes a path selection module, a dual-path neural network module, and a mental health assessment output module; the path selection module is used to receive system status information and window-level data integrity information and generate path selection results; The dual-path neural network module includes a main evaluation path, a backup evaluation path, and a state information combination layer. The main evaluation path is used to process EEG time series, the backup evaluation path is used to process dimensionality-reduced EEG features, and the state information combination layer is used to receive and combine hardware state features and dimensionality-reduced EEG features. The mental health assessment output module is used to receive the assessment results output by the dual-path neural network module and output at least one mental health indicator. System status information includes system frequency reduction information; the path selection module is used to determine the window status based on the system frequency reduction information and the window incompleteness information and time interval discrete information in the window-level data integrity information, and output the main evaluation path selection result or the backup evaluation path selection result according to the window status; The window states include stable state, warning state, and high-risk state. When the window state is stable or warning state, the main evaluation path is selected. When the window state is high-risk state, the backup evaluation path is selected, and the dimensionality-reduced EEG features and hardware status features are input into the backup evaluation path.

[0017] This embodiment discloses the basic architecture and processing path switching logic of a neural network-based EEG data mental health assessment system; The system is configured to simultaneously acquire the operating status of the hardware carrying the system during the continuous acquisition of EEG signals, and switch between inference paths of different complexities based on the hardware status and the reception of data sequences in order to maintain the stability of the evaluation output. The system's data flow steps include: establishing communication between the EEG data input interface and the external EEG acquisition device to continuously receive EEG data packets; the basic processing subsystem extracts the sampled values ​​and actual arrival times from the data packets, constructs a timestamped EEG time series, and slices the data according to a preset evaluation window; Within each evaluation window, the basic processing subsystem synchronously acquires the current operating status of the equipment, generates system status information, and calculates window incompleteness information and time interval discrete information based on the number of samples received within the current evaluation window and the time interval between adjacent samples. The aforementioned status and integrity information is transferred to the path selection module in the core evaluation subsystem; the path selection module determines the current window status based on the system frequency reduction information, window incompleteness information, and time interval discreteness information. The system presets warning thresholds and high-risk thresholds for the above three information items, and their values ​​are based on the benchmark fluctuation range extracted from the equipment's historical continuous operation under normal temperature and no-load conditions. When all three indicators are below the warning threshold, the system is considered to be in a stable state. When at least one of the indicators is greater than or equal to the warning threshold, but all are less than the high-risk threshold, the system is considered to be in a warning state. When any one of the indicators is greater than or equal to the high-risk threshold, the system is considered to be in a high-risk state. The path selection module generates backup evaluation path selection results. At this time, the system stops sending continuous tensor data to the main evaluation path. The hidden layer nodes of the Long Short-Term Memory Network perform hidden state self-decay freezing operation to retain historical temporal features. Meanwhile, the backup evaluation path receives the dimensionality-reduced EEG features for calculation. Meanwhile, the support subsystem converts the acquired raw hardware status information into fixed-dimensional hardware status features. The status information combination layer receives the hardware status features and combines them with the dimensionality-reduced EEG features. Finally, the backup assessment path outputs the assessment results, and the corresponding mental health indicators are output by the mental health assessment output module. If the EEG data input interface experiences a momentary communication interruption within a certain evaluation window, causing the basic processing subsystem to be unable to extract any valid timestamps, the system will directly set the window incomplete information of the current evaluation window to the maximum value, forcibly triggering a high-risk state judgment, thereby transferring the data flow to the backup evaluation path, and avoiding the main evaluation path from stopping the system due to the breakage of the input tensor dimension; The main assessment path adopts a one-dimensional convolutional neural network cascaded long short-term memory network structure. Its input vector dimension is equivalent to the theoretical number of sampling points in a single assessment window multiplied by the number of EEG channels, and the output is a single-dimensional fatigue value representing mental health indicators. The network topology of the main evaluation path consists of three consecutively stacked one-dimensional convolutional layers, each followed by a modified linear unit activation function and a max pooling layer with a stride of 2, used to extract local features in the temporal domain. The final output tensor of the convolutional layer, after being flattened, is input to the hidden layer nodes in the following number of nodes: Temporal global dependencies are obtained in a single-layer long short-term memory network. The output of the hidden state of the long short-term memory network in one step is obtained by including... A fully connected layer of neurons, mapped by a sigmoid activation function, outputs a single-dimensional fatigue value: The backup assessment path adopts a multilayer perceptron network structure, whose input vector dimension is equal to the sum of the dimension of the reduced EEG features and the dimension of the hardware state features. The training sample set of the model is composed of labeled data collected in advance from subjects under two working conditions: normal office and forced limitation of computing power. During the training process, the primary and backup assessment paths are optimized simultaneously and jointly, and the mean squared error is used as the loss function for regression training to ensure the benchmark consistency of the output values ​​under different hardware constraints. The system is deployed on lightweight edge devices used by office workers, such as tablets, to monitor focus and fatigue in real time. When the device is running normally, the system is in a stable state, and the complete EEG time series is sent to the main evaluation path. When the device runs for more than the preset time, causing the temperature to rise and the operating system to limit the processor frequency, the basic processing subsystem detects the increase in system frequency reduction information, and some EEG data packets are lost due to processing lag, resulting in an increase in window incomplete information. At this point, the path selection module determines the current evaluation window's window state as high-risk, and the system stops calling the high-computing-power main evaluation path. Instead, it sends the dimensionality-reduced EEG features and the hardware status features representing the device's frequency reduction to the backup evaluation path. Through this scheduling mechanism, the system can still continuously output fatigue indicators even under equipment-constrained operating conditions.

[0018] In a preferred embodiment of the present invention, the basic processing subsystem includes an EEG data receiving buffer module, a hardware state detection module, a window state determination module, and a data dimensionality reduction module; the EEG data receiving buffer module is used to receive EEG data packets, form EEG time series with timestamps, and provide information on the number of samples within the window, adjacent time intervals, and queue backlog. The hardware status detection module is used to read the real-time frequency, base frequency, temperature, scheduling delay information and queue backlog information of the central processing unit, and determine the central processing unit frequency decay information based on the real-time frequency and base frequency of the central processing unit; The window state determination module is used to generate system state information, window-level data integrity information and window state based on the output of the EEG data receiving buffer module and the hardware state detection module, and output them to the path selection module in the core evaluation subsystem. The data dimensionality reduction module is used to perform time-stamp-based dimensionality reduction processing on the EEG time series when the path selection result corresponds to the backup evaluation path, and generate fixed-dimensional dimensionality reduction features. The window state determination module uses CPU frequency attenuation information, temperature information, scheduling delay information and queue backlog information from the hardware state detection module to form system state information, wherein the system frequency reduction information includes at least CPU frequency attenuation information. The window state determination module also uses the number of samples within the window and adjacent time intervals from the EEG data receiving buffer module to determine window incompleteness information and time interval discreteness information; the window state determination module determines the window state based on system down-frequency information, window incompleteness information and time interval discreteness information. The data dimensionality reduction module divides the current evaluation window into multiple sub-segments and collects EEG samples in each sub-segment according to timestamps. The data dimensionality reduction module extracts dimensionality reduction features that characterize the EEG signals and time coverage of each sub-segment, and combines window incomplete information and time interval discrete information to generate fixed-dimensional dimensionality reduction features. The fixed-dimensional dimensionality reduction features are input to the backup evaluation path.

[0019] When the system determines that it is in a high-risk state, the basic processing subsystem executes the following data reduction and flow logic: the EEG data receiving buffer module stores the received data into the buffer queue, records the current queue backlog information of the module itself, and counts the number of samples in the current evaluation window and the time interval between adjacent samples. Meanwhile, the hardware status detection module periodically reads the real-time frequency, base frequency, temperature, scheduling latency information and queue backlog information of the central processing unit exposed by the underlying operating system, and derives the central processing unit frequency decay information based on the real-time frequency and base frequency of the central processing unit. The window state determination module receives the above two data streams, arranges the CPU frequency attenuation value, temperature value, scheduling delay time value, and queue backlog value in a fixed order, and concatenates them to generate a window state determination module. The one-dimensional feature vector is used as system state information; At the same time, the degree of insufficient sample size is quantified as window incompleteness information. The calculation formula is: ; in This represents the number of valid samples received within the current evaluation window. The total number of theoretically sampled samples; quantifying the fluctuation of time intervals into discrete time interval information. Calculate the variance of the sample time intervals: ; in, The actual time interval between adjacent samples. This is the theoretical average time interval; The index is the sequence number of the time interval; When the window state determination module outputs a high-risk status based on the above information, triggering the backup assessment path, the data dimensionality reduction module is activated. The data dimensionality reduction module does not perform any data interpolation or padding; instead, it employs a uniform time span division mechanism to divide the current assessment window into equal time segments. Sub-segments; For each sub-segment, valid samples are aggregated based on the actual timestamps of the samples, and the signal energy features within the sub-segment are extracted. The specific calculation rule is to extract the amplitude of all arriving samples within the sub-segment and sum the squares. Then, the sum of squares is divided by the number of arriving samples within the sub-segment to obtain the mean square energy value per unit sample, so as to ensure the physical consistency of the features. At the same time, the ratio of the time span of the samples contained in the sub-segment to the theoretical span is calculated as the time coverage feature. Suppose the current evaluation window is divided into For each sub-segment, the data dimensionality reduction module sequentially extracts the two features mentioned above for each sub-segment, forming a data structure of length [length missing]. The segment feature array; The data dimensionality reduction module will reduce the length to The segment feature array, single-dimensional window incomplete information, and single-dimensional time interval discrete information are sequentially concatenated. The concatenated tensor undergoes variance normalization based on statistical standards to eliminate differences in the physical dimensions of internal parameters, and a strictly generated tensor with a fixed length is generated. The fixed-dimensional reduced feature vector is then pushed into the subsequent inference queue to replace the original multidimensional time series tensor input; During the transfer process, if the data dimensionality reduction module finds that no valid EEG samples have arrived in a certain sub-segment when collecting samples according to timestamps, the system does not perform zero-value filling, but directly sets the signal characteristics of the sub-segment to the preset lower limit value. The specific value of the default minimum value is based on the mean value of the background thermal noise baseline pre-calibrated by the EEG acquisition device in the unloaded open circuit state in the constant temperature and humidity shielded chamber, and its time coverage characteristics are set to zero. The above feature extraction steps preserve the missing data states in the dimensionality-reduced features, and then process them through a subsequent neural network combined with the hardware status. When the device is in a high-temperature state, the operating system suspends some background tasks, causing the queue backlog information of the EEG data receiving buffer module to increase. Some EEG data packets are forcibly discarded by the system. The window state determination module detects that the number of samples has decreased and the central processing unit frequency decay exceeds the preset threshold, thus establishing a high-risk state. The data dimensionality reduction module divides the current evaluation window of a few seconds into several small segments according to the timestamp. It calculates the signal energy and time coverage of each segment only for the remaining EEG data that arrives, and strictly combines them according to the aforementioned splicing rules to generate a feature vector of a fixed length; thus suppressing the computational overhead caused by the device generating interpolated data under the condition of limited computing power.

[0020] In a preferred embodiment of the present invention, the support subsystem includes a hardware state feature generation module and a parameter calibration module; the hardware state feature generation module is used to align the CPU frequency attenuation information, temperature information, scheduling delay information and queue backlog information within the same evaluation window, and generate system state information and fixed-dimensional hardware state features accordingly. The parameter calibration module is used to update the window state determination parameters, path switching parameters, and output smoothing parameters based on the hardware status, path selection results, output fluctuations, and inference latency in the historical window, and writes the updated results back to the window state determination module, path selection module, and mental health assessment output module. The core assessment subsystem also includes a state information combination layer, which receives hardware state features and dimensionality-reduced EEG features, and combines and encodes the dimensionality-reduced EEG features based on the hardware state features to generate combined state features; the backup assessment path outputs mental health assessment results based on the combined state features. The mental health assessment output module is used to smooth the assessment results of the current assessment window with the assessment results of the previous assessment window. The mental health assessment output module determines a smoothing update coefficient based on the system down-frequency information and window incompleteness information of the current assessment window. The smoothing update coefficient is correlated with the system down-frequency information and window incompleteness information. The mental health assessment output module outputs at least one mental health indicator.

[0021] The supporting subsystem and the core evaluation subsystem execute the following feature fusion mechanism and result output steps: The hardware status feature generation module in the supporting subsystem extracts various hardware indicators aligned with the time axis of the current evaluation window to form an input vector. Convert it to a fixed dimension Hardware status characteristics Through the linear mapping formula: ; in, For the pre-trained weight matrix, For bias vectors, A preset integer constant is used to match the dimension of the reduced EEG features. Inside the dual-path neural network module, the state information combination layer simultaneously receives the hardware state features and the aforementioned fixed-dimensionality reduced features. Based on the hardware state features, the state information combination layer performs attention weighting or splicing operations on the reduced EEG features to generate the combined state features. When attention weighting is used, the state information combination layer uses the dimensionality-reduced EEG features as the value vector and the hardware state features as the query vector and key vector after linear mapping. The attention weight is obtained by calculating the inner product of the query vector and key vector and normalizing it through a normalized exponential function. The attention weight is then multiplied by the value vector, so that the model automatically focuses on reliable EEG feature segments that are not affected by severe sampling loss. The backup evaluation path calculates the initial evaluation result of the current evaluation window based on the combined features. The initial assessment results are transferred to the mental health assessment output module; this module reads the system down-frequency information and window incompleteness information of the current assessment window, and calculates the smoothing update coefficient based on the degree to which the two deviate from the normal range; The system extracts the excess difference between the current system frequency reduction information and window incomplete information and the preset normal baseline value, and then deducts these excess differences from the initial baseline value after multiplying them by the corresponding linear attenuation factor. When the excess difference of the system's down-frequency information or the excess difference of the window incomplete information exceeds the corresponding preset threshold, causing the cumulative increase of the deduction item, the system reduces the smoothing update coefficient accordingly. To avoid excessive deduction under extreme abnormal conditions, which could cause the smoothing update coefficient to fall below the zero lower limit and lead to erroneous calculations, the system is also configured with a bottom line constraint lower limit. If the deduction result is lower than the lower limit, the bottom line constraint lower limit will be forcibly taken. Its specific quantitative constraint calculation logic is as follows: using a preset initial baseline value Based on this, the excess difference in the current system's frequency reduction information is deducted respectively. With corresponding attenuation factor The product of the two, and the excess difference of window incomplete information. With corresponding attenuation factor The product of the products, and the deduction result is lower than the lower limit of the bottom line constraint. The lower limit of the baseline constraint must be taken at the time. Thus, the smooth update coefficients are calculated. The calculation formula is: ; in, The function to find the maximum value; the benchmark value. and the lower limit of bottom line constraints All are dimensionless normalized scalars; At the same time, attenuation factor The physical dimension is set to be the reciprocal of the dimension of the excess difference in the system's down-frequency information. For example, when the unit of the down-frequency information is MHz, The unit is 1 / MHz. Since the window incomplete information and its excess difference are both dimensionless proportional parameters, the attenuation factor... It is set as a dimensionless constant, thus in product operations and The original physical dimensions of the central processing unit frequency attenuation information are eliminated and uniformly transformed into dimensionless pure numerical deduction terms. The mental health assessment output module uses the calculated coefficients. The initial evaluation results of the current evaluation window will be displayed. Output results saved from the previous evaluation window Perform smoothing weighted calculations based on the smoothing formula: ; Generate the final mental health indicators It also outputs data externally; in addition, the parameter calibration module continuously collects the inference latency, output fluctuations and corresponding hardware status of multiple consecutive evaluation windows in the background, periodically calculates new judgment thresholds, and writes them back to the configuration records of the window status judgment module and the path selection module. During the parameter write-back process, if the parameter calibration module encounters storage read-write conflicts or thread lock occupation when attempting to update the configuration record, the system will abandon the write-back operation of the current cycle, keep the current switching parameters of the system unchanged, and temporarily store the newly calculated parameters in memory. The system will then retry after the next evaluation cycle ends to ensure the continuity of real-time evaluation data flow. When the device experiences frequency reduction, resulting in incomplete EEG data, the state information combination layer combines the dimensionality-reduced EEG features with the hardware state features characterizing the device's frequency reduction; the backup evaluation network then identifies that the current decrease in signal energy is caused by device sampling loss, rather than a sudden change in fatigue characteristics. Meanwhile, the mental health assessment output module detects that the window incomplete information of the current assessment window is greater than the preset threshold, and automatically reduces the smoothing update coefficient, so that the fatigue index output by the current assessment window relies more on the historical stable results of the previous second, thereby suppressing abnormal fluctuations in the index value on the output interface; in the background, the parameter calibration module uses the delayed data during this period to fine-tune the temperature limit for the next high temperature warning, so that the subsequent path switching parameters of the system match the heat dissipation performance threshold of the corresponding equipment.

Claims

1. A neural network-based EEG data-based mental health assessment system, characterized in that, It includes an EEG data input interface, a basic processing subsystem, a core evaluation subsystem, and a support subsystem; the EEG data input interface is used to communicate with the EEG acquisition device and receive EEG data packets; The basic processing subsystem is used to form a timestamped EEG time series based on the EEG data packet, and generate system status information and window-level data integrity information. The window-level data integrity information includes window incompleteness information and time interval discrete information. The core assessment subsystem includes a path selection module and a dual-path neural network module. The dual-path neural network module includes a main assessment path and a backup assessment path. The path selection module is used to switch between the main assessment path and the backup assessment path according to the system status information and the window-level data integrity information. The main assessment path is used to perform neural network assessment on the EEG time series, and the backup assessment path is used to perform neural network assessment on the dimensionality-reduced EEG features and output mental health assessment results. The support subsystem is used to generate hardware status features based on the original hardware status information, and the hardware status features are input to the core evaluation subsystem.

2. The EEG data-based mental health assessment system based on neural networks according to claim 1, characterized in that, The core assessment subsystem includes a path selection module, a dual-path neural network module, and a mental health assessment output module. The path selection module is used to receive the system status information and the window-level data integrity information and generate a path selection result; The dual-path neural network module includes a main evaluation path, a backup evaluation path, and a state information combination layer. The main evaluation path is used to process the EEG time series, the backup evaluation path is used to process the dimensionality-reduced EEG features, and the state information combination layer is used to receive and combine the hardware state features and the dimensionality-reduced EEG features. The mental health assessment output module is used to receive the assessment results output by the dual-path neural network module and output at least one mental health indicator.

3. The EEG data-based mental health assessment system according to claim 2, characterized in that, The system status information includes system frequency reduction information; the path selection module is used to determine the window status based on the system frequency reduction information and the window incomplete information and time interval discrete information in the window-level data integrity information, and output the main evaluation path selection result or the backup evaluation path selection result based on the window status. The window states include stable state, early warning state, and high-risk state; When the window status is stable or warning, the main evaluation path is selected; when the window status is high-risk, the backup evaluation path is selected, and the dimensionality-reduced EEG features and the hardware status features are input into the backup evaluation path.

4. The EEG data-based mental health assessment system according to claim 3, characterized in that, The basic processing subsystem includes an EEG data receiving buffer module, a hardware status detection module, a window status determination module, and a data dimensionality reduction module. The EEG data receiving buffer module is used to receive the EEG data packets, form a timestamped EEG time series, and provide information on the number of samples within the window, adjacent time intervals, and queue backlog. The hardware status detection module is used to read the real-time frequency, reference frequency, temperature, scheduling delay information and queue backlog information of the central processing unit, and determine the central processing unit frequency attenuation information based on the real-time frequency and the reference frequency of the central processing unit; The window state determination module is used to generate the system state information, the window-level data integrity information and the window state based on the output of the EEG data receiving buffer module and the hardware state detection module, and output them to the path selection module in the core evaluation subsystem; The data dimensionality reduction module is used to perform timestamp-based dimensionality reduction processing on the EEG time series when the path selection result corresponds to the backup evaluation path, and generate fixed-dimensional dimensionality reduction features.

5. The EEG data-based mental health assessment system according to claim 4, characterized in that, The window state determination module uses the CPU frequency attenuation information, temperature information, scheduling delay information and queue backlog information from the hardware state detection module to form the system state information, wherein the system frequency reduction information includes at least the CPU frequency attenuation information. The window state determination module also uses the number of samples within the window and the adjacent time intervals from the EEG data receiving buffer module to determine the window incomplete information and the time interval discrete information; The window state determination module determines the window state based on the system frequency reduction information, the window incompleteness information, and the time interval discreteness information.

6. The EEG data-based mental health assessment system according to claim 4, characterized in that, The data dimensionality reduction module divides the current evaluation window into multiple sub-segments and collects EEG samples in each sub-segment according to timestamps. The data dimensionality reduction module extracts dimensionality reduction features that characterize the EEG signals and time coverage of each sub-segment, and combines the window incomplete information and the time interval discrete information to generate the fixed-dimensional dimensionality reduction features. The fixed-dimensional dimensionality reduction features are input into the backup evaluation path.

7. The EEG data-based mental health assessment system according to claim 5, characterized in that, The support subsystem includes a hardware status feature generation module and a parameter calibration module; The hardware status feature generation module is used to align the CPU frequency decay information, temperature information, scheduling delay information and queue backlog information within the same evaluation window, and generate the system status information and fixed-dimensional hardware status features accordingly. The parameter calibration module is used to update the window state determination parameters, path switching parameters, and output smoothing parameters based on the hardware status, path selection results, output fluctuations, and inference latency in the historical window, and write the update results back to the window state determination module, the path selection module, and the mental health assessment output module.

8. The EEG data-based mental health assessment system according to claim 7, characterized in that, The core assessment subsystem also includes a state information combination layer, which receives the hardware state features and the dimensionality-reduced EEG features, and performs combination encoding on the dimensionality-reduced EEG features based on the hardware state features to generate combined state features; the backup assessment path outputs mental health assessment results based on the combined state features.

9. The EEG data-based mental health assessment system according to claim 3, characterized in that, The mental health assessment output module is used to smooth the assessment results of the current assessment window and the assessment results of the previous assessment window. The mental health assessment output module determines a smoothing update coefficient based on the system down-frequency information and window incompleteness information of the current assessment window. The smoothing update coefficient is associated with the system down-frequency information and the window incompleteness information. The mental health assessment output module outputs at least one mental health indicator.

Images