A brain age prediction method based on cross-electrode representation learning of neural spectral information

Abstract

The present invention discloses a brain age prediction method based on cross-electrode representation learning of neural spectrum information, comprising: preprocessing EEG data of each electrode channel in a subject's multi-electrode channels to obtain a power spectrum density PSD at a frequency of 1-20 Hz; using the PSD at a frequency of 1-20 Hz to calculate the EEG neural spectrum features of each electrode channel, including a non-periodic parameter AF, a periodic parameter PF, a power ratio PR, a relative energy RE and a kurtosis KUR; using the above-mentioned EEG neural spectrum features of each electrode channel as the spectrum information independently represented by a single electrode itself, and characterizing the cross-electrode mutual representation similarity relationship matrix of all electrode neural spectrum features through a sparse linear regression method; finally, constructing a fully connected multi-layer neural network, learning the mapping relationship between the subject's single-electrode neural spectrum features and the similarity matrix of the neural spectrum features between electrodes and the brain age, and predicting the subject's brain age based on the above-mentioned cross-electrode neural spectrum features and the neural network.

Description

Technical Field

[0001] The invention relates to a method for predicting brain age. Background Art

[0002] Brain Age (BA) refers to the age that is commensurate with the characteristics of an individual's brain structure, function, and metabolism, and is one of the important indicators of brain health. As one of the most complex organs in the human body, the brain's neural activity shows significant changes with age. These changes not only reflect the development and aging process of neural structure and function, but also directly affect the individual's cognition, emotion, and behavior.

[0003] Changes in brain function are closely associated with neural oscillations (periodic brain activity), which are an important and widely studied feature of EEG recordings, driven by fluctuations in the excitability of neuronal groups and have complex spatiotemporal patterns. However, EEG activity consists of a mixture of periodic and non-periodic signals, and if the frequency band power measurement is considered to directly reflect the periodic neural activity pattern, the effect of non-periodic activity on the power spectrum will be ignored. Even if there is no oscillatory activity, spectral analysis will show power within the frequency band, which is completely driven by non-periodic signals rather than any oscillatory activity. The spectrum of a non-periodic signal consists of two parameters: the exponent (spectral slope) and the offset. The exponent represents the power pattern across frequencies and reflects the steepness of the power spectrum decay, while the offset reflects the uniform displacement of power across frequencies. Existing studies have shown that these parameters change dynamically with age.

[0004] Therefore, integrating non-periodic and periodic neural activity characteristics helps to accurately assess brain age and its changes, which is of great significance for understanding the laws of brain development, preventing neurodegenerative diseases, and evaluating treatment effects.

[0005] According to the study by Dimitriadis and Salis et al. (Dimitriadis, SI, Salis, CI, 2017. Mining Time-Resolved Functional Brain Graphs to an EEG-Based Chronnectomic Brain Aged Index (CBAI). Front. Hum. Neurosci. 11, 423. https: / / doi.org / 10.3389 / fnhum.2017.00423), under resting conditions, reproducible accelerated patterns of brain aging can be observed in different frequency bands, which indicates the importance of intrinsic brain oscillations. They used feature selection techniques and extreme learning machine (ELM) classifiers to distinguish between young adults and middle-aged people, and used support vector regressors to establish a linear model of actual age based on the spatiotemporal features of EEG. They applied it to 194 resting EEGs of adults aged 19-67 years old using a single integrated dynamic functional connectivity map to predict brain age. The prediction process is divided into eyes open and eyes closed. In the eyes open state, R 2 is 0.6, R 2 It is 0.48.

[0006] According to the study of Al Zoubi et al. (Al Zoubi, O., Ki Wong, 2018. Predicting Age From Brain EEG Signals—A Machine Learning Approach. Front. Aging Neurosci. 10, 184. https: / / doi.org / 10.3389 / fnagi.2018.00184), the brain age of 468 subjects (average age 34.3 years old) was predicted using 5 sets of cross-channel and cross-band preprocessed EEG features combined with nested cross-validation and stacked ensemble learning methods. The result was a MAE of 6.87 and R 2 It is 0.37.

[0007] According to the study of Vandenbosch et al. (Vandenbosch, MMLJZ, Van 'T Ent, 2019. EEG-based age-prediction models as stable and heritable indicators of brain maturational level in children and adolescents. Human Brain Mapping 40, 1919–1926. https: / / doi.org / 10.1002 / hbm.24501), 1-24 Hz power was used as input features to predict the brain age of 702 underage subjects aged 5-18 years old using random forest (RF), support vector machine (SVM) and relevance vector machine (RVM) regression analysis, with MAE of 1.22, 1.46 and R 2 Are 0.547 and 0.448.

[0008] According to the study of Khayretdinova et al. (Khayretdinova, M., 2022. Predicting agefrom resting-state scalp EEG signals with deep convolutional neural networkson TD-brain dataset. Front. Aging Neurosci. 14, 1019869. https: / / doi.org / 10.3389 / fnagi.2022.1019869), data augmentation and deep convolutional neural network models were used to predict the resting EEG (including eyes open and eyes closed) age of 1,274 subjects aged 5-88 years. The participants included healthy controls and individuals with various mental illnesses. The mean absolute error of cross-validation was 5.96 years.

[0009] The difficulties of existing EEG (electroencephalogram) brain age estimation research are mainly reflected in the following aspects: 1. The key role of EEG non-periodic activity in age prediction is often overlooked. If non-periodic signals are not explicitly measured and controlled, it is impossible to determine whether the power change within the frequency band is caused by changes in oscillation bursts, changes in non-periodic signals, or both. 2. There is less attention paid to the similarity of neural activity characteristics between electrodes. Traditional EEG feature extraction focuses more on single-node spectra or multi-node functional connection patterns, and lacks understanding of the representation correlation of neural spectrum information of head electrode arrays. 3. The feature selection and model building process in deep learning is relatively complex, requiring a lot of computing resources, difficult to apply in clinical environments, and poorly interpretable. 4. The accuracy of the prediction model is not high.

[0010] Therefore, developing a prediction framework with better performance to accurately predict brain age remains a challenging task. To overcome this challenge, we first consider the problem of inaccurate spectral feature extraction caused by the mutual aliasing and masking of non-periodic and periodic spectral components in the passband. We use a spectral separation model to estimate the non-periodic and periodic features for the non-periodic and periodic components, as well as traditional spectral features such as power ratio, relative energy and band kurtosis, to obtain the complete spectral information of the single electrode itself. We use a sparse linear regression method to quantify the mutual representation relationship between EEG neural spectrum features between electrodes, so as to obtain the electrode spatial correlation matrix of neural spectrum features based on the full integration of EEG non-periodic and periodic feature information. Subsequently, a fully connected neural network is used to predict brain age. This method can provide a deeper understanding of the representational association of neural spectrum information between brain electrodes and explore its key role in improving the accuracy of EEG age prediction. Summary of the invention

[0011] The purpose of the present invention is to provide a brain age prediction method based on cross-electrode representation learning of neural spectrum information to improve the accuracy of EEG brain age estimation and reduce the resource requirements of model calculation.

[0012] To this end, the present invention provides a brain age prediction method based on cross-electrode representation learning of neural spectrum information, including: S1. Preprocessing the EEG data of each electrode channel in the subject's multi-electrode channels to obtain the power spectrum density PSD at a frequency of 1-20Hz; S2. Using the power spectrum density PSD at a frequency of 1-20Hz to calculate the EEG neural spectrum characteristics of each electrode channel, including non-periodic parameter AF, periodic parameter PF, power ratio PR, relative energy RE and kurtosis KUR; S3. Using the above-mentioned EEG neural spectrum characteristics of each electrode channel as electrode feature vectors, and quantifying the cross-electrode representation relationship of each electrode feature vector through a sparse linear regression method, to obtain a feature representation matrix of neural activity similarity between electrodes; S4. Constructing a fully connected multi-layer neural network to learn the relationship between the subject's neural spectrum characteristics and the spatial correlation matrix of the cross-electrode mutual representation of the neural spectrum characteristics and brain age, and finally obtaining the corresponding brain age based on the above-mentioned characteristics.

[0013] Compared with the prior art, the present invention has the following technical features / effects:

[0014] 1. Accurate characterization of single-electrode spectrum characteristics: The present invention emphasizes first eliminating the convex effect of non-periodic spectrum components on the characteristics of neural oscillation spectrum components. The latest neuroscience research believes that non-periodic spectrum characteristics, as the dominant key information of the neural spectrum, are of great relevance to brain development, aging, sleep, excitation-inhibition balance, etc. After separating the non-periodic and periodic spectrum components, the non-periodic and periodic spectrum characteristics, as well as the traditional relative energy, inter-band power ratio, band spectrum kurtosis, etc. are estimated, thereby eliminating the mutual aliasing and masking phenomenon between non-periodic and periodic components, and achieving more accurate spectrum characteristic characterization.

[0015] 2. Strengthen the consideration of the similarity of neural spectrum features across electrodes: Traditional neuroscience theories often only focus on the activities of single neurons or groups of neurons, while the present invention emphasizes the importance of the similarity of neural activity features across electrodes. Through the sparse linear regression method, the present invention can fully integrate and utilize this information.

[0016] 3. Optimize the feature selection and model building process: Deep learning is complex in feature selection and model building, but the present invention reduces the demand for computing resources and improves the interpretability of the model by combining sparse linear regression and fully connected neural network methods.

[0017] 4. Improve the accuracy of the prediction model: The present invention comprehensively considers the information of EEG non-periodic and periodic characteristics, as well as the importance of the similarity of neural activity characteristics between electrodes, and the constructed prediction model has been significantly improved in accuracy. This provides an effective way to solve the problem of low accuracy of the prediction model in existing EEG brain age estimation research.

[0018] In addition to the above-described purposes, features and advantages, the present invention has other purposes, features and advantages. The present invention will be further described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The drawings constituting a part of the present application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0020] Figure 1 A flowchart of the brain age prediction method based on cross-electrode representation learning of neural spectrum information of the present invention;

[0021] Figure 2 for Figure 1 Flowchart of data preprocessing module;

[0022] Figure 3 This is a structural diagram of a feature module for calculating similarity of neural activities between electrodes according to the present invention;

[0023] Figure 4 This is a scatter plot of the actual age and predicted age in the training set in a certain iteration of the ten-fold cross validation;

[0024] Figure 5 This is a scatter plot of the actual age and predicted age in the test set for a certain iteration of the ten-fold cross validation. DETAILED DESCRIPTION

[0025] The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.

[0026] Combined with reference Figures 1 to 3 The brain age prediction method based on cross-electrode representation learning of neural spectrum information of the present invention includes the following steps S1-S4.

[0027] S1. Data preprocessing

[0028] S11, Bandpass filtering. Apply a bandpass filter and set the passband frequency range to 1-20 Hz.

[0029] S12. Data capture. According to the experimental event markers, the continuous EEG data is divided into multiple time periods. Each time period is 1 minute of data before and after the stimulus appears.

[0030] S13. Independent component analysis (ICA). The ICA algorithm is used to decompose the EEG signal into independent components. The artifact components (eye movement, ECG artifacts, etc.) are separated and removed according to the ICLABEL algorithm, and the bad conductors are identified and marked.

[0031] S14, interpolating bad channels. In order to repair signal channels that are noisy due to electrode problems or other reasons, the signal of the bad channel is estimated by using the average value of data from adjacent electrodes.

[0032] S15, conversion reference. The average value of all electrode signals is calculated and then subtracted from each electrode signal to eliminate the influence of global potential changes.

[0033] S16. Calculate the power spectrum. Calculate the power spectrum density at a frequency of 1-20 Hz by Fourier transform.

[0034] is a time series signal, where t is time. Perform Fourier transform to obtain its spectrum ,in is the frequency. The formula for Fourier transform is as follows:

[0035]

[0036] Calculate the power spectral density (PSD): PSD represents the power spectrum density at each frequency. The power density under . The formula is as follows:

[0037]

[0038] in, is the total duration of the signal, is the frequency The amplitude of the Fourier transform at .

[0039] S2. Extracting neural spectrum features

[0040] Non-periodic spectral features and periodic spectral features: Spectral component decomposition and parameterization methods are used to calculate non-periodic and periodic spectral features. First, the Welch method (see Welch, P., 1967. The use of fast Fouriertransform for the estimation of power spectra: A method based on timeaveraging over short, modified periodograms. IEEE Trans. Audio Electroacoust.15, 70–73. https: / / doi.org / 10.1109 / TAU.1967.1161901) is used to estimate the EEG time series as a power spectrum, which is a combination of non-periodic components and periodic components. The non-periodic component is fitted as a function over the entire passband, and each oscillation peak is fitted separately with a Gaussian kernel function. The algorithm models the power spectrum as:

[0041]

[0042] in, is the non-periodic component, is the Gaussian component, is the vector of input frequencies;

[0043]

[0044] in, is the broadband offset, i.e. the non-periodic parameter offset, is the non-periodic parameter index;

[0045]

[0046] in, is the period parameter peak power, is the period parameter center frequency (Hz), is the Gaussian standard deviation. The periodic parameter bandwidth is .

[0047] Power ratio: After calculating the power spectrum, the power of each frequency band was extracted, including δ (1-4 Hz), θ (4-8 Hz), α (8-13Hz), and β (13-20 Hz), and then the power ratios of the four frequency bands were calculated: θ / β, δ / θ, δ / α, and θ / α.

[0048] Relative energy: Extract the relative energy of each frequency band (δ, θ, α, β). Taking the δ band as an example, the relative energy calculation formula is as follows:

[0049]

[0050] Kurtosis: describes the shape of the signal distribution. It measures the relative kurtosis or flatness of the signal distribution. The calculation formula is as follows:

[0051]

[0052] in is the sample size, For the Observations, is the sample mean, and s is the sample standard deviation.

[0053] In one embodiment, the subject uses 18 electrode channels, and each electrode channel extracts 17 features, for a total of 306 features. The extracted neural spectrum features are shown in the following table:

[0054]

[0055] S3. Extracting inter-electrode neural spectrum dependency features

[0056] Since pairwise correlations between electrodes (such as Pearson correlation) may ignore the potential influence of other electrodes, this paper uses sparse linear regression (Elastic Net algorithm) to quantify the cross-electrode relationships of various neural activity features, where the above-mentioned electrode features are regarded as target vectors and prediction vectors in the regression.

[0057] We represent each individual neural spectrum information as common electrodes, of which , which means the first All neural spectra of the electrodes are related to the characteristics of each is considered as a target vector and can be replaced by other The prediction vectors are represented as linear combinations.

[0058] The sparse regression model is defined as ,in is the regression coefficient, is the target vector. In the i-th regression, the data matrix Contains except All electrode vectors except . The sparse solution is achieved by L1-norm regularization and L2-norm regularization. The problem is expressed as ,in Represents the sparsity control parameter. The larger the value, the sparser the sparse matrix.

[0059] We do this by changing the , The similarity matrix of neural activity between electrodes is constructed by using the values ​​of the matrix, and each matrix group is evaluated according to its prediction performance. A brief overview of the method is shown in Figure 3 shown.

[0060] Repeat this process for each subject. Afterwards, a feature matrix of similarity of neural activities between electrodes was constructed.

[0061]

[0062] The matrix has m rows and m-1 columns, with consistent superscripts in each row and subscripts corresponding to the predicted electrodes. Each row of the matrix represents a set of regression coefficients from a single sparse regression model, with the corresponding node as the target variable (labeled on the right), and a non-zero value indicates a relatively strong relationship between the target electrode and all other electrodes included in the regression analysis, while a value of zero indicates a weak relationship.

[0063] Since each row in the matrix is ​​composed of sparse solutions of different regression processes, the coefficient matrix should be asymmetric. Different from the connectivity of information transfer involved in brain functional connectivity and brain network research, the asymmetric matrix does not reflect the brain functional dependency between electrodes, but only represents the mutual representation similarity relationship between spectral features across electrodes, reflecting the coefficients or weight information obtained from the mutual representation learning of spectral features in multi-electrode space. Therefore, it is based on the representation learning of multiple groups of neural spectral features across electrodes.

[0064] S4. Age prediction

[0065] Neural network is a model widely used in machine learning and data mining. Fully connected neural network builds a multi-layer neural network model to learn the complex nonlinear mapping relationship between input and output, thereby obtaining a function that can predict the output.

[0066] The fully connected neural network used in the present invention comprises an input layer, which receives the processed feature matrix, i.e., the features concatenated and fused in steps S2 and S3; three hidden layers, each containing 10 neurons, for further extracting and combining the complex features of the input data, and an output layer, containing one neuron, for outputting the predicted age.

[0067] The training process includes forward propagation, back propagation and weight update. In forward propagation, the input data is passed through the neurons and tansig activation functions of each layer to finally generate output. Back propagation calculates the error between the output and the target and uses the Levenberg-Marquardt (Ranganathan, nd) algorithm to adjust the network weights to minimize the error.

[0068] This paper uses a multinational cross-spectra dataset (MN-CS), which contains 1966 subjects. The gender distribution of the samples is almost uniform, and the age range is 5-97 years old. There are fewer middle-aged and elderly subjects. After removing outliers, 1628 subjects remain. The model is trained using the 10-fold cross-validation method.

[0069] The measurement indicators include the mean absolute error (MAE), root mean squared error (RMSE), coefficient of determination (R 2 ):

[0070]

[0071]

[0072]

[0073] in To predict age, is the actual age, is the average actual age, is the total number of samples.

[0074] By changing the value of the parameter, when λ1=2 -6 , the optimal MAE was obtained when λ2=0.7, training set MAE=2.25, RMSE=4.17, R 2 =0.85; test set MAE = 2.72, RMSE = 5.11, R 2 =0.73. Figure 4 and Figure 5 The scatter plot is of the actual age and predicted age of the training set and test set in a certain iteration process of the ten-fold cross validation.

[0075] The table below shows the results of our work compared with previous work on this topic. The datasets are all resting EEG data of healthy subjects.

[0076]

[0077] For the study by Dimitriadis and Salis, please refer to the literature "Dimitriadis, SI, Salis, CI, 2017. Mining Time-Resolved Functional Brain Graphs to an EEG-Based Chronnectomic Brain Aged Index (CBAI). Front. Hum. Neurosci. 11, 423. https: / / doi.org / 10.3389 / fnhum.2017.00423"; for the study by Al Zoubi, please refer to the literature "AlZoubi, O., Ki Wong, 2018. Predicting Age From Brain EEG Signals—A Machine Learning Approach. Front. Aging Neurosci. 10, 184. https: / / doi.org / 10.3389 / fnagi.2018.00184"; for the study by Khayretdinova, please refer to the literature "Khayretdinova, M., 2022. Predicting age from resting-state scalp EEG signals with deep convolutional neural networks on TD-brain dataset. Front. Aging Neurosci. 14, 1019869. https: / / doi.org / 10.3389 / fnagi.2022.1019869》.

[0078] In particular, for the MN-CS dataset used in this study, it is not directly based on the original EEG signals, but contains cross-spectral information. The current mainstream EEG age prediction methods generally rely on raw EEG signals, which limits the direct application of these methods on this dataset. So far, only Jarne et al. (Jarne C, Griffin B, Vidaurre D. Predicting subject traits from brain spectral signatures: an application to brain ageing. bioRxiv, 2023: 2023.11. 02.565261.) have attempted to perform EEG age prediction on the MN-CS dataset. The results obtained by their model have a MAE of about 10 years and R2 About 0.55.

[0079] For other EEG age prediction studies, we only found the TDBRAIN dataset with a wide age distribution. In order to verify the superiority of this study in the EEG age prediction task, we conducted comparative experiments on the TDBRAIN dataset. The experimental results (as shown in the table below) further demonstrated the significant superiority of this prediction model.

[0080]

[0081] The above description is only an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A brain age prediction method based on cross-electrode representation learning of neural spectrum information, comprising: S1. Preprocess the EEG data of each electrode channel in the subject's multi-electrode channels to obtain the power spectral density PSD at a frequency of 1-20 Hz; S2. Use the power spectral density PSD at a frequency of 1-20 Hz to calculate the EEG neural spectrum characteristics of each electrode channel, including the non-periodic parameter AF, the periodic parameter PF, the power ratio PR, the relative energy RE and the kurtosis KUR; S3, taking the above-mentioned EEG neural spectrum features of each electrode channel as neural spectrum features independently represented by a single electrode, quantifying the cross-electrode representation relationship of the neural spectrum features between electrodes by a sparse linear regression method, and obtaining the neural spectrum feature similarity representation matrix between electrodes of the neural spectrum features; S4. Construct a fully connected multi-layer neural network to learn the relationship between the EEG neural spectrum characteristics of the subjects and the neural spectrum characteristics similarity representation matrix between electrodes and brain age, and finally obtain the corresponding brain age based on the above characteristics. Step S3 includes: The EEG neural spectrum characteristics of the subjects There are m electrodes in total, , represents a target vector composed of the EEG neural spectrum features of the mth electrode, and the EEG neural spectrum features of other electrodes constitute m-1 prediction vectors; A set of regression coefficients between each target vector and other prediction vectors are calculated according to the sparse linear regression algorithm, where the sparse regression model is ,in is the regression coefficient, y is the target vector, and the data matrix Contains except All electrode vectors except ; For each subject, after repeating this process m times, m groups of regression coefficients are obtained, and the neural spectrum feature similarity representation matrix between electrodes is constructed based on them: Here, each row of the matrix represents a set of regression coefficients from a single sparse regression model.

2. The method for predicting brain age based on cross-electrode representation learning of neural spectrum information according to claim 1, characterized in that: Preprocessing the EEG data of each electrode channel in the subject's multi-electrode channels includes the following steps: S11, band-pass filter the EEG data through a band-pass filter, setting the passband frequency range to 1-20 Hz; S12. Divide the continuous EEG data into multiple time periods according to the experimental event markers, with each time period consisting of 1 minute of data before and after the stimulus. S13. Decompose the EEG signal into independent components using the ICA algorithm, separate and remove artifact components according to the ICLABEL algorithm, and identify and mark the bad leads; S14, for a signal channel generating noise, estimating a bad track signal by using an average value of data of adjacent electrodes; S15, by calculating the average value of all electrode signals, and then subtracting the average value from each electrode signal to eliminate the influence of global potential changes, a time series signal is obtained. , where t is time; S16. Time series signal Perform Fourier transform and calculate the power spectral density PSD at a frequency of 1-20 Hz.

3. The method for predicting brain age based on cross-electrode representation learning of neural spectrum information according to claim 2, characterized in that: The aperiodoc parameter AF (Aperiodoc Feature) includes two features, namely, index and offset; the periodic parameter PF (Periodoc Feature) includes three features, namely, center frequency, power and bandwidth; the power ratio PR (Power Ratio) includes four frequency band power ratio features, namely, θ / β, δ / θ, δ / α, and θ / α; the relative energy RE (Relative Energy) includes four frequency band relative energy features, namely, δ, θ, α, and β; the kurtosis KUR (Kurtosis) includes four frequency band kurtosis features, namely, δ, θ, α, and β, wherein the frequency band ranges are δ (1-4 Hz), θ (4-8 Hz), α (8-13 Hz), and β (13-20 Hz), respectively.

4. The method for predicting brain age based on cross-electrode representation learning of neural spectrum information according to claim 1, characterized in that: The fully connected multi-layer neural network consists of an input layer, three hidden layers and an output layer, wherein the number of neurons in the input layer is consistent with the number of matrix elements, the number of neurons in each hidden layer is 10, the output layer is one neuron, and each subject uses an 18-channel EEG signal.

5. The method for predicting brain age based on cross-electrode representation learning of neural spectrum information according to claim 1, characterized in that: The fully connected multi-layer neural network is trained by the multi-center data set MN-CS, and the samples after removing outliers are trained by the 10-fold cross validation method. The measurement indicators include the mean absolute error MAE between the actual age and the predicted age, the root mean square error RMSE, and the determination coefficient R 2 .

Images