Brain state estimation device, computer program, brain state estimation method, brain function testing system and method
A time series neural network-based system processes MEG and EEG data to classify and diagnose brain abnormalities by analyzing brain activity patterns, offering high precision in distinguishing healthy and diseased states.
Patent Information
- Application Number
- JP2021552465
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2019-10-18
- Filing Date
- 2020-10-16
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2040-10-16
AI Technical Summary
Existing brain function tests lack appropriate stimulation, measurement, and analysis methods to accurately classify and analyze brain activity data, particularly in response to stimuli, for diagnosing conditions like dementia.
A system utilizing a time series neural network to classify and estimate brain activity data, incorporating MEG and EEG data, which includes preprocessing, signal source estimation, and spectral analysis to differentiate between healthy and diseased brain states.
Enables precise classification and diagnosis of brain abnormalities, such as dementia, by accurately distinguishing brain activity patterns and functional connectivity, providing high spatial and temporal resolution.
Smart Images

Figure 0007591769000009 
Figure 0007591769000010 
Figure 0007591769000011
Abstract
Description
[Technical field]
[0001] The present invention relates to a brain state estimation device, a computer program, a brain state estimation method, and a brain function testing system and method. [Background technology]
[0002] Attempts are being made to examine the brain function of subjects (examinees, patients), including possible dementia, by capturing the state of the brain in response to stimuli via external signals. Summary of the Invention [Problem to be solved by the invention]
[0003] Brain function tests generally involve presenting stimuli to subjects and measuring their brain responses. It is necessary to select appropriate stimuli, measurements, and analyses for each test item, and to find appropriate solutions for the data obtained. [Means for solving the problem]
[0004] One aspect of the present invention is a system including a function of classifying brain activity data of a subject, which varies in a space including a time axis, using a time series model that has learned learning data including a large number of solutions related to brain function tests. An example of a time series model is a time series neural network including a recurrent neural network (RNN). Another aspect is a system including a function of estimating a spatial distribution including a time axis of a signal source indicated by activity data from brain activity data of a subject, which varies in a space including a time axis, using a model that has learned solutions related to a large number of inputs and outputs including brain activity regions and patterns. Another aspect is a system including a function of estimating a brain state of a subject using a model that has learned a spatial distribution including a time axis of a signal source estimated from brain activity data of a subject, which varies in a space including a time axis, and a case. These models may include a time series neural network. The brain activity data may be data at rest, and may include brain activity data in response to a stimulus including at least one of vision, hearing, touch, taste, and smell. The present invention includes a program having instructions for operating a computer as the above system.
[0005] Another embodiment of the present invention is a method including classifying brain activity data of a subject that varies in a space including a time axis using a time series model that has been trained with training data including a number of solutions related to brain function tests. This method may include estimating a spatial distribution including the time axis of a signal source indicated by the brain activity data of a subject that varies in a space including a time axis using a model that has been trained with a number of solutions related to inputs and outputs including brain activity regions and patterns, from the brain activity data of a subject that varies in a space including a time axis, and may include estimating a brain state of a subject using a model that has been trained with a spatial distribution including the time axis of a signal source estimated from the brain activity data of a subject that varies in a space including a time axis and cases. [Brief description of the drawings]
[0006] [Figure 1] A diagram showing the MEG measurement method and the technique for estimating the location of activated areas. [Diagram 2] FIG. [Diagram 3] A diagram showing the frequency band characteristics of EEG / MEG for reference. [Figure 4] 1 is a block diagram showing an example of the configuration of a brain state estimation device according to an embodiment of the present invention. [Diagram 5] FIG. 2 is a schematic diagram showing a configuration of epoch data. [Figure 6] FIG. 13 is a schematic diagram showing an example of an averaging process. [Figure 7] FIG. 1 is a schematic diagram showing the localization of brain surface activity distribution with head rotation relative to the center position of the MEG sensor map. [Figure 8] 1 is a schematic diagram showing a method of calculating a magnetic field distribution formed between an MEG sensor and a current dipole. FIG. [Figure 9] FIG. 2 is a schematic diagram showing an example of the results of the magnetic field distribution (Source Space Data) calculated according to the present embodiment. [Figure 10] FIG. 13 is a schematic diagram showing an example of an average result in the time domain of two samples with different shifts in periodicity. [Figure 11] FIG. 2 is a schematic diagram showing an example of the PSD of an autocovariance function calculated using the time domain T. [Figure 12] FIG. 1 is a schematic diagram showing an example of the difference in PSD between dementia patients and healthy individuals. [Figure 13] FIG. 1 is a schematic diagram showing an example of the phase synchronization distribution of magnetic fields in the head of a dementia patient and a healthy subject using cross-spectrum. [Figure 14] FIG. 1 is a schematic diagram showing an example of entropy distribution of subjects by age group. [Figure 15] FIG. 13 is a schematic diagram showing an example of the relationship between the Mahalanobis distance and the abnormality probability. [Figure 16] FIG. 1 is a schematic diagram showing the statistical properties (unbiasedness) of entropy sampled from each channel. [Figure 17] FIG. 13 is a schematic diagram showing an example of the result of an outlier test using entropy sampled from each channel. [Figure 18] FIG. 13 is a schematic diagram showing a first example of a brain state display screen. [Figure 19]FIG. 13 is a schematic diagram showing a second example of the brain state display screen. [Figure 20] 13 is a flowchart showing an example of a procedure of brain state estimation processing by the brain state estimation device. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0007] As an example, we will explain the classification of brain activity by neural network using magnetoencephalography. Magnetoencephalography (MEG) is the magnetic field activity in the brain recorded by a brain magnetometer (abbreviated as magnetoencephalography). In contrast to the widely used electroencephalography (EEG), MEG is essentially a different way of searching for the same phenomenon. When a part of the apical dendrite of a pyramidal cell in the cerebral cortex becomes excited and depolarized, a current flows outside and inside the cell. EEG records this extracellular current, while MEG records the magnetic field that surrounds the intracellular current. MEG records the response from the brain like EEG, and since there is no need to directly stimulate or load the brain, it can be said to be an extremely safe examination method.
[0008] The greatest advantage of MEG compared to EEG is its high spatial resolution. Between the brain and scalp there are three layers with significantly different electrical conductivities: cerebrospinal fluid, skull, and skin. Therefore, the electric field generated in the brain is greatly affected by these layers, so it is difficult to accurately estimate the active areas of the brain from EEG electrodes placed on the scalp unless special estimation methods (e.g., dipole tracking method) are used. However, in the case of MEG, the magnetic field is not affected by electrical conductivity, so if the recording conditions are good, it is possible to estimate the active areas with an accuracy of mm.
[0009] Similar to MEG and EEG, functional MRI (fMRI) is a method for measuring brain activation and its regions. However, compared to fMRI, MEG has the advantage that it directly records the electrical activity of cells rather than changes in local cerebral blood flow, and like EEG, it has a high time resolution of milliseconds.
[0010] This technology development proposes to use a time series model to classify, estimate, and analyze brain activity data obtained as data that varies in space including time. MEG is currently the best method for obtaining brain activity data, but data obtained by EEG may also be used. For example, a neural network is used to analyze MEG data, and based on brain activation information, it is possible to distinguish with high accuracy the state of various brain diseases and the state of brain activity in healthy people. It is also possible to create a method using a neural network for pre-analysis processing such as signal source estimation.
[0011] As shown in FIG. 1, a process 100 may be provided for classifying source level data using a time-series NN (neural network). The time-series NN includes a recurrent neural network, etc. Also, a signal source estimation method may be provided that includes a new process 200 for preparing a large number of solution models and learning a network that reproduces input and output. Also, a process 300 may be provided for discriminating signal source estimation data using a NN. The subject to be discriminated may include brain function, and the presence or absence of symptoms related to brain abnormality may be discriminated. For example, in the case of dementia, classification into NC / MCI / AD may be performed. The brain activity data may be data at rest, or may be brain activity data in response to a stimulus including at least one of vision, hearing, touch, taste, and smell.
[0012] Figure 2 shows an example of an experiment in which resting-state data from a healthy subject was classified into spontaneous eye closure and spontaneous eye open using a ResNet model. In this way, by using a model that has learned brain activity data related to eye open / closed states, data related to eye opening and closing can be extracted from the brain activity data and analyzed.
[0013] FIG. 3 shows, for reference, the radio wave intensity for each frequency band of EEG related to eye opening and closing, and also shows the difference in radio wave intensity for each frequency band between healthy subjects and dementia patients. Therefore, by acquiring brain activity data that varies depending on time and using artificial intelligence that has learned knowledge of appropriate frequency bands and knowledge of generation sources, a method and system can be provided that accurately judges the health of brain function or changes in brain function over time for each patient. In the present invention, the space including time is a concept that also includes a space in which a function of the time domain is converted into a frequency space by Fourier transform or the like, and relates to a system and method that handles brain activity data obtained by dynamic images or signals, rather than static images or signals.
[0014] By calculating the magnetic field changes and electric potential changes in the brain, it is possible to calculate the network (also called "functional connectivity") composed of electrophysiologically active regions. Functional connectivity is expressed as the change in the summed electrical phenomena (EPSP: excitatory postsynaptic potential, IPSP: inhibitory postsynaptic potential) generated by the network in the brain composed of neuronal groups. The functional connectivity is the result of calculating the magnetic field vector strength or electric potential impedance vector input and output for each region of the anatomical neural projection pathway. Functional connectivity is a type of classification of brain connectivity that expresses statistical relationships based on the time-dependent information of electrical activity projected to neuronal groups. Functional connectivity expresses the relationships between different regions in the brain as a network by capturing the entire brain from the perspective of brain function. To calculate functional connectivity, source-level analysis (particularly minimum norm estimation (MNE)) is used to estimate the activity distribution on the brain surface using the magnetic field measured from the MEG sensor (magnetic flux detection coil) of a biomagnetic measurement device (MEG measurement device). In MEG analysis, functional connectivity calculated from the brain magnetic field distribution observed when the subject is in a resting state is called the default mode network (DMC). DMC using MNE means the distribution of neural activity distributed synchronously in different brain regions. Below, we explain the method for estimating the brain state (a method for automatically classifying the difference from healthy subjects).
[0015] 4 is a block diagram showing an example of the configuration of brain state estimation device 50 of this embodiment. Brain state estimation device 50 includes a control unit 51 that controls the entire device, a data acquisition unit 52, a storage unit 53, a preprocessing unit 54, an SSP processing unit 55, a signal source position estimation unit 56, a signal distribution calculation unit 57, a spectrum analysis unit 58, an entropy calculation unit 59, a brain state estimation unit 60, a display panel 61, an operation unit 62, and a display processing unit 63.
[0016] The control unit 51 can be configured with a central processing unit (CPU), a read only memory (ROM), a random access memory (RAM), and the like. The storage unit 53 can be configured with a semiconductor memory, a hard disk, and the like. In the following, magnetic field data (hereinafter also referred to as "MEG data") measured by an MEG sensor (detection coil) of a biomagnetic measuring device will be described as an example of brain data of a subject that varies in space including a time axis. The brain data means brain activity data, electrooculogram data (EOG), electrocardiogram data (ECG), electromyogram data (EMG), reference noise data, and the like, and the brain data may include brain activity data, or may include at least one of EOG, ECG, EMG, and reference noise in addition to the brain activity data. For example, when analyzing a brain state, at least one of EOG, ECG, EMG, and reference noise can be used in addition to the brain activity data. Also, when analyzing using a difference in a brain state, brain activity data is essential. The brain activity data is not limited to magnetic field data, and may be data obtained by EEG.
[0017] The data acquiring unit 52 can acquire brain data of the subject that varies in space including a time axis. More specifically, the data acquiring unit 52 can acquire MEG data as brain activity data. The biomagnetic measurement device includes a large number (for example, 160 locations, etc.) of MEG sensors arranged on the scalp, each MEG sensor corresponding to each channel, and the data acquiring unit 52 can acquire MEG data for each of the multiple channels. The MEG data can be, for example, data obtained when the subject is in a recumbent position with eyes closed in a resting state. In addition, the data acquiring unit 52 can acquire all or a part of data other than the MEG data, such as an electrooculogram (EOG), an electrocardiogram (ECG), an electromyogram (EMG), and reference noise data. In addition to the MEG data, or instead of the MEG data, the data acquiring unit 52 can acquire an electroencephalogram (EEG).
[0018] The pre-processing unit 54 performs processes such as removing data from bad channels, baseline correction, hand-pass filtering, notch filtering, averaging, and independent component analysis (ICA). The measured magnetic field data includes not only data derived from the magnetic field, but also signals derived from environmental noise, head movement, and biological noise. The pre-processing unit 54 has a function of removing signals other than the magnetic field data.
[0019] To remove data from defective channels, MEG data from channels that could not be measured is removed from the data to be analyzed. Channels from which MEG data acquisition failed are identified by eigenvalue decomposition of the MEG data group, and data from channels that show a predetermined threshold value are removed. The removed channel data can be interpolated with data from channels that are not to be removed.
[0020] Baseline correction (also called "bias correction") uses the average value of the time series data recorded in each channel of MEG data and subtracts that average value from the time series data recorded in each channel. This is to convert MEG data into relative data, since the absolute values of signal data vary depending on temperature, air pressure, and individual differences.
[0021] The bandpass filter process uses a bandpass filter with a bandwidth of a desired frequency band to remove signals in an unnecessary frequency band from the MEG data. The desired frequency band can be, for example, 0 to 70 Hz. This frequency band includes electrophysiological electrical signals of the brain composed of neural cell populations, such as alpha waves (8 to 13 Hz), beta waves (14 to 30 Hz), theta waves (4 to 7 Hz), delta waves (0.5 to 3 Hz), and gamma waves (26 to 70 Hz).
[0022] The notch filter process uses a notch filter to remove power-source-derived periodic noise (50 Hz or 60 Hz) that is mixed in from the power sources of other devices used when operating the biomagnetic measurement device.
[0023] In order to remove noise components contained in the MEG data, the averaging process divides the recorded time series data into, for example, 5-second intervals (trial), and then performs an averaging of each divided data to attenuate the noise components (epoch data).
[0024] Fig. 5 is a schematic diagram showing the configuration of epoch data. As shown in Fig. 5, MEG data of each of the multiple channels is divided into multiple sections (trial1, trial2, ..., trialn). The duration of each section can be, for example, 5 seconds, but is not limited to this. The number of sections n can be determined appropriately.
[0025] FIG. 6 is a schematic diagram showing an example of averaging. In averaging, the signals of trials 1 to n are added for each trial. Signals observed in common to all trials are emphasized by the addition. On the other hand, waveforms observed randomly in each trial (for example, noise waveforms) are offset by the addition, and the noise waveform is reduced.
[0026] Independent component analysis is a method to separate a single signal into independent signal sources that generate it. The separated signals include MEG data, environmental noise, and EMG data. Independent component analysis uses two pieces of information, a reference channel that measures environmental noise originating from the indoor environment, and an EMG channel that measures eye movement and muscle activity, to remove noise components (e.g., noise synchronized with EMG and the reference) that are mixed into the data measured by each MEG sensor located on the head. Specifically, random signals, artifacts associated with head movement, and signals originating from biological noise are removed from magnetic field data measured for a single subject to obtain multiple signal components. Based on a correlation coefficient or a threshold, ICA elements of EMG-EMG and EMG-MEG combinations that are defined as artifacts are removed from each channel.
[0027] The SSP processing unit 55 reduces the influence of noise from the signal source from the MEG data by using the SSP (Signal Space Projection) method. The SSP method will be described below.
[0028] Assuming that the signal y(t) measured by each MEG sensor contains the influence of the original signal ys(t) and noise ε, it can be expressed by equation (1).
[0029]
number
[0030] At this time, the original signal ys(t) at time t is composed of various signal sources (signal sources with indefinite rotation and position) and the lead field l generated for each signal source. A signal source Sq shows a unique distribution for the MEG channel and is expressed by equation (2). Here, q is the channel index, Q is the number of channels, and lq is the signal state of channel q. The signal source Sq is unknown. The noise P that affects all of the original signal ys(t) and the effect of noise ε can be expressed by equation (3). When the effect of noise P is minimized, equations (1) to (4) can be defined.
[0031] The noise effect P can be minimized by considering the orthonormal basis of Uq shown in equation (5) and replacing it with signal source Sq = Cq and lead field lq = Uq as in equation (6). When the number of data samples is finite, the basis vector u consisting of the orthonormal basis of Uq becomes the maximum likelihood estimate of the eigenvector corresponding to the maximum eigenvalue consisting of Q elements of the data covariance matrix. To find equation (6), the original signal ys(t) can be found by using singular value decomposition from the data covariance matrix. This original signal ys(t) is called sensor space data.
[0032] Random data independent of the magnetic field distribution constructed from theoretical voltage values and EMG+ECG data measured in an independent environment were mixed into the magnetic field data, and the above-mentioned ICA and SSP processing was performed. The calculation result of GOF (Goodness of Fit), an index for evaluating the degree of agreement between the measured value and the theoretical value, was high at 97%. GOF is calculated as follows: GOF = {1-Σ(measured value-theoretical value) 2 / Σ(measurement value 2 )}×100[%]. Here, Σ means that the sum is taken for each sensor. When GOF is 100%, the two match perfectly, and the smaller the value is below 100%, the lower the match becomes. When ICA processing was used alone, the GOF was low at 60%.
[0033] As described above, in this embodiment, MEG data can be reconstructed after removing uncorrelated signals from the time series data of each channel and uncorrelated signals between channels. The applicant has focused on the fact that in a healthy person, functional connectivity in the brain is related and synchronized, whereas in a patient with a brain disease such as dementia, the synchrony of functional connectivity is gradually lost. According to this embodiment, by removing uncorrelated signals between channels, it is possible to emphasize the correlation components of the signals between channels, and to separate (screen) healthy people from patients with brain diseases by using the reconstructed MEG data. Specific examples of brain diseases include, for example, dementia (including AD, DLB, frontotemporal lobar degeneration (FTLD), normal pressure hydrocephalus (NPH), etc.), brain tumors, mental disorders (also called mental disorders, including schizophrenia, epilepsy, mood disorders, addiction disorders, higher-level functional disorders, etc.), Parkinson's disease, Asperger's syndrome, attention deficit hyperactivity disorder (ADHD), sleep disorders, pediatric diseases, ischemic brain disorders, mood disorders (including depression, etc.), etc.
[0034] The signal source location estimation unit 56 estimates the location of the signal source in the brain using the sensor space data obtained by the SSP processing unit 55. Here, the signal source is a current dipole that represents the synchronous activity of many neurons. The magnetic field depends on the distance function. The measurement data of the MEG sensor is strongly influenced by the distance between the MEG sensor and the head.
[0035] Figure 7 is a schematic diagram showing the localization of brain surface activity distribution with head rotation relative to the MEG sensor map center position. In the figure, the outline shows the head range rotated and scaled according to the head center position and range. The mesh within the outline shows the MNI template rotated and scaled according to the head center position and range. The left and right parts within the outline show the area with the shortest distance between the head and the MEG sensor position. The points around the outline show the positions of each MEG sensor.
[0036] To reduce the effect of the distance between the MEG sensor and the head, it is necessary to define the sensor position taking into account the head position. The center position of the head is set from the spatial positions of the nose center, right ear, and left ear recorded in the MEG data, and the head model (MNI template) is rotated and expanded using affine transformation so that the position of each MEG sensor is set corresponding to the set center position. Based on the shape of this head model and the sensor distance, the sensor space data obtained by the SSP processing unit 55 is used to estimate the position of the signal source in the brain using the Biot-Savart equation, which is a magnetic field model equation passing through a medium (dipole estimation by forward problem).
[0037] The signal source position estimation unit 56 reconstructs, using the Biot-Savart equation again from the current components at the estimated intracerebral signal source position, magnetic field data that is specifically attenuated or amplified depending on the distance to the sensor located outside the head, and generates reconstructed sensor space data.
[0038] The signal distribution calculation unit 57 calculates a magnetic field distribution (Source Space Data) by using the sensor space data reconstructed by the signal source position estimation unit 56. In this embodiment, the magnetic field distribution (Source Space Data) is calculated by using an adaptive beamformer (SAM).
[0039] Figure 8 is a schematic diagram showing a method for calculating the magnetic field distribution formed between the MEG sensor and the current dipole. When the position and rotation in the brain of the dipole Jo with a current component are given (r, rr′), the MEG magnetic field data can be calculated using the Biot-Savart equation shown in the figure, which is the distribution A(r) of the magnetic field until it reaches the MEG sensor that is measuring externally. SAM estimates the existence of the current dipole Jo from the magnetic field data recorded by each MEG sensor using the Biot-Savart equation, and then calculates the magnetic field distribution (Source Space Data) measured by the external sensor (Sensor Space Data) or at the boundary surface of the head surface (circular area in the figure) using mathematical programming. Using this method, it is possible to reduce signals other than the magnetic field distribution contained in the measured magnetic field data and improve the spatial SNR.
[0040] 9 is a schematic diagram showing an example of the result of the magnetic field distribution (Source Space Data) calculated by the present embodiment. As shown in the figure, the magnetic field distribution (Source Space Data) calculated using the adaptive beamformer (SAM) improves the spatial SNR and enables accurate evaluation of adjacent neural activity sources.
[0041] Next, the spectrum analysis will be described.
[0042] The spectrum analysis unit 58 can convert the brain data acquired through the data acquisition unit 52 to generate brain-related information. For example, the spectrum analysis unit 58 can generate brain-related information including rhythmic frequency band information by using frequency conversion. The rhythmic frequency band information is information that focuses on the fact that brain function information has a certain periodic structure. The spectrum analysis unit 58 can also generate brain-related information including fractal information by using conversion by nonlinear time series analysis. The fractal information is information that focuses on the randomness contained in the brain function information, and means the degree to which the similar structure of the brain function information is broken. Specifically, the spectrum analysis unit 58 can acquire magnetic field data (time series data of each MEG channel) obtained by the signal source position estimation unit 56 or the signal distribution calculation unit 57, and perform a single-channel spectrum analysis and a spectrum analysis between two channels. The spectrum analysis can be switched by specifying a setting file that includes calculation conditions. In the following, the single-channel spectrum analysis will be described first.
[0043] According to the Wiener-Finchin theorem (the Fourier transform of the autocovariance function is the spectrum, and the inverse Fourier transform of the spectrum is the autocovariance function), the spectrum is analyzed (correlation method) as shown in equation (8) using the Fourier transform of the autocovariance function ρ(τ) expressed in equation (7). Here, y(t) (t=0, 1, …, N) indicates the signal sequence, x(t) indicates the signal sequence discretely sampled for τ indexes, and k indicates the number of sampling indexes (k=0, 1, …, N-1).
[0044]
number
[0045] Generally, since spectrum estimation is performed on finite data, there is a problem that the influence of the edge of the data appears as spectral leakage from the frequency where the estimated spectrum peak exists to other frequencies. It is known that the spectrum does not completely represent the true physical phenomenon due to the spectral leakage contained in the observed spectrum. For this reason, although the error in the spectrum estimation can be reduced by the type of window function or smoothing process, the reality is that the error in the estimated spectrum is large.
[0046] By using the correlation method shown in equation (8), it is possible to quantitatively evaluate the correlation at each time point (t+τ) that is τ away from a certain time t in an arbitrarily set time domain. In other words, the Blackman-Tukey method, which obtains a spectrum from an autocorrelation function, is designed to obtain fluctuations in a finite time domain based on the error theory of the spectrum. A set element consisting of a random sequence in a finite time domain has an individual occurrence probability for each element. If the random variables in each time domain, which are components of a stochastic process, are distributed independently, all intertemporal autocorrelations will be zero. When performing averaging processing on MEG time series data, signals with shifted periodic phases are simultaneously averaged in order to set an arbitrary time domain, and the autocorrelation may be small and may not have a discrete-time Fourier transform (DTFT).
[0047] Figure 10 is a schematic diagram showing an example of the average result in the time domain of two samples with different shifts in periodicity. As shown in Figure 10, if statistics cannot be estimated by using a stochastic process that includes multiple samples, it is necessary to change the time domain treated as the sample. If all statistics can be estimated consistently and unbiasedly from any one sample process x(t), that is, if ergodicity is satisfied, the statistics can be shown. Therefore, it must be shown that ergodicity is guaranteed by changing the time domain treated as the sample.
[0048] In this embodiment, the time domain T that guarantees ergodicity for the mean μ and the autocovariance function ρ(τ) (autocovariance) is automatically determined. The mean in the time domain is expressed by equation (9), the mean ergodicity of the time series data is expressed by equation (10), and the zero convergence condition of the variance indicating a stationary process is expressed by equation (11). In equation (11), C x (t1-t2) denotes the autocovariance function in the frequency domains t1 and t2.
[0049]
number
[0050] In particular, a time domain T that satisfies the two conditions of convergence to the average value and variance convergence to zero is searched for, and the power spectral density (PSD) can be estimated using the autocovariance function ρ(τ) = P(ω) in that time domain T. PSD calculations using an automatically determined time domain T are performed as follows. If the rectangular sampling of time series data with a certain time domain T is given by equation (12), the PSD at this time can be found by equation (13). Here, E represents the time average of xτ(t)xτ(t+τ) of the time series data xτ in the τ index domain normalized by equation (12). τ represents the sampling width.
[0051]
number
[0052] Fig. 11 is a schematic diagram showing an example of the PSD of an autocovariance function calculated using the time domain T. In the figure, the horizontal axis indicates frequency, and the vertical axis indicates PSD (power spectral density). The solid line indicates the theoretical value, and the dashed line indicates the value obtained using the Blackman-Tukey method in the time domain T. It is possible to find the optimal time domain T regardless of the type of window function, such as a rectangular window, a Bartlett window, or a Blackman window.
[0053] Figure 12 is a schematic diagram showing an example of the difference in PSD between dementia patients and healthy individuals. In the figure, the horizontal axis indicates frequency, and the vertical axis indicates PSD (power spectral density). The solid line indicates healthy individuals (NC), and the dashed line indicates Alzheimer's disease patients (AD). As shown in the figure, characteristic differences can be seen in the alpha and theta bands, and differences in the clinical stages of dementia can be defined by the spectrum. Conversely, it is possible to classify and screen the clinical stages of dementia using spectral density. Although not shown in the figure, it is expected that the system will be able to classify, differentiate, diagnose, and screen for dementia (including AD, DLB, frontotemporal lobar degeneration (FTLD), normal pressure hydrocephalus (NPH), mild cognitive impairment (MCI), mild cognitive impairment due to Alzheimer's disease (MCIdue to AD), prodromal Alzheimer's disease, preclinical AD (pre-onset stage of Alzheimer's disease), etc.), multiple sclerosis brain tumors, mental disorders (also known as mental illnesses, including schizophrenia, epilepsy, mood disorders, addiction disorders, higher-level functional disorders, etc.), Parkinson's disease, Asperger's syndrome, attention-deficit hyperactivity disorder (ADHD), sleep disorders, pediatric diseases, ischemic brain disorders, and mood disorders (including depression, etc.).
[0054] Next, the spectral analysis between the two channels is explained.
[0055] To analyze the spectrum between the two channels (also called the "cross spectrum"), the cross spectrum is analyzed using a Fourier transform with the Multi-Taper Method (direct method), which is effective for relatively short time series data with frequency components below 70 Hz that represent brain activity.
[0056] As with single-channel spectral analysis, cross-spectral estimation is performed on finite data, and so there is a problem that the influence of the data edge appears as spectral leakage from the frequency where the estimated spectral peak exists to other frequencies. This spectral leakage occurs when the signal (average) and noise (standard deviation) of the estimated spectrum have the same magnitude, and the error of the estimated spectrum may not be improved even if the time domain T is changed. Therefore, in order to reduce the spectral leakage, a taper function can be used. This is called periodogram smoothing. The periodogram smoothing process in this embodiment is as follows.
[0057] Let y(t) be the time series data of MEG and h(t) be the taper function. The relationship with Y'(k) obtained by Fourier transform in this case is expressed by Equation (14).
[0058]
number
[0059] The inverse Fourier transform of Y'(k), which is the Fourier transform, is given by equation (15), and by substituting equation (15) into equation (14), equation (16) is obtained. H(kl) in equation (16) can be expressed by equation (17), which is the Fourier transform equation for the taper function h(t). The expected value {S'} of the Fourier transform using a taper is expressed by the convolution of the Fourier transform of the time series data y(t) and the Fourier transform of the taper function h(t). The expected value {S'} of the Fourier transform can be calculated by equation (18). Here, f N indicates the entire frequency band to be analyzed, where N = (0, 1, ..., N-1) [Hz]. f' indicates the sampling frequency. The spectrum data using the taper function h(t) is given by Equation (19).
[0060] Cross spectrum estimation by combination j between two channels can be expressed by equation (20), which is an extension of equation (16). When combination j between two channels is 1, 2, cross spectrum estimation is expressed by equation (21) based on equation (20). When the phase of data between the two channels for calculating the cross spectrum is synchronized, the cross spectrum shows a high value.
[0061] By calculating the modified periodogram using the taper function, the expected value of the cross spectrum is the same regardless of the Fourier window function (Hamming window, rectangular window, etc.), and the variation of the cross spectrum estimation error can be made extremely small.
[0062] Next, we will explain PSI (Phase Slope Index) as a feature of dementia. PSI is calculated using coherence, which is one of the metrics that localize the brain magnetic field distribution using a beamformer, and it has been reported that it shows stable statistics.
[0063] s k (t), s j Let (t) be a time series magnetic field signal at points k and j, and let a spectrum in the brain connecting points k and j be expressed by equations (22) and (23).
[0064]
number
[0065] The cross spectrum at this time can be expressed by equation (24). If the time it takes for the phase information to travel from the jth position to the kth position is τ, the synchronous phase state of the cross spectrum is proportional to the frequency f and the time difference τ of the signals. Based on this information, the Phase Slope (Ψ) is defined by equation (25). This can be calculated from the cross spectrum, but it can also be described as coherence. Coherence can be described by equation (26) from the cross spectrum Φ. Coherence represents the real part of the cross spectrum, which is a complex number, and can be seen as the cross-correlation coefficient of the Fourier frequency components of the two signals. The value shown in equation (26) takes a value between 0 and 1, and when it is close to 0, the correlation between the two is low, and when it is close to 1, the correlation between the two is high. The PSI is calculated by equation (27) using the coherence. Equations (25) and (27) are the same. PSI means synchronous information with a lag when the phase information travels from the jth position to the kth position.
[0066] As mentioned above, it is known that the coherence required for calculating PSI forms a false magnetic field distribution near the signal source. For this reason, the reliability of PSI using coherence is low. In the following, in this embodiment, the calculation of entropy using cross spectrum and power spectrum calculation without using coherence will be described.
[0067] By using the results of spectral analysis of MEG data, that is, by using both information from the spectral data about the frequency band of coordinated rhythms between each brain area related to brain activity (also called "synchronization component, coherence") and the power in a certain rhythm frequency band (also called "frequency component"), it is possible to calculate the degree of synchronization in the brain (the amount of synchronization phase difference indicated by the coherence seen in a certain frequency band) and the synchronization distribution (the amount of synchronization phase difference in coherence or cross spectrum), and thereby calculate features related to brain diseases (differences in the clinical stage of brain diseases, for example, entropy that determines the difference between dementia patients and healthy people).
[0068] The entropy calculation unit 59 calculates the entropy based on the information generated by the spectrum analysis unit 58. Specifically, in relation to the two channels, the entropy calculation unit 59 can calculate the entropy that characterizes the amount of synchronization phase difference in the brain based on frequency band information of rhythms related to brain activity. The frequency band information includes the frequency band of the rhythm and the power (also called "frequency component") in the frequency band of a certain rhythm. The amount of synchronization phase difference refers to the amount of synchronization indicated by coherence seen in a certain frequency band, and the synchronization component in coherence or cross spectrum. Coherence represents the amount of components synchronized between grids, and refers to signal components of the same phase and the same frequency. In addition, in relation to each channel, the entropy calculation unit 59 can calculate the entropy that characterizes the amount of synchronization in the brain based on frequency band information of rhythms related to brain activity. In addition, with respect to each channel, by focusing on the structure of the time series signal related to brain activity, the entropy calculation unit 59 can use nonlinear time series analysis to calculate the entropy that characterizes the disorder of the time series signal and the degree of breakdown of the phase structure. Entropy, which characterizes the randomness of a time series signal or the degree of breakdown of the phase structure, is calculated using nonlinear time series analysis and includes, for example, (1) appropriate entropy (ApEn), which is an index of randomness of a signal and a measure of the degree to which the difference between a specified combination of time series data is observed within a certain numerical range; (2) Sample Entropy, which is an index of randomness of a signal with autocorrelation and is an extension of ApEn that is independent of data size; (3) Lempel-Ziv complexity (LZC), which is a measure (randomness index) of time series data assuming a Kolmogorov process (uniform continuity); (4) multisacal entropy (MSE), which is a measure of time series data using autoregressive coefficients; (5) refined MSE (RMSE), which is a measure of time series data using autoregressive coefficients; and (6) Lopez Ruiz Mancini Calbet complexity (LMC), which is an extension of LZC and compares the measures of two time series data.
[0069] In the following, we will explain the metrics TSE, MF, SE, and SSE, which represent the entropy for a single channel, and the metric QSE, which represents the entropy between two channels.
[0070] The TSE can be calculated using equation (28) and represents the difference between a measure of observed spectral density (PSD) in the frequency band from 0.5 Hz to 70 Hz for each channel and the maximum likelihood function defined on the order of a q-exponential function.
[0071]
number
[0072] The QSE can be calculated using equation (29) and represents the difference between a measure of the cross-spectral density (PSD) observed in one channel and a maximum likelihood function defined in the order of a q-exponential function in the other channel in the frequency band from 0.5 Hz to 70 Hz.
[0073] The MF can be calculated using equation (30) and indicates the median of the cumulative sum of the MEG PSD in each frequency domain.
[0074] SE can be calculated by equation (31) and indicates the measure at which a certain spectral density is observed for each channel.
[0075] The SSE can be calculated using equation (32) and indicates the measure of the spectral density observed in the universe of MF, which is the frequency that is 50% of the cumulative sum of each frequency band. The metrics TSE, QSE, MF, SE, and SSE can be calculated using the spectral density PSD.
[0076] FIG. 13 is a schematic diagram showing an example of the phase synchronization distribution of magnetic fields in the heads of dementia patients and healthy individuals using cross-spectrum. In the figure, an Alzheimer's disease patient (AD) is given as an example of a dementia patient. As can be seen from the figure, there is a difference in the cross-spectrum distribution between healthy individuals and dementia patients. In other words, it is possible to classify and screen healthy individuals and dementia patients based on the cross-spectrum distribution. The difference in the cross-spectrum distribution means the difference in the distribution of cranial nerve activity. It can be seen that the cross-spectrum value is higher in each part of the brain in healthy individuals compared to dementia patients. A high cross-spectrum value means that cranial nerve activity is strongly activated.
[0077] Figure 14 is a schematic diagram showing an example of the entropy distribution of subjects by age. In the figure, the horizontal axis indicates age, and the vertical axis indicates the metric SE. The metric SE can be calculated using the power spectrum PSD. It can be seen that there is a probabilistic difference between the entropy representing healthy subjects (black dots) and the entropy representing dementia patients (X marks).
[0078] Next, the test using entropy will be described.
[0079] A physical quantity of interest is Q, and the expected value of that physical quantity obtained by measurement is QE. However, in systems where it is difficult to obtain other information, valid analysis is not possible. Therefore, the properties of the system are estimated by maximizing the uncertainty regarding the state of the system, using the prior probability distribution for the physical quantity Q as a constraint. This is called the maximum entropy principle, and the equilibrium probability distribution of a statistical system is determined based on this requirement. However, general brain activity is in an unbalanced state like a multifractal system, and conventional methods for analyzing balanced states are not valid. Therefore, the equilibrium probability distribution of a statistical system corresponding to the case where the probability distribution function behaves like a power law, such as a multifractal system, can be calculated using Tsallis statistics, which is a generalization of the Boltzmann-Shannon entropy, and index calculations can be performed using this equilibrium probability distribution.
[0080] For probability distributions such as normal distribution, if the mean and variance are known, the probability distribution of the values belonging to the distribution can be known. However, since only limited data exists for entropy, the probability distribution of entropy is unknown. Therefore, by using the balanced distribution (q-Gaussian distribution) and Mahalanobis distance as shown below, it is possible to classify healthy and disease groups from the entropy value.
[0081] This enables statistical analysis of multivariate two-sample problems using Hotelling's unbiased test for the entropy distribution of MEG channels. Specifically, the following procedure can be used: (1) Using equation (33) which represents the entropy TSE in the brain.
[0082]
number
[0083] (2) We use equation (34) to express that this TSE consists of a set of entropy elements (escort distribution). The entropy distribution consisting of the set of entropy elements is unknown. (3) We use equation (35), which shows that statistical analysis based on the probability distribution using q order is possible by making certain assumptions about the unknown entropy distribution. Equation (35) determines the value of q that determines the probability distribution to which the entropy set belongs. Equation (35) represents Jackson's q differential operator in Tsallis statistics. (4) Testing is possible by determining the abnormality probability using the Mahalanobis distance for a probability distribution using the q order as an index. The Mahalanobis distance refers to the information distance when the q-Gaussian distribution shows a zero-mean multivariate distribution and shows unbiasedness. The q-Gaussian distribution is a probability distribution obtained by maximizing the Tsallis entropy under the condition of constant variance.
[0084] As described above, the brain state estimation unit 60 can estimate the brain state of the subject based on the degree of deviation of the measure when the entropy calculated by the entropy calculation unit 59 is converted when it is assumed that the entropy has a certain probability measure. The probability distribution when it is assumed to have a certain probability measure includes, for example, a balanced probability distribution, a normal distribution, a discrete uniform distribution, a joint distribution, a negative binomial distribution, a geometric distribution, a binomial distribution, a Poisson distribution, a Cauchy distribution, and a gamma distribution. When the probability distribution is a balanced probability distribution, the brain state estimation unit 60 can estimate the brain state of the subject based on the degree of deviation of the probability set of the entropy when the probability distribution of the calculated entropy is assumed to be a balanced probability distribution. Specifically, by converting once to a q-Gaussian distribution, it is possible to define a probability distribution belonging to a measure called the Mahalanobis distance, and calculate the degree of deviation between the defined probability distributions. This allows the brain state estimation unit 60 to estimate the brain state by expressing the degree of discrepancy as the Mahalanobis distance of the brain disease group to the healthy subject group. The brain state estimation unit 60 can also distinguish between healthy subjects and brain disease patients and between them and the stimulation of the five senses. The brain state estimation unit 60 can be configured with a trained model generated by machine learning.
[0085] Figure 15 is a schematic diagram showing an example of the relationship between the Mahalanobis distance and the anomaly probability. When data is sampled from an m-dimensional normal distribution, it is known that the Mahalanobis distance follows a chi-square distribution with m degrees of freedom. If the 95% point α of the chi-square distribution is used as a threshold, then α<(x-μ)T Σ -1 When (x-μ), x can be determined to be an abnormal value. x is data sampled from a multivariate normal distribution N(μ,Σ). In the example of FIG. 15, for example, if the probability is smaller than the Mahalanobis distance corresponding to 95%, it can be determined to be abnormal.
[0086] Figure 16 is a schematic diagram showing the statistical properties (unbiasedness) of the entropy sampled from each channel. When using the Mahalanobis distance to find the abnormality probability, the sampled entropy must have statistical properties. As shown in Figure 16, it can be seen that the unbiased estimator (UMVU) calculated from the entropy sampled from each MEG channel of the whole brain distribution shows a uniform tendency even between subjects of dementia patients (AD: Alzheimer's disease patients) and healthy subjects (NC). In other words, the distribution of entropy is stationary. Here, TSE is used as a metric to represent entropy, but the same is true for other metrics. The statistical properties that indicate the universality of the entropy distribution sampled from a distribution with UMVU properties can be defined as a chi-square distribution. As a result, the entropy distribution of the MEG channel enables statistical analysis of multivariate two-sample problems using Hotelling's unbiased test.
[0087] FIG. 17 is a schematic diagram showing an example of the result of an outlier test using entropy sampled from each channel. The solid line shows the distance (metric) between two probability measures of NC for NC, and the dashed line shows the distance (metric) between two probability measures of AD for NC. The deviation of the probability set of entropy calculated individually for each MEG channel is expressed by Mahalanobis distance. As can be seen from the figure, it is possible to verify that the distinction between the AD group and the NC group is statistically valid. This embodiment can be used for screening, differentiation, diagnosis, and classification of brain diseases including dementia (including AD, DLB, frontotemporal lobar degeneration (FTLD), normal pressure hydrocephalus (NPH), etc.), brain tumors, mental disorders (also called mental illnesses, including schizophrenia, epilepsy, mood disorders, addiction disorders, higher functional disorders, etc.), Parkinson's disease, Asperger's syndrome, attention deficit hyperactivity disorder (ADHD), sleep disorders, pediatric diseases, ischemic brain disorders, mood disorders (including depression, etc.), etc.
[0088] The display panel 61 can be configured with a liquid crystal panel, an organic EL (Electro Luminescence) display, etc. In addition to or instead of the display panel 61, an external display device may be provided.
[0089] The operation unit 62 can be configured with, for example, a keyboard, a mouse, etc. The operation unit 62 may be configured with a touch panel incorporated in the display panel 61. The user can perform predetermined operations on the display panel 61.
[0090] The display processing unit 63 performs processing to display information such as the estimation result output by the brain state estimation unit 60 on the display panel 61. As exemplified in Fig. 13, the display processing unit 63 can display a cross-spectral distribution between two channels among a plurality of channels corresponding to a plurality of positions in the brain. Other display examples will be described below.
[0091] FIG. 18 is a schematic diagram showing a first example of a brain state estimation result. The display panel 61 displays the patient's name, diagnosis result, and reliability. The diagnosis result indicates, for example, normal (healthy subject), mild cognitive impairment (MCI), Alzheimer's disease (AD), other brain disease, etc. The reliability indicates the likelihood of the diagnosis result calculated based on the above-mentioned abnormality probability, etc. The display panel can display an image in which a cross-spectral distribution is superimposed on a brain volume model. The displayed image can be rotated in a desired direction using the "rotate" icon, and the image can be enlarged or reduced using the "zoom in / out" icon.
[0092] In this way, the display processing unit 63 can superimpose the cross-spectral distribution on the brain volume model and display it. The brain volume model can be constructed from MRI images, MI images, CT images, and the like.
[0093] Fig. 19 is a schematic diagram showing a second example of the brain state estimation result. On the display panel 61, a typical or representative cross-spectral distribution of a brain disease patient is displayed, and the cross-spectral distribution of the subject himself is displayed for comparison. In the example of the figure, the cross-spectral distributions of NC (healthy subject), MCI (mild cognitive impairment), AD (Alzheimer's disease), and other brain diseases are displayed as brain disease patients. In the example of the figure, the diagnosis of the subject is MCI, and the degree of deviation (certainty of abnormality probability) of the brain state of the brain disease group (MCI, AD, and others in the example of the figure) from the healthy subject group is displayed in comparison.
[0094] In this way, the display processing unit 63 can display the cross-spectral distributions of healthy subjects and patients with brain diseases, and can display the degree of deviation of the brain state of the brain disease group from the healthy subject group.
[0095] Next, the processing of the brain state estimation device 50 will be described.
[0096] 20 is a flowchart showing an example of the procedure of brain state estimation processing by brain state estimation device 50. In the following, for convenience, the processing will be described as being mainly performed by control unit 51. Control unit 51 acquires brain activity data of the subject (MEG data of each MEG channel) (S11). Note that in addition to MEG data, data such as electrooculogram (EOG), electroencephalogram (EEG), electrocardiogram (ECG), and electromyogram (EMG) can also be acquired.
[0097] The control unit 51 performs pre-processing including independent component analysis (ICA) on the acquired MEG data (S12). For pre-processing other than ICA, for example, at least one of bias correction for subtracting the average value of MEG data from the acquired MEG data, a band-pass filter for a predetermined frequency band, a notch filter for a predetermined frequency, and arithmetic averaging for each predetermined period can be used.
[0098] The control unit 51 performs SSP (signal space projection) processing (S13), estimates the signal origin (S14), and calculates the signal distribution (S15). Here, the signal distribution is a magnetic field distribution. The control unit 51 performs spectrum analysis for each channel and between channels using the MEG data obtained as a result of the processing of step S14 or S15 (S16). Here, the power spectrum and cross spectrum are calculated.
[0099] The control unit 51 calculates the entropy (S17), and uses the calculated entropy to estimate the subject's brain state based on the deviation of the probability set of entropy when the probability distribution of the entropy is assumed to be an balanced probability distribution (S18), outputs the estimation result (S19), and terminates the processing.
[0100] The brain state estimation device 50 can also be realized by using a computer equipped with a CPU (processor), RAM, etc. The brain state estimation device 50 can be realized on a computer by reading a recording medium, on which a computer program that defines the procedure of processing as shown in Fig. 20 is recorded, with a recording medium reading device, loading the read computer program into a RAM equipped in the computer, and executing the computer program with a CPU (processor).
[0101] In the above-described embodiment, MEG data has been used as an example of brain activity data of a subject, but the brain activity data is not limited to MEG data, and this embodiment can also be applied to EEG data.
[0102] The brain state estimation device 50 can acquire MEG data from a biomagnetic measurement device (MEG measurement device) via wireless communication or wired communication. The MEG measurement device places a number of MEG sensors (detection coils) on the subject's scalp and outputs time-series data recorded by the MEG sensors. Each of the many MEG sensors is associated with a channel.
[0103] The brain state estimation device of this embodiment includes an acquisition unit that acquires brain data of a subject that varies in space including a time axis, a generation unit that converts the acquired brain data to generate information related to the brain, a calculation unit that calculates entropy based on the generated information, and an estimation unit that estimates the brain state of the subject based on the deviation of the measure when the entropy is converted when the calculated entropy is assumed to have a certain probability measure.
[0104] In the brain state estimation device of this embodiment, the acquisition unit acquires brain data including brain activity data, or brain data including brain activity data, electromyogram data, electrocardiogram data, electrooculography data, and reference noise data, the generation unit generates brain-related information including rhythmic frequency band information or fractal information using frequency transformation or transformation by nonlinear time series analysis, the calculation unit calculates entropy characterizing the amount of synchronous phase difference in the brain based on the generated information including the rhythmic frequency band information, or entropy characterizing the disorder of the time series signal or the degree of breakdown of the phase structure based on the generated information including the fractal information, and the estimation unit estimates the brain state of the subject based on the deviation of the probability set of entropy when the probability distribution of the calculated entropy is assumed to be an balanced probability distribution.
[0105] In the brain state estimation device of this embodiment, the acquisition unit acquires time-series brain activity data for each of a plurality of channels corresponding to a plurality of locations in the brain, and the generation unit generates frequency band information of rhythms related to brain activity based on the brain activity data for each of the plurality of channels and the brain activity data between two of the plurality of channels.
[0106] In the brain state estimation device of this embodiment, the calculation section calculates the entropy for each channel based on the measure at which a certain spectral density is observed.
[0107] In the brain state estimation device of this embodiment, the calculation section calculates entropy based on a measure by which a certain spectral density is simultaneously observed in two channels.
[0108] In the brain state estimation device of this embodiment, the estimation unit estimates the brain state by expressing the degree of discrepancy as a Mahalanobis distance between a brain disease group and a healthy subject group.
[0109] The brain state estimation device of this embodiment includes a first noise removal unit that removes noise from the brain activity data acquired by the acquisition unit using at least one of a bias correction that subtracts an average value of the brain activity data from the brain activity data, a bandpass filter of a predetermined frequency band, a notch filter of a predetermined frequency, and an arithmetic average for each predetermined period.
[0110] The brain state estimation device of this embodiment includes a second noise removal unit that removes noise from the brain activity data acquired by the acquisition unit by applying independent component analysis and signal space projection.
[0111] The brain state estimation device of this embodiment includes a display processing unit that displays a cross-spectral distribution between two channels out of a plurality of channels corresponding to a plurality of positions in the brain.
[0112] In the brain state estimation device of this embodiment, the display processing unit displays the cross-spectral distribution superimposed on a brain volume model.
[0113] In the brain state estimation device of this embodiment, the display processing unit displays the cross-spectral distributions of each of a healthy subject and a patient with a brain disease.
[0114] In the brain state estimation device of this embodiment, the display processing unit displays the degree of deviation of the brain state of a brain disease group from that of a healthy subject group.
[0115] The computer program of this embodiment causes a computer to execute the following processes: acquire brain data of a subject that varies in space including a time axis, convert the acquired brain data to generate information related to the brain, calculate entropy based on the generated information, and estimate the brain state of the subject based on the deviation of the measure when the entropy is converted, assuming that the calculated entropy has a certain probability measure.
[0116] The brain state estimation method of this embodiment acquires brain data of a subject that varies in space including a time axis, converts the acquired brain data to generate information related to the brain, calculates entropy based on the generated information, and estimates the brain state of the subject based on the deviation of the measure when the entropy is transformed when it is assumed that the calculated entropy has a certain probability measure.
[0117] The system of this embodiment includes a function for classifying a subject's brain activity data, which varies in space including a time axis, using a model trained from learning data including a large number of solutions related to brain function testing.
[0118] The system of this embodiment includes a function for estimating the spatial distribution, including the time axis, of signal sources indicated by a subject's brain activity data, which varies in space, including the time axis, using a model that has learned solutions relating to a large number of inputs and outputs, including brain activation areas and patterns, from the brain activity data.
[0119] The system of this embodiment includes a function for estimating the brain state of a subject using a model that has learned cases and a spatial distribution including a time axis of a signal source estimated from the subject's brain activity data that varies in space including a time axis.
[0120] In the system of this embodiment, the model includes a time-series neural network.
[0121] In the system of this embodiment, the brain activity data includes brain activity data in response to stimuli including at least one of vision, hearing, touch, taste, and smell.
[0122] The method of this embodiment includes classifying a subject's brain activity data, which varies in space including a time axis, using a time series model trained on training data including a large number of solutions related to brain function testing.
[0123] The method of this embodiment includes estimating a spatial distribution, including a time axis, of signal sources indicated by a subject's brain activity data, which varies in space including a time axis, using a model that has learned solutions relating to a large number of inputs and outputs including brain activation areas and patterns.
[0124] The method of this embodiment includes estimating the brain state of a subject using a model that has learned cases and a spatial distribution including a time axis of a signal source estimated from the subject's brain activity data that varies in space including a time axis.
[0125] The computer program of this embodiment has instructions for operating as the above-mentioned system. [Explanation of symbols]
[0126] 50 Brain state estimation device 51 Control section 52 Data Acquisition Section 53 Memory section 54 Pretreatment section 55 SSP Processing Section 56 Signal source position estimation section 57 Signal distribution calculation section 58 Spectral Analysis Section 59 Entropy Calculation Unit 60 Brain state estimation unit 61 Display Panel 62 Operation section 63 Display processing unit
Claims
1. An acquisition unit for acquiring brain data of a subject that varies in space including a time axis; a generation unit that converts the acquired brain data to generate brain-related information; A calculation unit that calculates entropy based on the generated information; an estimation unit that estimates the brain state of the subject based on the degree of deviation of the measure when the calculated entropy is transformed assuming that the calculated entropy has a certain probability measure; A brain state estimation device comprising:
2. The acquisition unit is Acquire brain data including brain activity data, or brain data including brain activity data, electromyogram data, electrocardiogram data, electrooculography data, and reference noise data; The generation unit is generating brain-related information including rhythmic frequency band information or fractal information using frequency transformation or transformation by nonlinear time series analysis; The calculation unit Calculating an entropy that characterizes the amount of synchronous phase difference in the brain based on information including frequency band information of the generated rhythm, or an entropy that characterizes the disorder of the time series signal or the degree of breakdown of the phase structure based on information including the generated fractal information; The estimation unit is estimating the brain state of the subject based on the degree of deviation of the probability set of entropy when the calculated probability distribution of entropy is assumed to be an equilibrium probability distribution; The brain state estimation device according to claim 1 .
3. The acquisition unit is Acquire time-series brain activity data for each of multiple channels corresponding to multiple locations in the brain, The generation unit is generating frequency band information of rhythms associated with brain activity based on the brain activity data of each of the plurality of channels and the brain activity data between two of the plurality of channels; The brain state estimation device according to claim 2 .
4. The calculation unit is For each channel, calculate the entropy based on a measure of the observed spectral density. The brain state estimation device according to claim 3 .
5. The calculation unit is Calculate the entropy based on the measure of the simultaneous observation of a certain spectral density in two channels. The brain state estimation device according to claim 3 or 4.
6. The estimation unit is The degree of discrepancy is expressed as a Mahalanobis distance between the brain disease group and the healthy subject group to estimate the brain state. The brain state estimation device according to any one of claims 2 to 5.
7. a first noise removal unit that removes noise from the brain activity data acquired by the acquisition unit using at least one of a bias correction that subtracts an average value of the brain activity data from the brain activity data acquired by the acquisition unit, a band pass filter of a predetermined frequency band, a notch filter of a predetermined frequency, and an arithmetic average for each predetermined period; The brain state estimation device according to any one of claims 2 to 6.
8. a second noise removal unit that removes noise from the brain activity data acquired by the acquisition unit by using an independent component analysis and a signal space projection method; The brain state estimation device according to any one of claims 2 to 7.
9. a display processing unit that displays a cross-spectral distribution between two channels among a plurality of channels corresponding to a plurality of locations in the brain; The brain state estimation device according to any one of claims 1 to 8.
10. The display processing unit is The cross-spectral distribution is displayed superimposed on a brain volume model. The brain state estimation device according to claim 9.
11. The display processing unit is Displaying the cross-spectral distributions of each of a healthy subject and a patient with a brain disease. The brain state estimation device according to claim 9 or 10.
12. The display processing unit is Display the degree of discrepancy between the brain states of a brain disease group and a healthy group, The brain state estimation device according to any one of claims 9 to 11.
13. On the computer, Acquire brain data of subjects that varies in space, including time, Converting acquired brain data to generate brain-related information; Calculate entropy based on the generated information, estimating the brain state of the subject based on the degree of deviation of the measure when the calculated entropy is transformed, assuming that the calculated entropy has a certain probability measure; A computer program that executes a process.
14. Acquire brain data of subjects that varies in space, including time, Converting acquired brain data to generate brain-related information; Calculate entropy based on the generated information, estimating the brain state of the subject based on the degree of deviation of the measure when the calculated entropy is transformed, assuming that the calculated entropy has a certain probability measure; Brain state estimation method.
15. A system that includes a function for classifying a subject's brain activity data, which varies in space, including the time axis, using a model trained from training data including a large number of solutions related to brain function tests.
16. A system including a function for estimating a spatial distribution, including a time axis, of a signal source indicated by a subject's brain activity data, which varies in space, including a time axis, using a model that has learned solutions relating to a large number of inputs and outputs, including brain activity areas and patterns, from the brain activity data.
17. A system including a function for estimating a subject's brain state using a model that has learned cases and a spatial distribution, including a time axis, of a signal source estimated from the subject's brain activity data that fluctuates in space, including a time axis.
18. In any one of claims 15 to 17, The model includes a time series neural network.
19. In any one of claims 15 to 18, A system, wherein the brain activity data includes brain activity data in response to stimuli including at least one of vision, hearing, touch, taste, and smell.
20. A method comprising: classifying a subject's brain activity data, which varies in space including a time axis, using a time series model trained with training data including a large number of solutions relating to brain function testing.
21. A method comprising: estimating a spatial distribution, including a time axis, of a signal source indicated by a subject's brain activity data, the spatial distribution including the time axis, from the brain activity data, which varies in space, including a time axis, using a model that has learned solutions relating to a number of inputs and outputs, including brain activity areas and patterns.
22. A method comprising: estimating a brain state of a subject using a model that has learned cases and a spatial distribution including a time axis of a signal source estimated from brain activity data of the subject that varies in space including a time axis.
23. 20. A computer program comprising instructions for causing a computer to operate as a system according to any one of claims 15 to 19.
Citation Information
Patent Citations
Brain activity analysis method and device
JP2008229238A
Analysis of brain patterns using temporal scales
JP2009542351A
Systems and methods for perceptual and cognitive profiling.
JP2015533559A
Methods and systems for identifying brain disorders
US20180268942A1
