Electroencephalogram signal analysis method, system and equipment based on interactive concept bottleneck

Abstract

The invention provides an electroencephalogram signal analysis method, system and equipment based on interactive concept bottleneck, and relates to the technical field of electroencephalogram signal analysis, the system comprises an electroencephalogram acquisition and preprocessing module used for acquiring and preprocessing original electroencephalogram signals; the electroencephalogram feature extraction module is used for performing feature extraction on the preprocessed electroencephalogram signals to obtain symmetry features and index features, and calculating percentage feature values corresponding to the index features; the interactive concept layer module is used for mapping the percentage characteristic value into a low-order concept through a concept bottleneck layer; the health state classification module is used for mapping the low-order concepts into high-order concepts so as to realize the recognition of the high-order concepts; and the model optimization and transfer learning module is used for updating the high-order concept through transfer training. According to the scheme, the measurement result is accurate and stable, the analysis efficiency is high, real-time state evaluation is provided, and the cross-task multiplexing capability is good.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electroencephalogram signal analysis, and in particular to an electroencephalogram signal analysis method, system and device based on an interactive concept bottleneck. BACKGROUND

[0002] Electroencephalogram (EEG) is an important biological signal for recording brain activity, with the characteristics of real-time, non-invasive and high sensitivity, and is widely used in mental health assessment, emotional state monitoring, sleep quality analysis and other fields. In recent years, with the rapid development of artificial intelligence and biomedical signal processing technology, health status analysis based on electroencephalogram has gradually become a research hotspot. Common methods include frequency domain analysis (such as power spectral density calculation), time domain feature extraction (such as entropy and complexity analysis), time-frequency analysis (such as wavelet transform), and deep learning model end-to-end modeling of electroencephalogram in recent years.

[0003] EEG-based psychological / cognitive assessment technology is a method of using electroencephalogram signals to assess individual cognitive state and ability. According to the classification of electroencephalogram signal acquisition methods, it can be divided into two technical routes: invasive and non-invasive. Invasive electroencephalogram processing provides higher quality and resolution signals, but involves surgery and ethical issues. Non-invasive electroencephalogram (Non-invasive EEG) records and analyzes brain electrical activity by placing electrodes on the scalp. Compared with invasive methods (such as implanting electrodes in the brain), non-invasive electroencephalogram processing has the advantages of high safety, low risk, and relatively simple operation, and can be combined with wearable devices, and is widely used in brain science research and other fields. Non-invasive electroencephalogram processing technology has significant advantages in safety, convenience and cost. However, the quality and spatial resolution of non-invasive electroencephalogram signals are relatively low, and greatly rely on signal processing and feature extraction to achieve cognitive analysis and modeling, such as frequency domain features (power spectral density), time domain features (such as event-related potentials, ERP) and nonlinear features (such as electroencephalogram complexity); In addition, recent studies have shown that deep learning algorithms (such as convolutional neural networks, long short-term memory networks, and Transformer networks) can also achieve relatively high-quality electroencephalogram feature extraction.

[0004] In the mapping of non-invasive electroencephalogram features to individual cognitive states and abilities, according to the category difference of cognitive assessment tasks (local or global), typical local cognitive function assessment includes sensory and perception, attention, memory behavior, motor control, etc.; global cognitive function assessment involves reasoning and decision-making, complex semantic memory, psychological assessment and other more abstract aspects. Combined with the instantaneous electroencephalogram changes of the subject to a specific stimulus or the performance of a certain task, the local cognitive function can be directly defined, and the direct interpretation can be realized based on the amplitude or change of the quantitative features. Global cognitive function assessment requires a comprehensive analysis of a segment of electroencephalogram, and sometimes multiple possible feature dimensions need to be considered for collaborative analysis to obtain the comprehensive processing mode, type characteristics, ability hierarchy, etc. of the human brain for one or more cognitive tasks. Since global cognitive function assessment needs to model complex data (or features of electroencephalogram data) in multiple dimensions and nonlinearly, the existing technical route usually adopts machine learning methods to realize multi-stage modeling, but the existing technology at least has the following problems:

[0005] Problem 1: Lack of dynamic interaction capability: the existing model usually adopts fixed feature selection and classification strategy, which cannot dynamically adjust the analysis path according to the characteristics of the input signal, which makes it difficult for the user to add some deterministic information to the model decision to improve the friendliness of human-computer collaboration; on the other hand, the user cannot evaluate the model accuracy and bias by adjusting part of the features and dynamically verifying the expected results according to the specific situation of the input instance.

[0006] Problem 2: Lack of explainability: the "black box" nature of deep learning models makes it difficult to explain the prediction results, since the reasoning and calculation logic of the model is implicitly reflected in the internal parameters of the model, and there is a lack of method for embedding the decision path of human deterministic expert knowledge rules, which makes it challenging to align the decision path of the electroencephalogram output global cognitive function indicators with the existing application scenarios.

[0007] Problem 3: Poor generalization performance: the data distribution of different populations and application scenarios is significantly different, and the existing technical route obtains modeling for a specific downstream task through an end-to-end method, which has the problems of poor model reuse effect and low continuous learning ability, resulting in a decrease in prediction accuracy of the existing model in new scenarios. If new task migration is performed, retraining and fine-tuning are often required, but the parameter adjustment change controllability is poor, and there is inevitable performance loss of the original electroencephalogram model. SUMMARY

[0008] Therefore, the present application provides an electroencephalogram signal analysis method, system and device based on interactive concept bottleneck, to at least partially solve the problems existing in the prior art. Specifically, the present application provides the following technical features:

[0009] On the one hand, the present invention provides an EEG signal analysis system based on an interactive concept bottleneck, which includes an EEG acquisition and preprocessing module, an EEG feature extraction module, an interactive concept layer module, a health status classification module, and a model optimization and transfer learning module, which are arranged in sequence;

[0010] The EEG acquisition and preprocessing module is used to acquire the original EEG signal and preprocess it to obtain the preprocessed EEG signal; the EEG signal includes the symmetrically acquired left EEG signal and right EEG signal;

[0011] An EEG feature extraction module is used to extract features from the preprocessed EEG signal to obtain symmetry features, and calculate index features for different bands; and calculate the percentage eigenvalues ​​corresponding to the index features through quantile mapping;

[0012] The interactive concept layer module is used to map percentage feature values ​​to low-order concepts through the concept bottleneck layer based on interactive concept bottleneck models (ICBMs), and to screen the indicator features that contribute most to high-order concepts through dynamic intervention;

[0013] The health status classification module is used to map low-order concepts into high-order concepts and realize the recognition of high-order concepts;

[0014] The model optimization and transfer learning module is used to update high-level concepts through transfer training; in the transfer training, dynamic concept alignment is performed on newly added high-level concepts.

[0015] In a feasible embodiment, the symmetry feature includes:

[0016] The symmetry eigenvalue Asymmetry(t) is used to reflect the power difference or entropy difference between the left EEG signal and the right EEG signal at time point t;

[0017] Biased towards the global mean μ Asymmetry It is used to reflect the average difference trend between the left and right EEG signals in the entire time series, so as to measure the left and right brain activation bias;

[0018] Global average intensity μ Abs_Asymmetry It is used to reflect the trend of the average difference in the intensity between the left and right EEG signals in the entire time series, so as to measure the intensity of the difference between the left and right brain rather than the direction;

[0019] Symmetry index standard deviation σ Asymmetry Used to reflect the fluctuation of symmetry over time.

[0020] In a feasible implementation manner, the indicator features include: power spectrum density PSD, differential entropy DE, differential entropy asymmetry RASM and differential entropy asymmetry difference DASM.

[0021] In a feasible implementation manner, the method for calculating the percentage characteristic value includes:

[0022] For indicator features with sufficient sample size, calculate the cumulative distribution function F I Specific expressions of (x) include:

[0023]

[0024] Among them, n ≤v is the number of samples in the reference sample whose indicator characteristic value is less than or equal to v; N represents the total number of reference samples; v represents the preset threshold value of the indicator characteristic;

[0025] For indicator features with insufficient sample size, the cumulative distribution function F is calculated using normal distribution. I (x), specific expressions include:

[0026]

[0027] Among them, μ0 is the mean of the sampling points; σ is the standard deviation of the sampling points; x represents the indicator characteristics of the sampling points;

[0028] Based on the cumulative distribution function, calculate the percent eigenvalue P score Specific expressions of (x) include:

[0029] P score (x) = F I (x)×100,

[0030] In a feasible implementation, the health status classification module maps low-order concepts to high-order concepts through a high-order concept label predictor.

[0031] In a feasible implementation, the dynamic intervention method includes: screening out manual intervention values ​​by priority to find indicator features that contribute most to high-order concepts.

[0032] In a feasible implementation, in the model optimization and transfer learning module, dynamic concept alignment ensures the transfer training effect through the concept alignment loss function.

[0033] On the other hand, based on the same inventive concept, the present invention also provides an EEG signal analysis method based on an interactive concept bottleneck, the method comprising:

[0034] S1. Collecting original EEG signals and preprocessing them to obtain preprocessed EEG signals, wherein the EEG signals include symmetrically collected left and right EEG signals;

[0035] S2. Extracting features from the preprocessed EEG signal to obtain symmetry features, and calculating index features for different bands; calculating percentage eigenvalues ​​corresponding to the index features through quantile mapping;

[0036] S3. Based on the interactive concept bottleneck model, the percentage feature values ​​are mapped to low-order concepts through the concept bottleneck layer, and then the low-order concepts are mapped to high-order concepts to achieve the recognition of high-order concepts; the concept bottleneck layer selects the indicator features that contribute most to the high-order concepts through dynamic intervention;

[0037] S4. Updating high-level concepts through transfer training; in the transfer training, dynamic concept alignment is performed on the newly added high-level concepts.

[0038] In a feasible implementation manner, the pre-processing in S1 includes: high-pass and low-pass filtering, notch filtering, and artifact processing;

[0039] In the above S2, before extracting features from the pre-processed EEG signal, the process further includes: segmenting the pre-processed EEG signal according to fixed time windows, and then performing a standardization operation on the data of the EEG channel within the window before performing feature extraction.

[0040] On the other hand, based on the same inventive concept, the present invention also provides an EEG signal analysis device based on an interactive concept bottleneck, the device comprising EEG signal acquisition electrodes, a processor, a display, and a memory;

[0041] The EEG signal acquisition electrodes are used to collect or receive EEG signals; the display is used to display an interactive screen with the user based on instructions in the memory;

[0042] The processor can call the computer instructions in the memory to implement the above-mentioned EEG signal analysis system based on the interactive concept bottleneck.

[0043] In a feasible implementation manner, the interactive screen includes a specific task screen that needs to be completed by the user.

[0044] Compared with the existing technology, this solution has at least the following advantages:

[0045] The scheme maps input features (such as quantitative indicators of electroencephalogram signals) to high-level health concepts (such as brain stability, brain vitality, anxiety level, etc.), converts the underlying signals to high-level semantics, breaks through the limitations of traditional single feature analysis, and captures the complex relationships between features by using the interactive mechanism and learnable parameters of the concept bottleneck layer, thereby improving the accuracy and stability of the output results.

[0046] The scheme converts complex input features into interpretable high-level concepts through the concept bottleneck layer and supports dynamic intervention and updating. The evaluation results are based on clear high-level concepts, which are easy to understand and use, and the dynamic intervention strategy is combined to adjust the high-level concepts in real time according to the specific situation of individuals, thereby improving the practicality of the model.

[0047] The scheme dynamically selects the concept that has the greatest impact on the evaluation results through the intervention strategy, thereby optimizing the state evaluation. In the case of limited resources, the indicators that have the greatest impact on prediction are prioritized for intervention, reducing redundant calculations, saving time and cost, and reducing device power consumption and use cost. Through dynamic intervention, real-time state evaluation and suggestion support can be provided.

[0048] The scheme integrates multi-dimensional features through concept interaction to improve the overall effect of evaluation, and optimizes the prediction path in real time through the interactive concept bottleneck layer, thereby improving the efficiency and effectiveness of the model. It also avoids the problems of existing technologies that are difficult to capture complex feature interactions and cannot explain and adjust the results.

[0049] The scheme introduces an interactive concept layer and a dynamic distribution alignment mechanism to dynamically correct the high-level concept distribution and the source domain distribution in real time during new task migration, ensuring the semantic consistency of the new task and the source domain task, significantly improving the adaptation ability and generalization ability of the model in different application scenarios, reducing the performance decline caused by data distribution differences, and improving the cross-task reuse ability of the model in the few-shot scenario.

[0050] The scheme adopts a hierarchical transfer learning strategy, which only optimizes the high-level concepts (layers) and tasks (layers) during the fine-tuning stage, and fixes the low-level concepts (layers), thereby reducing the need for retraining during new task migration, reducing computational cost and training time, improving the applicability of the model, and improving the efficiency of cross-device and cross-scenario model deployment. BRIEF DESCRIPTION OF DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creating any inventive labor.

[0052] Figure 1 A schematic diagram of a brain electrical signal analysis system provided by an embodiment of the present application is shown in FIG. 1.

[0053] Figure 2 A schematic diagram of a model migration process provided by an embodiment of the present application is shown in FIG. 4.

[0054] Figure 3 A schematic diagram of a brain electrical signal analysis method provided by an embodiment of the present application is shown in FIG. 5. DETAILED DESCRIPTION

[0055] The embodiments of the present application will be described in detail below with reference to the accompanying drawings. It should be noted that the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0056] Those skilled in the art should know that the following specific embodiments or specific implementations are a series of optimized settings provided by the present application to further explain the specific inventive content, and these settings can be used in combination or in association with each other, unless the present application explicitly indicates that certain or a specific embodiment or implementation cannot be associated with other embodiments or implementations or cannot be used together. At the same time, the following specific embodiments or implementations are only the most optimized settings, and not as an understanding of the protection scope of the present application.

[0057] Embodiment 1:

[0058] In this embodiment, the content of the scheme is described in detail in combination with the system design of the scheme. Referring to FIG. 1, the overall system architecture design of the embodiment includes the brain electrical signal acquisition and preprocessing module, the brain electrical signal feature extraction module, the interactive concept layer module, the health state classification module, and the model optimization and migration learning module arranged in sequence. Figure 1-2

[0059] The main workflow is as follows: the original brain electrical signal is input into the brain electrical signal acquisition and preprocessing module for basic signal processing, and then input into the brain electrical signal feature extraction module to extract feature values; the extracted feature values are input into the interactive concept layer module for iterative processing, during which expert intervention adjustment is performed, and the output data is transmitted to the health state classification module for mapping and obtaining classification analysis results; the processing result data of the health state classification module is transmitted to the model optimization and migration learning module to realize the final optimization adjustment and migration of the model.

[0060] ​Furthermore, in the original EEG signal acquisition stage, multi-lead EEG data can first be collected for a certain period of time (for example, about 6 minutes) for the person to be tested or evaluated. According to the actual hardware device parameters, including but not limited to using the universal 512Hz acquisition frequency standard, EEG signals in categories such as resting state, eyes opening and closing process, stress induction task, and cognitive load task can be fully collected.

[0061] Each module in the system architecture is introduced as follows with examples:

[0062] EEG acquisition and preprocessing module:

[0063] Input raw EEG signals Where C is the total number of EEG channels, T is the length of time for a single individual measurement, The data set is represented by a set representation method commonly used in the art. Conventional EEG equipment is used to collect multi-channel EEG signals, and electrodes are arranged according to the 10-20 international electrode positioning system.

[0064] In practical EEG measurement applications, it's important to balance user-friendliness and convenience with data accuracy and complexity. Because EEG channels near the temporal and frontal lobes are most important for modeling, and the accuracy of modeling based on four key lead data from relevant brain regions can match or even exceed the accuracy of using all 62 EEG channels, and to facilitate the design and deployment of portable devices, this embodiment uses two electrodes to symmetrically collect the left and right frontal lobe signals, Fp1 and Fp2.

[0065] It should be noted that this solution can also add more pairs of symmetrical electrodes, and these methods of adding symmetrical electrodes should also be considered to fall within the scope of protection of the present invention.

[0066] The symmetry or asymmetry of the signal can reflect the lateralization of brain activity and provide support for comparative analysis of left and right brain activity.

[0067] For the left frontal lobe signal Fp1 and the right frontal lobe signal Fp2 in this embodiment, the collected original EEG signals can be expressed as follows:

[0068]

[0069] First, the original EEG signal X needs to be filtered. In this embodiment, the filtering method is as follows:

[0070] X Filtered =H LP (f,fcut，LP )·H HP (f,f cut，HP )·X

[0071] Among them, H LP Indicates low-pass filter calculation; H HP Indicates high-pass filter calculation; X Filtered Represents the EEG signal after high-pass and low-pass filtering.

[0072] Preferably, in order to adapt to the EEG signal characteristics of different subjects, the high-pass cutoff frequency in the present invention can be set between 0.1 Hz and 1.0 Hz, and the low-pass cutoff frequency can be set between 30 Hz and 45 Hz, so as to retain complete and effective EEG activity band information while filtering out artifacts and non-brain source signals.

[0073] Preferably, the high-pass cutoff frequency f cut，HP Set to 0.1Hz, low-pass cutoff frequency f cut，LP Set to 40Hz.

[0074] Furthermore, the EEG signal after high-pass and low-pass filtering must be passed through a notch filter to remove power frequency interference. The specific processing method is as follows:

[0075] X Notch =H Notch (f, f0, Q)·X Filtered

[0076] Among them, X Notch represents the EEG signal after notch filtering; f represents the selected frequency; f0 represents the center frequency of the notch filter, which is used to specify the interference frequency that needs to be suppressed and can be set to 50Hz to filter out external power grid frequency noise; H Notch Represents the notch filter calculation, the specific formula is:

[0077]

[0078] The quality factor Q can be set to 30; the notch depth γ ranges from [0.8, 1.2] and can be set to 1.

[0079] Next, artifacts such as blink artifacts and motion artifacts in the EEG signal are processed, that is, the large-amplitude low-frequency spikes in Fp1 and Fp2 are identified and eliminated. The specific expressions include:

[0080] t blink ={t||X Notch (c, t)|>τ}

[0081] Wherein, τ represents the amplitude threshold for artifact detection, which is preferably 80 μV in this embodiment and is used to determine the amplitude abnormality of eye movement or other artifact signals; c represents the EEG channel (number), namely Fp1 and / or Fp2, which are most sensitive to blinking and forehead EMG artifacts; t represents the sampling time point of the EEG signal; t blink Indicates the detected artifact points.

[0082] Directly removing the detected artifact points will cause time breaks in the EEG signal, destroying the temporal continuity and hindering frequency domain analysis and feature extraction. Therefore, in this embodiment, after removing these artifact points, signal repair methods (such as linear interpolation, spline interpolation, independent component analysis, etc.) are used for processing. Linear interpolation or spline interpolation can also be used to reconstruct the signal within a ±200ms time window, and independent source denoising is performed using the ICA component stripping method. These methods can adopt well-known algorithms well known to those skilled in the art, so they will not be described in detail.

[0083] EEG feature extraction module:

[0084] For the EEG signals processed by the EEG acquisition and preprocessing module, a variety of methods can be used to obtain predefined EEG features (for example, using a pre-trained deep learning model). After the model structure is calculated, high-dimensional feature quantification results related to psychological and physiological states are output (such as stress level scores, sleep quality scores, emotion classification labels, disease risk scores, etc.). The acquisition of these features is related to the specific deep model training tasks and label design. This type of method can be implemented through the commonly used neural network models in the existing technology. In addition, EEG analysis can also be carried out using methods such as time domain features (such as amplitude, mean) and frequency domain features (such as band energy).

[0085] In this embodiment, the specific method of performing feature extraction is as follows:

[0086] First, the EEG signal processed by the EEG acquisition and preprocessing module is segmented according to a fixed time window for subsequent feature extraction or modeling. The window segmentation method can be expressed as:

[0087] X Segmented ={X[:,t s :t s +W]|t s =1, 2, ..., TW}

[0088] Among them, X[:,t s :t s +W] represents the EEG signal obtained after EEG acquisition and preprocessing module processing; X SegmentedX(c, t) represents the segmented electroencephalogram; T represents the length of electroencephalogram collection time; W represents the length of the window; for example, for an electroencephalogram with a sampling rate of 256 Hz, the window length can be set to 1 second, and the electroencephalogram involves 256 data points;

[0089] The electroencephalogram X(c, t) in the window is standardized, and the specific expression is:

[0090]

[0091] Wherein, μ c represents the mean of the sampling points in the electroencephalogram channel c, which is used to estimate the baseline potential of the electroencephalogram channel in the window; σ c represents the standard deviation of the sampling points in the electroencephalogram channel c, which is used to measure the fluctuation amplitude of the electroencephalogram channel signal in the window; X Norm (c, t) represents the electroencephalogram after standardization.

[0092] The following takes the symmetry evaluation of left and right brain activities as an example to illustrate that the difference in brain activity of the left and right frontal lobes can be reflected by calculating the symmetry feature between Fp1 and Fp2:

[0093] In this embodiment, the specific calculation formula of the symmetry feature includes:

[0094] Asymmetry(t) = X Fp1 (t) - X Fp2 (t)

[0095]

[0096] Wherein, Asymmetry(t) represents the symmetry feature value, which is used to reflect the power difference between the left and right electroencephalograms (in a certain wave band) at time point t; μ Asymmetry represents the bias global average, which is used to reflect the average difference trend between the left and right electroencephalograms in the entire time sequence, so as to measure the left and right brain activation bias; μ Abs_Asymmetry represents the intensity global average, which is used to reflect the average difference trend in intensity between the left and right electroencephalograms in the entire time sequence, so as to measure the intensity of the left and right brain difference rather than the direction; σ Asymmetry represents the standard deviation of the symmetry index, which is used to reflect the fluctuation of the symmetry with time, the smaller the value, the more stable the emotion, and the larger the value, the more likely the attention fluctuation, emotional instability, etc.; X Fp1 (t) represents the value of Fp1; X Fp2 (t) represents the value of Fp2.

[0097] In this way, using power difference for calculation can not only effectively evaluate psychological states (such as anxiety and depression), analyze cognitive behaviors, and monitor emotional stability, but also reduce the computing power requirements of hardware.

[0098] Furthermore, the indicator features are set as: energy spectrum density PSD, differential entropy DE, differential entropy asymmetry RASM and differential entropy asymmetry difference DASM, and the corresponding calculation methods are as follows:

[0099]

[0100] DASM=DE left -DE right ;

[0101] Where f represents the selected frequency, which can be determined by the band division; DE left Denotes the differential entropy of Fp1, that is, the differential entropy eigenvalue of Fp1 in a certain band; DE right represents the differential entropy of Fp2; σ represents the standard deviation of the sampling points to measure volatility; e represents a natural constant.

[0102] Furthermore, the characteristics of each indicator are calculated in the delta, theta, alpha, beta, gamma and other bands, which can be used to evaluate the psychological state of different subjects, for example:

[0103] The delta (0.5-4 Hz) band is associated with deep sleep, recovery, brain damage, or fatigue. When delta wave activity is too strong, it may reflect drowsiness, fatigue, or decreased cognitive ability.

[0104] The theta (4-8 Hz) band is associated with memory, meditation, and early sleep states. High theta wave activity may reflect a state of relaxation or meditation, while low theta wave activity may indicate anxiety or inattention.

[0105] The alpha (8-13 Hz) band is associated with states of relaxation, wakefulness, and concentration. Increased alpha wave activity generally indicates relaxation, while decreased alpha activity may indicate anxiety or emotional stress.

[0106] The beta (13-30 Hz) band is associated with alertness, cognitive tasks, and anxiety. High beta activity may be associated with anxiety or high cognitive load, while low beta activity may reflect inattention.

[0107] The gamma (30-42 Hz) band is associated with higher cognitive functions, learning, and memory, and high gamma wave activity may be associated with deep concentration, emotional activation, or excitement.

[0108] The following examples illustrate the specific calculation process of each band:

[0109] Calculate the asymmetry characteristics of the alpha band and quantify them using RASM and DASM values. α Indicates the DASM value of the alpha band, RASM α The RASM value representing the alpha band can be used to calculate the quantitative emotional state characteristic value through standard interval mapping:

[0110] DASM α =DE α，left -DE α，right ;

[0111] Among them, DE α，left Indicates the differential entropy value calculated from the left EEG signal (such as Fp1) in the alpha band, DE α，right It represents the differential entropy value calculated from the right EEG signal (such as Fp2) in the alpha band;

[0112] Calculate the PSD and DE values ​​of the theta band, that is, PSD θ 、DE θ , as the characteristic value of attention or cognitive function, the corresponding calculation method is as follows:

[0113]

[0114] Calculate the RASM and PSD values ​​of the beta band as the pressure characteristic value, namely RASM β and PSD β , which is calculated as follows:

[0115]

[0116] Calculate the DASM value of the alpha band as the emotion regulation feature value, namely DASM α , a positive value can reflect good emotion regulation ability, and its calculation method is:

[0117] DASM α =DE α，left -DE α，right ;

[0118] Calculate the PSD and DE values ​​of the gamma band, that is, PSD γ and DE γ , as a characteristic value of the load state, a high value can reflect deep concentration and high workload. The corresponding calculation method is as follows:

[0119]

[0120] This embodiment provides specific applications of quantitative values ​​in several EEG data analysis scenarios. Since the various numerical values ​​of EEG data analysis themselves have different numerical ranges and distributions, in order to improve the applicability of different numerical values, this embodiment further maps the position of the target value in the reference measured person data to a percentile score through quantile mapping, and obtains the percentage characteristic value of a specific range, so as to facilitate more detailed quantification and analysis applications of EEG data.

[0121] In this embodiment, the specific calculation method of the percentage characteristic value is:

[0122] First, I(X) represents the individual (i.e., the original indicator characteristics of EEG features, such as specific indicator characteristics based on PSD, DE, RASM, etc.);

[0123] Secondly, when the sample size is sufficient and the empirical distribution can be estimated, the cumulative distribution function is estimated based on the reference population distribution. The specific calculation formula is: Among them, n ≤v is the number of samples in the reference population whose indicator value is less than or equal to v; N is the total number of reference samples; v represents the preset threshold value of the indicator feature (a certain dimension);

[0124] Since some task-driven indicators require specific high-quality EEG acquisition and labeling, there may be a situation where the available samples are insufficient to estimate the empirical distribution. For this situation, this embodiment uses the normal distribution assumption to estimate the cumulative distribution function, namely: Among them, μ0 is the mean of the sampling points; σ is the standard deviation of the sampling points; x represents the indicator characteristics of the sampling points;

[0125] It should be noted that whether the sample size is sufficient can be determined by setting an empirical threshold. For example, when the reference sample size N is less than 50, the sample size is considered insufficient.

[0126] Finally, calculate the percentage eigenvalue P score (x), the specific formula is: P score (x) = F I (x)×100, where P score (x)∈[0,100] represents the relative position of individual I(X) in the reference distribution.

[0127] Interactive concept layer module and health status classification module:

[0128] Since the percentage characteristic value can reflect the relative position of the individual index in the reference population, but it is difficult to provide an explanatory overall state of the measured personnel by directly outputting multiple discrete percentage characteristic values. Therefore, based on the interactive concept bottleneck model, the percentage characteristic value calculated is integrated and mapped into the corresponding high-order concept by the interactive concept layer module and the health state classification module, such as brain stability, left-right brain bias, and resistance to interference ability under the perspective of brain cognition, so that the result is more semantically clear and understandable; in addition, there may be complex mutual influence between different electroencephalogram characteristic values, for example, the synergistic change of PSD and DE in a specific wave band may be related to a certain specific state, and the interactive concept layer module captures these implicit feature relationships through a dynamic mapping mechanism, thereby improving the prediction ability of the model; and directly outputting the percentage characteristic value is difficult to provide effective intervention support.

[0129] For example Figure 1 As shown in the figure, the interactive concept layer module designed in this embodiment can dynamically select and adjust high-order concepts through a feedback mechanism, such as selecting the feature that has the greatest impact on the anxiety index for intervention, thereby achieving more efficient personalized health management and health state simulation; in practical applications, mental health and physical health may need to be evaluated at the same time, and by introducing a health state classification module, the high-order concept can be used as a bridge to support a multi-task learning framework, while predicting the state of multiple dimensions of psychology, physiology, and cognition.

[0130] The pre-sequencing input of the interactive concept layer module is the P score generated by the electroencephalogram feature extraction module, which first obtains low-order concepts (i.e., intermediate layer interactive concepts) through a concept bottleneck layer function, and then obtains the final model output through high-order concept mapping of the health state classification module, the process of which can be represented as follows:

[0131] y = f(g(P score , μ, c exp ))

[0132] Where f() is a high-order concept label predictor for realizing prediction classification, and can use common classifiers or network models in the prior art according to specific task requirements; μ represents an intervention mask vector; y represents a high-order concept label; and c exp represents a set of artificial intervention value vectors.

[0133] The concept bottleneck layer function includes a dynamic intervention method to assist users in preferentially selecting low-order concepts that have the greatest influence on the final prediction results for intervention, and the specific function setting is as follows:

[0134]

[0135] Where g iRepresents the concept bottleneck layer function of the i-th low-order concept; μ i Represents the intervention gate of the i-th low-level concept, which is used to control whether to use manual intervention value, and the value range is [0, 1]. When the value is 0, it means no manual intervention is required, and when the value is 1, it means full manual intervention; represents the manual intervention value of the i-th low-level concept, which is usually set by domain experts (such as doctors, testing researchers, etc.) based on experience. For example, for the concept of "anxiety level", it can be determined by domain experts based on the reference value provided by questionnaires or other comprehensive data; i (P score ) represents the low-order concept mapping function of the i-th low-order concept, that is, the mapping from input features to low-order concepts, which is usually the first layer output in the neural network. The specific expression is:

[0136] φ i (P score )=σ(W i P score +b i )

[0137] Among them, W i represents the weight matrix corresponding to the i-th low-order concept, which is used for linear transformation; b i represents the bias term corresponding to the i-th low-order concept; σ() represents the standard deviation;

[0138] Furthermore, the dynamic intervention method includes: screening out manual intervention values ​​through priority calculation to find the indicator features that contribute most to high-level concepts. The specific calculation formula includes:

[0139]

[0140] Among them, i * Indicates the number of the high-level concept selected by the strategy as the optimal intervention candidate in the current step; g(P score , μ∨1 i , c exp ) represents the bottleneck activation state after intervention; f(g(P score , μ∨1 i , c exp )) y Represents the probability of correct prediction of the updated high-order concept label y; μ∨1 i It means that based on μ, manual intervention is enabled for the i-th high-order concept; k represents the total number of high-order concepts.

[0141] Due to the above dynamic intervention method, it is necessary to traverse and simulate the manual intervention path in order to solve i *In order to reduce the computational cost and improve the effect of immediate manual intervention, this embodiment simulates the dynamic selection process through Gumbel-Softmax sampling. The process specifically includes:

[0142] First, set the initial value of the intervention mask vector μ (0) (For example, all are set to 0), update the intervention trajectory, and generate the next round of intervention distribution through the intervention strategy (calculate the intervention influence gain of each high-order concept, build the high-order concept intervention priority distribution, and perform strategy sampling on the distribution to achieve an efficient human-machine collaborative decision path with low computational cost). The intervention mask vector and model state are updated based on the distribution sampling. The specific calculation formula includes:

[0143]

[0144] in, Represents the high-order concept intervention priority distribution, which is used to construct the softmax sampling strategy distribution for all high-order concepts; It represents the intervention effect gain, that is, the change in the predicted value after the intervention of the i-th high-order concept. The specific calculation formula includes:

[0145]

[0146] Then, update the intervention mask vector. The specific expression includes:

[0147] μ (t′) =μ (t′-1) +η (t′) , η (t′) ~Categorical(ψ(c (t′-1) , μ (t′-1) ))

[0148] Among them, μ (t’) represents the intervention mask vector at step t', and μ (t’) ∈{0, 1} K , K represents the total number of low-level concepts;

[0149] In this way, the aforementioned sampling strategy replaces a full traversal, performing controlled random sampling of possible intervention paths to approximate the impact of each high-level concept intervention on the model output. This sampling method can be flexibly configured in practice based on task complexity, model structure, and system resources, supporting a variety of candidate sampling mechanisms, such as gain-weighted prioritized sampling, probability-driven concept perturbation, and lightweight path estimation based on a priori strategies. This sampling mechanism significantly reduces computational latency while ensuring intervention selection accuracy, improving the model's responsiveness in interactive decision-making scenarios.

[0150] The whole data processing procedure experiences a step-by-step mapping process of electroencephalogram signal-electroencephalogram feature-low-order concept label-high-order concept label, and the overall system is a deep learning model that can be cooperatively parameter trained and optimized, and the loss function to be optimized involves high-order task loss (used for fitting the consistency of high-order concepts), low-order concept consistency loss (used for ensuring the reliability of low-order concepts), intervention regularization loss (used for controlling the complexity of intervention and improving the friendliness and understandability of interaction) three parts, in the embodiment, the loss function is set as follows:

[0151]

[0152] Finally, the above loss terms are combined to form the final optimization target, i.e., the total loss function

[0153]

[0154] In the above formula, λ1 represents the importance weighting value of the loss; λ2 represents the importance weighting value of the loss; y ij represents the one-hot encoding of the true label of the i-th sample in the j-th class (1 represents belonging to the class, and 0 represents not belonging to the class); represents the probability value of the i-th sample being predicted by the model as the j-th class, which is usually from the softmax output; N represents the total number of training samples (i.e., batch size or total number of dataset samples).

[0155] Model optimization and transfer learning module:

[0156] The existing electroencephalogram signal analysis method faces significant data distribution difference problems in different populations and scenarios, and the traditional end-to-end modeling method has insufficient generalization ability in new tasks or new scenarios. Each time the model is transferred to a new task, it needs to be retrained, which not only increases the computational cost, but also causes uncontrollable loss of the performance of the original model, especially under the condition that the labeled sample amount of the new high-order concept task is limited, the model performance may be severely degraded. The embodiment introduces an interactive concept layer, which can constrain the model parameters of the new task transfer by using the distribution consistency of the concept layer, and realize the efficient transfer and stability improvement of the model.

[0157] Assuming that the aforementioned EEG model can process high-level concepts across multiple physiological, psychological, and cognitive dimensions, we collect a new batch of specific sample data (new sample type A labels). We hope to fine-tune the model so that it can gain predictive power for these new labels. Because high-level concepts have universal capabilities for evaluating samples, a good evaluation system should be able to output the risk probability of sample type A while also maintaining the evaluation capabilities of the original high-level concepts.

[0158] Specifically, the method of this embodiment is to fix the parameters of the model g(*) before the concept bottleneck layer, and only adjust the parameters in the f(*) model during the new task training (transfer) process, while using the concept alignment loss to ensure the transfer effect. The calculation methods include:

[0159]

[0160] in, is the distribution of the i-th low-level concept in the new task data; is the distribution of the i-th low-order concept in the source domain data; is the concept alignment loss function, which is used to measure the degree of difference between the concept distribution generated by the model under the new task and the concept distribution in the original task, so as to utilize the additional supervision information of the concept bottleneck layer to improve the EEG data migration effect across populations / tasks; KL (q||p) represents the calculation of the divergence (Kullback-Leibler) between q and p, which is used to measure the degree of difference between the two probability distributions. The smaller the divergence value, the more consistent the concept distribution learned by the model in the new task is with the source task. On the contrary, if the divergence is too large, it indicates that the concept semantics have shifted, which may cause the model to have problems such as decreased predictive performance or reduced interpretability in the new task. The divergence can be used as a regularization constraint term in the training process of new tasks to suppress the drift of high-order concept semantics and improve cross-task generalization performance.

[0161] In this way, the migration loss problem caused by data distribution differences can be effectively alleviated, and the adaptability and stability of EEG data across populations and scenarios can be improved. As an evaluation of the migration effect, after the model is migrated, the distribution of each concept in the new task is calculated. Distribution in the source domain The sum of KL divergences It can be used as an important quantitative indicator of the degree of semantic preservation of transfer, and together with task loss (such as classification accuracy), it constitutes a transfer evaluation standard.

[0162] Example 2:

[0163] Combine Figure 3As shown, in the analysis of the electroencephalogram signal, the analysis process can be as follows:

[0164] S1, collect the original electroencephalogram signal and pre-process to obtain the pre-processed electroencephalogram signal, the electroencephalogram signal including left frontal lobe signal Fp1 and right frontal lobe signal Fp2;

[0165] S2, feature extraction is performed on the pre-processed electroencephalogram signal to obtain symmetry features, and index features are calculated for different wave bands; through quantile mapping, the percentage feature value corresponding to the index feature is calculated;

[0166] S3, based on the interactive concept bottleneck model, the percentage feature value is mapped to a low-order concept through the concept bottleneck layer, and then the low-order concept is mapped to a high-order concept, so as to realize the recognition of the high-order concept; the concept bottleneck layer filters the index feature that contributes most to the high-order concept through a dynamic intervention mode;

[0167] S4, update the high-order concept through transfer training; in the transfer training, dynamic concept alignment is performed on the newly added high-order concept.

[0168] Further, in S1, the pre-processing includes high-pass filtering, notch filtering and artifact processing;

[0169] The notch filtering mode is:

[0170]

[0171] Where H Notch () represents the notch filtering function; γ represents the notch depth; Q represents the quality factor; f0 represents the center frequency of the notch filter; f represents the selected frequency.

[0172] Further, in S2, before the feature extraction of the pre-processed electroencephalogram signal, it further includes:

[0173] The pre-processed electroencephalogram signal is segmented according to a fixed time window, and the specific expression includes:

[0174] X Segmented ={X[:,t s :t s +W]|t s =1,2,…,T-W}

[0175] Where X[:,t s :t s +W] represents the electroencephalogram signal obtained after the electroencephalogram acquisition and pre-processing module is processed; W represents the window length; T represents the electroencephalogram signal acquisition time length;

[0176] The data of the electroencephalogram channel c in the window is normalized, and feature extraction is performed on the normalized data.

[0177] Preferably, the normalization is performed in the following manner: wherein μ c represents the mean value of the sampling points in the electroencephalogram channel c; and σ c represents the standard deviation of the sampling points in the electroencephalogram channel c, reflecting the fluctuation amplitude or energy distribution of the electroencephalogram channel signal.

[0178] In another implementation form of the present solution, the above-described steps or some of them can be implemented by means of a device, which can comprise corresponding modules for performing the above-described steps or some of them. Thus, each of the steps of the above-described embodiments or some of them can be performed by corresponding modules, and the electronic device can comprise one or more of these modules. The modules can be one or more hardware modules specially configured to perform the corresponding steps, or implemented by a processor configured to perform the corresponding steps, or stored in a computer-readable medium for implementation by a processor, or implemented by some combination.

[0179] The device can further comprise an electroencephalogram signal acquisition electrode for acquiring or receiving the electroencephalogram signal; and the device can further be provided with a display for displaying an interaction picture with the user based on the instructions in the memory, the interaction picture comprising a specific task picture to be completed by the user.

[0180] The device can be implemented by means of a bus architecture. The bus architecture can comprise any number of interconnecting buses and bridges, depending on the specific application of the hardware and on overall design constraints. The bus will connect various circuits of the one or more processors, the memory, and / or the hardware modules together.

[0181] Any process or method described in flowcharts or otherwise described herein can be understood as representing code modules, segments, or portions of code that include one or more executable instructions for implementing specific logic functions or steps, and the various embodiments of the present solution can include additional implementation in which the functions described are performed in a different order, including substantially simultaneously, or in reverse order, as will be understood by those skilled in the art of the present solution. The processors perform the various methods and processes described above. For example, the method embodiments in the present solution can be implemented as a software program that is tangibly embodied in a machine-readable medium, such as a memory. The logic and / or steps represented in the flowcharts or otherwise described herein can be embodied in any readable medium for use by, or in connection with, an instruction execution system, apparatus, or device.

[0182] 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 changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. An EEG signal analysis system based on interactive concept bottleneck, characterized by: The system includes an EEG acquisition and preprocessing module, an EEG feature extraction module, an interactive concept layer module, a health status classification module, and a model optimization and transfer learning module, which are arranged in sequence; The EEG acquisition and preprocessing module is used to acquire the original EEG signals and preprocess them to obtain preprocessed EEG signals; The EEG signals include symmetrically collected left and right EEG signals; An EEG feature extraction module is used to extract features from the preprocessed EEG signal to obtain symmetry features, and calculate index features for different bands; and calculate the percentage eigenvalues ​​corresponding to the index features through quantile mapping; The interactive concept layer module is used to map the percentage feature values ​​to low-order concepts through the concept bottleneck layer based on the interactive concept bottleneck model, and screen the indicator features that contribute most to high-order concepts through dynamic intervention; Health status classification module, used to map low-level concepts to high-level concepts; The model optimization and transfer learning module is used to update high-level concepts through transfer training; in the transfer training, dynamic concept alignment is performed on newly added high-level concepts.

2. The system according to claim 1, wherein: The left EEG signal is specifically the left frontal lobe signal Fp1, and the right EEG signal is specifically the right frontal lobe signal Fp2.

3. The system according to claim 2, characterized in that The specific calculation formula of the symmetry feature includes: Asymmetry(t)=X Fp1 (t)-X Fp2 (t); Among them, Asymmetry(t) represents the symmetry eigenvalue; X Fp1 (t) represents the value of Fp1; X Fp2 (t) represents the value of Fp2; t represents the time point; μ Asymmetry Indicates the bias towards the global average; T represents the length of EEG signal acquisition time; μ Abs_Asymmetry represents the global average intensity; σ Asymmetry represents the standard deviation of the symmetry index.

4. The system according to claim 1, wherein: The calculation method of the percentage characteristic value includes: For indicator features with sufficient sample size, calculate the cumulative distribution function F I Specific expressions of (x) include: Among them, n ≤v is the number of samples in the reference sample whose indicator characteristic value is less than or equal to v, N represents the total number of reference samples, and v represents the preset threshold of the indicator characteristic; For indicator features with insufficient sample size, the cumulative distribution function F is calculated using normal distribution. I (x), specific expressions include: Among them, μ0 represents the mean of the sampling point; σ represents the standard deviation of the sampling point; x represents the indicator characteristics of the sampling point; Based on the cumulative distribution function, calculate the percent eigenvalue P score Specific expressions of (x) include: P score (x)=F I (x)×100。 5. The system according to claim 4, characterized in that The function of the conceptual bottleneck layer is set as: Among them, g i represents the concept bottleneck layer function of the i-th low-order concept; μ represents the intervention mask vector; μ i represents the intervention gate of the i-th low-level concept; represents the manual intervention value of the i-th low-level concept; c exp represents the set vector of artificial intervention values; φ i (P score ) represents the low-order concept mapping function of the i-th low-order concept, and the specific expression is: f i (P score )=σ(W i P score +b i ) Among them, W i represents the weight matrix corresponding to the i-th low-order concept; b i represents the bias term corresponding to the i-th low-order concept; σ() represents the standard deviation.

6. The system according to claim 5, characterized in that The specific expression of the health status classification module includes: y=f(g(P score ,μ,C exp )); Among them, f() is the high-order concept label predictor; y represents the high-order concept label.

7. The system according to claim 5, characterized in that The dynamic intervention method includes: screening out the manual intervention value through priority calculation; the specific calculation formula includes: Among them, i * Indicates the number of the high-level concept selected by the strategy as the optimal intervention candidate in the current step; g(P score , μ∨1 i , c exp ) represents the bottleneck activation state after intervention; f(g(P score , μ∨1 i , c exp )) y Represents the probability of correct prediction of the updated high-order concept label y; μ∨1 i It means that based on μ, manual intervention is enabled for the i-th high-order concept; k represents the total number of high-order concepts.

8. The system according to claim 7, characterized in that The dynamic intervention method also includes: simulating a dynamic selection process through Gumbel-Softmax sampling, which specifically includes: First, set the initial value of the intervention mask vector μ (0) , update the intervention trajectory, generate the next round of intervention distribution through the intervention strategy, and update the intervention mask vector and model state according to the distribution sampling. The specific calculation formula includes: in, Represents the high-order concept intervention priority distribution, which is used to construct the softmax sampling strategy distribution for all high-order concepts; It represents the intervention impact gain. The specific calculation formula includes: Then, update the intervention mask vector. The specific expression includes: m (t′) =μ (t′-1) +n (t′) ,or (t′) ~Categorical(ψ(c (t′t-1) ,m (t′1) )); Among them, μ (t′) represents the intervention mask vector at step t', and μ (t’) ∈{0, 1} K , K represents the total number of low-order concepts.

9. The EEG signal analysis method based on interactive concept bottleneck is characterized by: include: S1, collecting original EEG signals and preprocessing them to obtain preprocessed EEG signals; The EEG signals include symmetrically collected left and right EEG signals; S2. Extracting features from the preprocessed EEG signal to obtain symmetry features, and calculating index features for different bands; calculating percentage eigenvalues ​​corresponding to the index features through quantile mapping; S3. Based on the interactive concept bottleneck model, the percentage feature values ​​are mapped to low-order concepts through the concept bottleneck layer, and then the low-order concepts are mapped to high-order concepts; the concept bottleneck layer selects the indicator features that contribute most to the high-order concepts through dynamic intervention; S4. Updating high-level concepts through transfer training; in the transfer training, dynamic concept alignment is performed on the newly added high-level concepts.

10. An EEG signal analysis device based on an interactive concept bottleneck, characterized in that The device includes EEG signal acquisition electrodes, a processor, a display and a memory; The EEG signal acquisition electrodes are used to acquire or receive EEG signals; The display is used to display an interactive screen with the user based on the instructions in the memory; The processor is capable of calling the computer instructions in the memory to implement the system according to any one of claims 1-8.

Images