A memory disorder assessment system based on electroencephalogram signals

Abstract

The application provides a memory disorder evaluation system based on an electroencephalogram, comprising: a signal acquisition module configured to acquire an electroencephalogram of a target object; a preprocessing module configured to perform a band-pass filtering of 0.5Hz-40Hz on the electroencephalogram of the target object; a signal segmentation module configured to segment each channel and determine a plurality of non-overlapping Epochs corresponding to each channel; a feature extraction module configured to perform linear feature extraction and nonlinear feature extraction on each Epoch in each channel, and determine an electroencephalogram fusion feature of the target object based on the linear feature and the nonlinear feature of each Epoch; and a memory disorder evaluation module configured to input the electroencephalogram fusion feature into a trained memory disorder evaluation model, use the memory disorder evaluation model to evaluate the memory disorder of the target object, and obtain a memory disorder evaluation result of the target object. The scheme can efficiently evaluate the memory disorder of the target object.

Description

Technical Field

[0001] This application relates to the field of electroencephalogram (EEG) signal processing technology, and more specifically, to a memory impairment assessment system based on EEG signals. Background Technology

[0002] Memory impairment is a typical symptom of neurodegenerative diseases such as Alzheimer's disease, mild cognitive impairment, and vascular dementia. Currently, standardized psychological scales remain an indispensable core tool for the diagnosis and assessment of memory impairment, especially in primary care hospitals and initial screening. Standardized psychological scales provide uniform instructions, scoring criteria, and interpretation guidelines, making assessment results comparable between different doctors and hospitals and reducing variability in subjective judgment. Screening tools include the Brief Mental Status Examination and the Montreal Cognitive Assessment, while detailed assessment scales include the Clinical Dementia Rating Scale and the Alzheimer's Disease Assessment Scale.

[0003] Scale-based assessments are cumbersome, susceptible to subjective interference, and inefficient. Electroencephalography (EEG), with its high temporal resolution (milliseconds), non-invasiveness, relatively low cost, and ease of clinical deployment, has gradually become an ideal tool for studying the dynamic neural processes of memory. Therefore, how to utilize EEG signal analysis to provide a memory impairment assessment system for efficient evaluation of target subjects is a technical problem that needs to be solved in this field. Summary of the Invention

[0004] The purpose of this application is to provide a memory impairment assessment system based on electroencephalogram (EEG) signals, so as to efficiently assess the memory impairment of the target subject.

[0005] To achieve the above objectives, the embodiments of this application are implemented in the following manner:

[0006] This application provides a memory impairment assessment system based on electroencephalogram (EEG) signals, comprising: a signal acquisition module for acquiring EEG signals of a target object, wherein the EEG signals are 64-channel high-density scalp EEG signals; a preprocessing module for preprocessing the EEG signals of the target object, wherein the data preprocessing includes bandpass filtering from 0.5Hz to 40Hz; a signal segmentation module for segmenting each channel and determining several non-overlapping epochs corresponding to each channel, wherein an epoch is a time period of uniform length; a feature extraction module for performing linear and nonlinear feature extraction on each epoch in each channel, and determining the EEG signal fusion features of the target object based on the linear and nonlinear features of each epoch, wherein the linear features include frequency domain power spectrum and the nonlinear features include multi-scale entropy; and a memory impairment assessment module for inputting the EEG signal fusion features into a trained memory impairment assessment model, using the memory impairment assessment model to assess the memory impairment of the target object, and obtaining the memory impairment assessment result of the target object.

[0007] Furthermore, the feature extraction module is specifically used for: for each epoch in each channel: calculating the power spectral density of the epoch, extracting the relative power of the epoch in each frequency band, wherein each frequency band is divided into: 1-4 Hz, 4-8 Hz, 13-30 Hz, and 30-40 Hz; for each epoch in each channel: coarsely processing the epoch to obtain coarse-grained sequences at multiple scales, and then calculating the sample entropy for each coarse-grained sequence to obtain the multi-scale entropy sequence features corresponding to the epoch; based on the relative power of each frequency band and the multi-scale entropy sequence features corresponding to each epoch in each channel, determining the EEG signal fusion features of the target object.

[0008] Furthermore, the feature extraction module is specifically used for: for each Epoch in each channel: dividing the Epoch into K data segments, where the overlap between any two adjacent data segments is 50%; adding a Hanning window to each data segment, and performing a Fast Fourier Transform on each windowed data segment to calculate the periodogram corresponding to each data segment; and calculating the mean of the periodogram corresponding to each data segment to determine the power spectral density corresponding to the Epoch.

[0009] Furthermore, the feature extraction module is specifically used to: calculate the integral of each frequency band based on the power spectral density corresponding to the Epoch, and determine the absolute power of each frequency band; sum the absolute power of all frequency bands; and calculate the relative power of each frequency band based on the absolute power of each frequency band and the sum of absolute power.

[0010] Furthermore, the feature extraction module is specifically used for: for each Epoch in each channel: constructing multiple coarse-grained sequences under different scale factors based on the original sequence of the Epoch; calculating the sample entropy for each scale of the coarse-grained sequence to obtain the multi-scale entropy sequence features corresponding to the Epoch.

[0011] Furthermore, the scale factor is taken as each positive integer from 1 to 20, for a total of 20 scale factors. The feature extraction module is specifically used for: for each epoch in each channel: determining the first feature component based on the relative power of the frequency band 1-4 Hz and the sample entropy of the multi-scale entropy sequence features with a scale factor of 16-20; determining the second feature component based on the relative power of the frequency band 4-8 Hz and the sample entropy of the multi-scale entropy sequence features with a scale factor of 5-12; determining the third feature component based on the relative power of the frequency band 13-30 Hz and the sample entropy of the multi-scale entropy sequence features with a scale factor of 2-5; determining the fourth feature component based on the relative power of the frequency band 30-40 Hz and the sample entropy of the multi-scale entropy sequence features with a scale factor of 1-3; and determining the EEG signal fusion features of the target object based on the first, second, third, and fourth feature components corresponding to each epoch in each channel.

[0012] Furthermore, the feature extraction module is specifically used for: for each channel: integrating the first feature component corresponding to each Epoch in this channel to form a first feature sequence; integrating the second feature component corresponding to each Epoch in this channel to form a second feature sequence; integrating the third feature component corresponding to each Epoch in this channel to form a third feature sequence; integrating the fourth feature component corresponding to each Epoch in this channel to form a fourth feature sequence; and organizing the first feature sequence, second feature sequence, third feature sequence, and fourth feature sequence corresponding to each channel according to the channel number to form the first input feature, second input feature, third input feature, and fourth input feature, respectively, as the EEG signal fusion features of the target object.

[0013] Furthermore, the memory impairment assessment model includes an input layer, an LSTM module, and a multilayer perceptron classification head. The multilayer perceptron classification head contains several fully connected layers, and the last fully connected layer outputs the classification probability through a softmax function, which serves as the memory impairment assessment result for the target object.

[0014] Furthermore, the LSTM module contains four parallel LSTM units: an input layer for receiving EEG signal fusion features carrying channel and frequency band identifiers and inputting the corresponding input features into the corresponding LSTM unit; each LSTM unit of the LSTM module for extracting features from the corresponding input features to obtain the temporal features of the corresponding channel and frequency band; and a multilayer perceptron classification head for performing fully connected processing on the temporal features of all channels and frequency bands, and outputting the classification probability through the softmax function in the last fully connected layer as the assessment result of the target object's memory impairment.

[0015] Furthermore, the multilayer perceptron classification head has three fully connected layers. Beneficial effects

[0016] Memory impairment is essentially a disorder of the brain's neural network function and structure. While traditional power spectrum analysis (i.e., linear features) can effectively capture synchronized intensity changes in neural oscillations across different frequency bands (e.g., Alzheimer's patients typically exhibit decreased power in posterior alpha rhythms and abnormally increased power in anterior theta rhythms), it is insufficient to capture the complexity of neural activity changes. This approach extracts both linear features (frequency domain power spectrum) and nonlinear features (multi-scale entropy) simultaneously to perform power spectrum analysis on signal characteristics across different frequency bands and calculate the complexity of time series at different scales (the normal brain generates complex and ordered electrical signals with high multi-scale entropy values), thereby measuring the efficiency of brain information processing and the integrity of the neural network. In patients with memory impairment, due to neural network degeneration and neuronal loss, their electrical activity becomes either overly regular (e.g., periodic firing) or completely random, manifested as a decrease in multi-scale entropy values. Therefore, combining power spectrum analysis with multi-scale entropy allows for the assessment of the "intensity" and "quality" of brain activity, providing a more comprehensive and robust approach for memory impairment evaluation. Furthermore, model-based identification can effectively improve the efficiency of memory impairment assessment.

[0017] This scheme associates specific frequency bands with specific scale ranges of multi-scale entropy during feature extraction. This is achieved through frequency band division (δ: 1-4Hz, θ: 4-8Hz, β: 13-30Hz, γ: 30-40Hz) and the corresponding multi-scale entropy scale ranges for each frequency band (e.g., δ band corresponds to sample entropy at scales 16-20, and γ band corresponds to sample entropy at scales 1-3). This association reflects the spatiotemporal characteristics of impaired neural oscillations at different frequency bands under pathological conditions: lower-frequency oscillations (such as δ) are typically associated with large-scale, slowly varying neural modulation and long-range communication, and their complexity requires a larger time scale (scale 16-20) to be effectively captured; while higher-frequency oscillations (such as γ) are associated with rapid information processing in local neural microcircuits, and their complexity changes at a smaller scale (scale 1-3). The pathological process of memory impairment begins in specific brain regions (such as the entorhinal cortex and hippocampus), gradually damaging different levels of neural networks. This "frequency-scale" mapping relationship allows the model to accurately pinpoint specific spatiotemporal neural activity patterns affected by pathological processes. For example, the theta rhythms generated by the hippocampus are closely related to memory encoding, and their reduced complexity (reflected in sample entropy changes at scales 5-12) may occur at various stages of memory impairment, exhibiting varying degrees at different times. Therefore, by combining power spectrum analysis of each frequency band with multi-scale entropy analysis and establishing corresponding correlations to construct fusion features, this approach can more effectively enhance the differential diagnostic ability for memory impairments and improve the reliability of memory impairment assessments.

[0018] In designing the memory impairment assessment model, to adapt to the characteristics of different frequency bands and multiple channels, a hybrid model architecture of LSTM+MLP (actually several parallel LSTM units + MLP classification head) is adopted. This cleverly addresses the processing difficulties brought about by the high-dimensional, temporal, and multi-channel characteristics of EEG signals, enabling efficient feature extraction and accurate classification from complex data (actually outputting classification probabilities). This architecture can improve feature utilization efficiency and model performance by extracting local temporal features and fusing global features for decision-making. Compared to scales that rely entirely on doctors for manual administration, scoring, and interpretation, this system can automatically complete the assessment of a single EEG recording in a short time, achieving efficient assessment and providing an effective solution for large-scale screening and long-term follow-up monitoring.

[0019] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of a memory impairment assessment system based on electroencephalogram (EEG) signals provided in an embodiment of this application.

[0022] Figure 2 This is a schematic diagram of a memory impairment assessment model. Detailed Implementation

[0023] The technical solutions in the embodiments of this application will now be described with reference to the accompanying drawings.

[0024] Please see Figure 1 , Figure 1 This is a schematic diagram of a memory impairment assessment system based on electroencephalogram (EEG) signals provided in an embodiment of this application. In this embodiment, the memory impairment assessment system includes a signal acquisition module, a preprocessing module, a signal segmentation module, a feature extraction module, and a memory impairment assessment module.

[0025] The signal acquisition module is used to acquire the EEG signal of the target object. In this embodiment, the EEG signal is a 64-channel high-density scalp EEG signal with a sampling frequency of more than 500Hz and a length of 15 minutes (or 10 minutes). In this embodiment, 2kHz is selected.

[0026] The preprocessing module is mainly used to preprocess the EEG signals of the target object. The data preprocessing mainly involves bandpass filtering from 0.5Hz to 40Hz. A power frequency notch filter can be applied before the bandpass filter (to remove the influence of the 50Hz power frequency) to avoid a very small part of the power frequency interference.

[0027] After preprocessing, the signal segmentation module can segment each channel and determine several non-overlapping epochs corresponding to each channel. Here, an epoch is a time period of uniform length. In this embodiment, 2 seconds (or 4 seconds) is used as an example. Then, the signal of each channel can be divided into 450 non-overlapping epochs, and the number of sampling points in each epoch is 4000.

[0028] After completing the Epoch division, the feature extraction module can extract linear features (such as frequency domain power spectrum) and nonlinear features (such as multi-scale entropy) for each Epoch in each channel, and determine the EEG signal fusion features of the target object based on the linear and nonlinear features of each Epoch.

[0029] In this embodiment, for each Epoch in each channel, the feature extraction module can calculate the power spectral density of the Epoch and extract the relative power of the Epoch in each frequency band, wherein each frequency band is divided into: 1-4 Hz, 4-8 Hz, 13-30 Hz, and 30-40 Hz.

[0030] For example, for each Epoch in each channel: the feature extraction module can divide the Epoch into K data segments (with 1 second as a segment, a step of 500ms, and each Epoch divided into 3 data segments; of course, it can also be 500ms as a segment, a step of 250ms, and each Epoch divided into 7 data segments; this embodiment takes the former as an example). Then, the overlap between any two adjacent data segments is 50%.

[0031] Then, the feature extraction module can add a Hanning window to each data segment (to reduce spectral leakage) and perform a fast Fourier transform on each windowed data segment to calculate the periodogram corresponding to each data segment.

[0032] Specifically, the data segment can be multiplied by a window function:

[0033] (1)

[0034] in, After adding a window to the data segment, the first The value of each sampling point, For the first data segment The value of each sampling point, , The number of sampling points in the data segment (in this embodiment) (For 2000).

[0035] Therefore, for each data segment after windowing, the feature extraction module can perform a Fast Fourier Transform (FFT) on it:

[0036] (2)

[0037] in, For the Fourier transform of the data segment, For frequency index, The imaginary unit, This represents the number of sampling points in the data segment.

[0038] Therefore, the feature extraction module can calculate the periodic graph corresponding to each data segment:

[0039] (3)

[0040] in, Data segment The corresponding periodic chart, For data segment Fourier transform, express The square of the amplitude of the FFT result represents the energy spectrum. It is the sampling frequency. This is a window function compensation factor used to correct the signal energy attenuation caused by adding the Hanning window. This represents the number of sampling points in the data segment (in this embodiment, the number of sampling points in the data segment is also 2000). Representing window functions Energy, window function .

[0041] After calculating the periodic chart corresponding to each data segment, the mean of the periodic chart corresponding to each data segment can be calculated (simply calculate the arithmetic mean):

[0042] (4)

[0043] in, This is the power spectral density estimate for the Epoch, i.e., the power spectral density corresponding to the Epoch.

[0044] Based on this, the feature extraction module can extract the relative power of Epoch in each frequency band. The frequency bands are divided into δ: 1-4Hz, θ: 4-8Hz, β: 13-30Hz, and γ: 30-40Hz. Each frequency band interval uses a left-closed and right-open interval when taking values.

[0045] Therefore, the feature extraction module can calculate the integral for each frequency band based on the power spectral density corresponding to the Epoch:

[0046] (5)

[0047] (6)

[0048] (7)

[0049] (8)

[0050] in, , , and These represent the absolute power for the frequency bands 1-4Hz, 4-8Hz, 13-30Hz, and 30-40Hz, respectively.

[0051] Furthermore, the sum of the absolute power of all frequency bands corresponding to the Epoch can be added together (since the 8-13Hz alpha wave is mainly related to relaxation, closed eyes, and awake but no task when assessing memory impairment in this scheme, its impact on memory impairment analysis is relatively small and indirect, so it was not selected. However, this frequency band can also be used for analysis in other embodiments. This embodiment provides a better implementation scheme, so the analysis of the frequency band where the alpha wave is located is not selected, but it should not be regarded as a limitation of this application. If it is to be introduced, the calculation method is similar to the waveform of other frequency bands):

[0052] (9)

[0053] in, It is the sum of the absolute power of all frequency bands corresponding to the Epoch (actually the sum of the absolute power of the entire frequency band, which is more than the number of frequency bands divided, such as the 0.5-1Hz band and the 8-13Hz band).

[0054] After that, it can be done through , , , Calculate the relative power for each frequency band.

[0055] In this way, the feature extraction module can calculate the relative power for each channel at each epoch. , , , .

[0056] In this embodiment, the feature extraction module can also perform nonlinear feature extraction for each Epoch in each channel.

[0057] First, the feature extraction module can coarse-grain the Epoch to obtain coarse-grained sequences at multiple scales.

[0058] For each epoch in each channel: the feature extraction module can construct multiple coarse-grained sequences at different scale factors based on the original sequence of the epoch.

[0059] Original sequence ,in, For the first epoch The signal value at each sampling point This represents the total number of sampling points in the Epoch, as described in this embodiment. Take 4000.

[0060] For the original sequence, different scale factors are constructed using non-overlapping arithmetic means. The new sequence below For scale factor New sequence of time The Middle There are several parameters:

[0061] (10)

[0062] in, scale factor New sequence of time The first in One parameter, , where is the scale factor and is a positive integer. The value satisfies ,certainly, That is, the original sequence. Indicates to Round down to the nearest integer.

[0063] Therefore, the feature extraction module can construct multiple (e.g., 20) coarse-grained sequences at different scale factors based on the original sequence of the Epoch. :

[0064] (11)

[0065] Accordingly, the feature extraction module can calculate the sample entropy for each scale of coarse-grained sequence. Before the calculation, it is necessary to obtain the pre-set embedding dimension m (the length of the vector used for comparison, where m is a general value of 2) and similarity tolerance r (the threshold used to determine whether two data points are similar, where r is traditionally between 0.15σ and 0.25σ, and in this embodiment r is 0.2σ, where σ is the standard deviation of the original sequence).

[0066] Therefore, the feature extraction module can extract features from the currently processed coarse-grained sequence (e.g., a coarse-grained sequence with a scale factor of 2). Construct all vectors of length m in the given array:

[0067] (12)

[0068] A total of There are m vectors (the maximum number of vectors here is...). This is because a vector of length m+1 will be constructed subsequently to maintain a consistent number of vectors. Then, the distance between each vector and other vectors is calculated (using Chebyshev distance, which measures the maximum coordinate difference). The number of vectors whose distance is less than the similarity tolerance is determined, resulting in a total of... There are 10 data points, and the value of each data point will not exceed 100. When the scale factor is 2, due to The number of vectors with length m=2 is 1998, and the value of each data point will not exceed 1997.

[0069] The matching probability is then calculated using the following method:

[0070] (13)

[0071] in, This represents the probability that a vector of length m will match in a coarse-grained sequence under a tolerance of r. This represents the number of vectors whose distance to the i-th vector of length m is less than the similarity tolerance r.

[0072] In addition, the feature extraction module needs to extract features from the currently processed coarse-grained sequence (e.g., a coarse-grained sequence with a scale factor of 2). Construct all vectors of length m+1 in the given array:

[0073] (14)

[0074] A total of Given a vector of length m+1, calculate the distance between each vector and all other vectors (using Chebyshev distance, measuring the maximum coordinate difference). Determine the number of vectors whose distance is less than the similarity tolerance. Similarly, obtain... There are 10 data points, and the value of each data point will not exceed 100. When the scale factor is 2, due to The number of vectors with length m+1=3 is also 1998, and the value of each data point will not exceed 1997.

[0075] The matching probability is then calculated using the following method:

[0076] (15)

[0077] in, This represents the probability that a vector of length m+1 matches in the coarse-grained sequence under a tolerance of r. In order to be with the first The number of vectors of length m+1 whose distance is less than the similarity tolerance r.

[0078] The sample entropy is then calculated using the following method:

[0079] (16)

[0080] in, is the sample entropy corresponding to the coarse-grained sequence of the Epoch at a scale factor of .

[0081] Accordingly, the feature extraction module can calculate the sample entropy corresponding to each scale factor in this Epoch, thus obtaining a multi-scale entropy sequence feature of length 20 (number of scale factors). Then, the feature extraction module can calculate the multi-scale entropy sequence feature for each channel in each Epoch. Since this process involves a large amount of computation, mainly concentrated on sample entropy, the core of which is a double loop used to compare all possible vector pairs. This can be accelerated using optimized NumPy vectorization or GPU acceleration, thereby controlling the multi-scale entropy calculation time for each Epoch to the millisecond level. Without GPU acceleration, the calculation time for each Epoch is expected to be within 500 milliseconds (between 100 and 500 milliseconds). Without considering parallel processing of channels and Epochs, the analysis of a single target object is expected to take significantly longer, resulting in low processing efficiency. Further optimization of some features can be considered to improve system efficiency, but this embodiment will not elaborate on this.

[0082] After extracting linear features (power spectral density) and nonlinear features (multi-scale entropy), the feature extraction module can determine the EEG signal fusion features of the target object based on the relative power and multi-scale entropy sequence features of each frequency band corresponding to each Epoch in each channel.

[0083] Because low-scale (τ=1-5) sample entropy reflects rapid, localized neural microcircuit activity, with a timescale of tens to hundreds of milliseconds, it is closely related to the synchronization and desynchronization of high-frequency oscillations such as gamma waves (corresponding to the frequency band 30-40Hz) and beta waves (corresponding to the frequency band 13-30Hz). Mid-to-high-scale (τ=5-12) sample entropy reflects synchronous oscillations at the neural cluster level, with a timescale of hundreds of milliseconds to seconds, and is directly related to the generation and modulation of theta wave rhythms (corresponding to the frequency band 4-8Hz). High-scale (τ>15) sample entropy reflects slower, global brain state fluctuations, with a timescale of seconds, and is closely related to slow cortical potentials, delta waves (corresponding to the frequency band 1-4Hz), and the slow dynamic fluctuations of large-scale brain networks. Therefore, this scheme establishes a connection between power spectral density and multi-scale entropy, achieving deep feature-level fusion.

[0084] For each Epoch in each channel:

[0085] The feature extraction module can determine the first feature component based on the relative power of the frequency band 1-4 Hz and the sample entropy with a scale factor of 16-20 in the multi-scale entropy sequence features. For example, the relative power of the delta wave (1-4 Hz frequency band) and the sample entropy with scale factors of 16, 17, 18, 19, and 20 are calculated as follows:

[0086] (17)

[0087] Where is the fused eigenvalue of relative power at the scale factor. Then, the first eigencomponent is [ , , , , ].

[0088] Similarly, the feature extraction module can determine the second feature component based on the relative power in the 4-8 Hz frequency band and the sample entropy with a scale factor of 5-12 in the multi-scale entropy sequence features. For example, the relative power of the theta wave (4-8 Hz frequency band) can be used to determine the second feature component. The entropy of samples with scale factors of 5-12 is calculated as follows:

[0089] (18)

[0090] in, relative power In scale factor The fusion feature value at time. Then, the second feature component is [ , , , , , , , ].

[0091] Similarly, the feature extraction module can determine the third feature component based on the relative power in the 13-30Hz frequency band and the sample entropy with a scale factor of 2-5 in the multi-scale entropy sequence features. For example, the relative power of the β wave (13-30Hz frequency band) can be used to determine the third feature component. The entropy of samples with a scale factor of 2-5 is calculated as follows:

[0092] (19)

[0093] in, relative power In scale factor The fusion feature value at time. Then, the third feature component is [ , , , ].

[0094] Furthermore, the feature extraction module can determine the fourth feature component based on the relative power in the 30-40Hz frequency band and the sample entropy with a scale factor of 1-3 in the multi-scale entropy sequence features. For example, the relative power of the γ wave (13-30Hz frequency band) can be used to determine the fourth feature component. The entropy of samples with scale factors of 1-3 is calculated as follows:

[0095] (20)

[0096] in, relative power In scale factor The fusion feature value at time. Then, the fourth feature component is [ , , ].

[0097] Subsequently, the feature extraction module can determine the EEG signal fusion features of the target object based on the first, second, third, and fourth feature components corresponding to each Epoch in each channel. The first, second, third, and fourth feature components corresponding to each Epoch are associated with the Epoch number.

[0098] For each channel: the feature extraction module can integrate the first feature component corresponding to each Epoch in this channel to form the first feature sequence.

[0099] Specifically, the feature extraction module can sequentially integrate the first feature components belonging to the same channel according to the channel number (each channel has a channel identifier) ​​and the Epoch sequence number, forming 450 channel feature sequences of length 5 (here, 5 is the number of feature values ​​of the first feature component, and 450 is 450 Epochs divided into one channel), which serve as the first feature sequence. The first feature sequence carries the corresponding frequency band identifier (which can be represented by frequency band or number, such as 1-4 Hz, or δ).

[0100] Similarly, for each channel: the feature extraction module can integrate the second feature components corresponding to each Epoch in this channel to form a second feature sequence.

[0101] Specifically, the feature extraction module can sequentially integrate the second feature components belonging to the same channel according to the channel number (each channel has a channel identifier) ​​and the Epoch sequence number, forming 450 channel feature sequences of length 8 (here, 8 is the number of feature values ​​of the second feature component, and 450 is 450 Epochs divided into one channel), which serve as the second feature sequences. The second feature sequences carry the corresponding frequency band identifier (which can be represented by frequency band or number, such as 4-8Hz, or θ).

[0102] Similarly, for each channel: the feature extraction module can integrate the third feature component corresponding to each Epoch in this channel to form a third feature sequence.

[0103] Specifically, the feature extraction module can sequentially integrate the third feature components belonging to the same channel according to the channel number (each channel has a channel identifier) ​​and the Epoch sequence number, forming 450 channel feature sequences of length 4 (where 4 is the number of feature values ​​of the third feature component, and 450 is 450 Epochs divided into one channel), which serve as the third feature sequences. The third feature sequences carry the corresponding frequency band identifier (which can be represented by frequency band or number, such as 13-30Hz, or β).

[0104] Furthermore, for each channel: the feature extraction module can integrate the fourth feature component corresponding to each Epoch in this channel to form a fourth feature sequence.

[0105] Specifically, the feature extraction module can sequentially integrate the fourth feature components belonging to the same channel according to the channel number (each channel has a channel identifier) ​​and the Epoch sequence number, forming 450 channel feature sequences of length 3 (here, 3 is the number of feature values ​​of the fourth feature component, and 450 is 450 Epochs divided into one channel), which serve as the fourth feature sequences. The fourth feature sequences carry the corresponding frequency band identifier (which can be represented by frequency band or number, such as 30-40Hz, or γ).

[0106] After integrating the sequence features of each channel, the first feature sequence, second feature sequence, third feature sequence, and fourth feature sequence of each channel can be associated with the channel identifier according to the channel number (channel identifier, a total of 64). The first feature sequence, second feature sequence, third feature sequence, and fourth feature sequence corresponding to each channel are organized to form the first input feature, second input feature, third input feature, and fourth input feature, respectively, which serve as the EEG signal fusion features of the target object.

[0107] In this embodiment, the memory impairment assessment module can input the fusion features of EEG signals into a trained memory impairment assessment model, and use the memory impairment assessment model to assess the memory impairment of the target object, thereby obtaining the memory impairment assessment result of the target object.

[0108] like Figure 2As shown, in this embodiment, the memory impairment assessment model includes an input layer, an LSTM module, and a multilayer perceptron classification head. The LSTM module contains four parallel LSTM units (e.g., denoted as the first LSTM unit, the second LSTM unit, the third LSTM unit, and the fourth LSTM unit. The first LSTM unit processes the first input feature corresponding to the delta wave at 1-4 Hz, the second LSTM unit processes the second input feature corresponding to the theta wave at 4-8 Hz, the third LSTM unit processes the third input feature corresponding to the beta wave at 13-30 Hz, and the fourth LSTM unit processes the fourth input feature corresponding to the gamma wave at 30-40 Hz). The multilayer perceptron classification head contains several fully connected layers (this embodiment uses three layers as an example; other embodiments may increase or decrease the number of fully connected layers as needed, but the minimum number must be two). The last fully connected layer outputs the classification probability through a softmax function, which serves as the memory impairment assessment result for the target object.

[0109] The input layer receives EEG signal fusion features carrying channel and frequency band identifiers, and inputs the corresponding input features into the corresponding LSTM units according to the frequency band identifier. For example, the EEG signal fusion features carrying channel and frequency band identifiers are organized, and the first feature sequence of the same channel number is input into the corresponding first LSTM unit, the second feature sequence into the corresponding second LSTM unit, the third feature sequence into the corresponding third LSTM unit, and the fourth feature sequence into the corresponding fourth LSTM unit.

[0110] The first LSTM unit is processed into a feature vector with a receive length of 450 (450 epochs divided in the same channel, the data receive length of the first LSTM unit can be adjusted according to the number of epochs) × 5 (the number of eigenvalues ​​of the first feature component) for feature extraction. The second LSTM unit is processed into a feature vector with a receive length of 450 (450 epochs divided in the same channel, the data receive length of the second LSTM unit can be adjusted according to the number of epochs) × 8 (the number of eigenvalues ​​of the second feature component) for feature extraction. The third LSTM unit is processed into a feature vector with a receive length of 450 (450 epochs divided in the same channel, the data receive length of the third LSTM unit can be adjusted according to the number of epochs) × 4 (the number of eigenvalues ​​of the third feature component) for feature extraction. The fourth LSTM unit is processed into a feature vector with a receive length of 450 (450 epochs divided in the same channel, the data receive length of the fourth LSTM unit can be adjusted according to the number of epochs) × 3 (the number of feature values ​​of the fourth feature component) (i.e., processing the fourth feature component over 450 time steps) for feature extraction.

[0111] After the first LSTM unit completes feature extraction for one channel, it stores the hidden state (e.g., 128-dimensional parameters) of the last time step of the first LSTM unit, associates it with the channel identifier, and then stores it for later use. Similarly, the other LSTM units (i.e., the second LSTM unit, the third LSTM unit, and the fourth LSTM unit) perform the same operation, taking the hidden state (also 128-dimensional parameters) of the last time step, associating it with the channel identifier, and then storing it for later use.

[0112] After processing the 64 channels, 4×64 hidden layer parameters are obtained. According to the channel identifier and frequency band identifier, the hidden layer parameters are concatenated in a conventional way (e.g., from small to large channel number, and according to the frequency band identifier order of δ, θ, β and γ for the same channel number) to obtain a one-dimensional vector with a length dimension of 4×64×128, which is used as the input of the multilayer perceptron classification head.

[0113] The multilayer perceptron classification head can perform fully connected processing through its several fully connected layers. Its last fully connected layer outputs the classification probability through a softmax function, which serves as the assessment result of the target object's memory impairment.

[0114] Construction of the Memory Impairment Assessment Model: After establishing the model framework, the sample set required for model training is constructed using normal samples (several sets of EEG data from subjects without memory impairment, such as three sets of EEG data collected at different time periods, labeled as "no memory impairment") and memory impairment samples (several sets of EEG data from subjects diagnosed with memory impairment through comprehensive assessments such as the Clinical Dementia Rating Scale and the Alzheimer's Disease Assessment Scale, such as three sets of EEG data collected at different time periods, labeled as "memory impairment"). The selection of memory impairment samples should be as diverse as possible to cover different degrees of memory impairment. The model is trained using the sample set (split into training, testing, and validation sets in a 7:2:1 ratio). Once training is complete, the memory impairment assessment model is obtained.

[0115] In summary, this application provides a memory impairment assessment system based on electroencephalogram (EEG) signals. Since memory impairment is essentially a disorder of the brain's neural network function and structure, traditional power spectrum analysis (i.e., linear features), while effectively capturing the synchronization intensity changes of neural oscillations at different frequency bands (e.g., Alzheimer's patients typically exhibit decreased power in the posterior alpha rhythm and abnormally increased power in the anterior theta rhythm), struggles to capture the complexity of neural activity changes solely through power spectrum analysis. This approach simultaneously extracts linear features (frequency domain power spectrum) and nonlinear features (multi-scale entropy) to perform power spectrum analysis on signal characteristics at different frequency bands. It also calculates the complexity of time series at different scales (a normal brain can generate complex and ordered electrical signals with high multi-scale entropy values), thereby measuring the efficiency of brain information processing and the integrity of the neural network. In patients with memory impairment, due to neural network degeneration and neuron loss, their electrical activity becomes either overly regular (e.g., periodic firing) or completely random, manifesting as a decrease in multi-scale entropy values. Therefore, combining power spectrum with multi-scale entropy can assess the "intensity" and "quality" of brain activity, providing a more comprehensive and robust approach for memory disorder assessment. Furthermore, model-based identification can effectively improve the efficiency of memory disorder assessment.

[0116] This scheme associates specific frequency bands with specific scale ranges of multi-scale entropy during feature extraction. This is achieved through frequency band division (δ: 1-4Hz, θ: 4-8Hz, β: 13-30Hz, γ: 30-40Hz) and the corresponding multi-scale entropy scale ranges for each frequency band (e.g., δ band corresponds to sample entropy at scales 16-20, and γ band corresponds to sample entropy at scales 1-3). This association reflects the spatiotemporal characteristics of impaired neural oscillations at different frequency bands under pathological conditions: lower-frequency oscillations (such as δ) are typically associated with large-scale, slowly varying neural modulation and long-range communication, and their complexity requires a larger time scale (scale 16-20) to be effectively captured; while higher-frequency oscillations (such as γ) are associated with rapid information processing in local neural microcircuits, and their complexity changes at a smaller scale (scale 1-3). The pathological process of memory impairment begins in specific brain regions (such as the entorhinal cortex and hippocampus), gradually damaging different levels of neural networks. This "frequency-scale" mapping relationship allows the model to accurately pinpoint specific spatiotemporal neural activity patterns affected by pathological processes. For example, the theta rhythms generated by the hippocampus are closely related to memory encoding, and their reduced complexity (reflected in sample entropy changes at scales 5-12) may occur at various stages of memory impairment, exhibiting varying degrees at different times. Therefore, by combining power spectrum analysis of each frequency band with multi-scale entropy analysis and establishing corresponding correlations to construct fusion features, this approach can more effectively enhance the differential diagnostic ability for memory impairments and improve the reliability of memory impairment assessments.

[0117] In designing the memory impairment assessment model, to adapt to the characteristics of different frequency bands and multiple channels, a hybrid model architecture of LSTM+MLP (actually several parallel LSTM units + MLP classification head) is adopted. This cleverly addresses the processing difficulties brought about by the high-dimensional, temporal, and multi-channel characteristics of EEG signals, enabling efficient feature extraction and accurate classification from complex data (actually outputting classification probabilities). This architecture can improve feature utilization efficiency and model performance by extracting local temporal features and fusing global features for decision-making. Compared to scales that rely entirely on doctors for manual administration, scoring, and interpretation, this system can automatically complete the assessment of a single EEG recording in a short time, achieving efficient assessment and providing an effective solution for large-scale screening and long-term follow-up monitoring.

[0118] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A memory impairment assessment system based on electroencephalogram (EEG) signals, characterized in that, include: The signal acquisition module is used to acquire the electroencephalogram (EEG) signals of the target object, wherein the EEG signals are 64-channel high-density scalp EEG signals; The preprocessing module is used to preprocess the EEG signals of the target object, including bandpass filtering from 0.5Hz to 40Hz. The signal segmentation module is used to segment each channel and determine several non-overlapping Epochs corresponding to each channel, where each Epoch is a time period of uniform length. The feature extraction module is used to perform linear and nonlinear feature extraction for each epoch in each channel, and to determine the EEG signal fusion features of the target object based on the linear and nonlinear features of each epoch. The linear features include frequency domain power spectrum, and the nonlinear features include multi-scale entropy. The memory impairment assessment module is used to input the fusion features of EEG signals into a trained memory impairment assessment model, and then use the memory impairment assessment model to assess the memory impairment of the target object and obtain the memory impairment assessment results of the target object.

2. The memory impairment assessment system based on the time-frequency characteristics and multi-scale entropy of electroencephalogram (EEG) signals according to claim 1, characterized in that, The feature extraction module is specifically used for: For each epoch in each channel: calculate the power spectral density of the epoch, and extract the relative power of the epoch in each frequency band, wherein each frequency band is divided into: 1-4 Hz, 4-8 Hz, 13-30 Hz, and 30-40 Hz; For each Epoch in each channel: coarse-grained processing of the Epoch is performed to obtain coarse-grained sequences at multiple scales, and then the sample entropy is calculated for each coarse-grained sequence to obtain the multi-scale entropy sequence features corresponding to the Epoch; Based on the relative power and multi-scale entropy sequence features of each frequency band corresponding to each Epoch in each channel, the EEG signal fusion features of the target object are determined.

3. The memory impairment assessment system based on the time-frequency characteristics and multi-scale entropy of electroencephalogram (EEG) signals according to claim 2, characterized in that, The feature extraction module is specifically used for: For each Epoch in each channel: The Epoch is divided into K data segments, where the overlap between any two adjacent data segments is 50%. Add a Hanning window to each data segment, and perform a fast Fourier transform on each windowed data segment to calculate the periodogram corresponding to each data segment; The power spectral density corresponding to each epoch is determined by averaging the periodograms corresponding to each data segment.

4. The memory impairment assessment system based on time-frequency characteristics and multi-scale entropy of electroencephalogram (EEG) signals according to claim 3, characterized in that, The feature extraction module is specifically used for: Based on the power spectral density corresponding to the Epoch, the integral of each frequency band is calculated to determine the absolute power of each frequency band. Sum the absolute power of all frequency bands; The relative power of each frequency band is calculated based on the absolute power of each band and the sum of the absolute powers.

5. The memory impairment assessment system based on the time-frequency characteristics and multi-scale entropy of electroencephalogram (EEG) signals according to claim 2, characterized in that, The feature extraction module is specifically used for: For each Epoch in each channel: Multiple coarse-grained sequences with different scale factors were constructed based on the original Epoch sequences. The sample entropy is calculated for each coarse-grained sequence at each scale to obtain the multi-scale entropy sequence features corresponding to the Epoch.

6. The memory impairment assessment system based on the time-frequency characteristics and multi-scale entropy of electroencephalogram (EEG) signals according to claim 2, characterized in that, The scaling factor takes any positive integer from 1 to 20, for a total of 20 scaling factors. The feature extraction module is specifically used for: For each Epoch in each channel: The first feature component was determined based on the relative power in the frequency band 1-4 Hz and the sample entropy with a scale factor of 16-20 in the multi-scale entropy sequence features. The second feature component was determined based on the relative power in the frequency band 4-8 Hz and the sample entropy of the multi-scale entropy sequence features with a scale factor of 5-12. Based on the relative power in the frequency band 13-30 Hz and the sample entropy with a scale factor of 2-5 in the multi-scale entropy sequence characteristics, the third feature component was determined. The fourth feature component was determined based on the relative power in the 30-40 Hz frequency band and the sample entropy with a scale factor of 1-3 in the multi-scale entropy sequence features. Based on the first feature component, second feature component, third feature component, and fourth feature component corresponding to each Epoch in each channel, the EEG signal fusion features of the target object are determined.

7. The memory impairment assessment system based on time-frequency characteristics and multi-scale entropy of electroencephalogram (EEG) signals according to claim 6, characterized in that, The feature extraction module is specifically used for: For each channel: Integrate the first feature components corresponding to each Epoch in this channel to form the first feature sequence; Integrate the second feature components corresponding to each Epoch in this channel to form the second feature sequence; Integrate the third feature components corresponding to each Epoch in this channel to form a third feature sequence; Integrate the fourth feature component corresponding to each Epoch in this channel to form the fourth feature sequence; According to the channel number, the first feature sequence, second feature sequence, third feature sequence, and fourth feature sequence corresponding to each channel are organized to form the first input feature, second input feature, third input feature, and fourth input feature, respectively, which serve as the EEG signal fusion features of the target object.

8. The memory impairment assessment system based on the time-frequency characteristics and multi-scale entropy of electroencephalogram (EEG) signals according to claim 7, characterized in that, The memory impairment assessment model includes an input layer, an LSTM module, and a multilayer perceptron classification head. The multilayer perceptron classification head contains several fully connected layers, and the last fully connected layer outputs the classification probability through a softmax function, which serves as the memory impairment assessment result for the target object.

9. The memory impairment assessment system based on time-frequency characteristics and multi-scale entropy of electroencephalogram (EEG) signals according to claim 8, characterized in that, The LSTM module contains four LSTM units arranged in parallel. The input layer is used to receive EEG signal fusion features carrying channel identifiers and frequency band identifiers, and input the corresponding input features into the corresponding LSTM unit; Each LSTM unit in the LSTM module is used to extract features from the corresponding input features to obtain the time-series features of the corresponding channel and frequency band. The multilayer perceptron classification head is used to perform fully connected processing on the temporal features of all channels and frequency bands, and outputs the classification probability through the softmax function in the last fully connected layer as the assessment result of the memory impairment of the target object.

10. The memory impairment assessment system based on time-frequency characteristics and multi-scale entropy of electroencephalogram (EEG) signals according to claim 9, characterized in that, The multilayer perceptron classification head has 3 fully connected layers.

Images