A method and apparatus for predicting brain age based on electroencephalogram

Abstract

The present application provides a method and device for predicting brain age. The method comprises: acquiring an electroencephalogram (EEG) signal of a subject in a wakeful state; performing artifact detection and segmentation on the EEG signal to select a plurality of EEG signal segments meeting signal quality requirements; converting the plurality of EEG signal segments meeting the signal quality requirements into time-frequency map data; and using a neural network model to predict the brain age of the subject corresponding to the plurality of EEG signal segments meeting the signal quality requirements based on the time-frequency map data. Compared with using the EEG signal in a sleep state, the present application can use the EEG signal of the subject in a wakeful state to monitor the brain age and its changes in a shorter time period or in real time.

Description

Technical Field

[0001] This application generally relates to electronic signal processing, and in particular to methods, apparatus, and computer-readable storage media for predicting the brain age of a subject based on electroencephalogram (EEG) signals. Background Technology

[0002] Human brain development is a highly complex process, primarily manifested in the varying order and trends of development across different brain regions. Poor brain development can negatively impact a child's overall development. Certain brain disorders can also lead to regional differences in brain age. Therefore, accurately examining and determining the brain's age is crucial for assessing its development or condition.

[0003] Clinically, methods used to determine brain age when examining brain development include physical examination, mental development scales, CT scans, electroencephalograms (EEGs), MRI scans, and genetic testing. When diagnosing brain dysplasia, doctors also need to consider the child's birth history, the parents' marital and reproductive history, and clinical manifestations. Completing the entire examination process often takes tens of hours or even several days, and the determination of brain age is highly dependent on the doctor's experience.

[0004] Some existing computer-aided brain age prediction schemes utilize deep learning and magnetic resonance imaging (MRI) data to build models. The best predictive performance of this MRI-based image processing scheme in adult subjects is a mean absolute error (MAE) of around 5 years, while in children and adolescents, the best performance is an MAE of around 1 year. Due to the complex and time-consuming nature of MRI examinations, their relatively high cost, and the limited number of patients that can be examined daily, the promotion and application of MRI-based brain age prediction schemes in the field of brain age screening for developmental delays are somewhat limited.

[0005] Brain age prediction schemes based on electroencephalography (EEG) also exist. However, most of these EEG-based schemes use EEG signals from sleep states. The complexity of sleep-state EEG signals significantly hinders algorithm development, leading to unsatisfactory accuracy. Furthermore, the long acquisition time for sleep-state EEG signal data also severely limits the application of these schemes.

[0006] Therefore, there is a need to improve the schemes used to predict brain age. Summary of the Invention

[0007] This application proposes a signal processing method and device based on electroencephalogram (EEG) signals, which aims to address at least one of the shortcomings and deficiencies of the existing brain age prediction schemes mentioned above. It can predict the brain age of subjects more quickly, accurately, and economically, and provide basic data for doctors to assess brain development levels.

[0008] According to one aspect of this application, a method for predicting brain age is proposed, comprising: acquiring an electroencephalogram (EEG) signal of a subject in a conscious state; performing artifact detection and segmentation on the EEG signal to select multiple EEG signal segments that meet signal quality requirements; converting the multiple EEG signal segments that meet the signal quality requirements into time-frequency plot data; and using a neural network model to predict the brain age of the subject corresponding to the multiple EEG signal segments that meet the signal quality requirements based on the time-frequency plot data.

[0009] According to another aspect of this application, a device for predicting brain age is proposed, comprising: an acquisition unit configured to acquire electroencephalogram (EEG) signals of a subject in a conscious state; a prediction unit configured to perform artifact detection and segmentation on the EEG signals to select multiple EEG signal segments that meet signal quality requirements; convert the multiple EEG signal segments that meet the signal quality requirements into time-frequency plot data; and use a neural network model to predict the brain age of the subject corresponding to the multiple EEG signal segments that meet the signal quality requirements based on the time-frequency plot data; and an output unit configured to output the brain age.

[0010] According to another aspect of this application, a computer-readable storage medium is provided that stores a computer program thereon, the computer program including executable instructions that, when executed by a processor, implement the method as described above.

[0011] According to another aspect of this application, an electronic device is provided, comprising: a processor and a memory for storing executable instructions of the processor, wherein the processor is configured to execute the executable instructions to implement the method described above.

[0012] By employing the method and device proposed in this application for predicting brain age, compared to using EEG signals during sleep, this application can dynamically monitor brain age and its changes in a shorter timeframe or in real-time using EEG signals from subjects in a waking state. This brain age prediction scheme has high sensitivity, and while improving the accuracy of brain age prediction, it can further shorten the detection time and cost, and make brain age prediction more compatible with digital electroencephalography (EEG) and video EEG examination methods, thereby improving the clinical usability of the brain age prediction scheme. This brain age assessment method can not only be used to assess brain development levels, but also serve as a reference for rapid evaluation of the efficacy of drug treatments for neurological diseases. Attached Figure Description

[0013] The above and other features and advantages of this application will become more apparent from a detailed description of exemplary embodiments thereof with reference to the accompanying drawings.

[0014] Figure 1 This is a flowchart illustrating a system algorithm for predicting brain age according to an embodiment of this application.

[0015] Figure 2 This is a schematic structural diagram of a neural network model for predicting brain age based on time-frequency graph data, according to an embodiment of this application.

[0016] Figure 3 This is a schematic flowchart of a method for predicting brain age according to an embodiment of this application.

[0017] Figure 4 This is a schematic structural diagram of a device for predicting brain age according to an embodiment of this application.

[0018] Figure 5 This is a schematic diagram illustrating the relative positional relationship between a brain age sample predicted by a method or device for predicting brain age according to an embodiment of this application and the range of the normal population.

[0019] Figure 6 This is a schematic block diagram of an electronic device for predicting brain age according to an embodiment of this application. Detailed Implementation

[0020] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided to make the content of this application comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art. In the drawings, the dimensions of some elements may be exaggerated or modified for clarity. The same reference numerals in the drawings denote the same or similar structures, and therefore their detailed descriptions will be omitted.

[0021] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a full understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details described, or other methods, elements, etc., can be employed. In other instances, well-known structures, methods, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0022] As mentioned above, computer-aided brain age prediction schemes can be categorized into two types based on data sources: image processing based on MRI images and signal processing based on EEG signals during sleep. MRI image-based brain age prediction schemes primarily rely on static image structural information of changes in brain regions. Since MRI images of the human brain do not change significantly over a short period, the lack of time-related data means that the brain age assessed and predicted by this method is relatively stable over a certain time range, making it unsuitable for dynamically detecting brain changes. Even with the use of deep learning models to improve the image processing accuracy of MRI images, the inherent limitations of this approach cannot be overcome. Furthermore, MRI image-based brain age prediction schemes suffer from expensive equipment and time-consuming processes, thus limiting their widespread adoption and making them unsuitable for most screening scenarios.

[0023] The primary reason for choosing EEG signals over MRI image-based processing is their dynamic nature. Using EEG signals to predict brain age allows for sampling at shorter time intervals, such as every 20 or 30 minutes. For example, in epilepsy patients undergoing drug treatment, the EEG signals can undergo significant changes within a short period, which can be reflected in the morphology and characteristics of the EEG signals—changes that MRI images cannot capture. The high dynamic range of EEG signals improves the sensitivity of brain age prediction compared to MRI images.

[0024] In EEG signal-based methods, brain age prediction based on sleep-state EEG signals faces the challenge of the inherent complexity of these signals. Generally, the complex characteristics of EEG signals from sleeping individuals make it difficult to develop suitable feature extraction algorithms, resulting in low accuracy. Another issue is that screening scenarios typically require short EEG signal acquisition periods of around 10 minutes, which makes it impossible to allow subjects to sleep for extended periods. The relatively long detection time during sleep, coupled with the need for uninterrupted sleep, results in a prolonged signal duration that does not meet the requirements for rapid (e.g., tens of minutes) and real-time screening in various scenarios.

[0025] Therefore, methods for predicting brain age based on data processing of MRI images and sleep EEG signals are unsuitable for rapid screening. The primary technical challenge this application aims to address is finding a solution specifically designed for rapid and real-time screening scenarios, without compromising prediction accuracy, and even improving it further. Real-time screening requires simple signal characteristics for rapid extraction, and short signal data duration for rapid real-time acquisition to meet the requirements of real-time dynamic analysis and monitoring. Furthermore, low operating costs and good compatibility of the signal sampling equipment are also important considerations. Existing EEG signal machines and devices are digital and can provide more intuitive output methods such as image / video EEG. Compared to MRI equipment, which cannot provide real-time digital presentation, EEG signals are more suitable for clinical use.

[0026] Therefore, the applicant proposes using short-duration EEG signals from awake subjects as data to predict brain age and dynamically monitor brain changes, thus avoiding the complex characteristics of EEG signals during sleep. Furthermore, since EEG signal acquisition can be started and stopped at any time while awake, there is no issue of excessively long detection durations. The improved accuracy of brain age prediction also stems from improvements in the signal data processing algorithm. The signal data processing algorithm used in this application's brain age prediction scheme is based on a neural network model and features a network architecture conducive to EEG signal feature extraction, making it faster and more efficient than traditional signal processing algorithms (which often use common features such as power, amplitude, and frequency for routine processing and calculation).

[0027] Below, we will combine Figure 1 The system algorithm flow shown illustrates the solution of this application.

[0028] Figure 1 This illustration shows the process of predicting the brain age of a subject based on short-duration (shown as 10 minutes or more in the figure) EEG signals 101 acquired by an EEG instrument while the subject is awake (e.g., a child to be evaluated or a patient with epilepsy). EEG signals can be acquired using an EEG instrument according to clinical standards, with a sampling rate of 200 Hz or higher and at least 16 lead channels. According to an embodiment of this application, scalp EEG signals from 21 lead channels are used, including channels Fp1, Fp2, F3, F4, C3, C4, P3, P4, O1, O2, F7, F8, T3, T4, T5, T6, Fz, Cz, Pz, A1, and A2. EEG signals from multiple lead channels are synchronously recorded on the same EEG waveform graph on the horizontal axis of time (in seconds) for processing and comparison.

[0029] The EEG signal undergoes preprocessing at position 110. The purpose of preprocessing is to reduce the computational load to improve efficiency, eliminate invalid data and interference from affecting detection, and improve the resolution and accuracy of the EEG signal waveform.

[0030] Preprocessing operations include at least one of rereference, downsampling, notch filtering to power frequency, and bandpass filtering.

[0031] The rereference operation uses the average value AV of the raw EEG signals from 21 leads as the reference signal. The EEG signal is calibrated by subtracting the reference signal (AV value) from the signal of each lead. EEG signals from 19 leads excluding A1 and A2 are selected and denoted as... .

[0032] Resampling is used to improve computational efficiency. The resampling frequency should be at least twice the sampling frequency of the EEG instrument, generally above 150Hz, for example, 200Hz is chosen in this application to ensure detection accuracy. Resampling can be achieved through downsampling or upsampling through interpolation. In the embodiments of this application, Downsampled to 200Hz, the unit of EEG signal amplitude is V (volts). Generally, the horizontal axis of the raw EEG signal is time t (seconds). The vertical axis, voltage amplitude, is measured in volts / V and has a value in the range of 10. -6 The voltage amplitude is on the order of magnitude of an exponent, but conventional graph reading software converts it to microvolts for display. Therefore, the voltage amplitude can be multiplied by 1.0*e. 6 Transform into The unit is microvolt (µV), which facilitates the learning and updating of neural network parameters.

[0033] Next, a 50Hz notch filter and a bandpass filter are sequentially applied to the EEG signal to remove noise such as power frequency noise, baseline drift noise, and / or high-frequency noise. Filtering is used to avoid the negative impact of various interference signals on the detection. Power frequency noise is generally caused by the inherent frequency of the power system. In my country, the operating frequency of AC power is 50Hz (or 60Hz). Power frequency noise can interfere with the normal operation of electronic equipment and may also overwrite the valid information related to brain age prediction contained in the EEG signal. Notch filters with operating frequencies of 50Hz and / or multiples of 50Hz can be designed to eliminate power frequency noise interference in the EEG signal. In this application, a 50Hz notch filter is designed to remove power frequency noise. A high-pass filter greater than or equal to 1.5Hz is used to remove baseline drift noise from the signal, and a low-pass filter less than or equal to a set frequency is used to remove high-frequency noise from the signal. Baseline drift mainly comes from changes in circuit component parameters caused by the physical characteristics of the channel sensor and environmental factors (such as changes in ambient temperature), which results in the superposition of components unrelated to the measured signal in the signal input. High-frequency noise originates from electromyography (EMG) interference and environmental noise. In this application, a bandpass filter with an operating frequency range of 1.5Hz or higher, preferably within the 1.5Hz-100Hz frequency range, is used to filter and denoise the EEG signal. Both notch filtering and bandpass filtering can employ an FIR filter with parameters selected as a passband bandwidth of 1.0Hz and a Hanning window function, resulting in a pre-processed EEG signal. .

[0034] Next, artifact detection is performed in step 120. The purpose of artifact detection is to select signal regions with high signal quality (also known as clean regions) in the EEG signal. Its basic principle is to select clean portions of the EEG signal that meet predetermined signal quality requirements based on two conditions: the amplitude (amplitude and / or average amplitude) and frequency (line length or signal duration for high-frequency signals). In contrast to clean EEG signal regions (or time intervals), there are EEG signal regions or time intervals containing artifacts. In this application, signal quality requirements mainly include conditions and requirements related to the signal amplitude, average signal amplitude, signal duration, and artifact percentage of the EEG signal. Among these conditions, the signal amplitude and average signal amplitude are related to signal fluctuations, the signal duration (also known as signal line length) is related to the signal frequency, and the artifact percentage represents the proportion of signals with artifacts in the EEG signal, which can also characterize the quality of the EEG signal.

[0035] The artifact detection process includes four operations: quartile normalization, high-pass filtering, calculation of average amplitude, and merging of artifact signals.

[0036] Different EEG instruments and the techniques used by technicians when installing electrodes on the subject's scalp can lead to systematic errors in the amplitude of the recorded EEG signals, resulting in variations in signal amplitude. Quartile normalization can mitigate these errors caused by equipment differences and technician / physician procedures.

[0037] In the quartile normalization operation, the EEG signal will be normalized. The signals from the 19 lead channels were statistically analyzed to obtain the EEG signal. The four interquartiles are denoted as follows: and In one embodiment of this application, the corresponding target quartile is: and According to the following formula (1), the EEG signal after quartile normalization can be calculated. :

[0038] (1)

[0039] Based on this, EEG signals are obtained. The time interval corresponding to the portion of the signal whose absolute amplitude is ≥ 300μV is denoted as . Generally, the maximum amplitude of an EEG signal from the scalp will not exceed 300 μV. Therefore, 300 μV is set as the normal amplitude limit for EEG signals. Exceeding this value indicates that the amplitude of the EEG signal in the corresponding time interval is outside the normal recording range, i.e., it is an EEG signal portion with artifacts. The time interval corresponding to this signal portion... This is the first type of pseudo-EEG signal segment found based on conditions related to signal amplitude. In other words, the EEG signal segment with an amplitude not exceeding 300μV is a normal and clean EEG signal segment that meets the signal amplitude-related conditions in the signal quality requirements.

[0040] In the subsequent high-pass filtering operation, the EEG signal after quartile normalization is... A high-pass filtering operation is performed. In the embodiments of this application, the filtering parameters are selected as 60Hz high-pass filtering, using an FIR filter with a passband bandwidth of 1.0Hz and a Hanning window. Here, the purpose and parameters of the high-pass filtering operation are different from those in the signal preprocessing 110 process. The filtering frequency of the high-pass filtering here is greater than or equal to 60Hz, which is different from the previous filtering parameters (greater than or equal to 1.5Hz, for example, 1.5Hz-100Hz). The reason is that the high-pass filtering in signal preprocessing is to eliminate the baseline drift of the signal, while this high-pass filtering operation only needs to consider the high-frequency components above 60Hz (for example, the frequency range of 60-100Hz), because the artifacts of the EEG signal originating from the action will be more obvious in this high-frequency range. If the calculation of the signal time length (line length) mentioned below is further combined, these artifacts can be captured well, thereby further improving the accuracy of artifact detection. The processed EEG signal after high-pass filtering. .

[0041] In calculating the average amplitude, the following formula (2) is used for the EEG signal. Calculate the average amplitude of the signal over a smoothed line length (the time length of the EEG signal) with a window size of N. The sliding step size is taken from one data point:

[0042] (2)

[0043] Taking a sliding step of 1 data point is equivalent to calculating the EEG signal between adjacent time points within a time interval of N time steps (unit of time length) for a time window. The average of the differences in the absolute values ​​of the amplitudes. When the sampling frequency is set to 200Hz during data preprocessing 110, the sliding step size is equivalent to a time step of 1 / 200th of a second. In the embodiments of this application, N in the N time steps included in the time window is 40, equivalent to a time length of 40 * 1 / 200 seconds (equivalent to a time window width of 40 / 200 seconds).

[0044] Generally, the average amplitude of EEG signals from the scalp does not exceed 15 μV. Therefore, 15 μV is set as the normal average amplitude limit for EEG signals. The average amplitude in Exceeding this value indicates that the average signal amplitude (or amplitude fluctuation) of the EEG signal in the corresponding time interval is outside the normal recording range, i.e., another part of the EEG signal with artifacts, and the time interval corresponding to this signal part... This is the second type of pseudo-EEG signal segment found based on conditions related to the average signal amplitude. In other words, the EEG signal segment with an average amplitude not exceeding 15μV is a normal and clean EEG signal segment that meets the signal quality requirements related to the average signal amplitude.

[0045] Next, the two types of EEG signal components with artifacts determined in the quartile normalization operation and the average amplitude calculation operation above are combined to determine the time interval of the EEG signal with artifacts. The time interval is detected based on two conditions. and The EEG signal portions shown are all artifact-laden and should be removed from the EEG signal in subsequent processing. This merging operation can be performed using a union operation. The time intervals corresponding to the two artifact-laden signal portions are not necessarily in an inclusive relationship; for example, they may or may not overlap. The union operation is equivalent to taking the time positions (time intervals) that do not satisfy at least one of the two conditions—signal amplitude and average signal amplitude—as the base positions (time intervals) for artifacts, and merging these base positions to form the final time interval corresponding to the artifact-laden EEG signal portion.

[0046] Specifically, take and union As the base position of the artifact, then for If the time interval between any two adjacent pseudopoints is less than 0.1 seconds, then these two adjacent time intervals are merged into one time interval to reduce the number of discrete pseudopoints or fragmented pseudopoint segments, thus obtaining the final pseudopoint time interval. .

[0047] Here, the time interval can be understood as the time difference between the two closest time points in two time intervals, i.e., the length of time the two time intervals are separated. Because EEG signals are fluctuating, the artifact time intervals detected based on quartile normalization are... There are many fragmented regions. For example, the time interval [t1, t2] defined by time points t1 and t2 is the pseudo-time interval, and the time interval [t3, t4] defined by time points t3 and t4 is also a pseudo-time interval. However, often the time interval between t2 and t3 may only be tens to hundreds of milliseconds. Therefore, such a short time interval as [t2, t3] can be directly regarded as a pseudo-time interval. After the merging operation, the time interval corresponding to the pseudo-EEG signal portion is determined to be [t1, t4]. In other words, the normal and clean EEG signal portion that meets the signal quality requirements related to signal amplitude and average signal amplitude is the entire signal minus the pseudo-time interval. The remaining signal portion after the corresponding signal portion.

[0048] After artifact detection, signal segmentation is performed at point 130 to select multiple continuous, conscious EEG signal segments with low artifact ratios that meet signal quality requirements from the EEG signal. According to an embodiment of this application, the selection parameters are... The width of the time window for the predetermined time length, A sliding step of 4 times the EEG signal after quartile normalization. Swipe up to capture a series of EEG signal segments. Duration The width of the time window is equivalent to... The duration of continuous EEG signals Segmentation is performed to select multiple EEG signal segments. Among them, The signal duration requirement, as part of the signal quality specifications, is set to either 1 minute or 2 minutes. This is combined with the artifact time interval. The information allows us to calculate the proportion of the time interval containing artifacts in each EEG signal segment to the total time interval of that segment, i.e., the artifact proportion level. We select segments with an artifact proportion level less than [a certain value]. The EEG signal segments that meet the artifact ratio requirement in the signal quality requirements constitute the EEG signal segments used for the next step of processing. Here, Known as a hyperparameter related to the quality of EEG signals, its possible values ​​range from [value missing]. .

[0049] After signal segmentation operation 130, each segment that meets the signal quality requirements is obtained. Each EEG signal segment of this duration includes 19 lead channels. According to one embodiment of this application, the EEG signal segment in each lead channel has a frequency of 12000*... 10 data points. These EEG signal segments consist of each signal segment with a length of 1. The proportion of artifacts in subjects who are awake for 30 minutes is less than The N processed EEG signal segments are shown in Figure 102.

[0050] Next, the system algorithm performs a Fourier transform on the EEG signal segments to separate them into N signal segments in the time dimension, each segment comprising D (D=12000). ) continuous signal data of data points are converted into time-frequency plot data represented by data in the frequency and time dimensions of the time-frequency plot 103.

[0051] According to embodiments of this application, each continuous time period can be... (For example A 1-minute signal segment undergoes a Short-Time Fourier Transform (STFT) to convert the clean EEG signal segment (continuous signal) after artifact removal into time-frequency plot data (discrete signal) for processing by a neural network model. Each of these segments... The dimensions of the 1-minute EEG signal segment are [C, D], where C is the number of lead channels in the EEG signal, and D = 12000. Let be the number of data points in the EEG signal segment along the time dimension. After a short-time Fourier transform, information from the frequency band 1.0Hz-40.0Hz is extracted, resulting in a dimension of . The time-frequency plot data, where F=157 is the size of the frequency dimension of the time-frequency plot data (also known as the number of data components in the frequency dimension of the time-frequency plot data), and K is the size of the time dimension of the time-frequency plot data (also known as the number of data components in the time dimension of the time-frequency plot data, which is also the number of data points in the signal segment or window).

[0052] According to embodiments of this application, the following STFT parameters can be selected: resampling frequency fs = 200Hz; number of samples per frame n_fft = fs*4; length of each FFT segment n_per_seg = fs*4; overlap between adjacent FFT segments n_overlap = fs*2; the window function can be a Hanning window; a window can be used; or an FIR filter can be used. Without zero-padding at the signal ends, K can be calculated using K = (D - n_per_seg) / n_overlap + 1. When =1, K= 29; When K = 2, K = 59. The calculation of the Hanning window is as follows:

[0053] (3)

[0054] Where N is the total length of the Hanning window function, that is, the number of EEG signal segments extracted from the EEG signal that meet the signal quality requirements. In the neural network model introduced below, the number of EEG signal segments N is also used as the batch size of the neural network model, that is, the input data to the neural network model is N EEG signal segments, and the brain age predicted by the model is the brain age prediction value based on N EEG signal segments.

[0055] The time-frequency plot data 103 was then input into the neural network model 140. The neural network model in... Figure 1 The brain age prediction target of this application was selected as BANet. Those skilled in the art should understand that other types of neural network models can also be used to achieve the brain age prediction target of this application.

[0056] Figure 2 The structure of the neural network model 140 is shown using BANet as an example. It includes an input layer 200, a first sub-network 210 with a trainable triangular filter layer, a second sub-network 220 (shown in dashed boxes) with at least two (unidirectional) GRU layers, a third sub-network 230 with a self-attention layer, linear and SoftPlus layers 240, and an output layer 250.

[0057] Input layer 200 is used to receive the time-frequency diagram data 103 processed by STFT. The dimension of the input data is... ,in The batch size is the number of clean EEG signal segments that meet the signal quality requirements. The definitions of C, F, and K are as described above.

[0058] Before the input data received by the input layer 200 is sent to the first sub-network 210, the dimensions of the input data can be rearranged, and the dimensions of the rearranged input data become... As shown in 201, the purpose of dimensional rearrangement is to adjust the data format to a form that is easier for subsequent network layers to process.

[0059] The first subnetwork 210 mainly includes a nonlinear trainable triangular filter layer. Figure 2 Figure 211 shows a schematic diagram of the triangular filter, where the white area represents trainable parameters and the black area has a constant value of 0.

[0060] The trainable triangular filter layer performs dimensionality reduction on the frequency dimension data components of the input time-frequency plot data, making the input time-frequency plot data more dimensional. The dimensionality-reduced time-frequency plot data is obtained as the output. :

[0061] (4)

[0062] The transformation of formula (4) changes the frequency dimension from Transform into and satisfy .

[0063] in, F0 is the size of the frequency dimension of the time-frequency plot data as defined above (i.e., the number of data components in the frequency dimension), and K is the size of the time dimension of the time-frequency plot data as defined above (i.e., the number of data components in the frequency dimension). This refers to the number of lead channels for the EEG signal as defined above. The transformation matrix of a trainable nonlinear triangular filter layer:

[0064] (5)

[0065] in, It is a non-negative function that operates on the elements of a matrix, used to transform a trainable parameter matrix. Transform into a nonnegative constant matrix Here, use function. It is a non-negative constant matrix used to constrain the triangular filter within a finite bandwidth to reduce the dimensionality of the input time-frequency plot data. This represents matrix multiplication, which multiplies corresponding elements of two matrices.

[0066] The design philosophy is that the bandwidth of the triangular filter is small in the low-frequency region and large in the high-frequency region. Formula (6) can be used to calculate the frequency of the time-frequency plot data. Frequency converted to logarithmic space :

[0067] (6)

[0068] Then, based on formulas (7) and (8), the frequency intervals in the logarithmic space are... Divide into equal intervals Sub-intervals are obtained :

[0069] (7)

[0070] (8)

[0071] Formulas (6) to (8) are equivalent to changing the intervals corresponding to F frequencies to correspond to F+1 frequencies.

[0072] Determine the constant matrix according to Formula 9. :

[0073] (9)

[0074] The bandwidth range constraint defined in formula (9) makes the calculated It is smaller than the original frequency dimension, thus achieving dimensionality reduction.

[0075] Before inputting the dimension-reduced time-frequency plot data output from the first subnetwork 210 into the second subnetwork 220, the data components related to the lead channels of the EEG signal (i.e., dimension C) and the dimension-reduced frequency dimension data components (i.e., dimension ...) in the time-frequency plot data can be ... This is achieved by merging the data, thereby further reducing the dimensionality of the data input to the second sub-network 220 to improve feature extraction speed and efficiency. Figure 2 In this context, the dimension merging operation is equivalent to merging the last two dimensions C and C of the time-frequency plot data. ( The dimensions of the new time-frequency plot data are obtained as follows: ,in .

[0076] The second subnetwork 220 is used for the dimension-reduced time-frequency plot data output from the first subnetwork 210 (e.g., in data with a dimension of [missing value] after data component merging). Extract those features from EEG signals that are related to the subject's brain age.

[0077] The second subnetwork 220 may include at least two sequentially connected GRU subnetworks, each GRU subnetwork may include a sequentially connected unidirectional GRU layer and a linear layer. For example, Figure 2The diagram illustrates a first GRU subnetwork consisting of GRU layer 221 and linear layer 222, and a second GRU subnetwork consisting of GRU layer 223 and linear layer 224. Each unidirectional GRU layer used to extract relevant features from its input data is followed by a linear layer. Since the time-dimensional data components in the time-frequency plot data are sequential, GRU networks or similar network structures (e.g., LSTM and RNN networks) particularly suitable for extracting features from time-related data sequence inputs can be used. Compared to other similar neural network structures, GRU networks have a relatively small number of parameters, thus processing faster. Furthermore, GRU networks can effectively avoid problems such as gradient vanishing and exploding present in RNN networks. Under supervised training, GRU networks can automatically extract effective information related to brain age prediction more quickly and accurately. According to embodiments of this application, each GRU layer has 64 hidden layers.

[0078] As needed, a greater number of GRU subnetworks can be used, where the linear layer output of the last GRU subnetwork outputs the extracted feature data. Figure 2 The last linear layer in the output feature data is a linear layer 224. Feature extraction in these GRU subnetworks does not modify the dimensionality of the data, therefore Figure 2 The dimensions of all feature data in GRU 221 to linear layer 224 shown are 1 / 2 dimensional. .

[0079] Before the feature data output from the second sub-network 220 is input to the third sub-network 230, the time dimension feature components (which correspond to the time dimension data components of the time-frequency plot data, and number K) and the feature components related to the number of EEG signal segments that meet the signal quality requirements (number N) can be exchanged, thereby changing the dimension of the input feature data received by the third sub-network 230 from... Convert to Since the third sub-network 230 requires weighted dimensionality reduction of the feature data based on the dimension (dimensionality value K) of the feature components along the time dimension, this dimension swapping operation facilitates computation. For example... Figure 2 As shown, the 0th and 1st dimensions (corresponding to dimension values ​​K and N) of the feature data output by the linear layer 224 are swapped.

[0080] The third sub-network 230 mainly includes a self-attention layer, used for dimensionality reduction based on time-related feature components in the feature data. The detailed structure of the self-attention layer is described in... Figure 2 As shown in Figure 231, it includes a linear layer, a tanh activation function, and scale and softmax layers.

[0081] In the self-attention layer, the dimension of the input feature data matrix P is... The dimensions are numbered starting from 0. The feature data is weighted based on the first dimension (time dimension) to obtain the output feature data matrix Q, whose dimension size is... Let the elements of matrix P be... ( j v The elements of matrix Q are The parameters of the linear layers in the self-attention layer include weight parameters. Bias parameter B, where The dimension is The dimension of B is Thus, after the input feature data matrix P passes through a linear layer, its output... The data dimensions are still .

[0082] The transformation relationship of the tanh activation function is shown in formula (10):

[0083] (10)

[0084] The dimension of the feature data matrix G is Let its elements be ,in j v .

[0085] In the Scale and Softmax layers, the weighting coefficient matrix is ​​calculated. Its dimensions are Let its elements be Where i, j, v are defined and have the same range of values ​​as above. According to the following formulas (11) and (12), the feature components in the time dimension (with a dimension of K) are weighted and incorporated into the feature components (with a dimension of N) that are related to the number of EEG signal segments that meet the signal quality requirements, so as to obtain the feature quantity output Q after dimensionality reduction (with a dimension of N). ):

[0086] (11)

[0087] (12)

[0088] In this way, after Scale performs a numerical transformation on the feature data from the tanh activation function, the softmax function ensures that the output Q focuses on the feature parts with the highest probabilities. The purpose of weighted dimensionality reduction is to ensure that the final output of the neural network model has a dimension of 1. The brain age prediction data (i.e., the predicted brain age values ​​corresponding to N EEG signal segments that meet the signal quality requirements), the feature components in the time dimension of the feature data from the time-frequency plot data of the EEG signal, i.e., the number of data points in the signal segment (of size K), are intermediate variables and do not need to be output. In fact, the dimension (of size H) obtained by merging C and F0 does not need to be included in the output of the neural network model.

[0089] The output of the third subnetwork 230 is fed into the linear and Softplus layer 240. The transfer function of this layer is expressed by equation (13):

[0090] (13)

[0091] Where z represents the model's output data, which is the brain age prediction data consisting of the predicted brain age values ​​corresponding to N EEG signal segments that meet the signal quality requirements. Its dimension is... x represents the dimensionality-reduced feature data from the third subnetwork 230. The purpose of the linear and Softplus layers is to ensure that the element values ​​of the brain age output of the neural network model are positive.

[0092] Finally, the output data z of the neural network model is output through the output layer 250, i.e. Figure 1 The predicted brain age value of 104 is shown, corresponding to N EEG signal segments.

[0093] Before using neural network model 140, the model's parameters need to be trained using a training dataset consisting of training samples.

[0094] Subject Physiological age This indicates that further interaction with subjects... The first EEG signal that meets the signal quality requirements in a conscious state. A segment of EEG signal Composition of data pairs This constitutes a training sample. During the training of neural network model 140, two hyperparameters are used to expand the amount of training sample data, namely, to... Used for random jitter and against Used for random scaling . and It is only used during model training to increase the robustness of model training, and does not affect the prediction results of the trained neural network model 140 during use.

[0095] Hyperparameters In actual clinical records, due to the subject's self-reporting (e.g., epilepsy patients) or the doctor's rough recording, the true age is often inaccurate, sometimes differing from the actual age by days, months, or even half a year. To correct the subject's age during the EEG signal data preprocessing stage, the birth date recorded in the report and the EEG acquisition date can be used. While the EEG signal acquisition date is relatively accurate, the registered subject's birth date may be inaccurate or incomplete. However, overall, the corrected age is more accurate than the age recorded in the report. Therefore, during the training of the neural network model 140 used for brain age prediction, it is necessary to randomly jitter the corrected subject's age data to a certain extent to achieve data amplification. According to embodiments of this application, the amplification strategy includes, each time sample data of a subject is read, adding data to their physiological age data... Add a small float. ,in Age data in the sample data after random shaking Determined by formula (14):

[0096] (14)

[0097] Among them, the hyperparameters used to characterize age error information from patients .

[0098] Hyperparameters In clinical practice, the acquisition of EEG signals can be affected by numerous factors, including differences in EEG equipment, environmental variations, and electrode fit. Even for the same subject, under essentially identical testing conditions for a short period, using the same machine and the same technician with the same parameters to acquire two EEG signal segments, unavoidable systematic errors may still exist between these two segments. To address this, during model training, data corresponding to EEG signal segments in the sample data (e.g., time-frequency plot data generated via STFT) can be analyzed. conduct Data augmentation is achieved through random scaling by multiples, as shown in formula (15), where, :

[0099] (15)

[0100] Hyperparameters characterizing sampling error-related EEG signals from EEG instruments .

[0101] In the brain age prediction scheme of this application, there is another hyperparameter, namely the artifact percentage level of the signal segment used in artifact detection 120. .

[0102] Hyperparameters In the process of building a neural network model, the cleaner the EEG signal data in the sample data, the better the accuracy of brain age prediction on that EEG signal data set. However, the robustness of the model will decrease accordingly, making it possible for the model to have larger prediction errors for EEG signal data outside the training sample data set or EEG signal data with a high proportion of artifacts. To mitigate this phenomenon, the model can be designed to allow a certain degree of artifacts in the input EEG signal segment data, which can improve the robustness of the neural network model to some extent. According to the embodiments of this application, hyperparameters characterizing the impact of EEG signal data quality are introduced. As mentioned above, the proportion of artifacts With hyperparameters and The difference is in hyperparameters It is used in both the training and prediction processes of the model (specifically in the signal segmentation process).

[0103] According to embodiments of this application, the training process can be configured using the following parameters:

[0104] Use the Adam optimizer , Learning rate The batch size (N) is 128. The learning rate is halved when the training loss drops to 4.0; it is halved again when the training loss continues to drop to 2.0; this process continues for another 100 epochs to complete the training. The loss function can be MAE Loss with L2 regularization, where the regularization coefficients are... The loss function is calculated using formula (16):

[0105] (16)

[0106] in, For the parameter set of a neural network model (e.g., BANet), y true y represents the actual physiological age, and y represents the predicted brain age. .

[0107] After the brain age data 104 is predicted by the neural network model 140, joint decision-making 150 can be used to select all EEG signals that meet the signal quality requirements. The predicted brain age values ​​corresponding to N segments of awake EEG signal data (each lasting N minutes) are statistically analyzed to obtain the final brain age prediction. The joint decision-making process 150 can fully utilize the information contained in the EEG signal data to improve the robustness of the brain age prediction results. For example, the statistical median of the predicted brain age values ​​corresponding to these signal segments can be used as the final brain age prediction result. Those skilled in the art will also realize that other statistical values, such as the average, can be used as the final brain age prediction result.

[0108] Figure 5 The figure shows the relative position of brain age values ​​predicted by a trained neural network model 140 based on the EEG signals of new subjects as sample data (sample points) to the brain age range of the normal population. The horizontal axis in the figure represents the physiological age of the subjects, and the vertical axis represents the predicted brain age of the subjects.

[0109] It should be noted that there is no objective evidence to suggest a necessary relationship between brain age and brain maldevelopment. However, a more objective and reliable statistical result can be obtained by statistically comparing the brain age distribution of subjects with brain maldevelopment with the brain age range of the normal population. That is, the statistical comparison results of this application can demonstrate the relative relationship between the predicted brain age of the sample and the normal brain age range.

[0110] Compared to healthy subjects, the brain age of subjects with brain diseases is more correlated with their chronological age, while the correlation between brain age and chronological age decreases in disease patients, for example, brain age usually decreases and becomes irrelevant. Current literature studies have found significant differences in brain age before and after medication in epilepsy patients. The applicant also conducted statistical analysis on epilepsy patients with different degrees of severity and found that brain age tends to decrease with increasing severity. In addition, the brain age of patients in the febrile seizure group and the developmental delay group was significantly older. That is to say, brain age is related to the type of subject (especially the patient) (disease type, disease duration) and also to the patient's condition. For example, medication can affect the patient's condition, and EEG signals can capture changes in condition. Brain age can be given a quantitative value based on the changed EEG signal morphology. This result is convenient for doctors to refer to, rather than being directly used for the diagnosis of brain diseases. This type of research requires the additional collection of EEG datasets from subjects with the corresponding diseases. The proposed solution can provide auxiliary tools for this research method and assessment technology. In short, different types of brain-related diseases have different patterns of brain age distribution. This morphological difference has been studied relatively extensively in subjects such as epilepsy patients. For example, the comparison and statistics of brain age data before and after medication, and before and after surgery, have shown significant differences in statistical results, which have a good inspirational effect.

[0111] Figure 3An illustrative flow diagram of a method for predicting brain age according to an embodiment of this application is shown.

[0112] The method mainly includes steps S310 for acquiring EEG signals of a subject in a conscious state, S320 for preprocessing the EEG signals, S330 for performing artifact detection and segmentation on the signals to select multiple EEG signal segments that meet the signal quality requirements, S340 for converting the multiple EEG signal segments that meet the signal quality requirements into time-frequency plot data, S350 for using a neural network model (e.g., BANet) to predict the brain age of the subject corresponding to the multiple EEG signal segments that meet the signal quality requirements based on the time-frequency plot data, and S360 for joint decision-making based on statistical data (e.g., statistical median) of the determined brain age.

[0113] In step S310, the method uses an electroencephalography (EEG) machine to acquire EEG signals from a conscious subject for a relatively short duration (e.g., more than 10 minutes).

[0114] Step S320 performs signal preprocessing on the raw EEG signal, as detailed above. Figure 1 The introduction section 110. Also in the following text, in conjunction with... Figure 1 and Figure 2 The process of the system algorithm described in this application will not be repeated.

[0115] Next, step S330 converts multiple EEG signal segments that meet the signal quality requirements into time-frequency plot data through artifact detection and signal segmentation. Specifically, step S330 includes a sub-step S331 for artifact detection and a sub-step S332 for signal segmentation (corresponding to respectively). Figure 1In sub-steps 120 and 130, signal quality requirement 301 is required. Signal quality requirement 301 mainly includes conditions such as signal amplitude, average signal amplitude, signal duration, and artifact percentage. Specifically, artifact detection sub-step S331 further includes sub-steps S3311 (quartile normalization), S3312 (determining signal amplitude to extract EEG signals related to the first type of artifact EEG signal), S3313 (high-pass filtering), S3314 (calculating average amplitude to extract EEG signals related to the second type of artifact EEG signal), and S3315 (merging artifact signals), corresponding to the processing described above for obtaining the time interval information corresponding to the EEG signal portion with artifacts in artifact detection. The sub-step S332 of signal segmentation selects multiple EEG signal segments from the EEG signal according to a time window with a predetermined time length and a sliding step size. Then, based on the time interval information and the proportion of artifacts in the EEG signal portion determined in sub-step S331, a clean EEG signal segment that meets the signal quality requirements is selected.

[0116] In step S340, the segmented EEG signal segments are processed by STFT to obtain signal segments in the form of time-frequency diagram data.

[0117] Next, the time-frequency graph data is used as input to the neural network model, and in step S350, methods such as... Figure 2 The model with the illustrated architecture predicts the brain age of the subjects. Specifically, this step includes, for example, using... Figure 2 The sub-step S351, which involves the input layer 200 receiving time-frequency plot data, and the optional sub-step S352, which performs dimensional rearrangement, for example using... Figure 2 The first sub-network 210 shown includes sub-step S353 for dimensionality reduction of frequency components and optional sub-step S354 for dimensionality merging, for example, using... Figure 2 The second subnetwork 220 shown in the figure extracts brain age-related features in sub-step S355, and performs an optional sub-step S356 of dimension exchange, for example, using... Figure 2 The third subnetwork 230 performs dimensionality reduction on the time-related feature components in substep S357, for example, using... Figure 2 The sub-step S358 of the linear and Softplus layers 240 shown performs numerical transformation and probability sorting of the feature data, and uses, for example... Figure 2 The output layer 250 will output the predicted brain age results in sub-step S359. For specific details of the above steps, please refer to the relevant sections of the model described above.

[0118] After obtaining the model prediction results, the method may also include an optional joint decision step S360, as described above. Figure 1As described in section 150.

[0119] Figure 4 This illustrates a device 400 for predicting brain age according to an embodiment of this application. The device 400 mainly includes an acquisition unit 410 for acquiring EEG signals of a subject in a conscious state, a prediction unit 420 for performing artifact detection and segmentation on the EEG signals to select multiple EEG signal segments that meet signal quality requirements, converting the multiple EEG signal segments that meet the signal quality requirements into time-frequency plot data, and using a neural network model to predict the subject's brain age corresponding to the multiple EEG signal segments that meet the signal quality requirements based on the time-frequency plot data, and an output unit 430 for outputting the brain age prediction result.

[0120] Furthermore, the prediction unit 420 can also implement the above-described... Figure 3 The specific steps of the method and / or at least include signal preprocessing, artifact detection, signal segmentation, Fourier transform, Figure 2 The related modules of the neural network model shown will not be described in detail here.

[0121] The proposed brain age prediction scheme is highly sensitive and can dynamically monitor brain changes. It can be used not only to assess brain development levels but also as a reference for rapid evaluation of drug efficacy in neurological diseases. This scheme improves the accuracy of brain age prediction by using EEG signals from awake subjects in a shorter timeframe and in real-time, compared to using EEG signals during sleep. This significantly reduces the detection time and cost, and makes the brain age prediction method more compatible with existing digital EEG and video EEG equipment and protocols. This enhances the clinical usability and applicability of the brain age prediction method in screening scenarios, making it particularly suitable for rapid screening applications.

[0122] It should be noted that although several modules or units of the system for predicting brain age have been mentioned in the detailed description above, this division is not mandatory. In fact, according to embodiments of this application, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided into multiple modules or units. Components shown as modules or units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without any inventive effort.

[0123] In exemplary embodiments of this application, a computer-readable storage medium is also provided, on which a computer program is stored, the program including executable instructions that, when executed by, for example, a processor, can implement the steps of the method for predicting brain age described in any of the above embodiments. In some possible implementations, various aspects of this application can also be implemented as a program product including program code that, when run on a terminal device, causes the terminal device to perform the steps described in this specification for the method of predicting brain age according to various exemplary embodiments of this application.

[0124] The program product for implementing the above-described method according to embodiments of this application may employ a portable compact disc read-only memory (CD-ROM) and include program code, and may run on a terminal device, such as a personal computer. However, the program product of this application is not limited thereto. In this document, a readable storage medium may be any tangible medium that contains or stores a program that may be used by or in conjunction with an instruction execution system, apparatus, or device.

[0125] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0126] The computer-readable storage medium may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The readable storage medium may also be any readable medium other than a readable storage medium, capable of transmitting, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0127] Program code for performing the operations of this application can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0128] In an exemplary embodiment of this application, an electronic device is also provided, which may include a processor and a memory for storing executable instructions of the processor. The processor is configured to perform the steps of the method for predicting brain age in any of the above embodiments by executing the executable instructions.

[0129] Those skilled in the art will understand that various aspects of this application can be implemented as a system, method, or program product. Therefore, various aspects of this application can be specifically implemented in the following forms: a completely hardware implementation, a completely software implementation (including firmware, microcode, etc.), or a combination of hardware and software implementations, collectively referred to herein as a "circuit," "module," or "system."

[0130] The following reference Figure 6 To describe an electronic device 600 according to this embodiment of the present application. Figure 6 The electronic device 600 shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0131] like Figure 6 As shown, the electronic device 600 is presented in the form of a general-purpose computing device. The components of the electronic device 600 may include, but are not limited to: at least one processing unit 610, at least one storage unit 620, a bus 630 connecting different system components (including storage unit 620 and processing unit 610), a display unit 640, etc.

[0132] The storage unit stores program code that can be executed by the processing unit 610, causing the processing unit 610 to perform the steps described in the method for predicting brain age according to various exemplary embodiments of this application. For example, the processing unit 610 can perform actions such as... Figure 3The steps are shown in the figure.

[0133] The storage unit 620 may include a readable medium in the form of a volatile storage unit, such as a random access memory unit (RAM) 6201 and / or a cache storage unit 6202, and may further include a read-only memory unit (ROM) 6203.

[0134] The storage unit 620 may also include a program / utility 6204 having a set (at least one) program module 6205, such program module 6205 including but not limited to: an operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0135] Bus 630 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the various bus structures.

[0136] Electronic device 600 can also communicate with one or more external devices 700 (e.g., keyboard, pointing device, Bluetooth device, etc.), and with one or more devices that enable a user to interact with electronic device 600, and / or with any device that enables electronic device 600 to communicate with one or more other computing devices (e.g., router, modem, etc.). This communication can be performed via input / output (I / O) interface 650. Furthermore, electronic device 600 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 660. Network adapter 660 can communicate with other modules of electronic device 600 via bus 630. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 600, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0137] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, or network device, etc.) to execute the method for predicting brain age according to the embodiments of this application.

[0138] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the appended claims.

Claims

1. A method for predicting brain age, comprising: Acquire electroencephalogram (EEG) signals of subjects while they are awake; The EEG signal is subjected to artifact detection and segmentation to select multiple EEG signal segments that meet the signal quality requirements; The multiple EEG signal segments that meet the signal quality requirements are converted into time-frequency plot data; and The brain age of the subject is predicted using a neural network model based on the time-frequency plot data, corresponding to multiple EEG signal segments that meet the signal quality requirements. The neural network model includes an input layer, a first sub-network including a trainable nonlinear triangular filter layer, a second sub-network, a third sub-network including a self-attention layer, and an output layer. The neural network model is used to predict the brain age corresponding to multiple EEG signal segments that meet the signal quality requirements based on the time-frequency plot data, including: The time-frequency graph data is received using the input layer; The first sub-network is used to perform dimensionality reduction on the frequency dimension data components in the time-frequency graph data. The second sub-network is used to extract feature data related to the brain age of the subject from the dimensionality-reduced time-frequency plot data; The third sub-network is used to perform dimensionality reduction on the time dimension feature components in the feature data. The output layer outputs a predicted brain age corresponding to each of the multiple EEG signal segments that meet the signal quality requirements; and The statistical median of the predicted brain age corresponding to each of the plurality of EEG signal segments is determined as the final predicted brain age of the subject.

2. The method according to claim 1, characterized in that, The signal quality requirements include at least one of the following conditions: signal amplitude, average signal amplitude, signal duration, and artifact percentage.

3. The method according to claim 1, characterized in that, The artifact detection of the EEG signal includes: The EEG signal is subjected to quartile normalization to generate a first EEG signal; Based on the signal amplitude of the first EEG signal, extract the second EEG signal that satisfies the signal amplitude condition; The first EEG signal is high-pass filtered to generate the third EEG signal; Calculate the average amplitude of the third EEG signal over the signal time length to extract the fourth EEG signal that satisfies the average signal amplitude condition; The time interval of the EEG signal with artifacts is determined based on the second EEG signal and the fourth EEG signal.

4. The method according to claim 3, characterized in that, Segmenting the EEG signal includes: Multiple EEG signal segments are selected from the EEG signal based on a time window of predetermined length and a sliding step size; Based on the determined time interval and artifact ratio of the EEG signal with artifacts, select the multiple EEG signal segments that meet the signal quality requirements from the multiple EEG signal segments.

5. The method according to claim 1, characterized in that, Converting the multiple EEG signal segments that meet the signal quality requirements into time-frequency plot data includes: The multiple EEG signal segments that meet the signal quality requirements are converted into time-frequency plot data using a short-time Fourier transform (STFT), wherein a Hanning window function is used in the STFT.

6. The method according to claim 1, characterized in that, The first subnetwork is defined by the following formula: （1） （2） Where x is the output of the first sub-network. S represents the time-frequency graph data input to the first sub-network. , Let be the transformation matrix of the trainable nonlinear triangular filter layer. , For a trainable parameter matrix, , Let be a nonnegative function that operates on the elements of a matrix. It is a non-negative constant matrix. , The expression represents matrix multiplication, where C is the number of lead channels in the EEG signal, N is the number of multiple EEG signal segments that meet the signal quality requirements, F is the number of data components in the frequency dimension of the time-frequency plot data, F0 is the number of data components in the frequency dimension of the dimensionality-reduced time-frequency plot data, and K is the number of data components in the time dimension of the time-frequency plot data. .

7. The method according to claim 6, characterized in that, The nonnegative constant matrix is ​​calculated as follows: : Based on formula (3), the frequency of the time-frequency graph data is... Frequency converted to logarithmic space ,in and These represent the minimum and maximum values ​​of the frequencies in the time-frequency graph data, respectively. and These are the minimum and maximum values ​​of the frequencies in the logarithmic space, respectively: （3） Based on formulas (4) and (5), the frequency range in logarithmic space is... Divide the intervals into F+1 equal subintervals: （4） （5） Determine according to formula (6) : （6） Where m is the triangular filter index, corresponding to the data components of the m original frequency dimensions; k is a discrete index variable in the frequency dimension, representing a nonnegative constant matrix. Column index; f(m) represents the center frequency of the m-th triangular filter, and f(m-1) and f(m+1) represent the center frequencies of the triangular filters adjacent to the m-th triangular filter, respectively. Represents a nonnegative constant matrix The element in the m-th row and k-th column.

8. The method according to claim 1, characterized in that, Before inputting the time-frequency graph data into the first sub-network, the method further includes: The time-frequency graph data is rearranged in dimensions.

9. The method according to claim 1, characterized in that, Before inputting the dimensionality-reduced time-frequency plot data into the second sub-network, the following steps are also included: The data components related to the lead channels of the EEG signal and the data components of the frequency dimension in the dimensionality-reduced time-frequency plot data are merged.

10. The method according to claim 1, characterized in that, The second subnetwork includes at least two sequentially connected GRU subnetworks, each GRU subnetwork including a sequentially connected GRU network layer and a linear layer.

11. The method according to claim 1, characterized in that, Before inputting the extracted brain age-related feature data of the subject into the third sub-network, the method further includes: The time dimension feature components in the feature data are swapped with the feature components associated with the multiple EEG signal segments that meet the signal quality requirements.

12. The method according to claim 1, characterized in that, The dimensionality reduction operation of the time dimension feature components in the feature data from the second subnetwork using the third subnetwork includes: The self-attention layer weights the feature data based on the feature components of the time dimension of the feature data.

13. The method according to claim 1, characterized in that, The neural network model also includes linear and Softplus layers between the third sub-network and the output layer.

14. The method according to claim 1, characterized in that, During the training phase of the neural network model, the amount of training sample data is increased by random jittering and / or random scaling.

15. The method according to claim 14, characterized in that, The random jitter includes adding a floating amount to the age of the sample data. ,and ,in The hyperparameters related to the age error of the subjects, and The "~" symbol indicates obedience. Indicates interval [ A uniform distribution on [the surface].

16. The method according to claim 14, characterized in that, The random scaling includes scaling the time-frequency plot data of the EEG signal segments of the sample data. times, and ,in The hyperparameters are related to the sampling error of the EEG signal, and The "~" symbol indicates obedience. Indicates interval [ A uniform distribution on [the surface].

17. The method according to claim 4, characterized in that, The artifact proportion condition includes: in an EEG signal segment, the artifact proportion of the time interval of the EEG signal with artifacts to the entire time interval of the EEG signal segment is less than 1%. ,in These are hyperparameters related to the quality of the EEG signal, and .

18. The method according to claim 1, characterized in that, Before performing artifact detection and segmentation on the EEG signal, the following steps are also included: The EEG signal is preprocessed, and the preprocessing includes at least one of downsampling, rereference, notch filtering to remove power frequency, and bandpass filtering.

19. The method according to claim 18, characterized in that, The rereference in the preprocessing uses the average value of the original EEG signals from the 21 lead channels as the reference signal.

20. A device for predicting brain age, comprising: The acquisition unit is configured to acquire electroencephalogram (EEG) signals of a subject in a conscious state; The prediction unit is configured to perform artifact detection and segmentation on the EEG signal to select multiple EEG signal segments that meet signal quality requirements; convert the multiple EEG signal segments that meet signal quality requirements into time-frequency plot data; and use a neural network model to predict the brain age of the subject corresponding to the multiple EEG signal segments that meet signal quality requirements based on the time-frequency plot data. An output unit, configured to output the brain age, The neural network model includes an input layer, a first sub-network including a trainable nonlinear triangular filter layer, a second sub-network, a third sub-network including a self-attention layer, and an output layer. The neural network model is used to predict the brain age corresponding to multiple EEG signal segments that meet the signal quality requirements based on the time-frequency plot data, including: The time-frequency graph data is received using the input layer; The first sub-network is used to perform dimensionality reduction on the frequency dimension data components in the time-frequency graph data. The second sub-network is used to extract feature data related to the brain age of the subject from the dimensionality-reduced time-frequency plot data; The third sub-network is used to perform dimensionality reduction on the temporal dimension feature components in the feature data; and The output layer outputs a predicted brain age corresponding to each of the multiple EEG signal segments that meet the signal quality requirements; and The statistical median of the predicted brain age corresponding to each of the plurality of EEG signal segments is determined as the final predicted brain age of the subject.

21. The device according to claim 20, characterized in that, The prediction unit is configured to perform the steps of the method according to any one of claims 2 to 19.

22. A computer-readable storage medium having a computer program stored thereon, the computer program including executable instructions that, when executed by a processor, implement the method according to any one of claims 1 to 19.

23. An electronic device, characterized in that, include: processor; as well as Memory for storing the executable instructions of the processor; The processor is configured to execute the executable instructions to implement the method according to any one of claims 1 to 19.

Images