Biological information measurement system and biological information measurement method

Abstract

To provide a biological information measuring method capable of improving detection accuracy for occurrence of a phenomenon such as neural firing in a living body using magnetic measuring means.SOLUTION: A dataset of output signals is generated, the output signals consisting of a first output signal i-x and a second output signal i-y of an in-phase component i responding to a high frequency excitation signal for exciting a ferromagnetic resonance of a magnetic sensor 1, and a third output signal q-x and a fourth output signal q-y of an orthogonal component q responding to the high frequency excitation signal. The output signal responding to a reference signal is identified from the output signals of the magnetic sensor 1 and a state of a test object is estimated, using a learning model that has executed machine learning with a first dataset consisting of a time series of the output signal in a first signal section T1 that can respond to the reference signal, and a second dataset consisting of a time series of the output signal in a second signal section T2 that cannot respond to the reference signal as input data.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a system for measuring biological information using a magnetic measurement means. [Background technology]

[0002] The inventors have proposed a technology that uses sensors to detect signals corresponding to the location, direction, and strength of the movement that occurs when a person with hemiplegia due to stroke or quadriplegia due to spinal cord injury moves a muscle or joint, even slightly, through their own will and effort (see Patent Document 1). For example, the positions of multiple reflectors attached to specific parts of the human body, such as the fingers, are measured by an optical sensor using multiple CCD cameras. Pulse magnetic fields with different intensities and repetition rates are generated depending on the sensor signal level, and muscle contraction is induced by magnetic stimulation. This process is repeated, actively utilizing the brain's plasticity and building new neural networks, resulting in effective rehabilitation.

[0003] The present inventors have also proposed a bioinformation measurement system that can improve the measurement accuracy of bioinformation using a magnetic measurement means (see Patent Document 2). The magnetic sensor of this bioinformation measurement system includes a signal line layer laminated on a dielectric substrate, a transmission line composed of a ground conductor layer, an input / output terminal formed at one end of the transmission line, and a soft magnetic film laminated on a part of the other end of the transmission line. An incident signal is transmitted to the transmission line via the input / output terminal, and a reflected signal received from the transmission line via the input / output terminal is analyzed, thereby detecting with high accuracy the distribution of an external magnetic field that represents bioinformation according to the arrangement of multiple magnetic sensors. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent No. 5893367 [Patent Document 2] Patent application 2019-158120 Summary of the Invention [Problem to be solved by the invention]

[0005] However, the output signal of the magnetic sensor corresponding to the occurrence of a phenomenon such as neural firing in a living body is weak, and it is often difficult to distinguish the signal from noise.

[0006] Therefore, an object of the present invention is to provide a biological information measurement method that can improve the accuracy of detecting the occurrence of phenomena such as neural firing in a living body using magnetic measurement means. [Means for solving the problem]

[0007] The biological information measuring system of the present invention comprises: A system for measuring biological information, an element for generating a data set comprising a time series of at least one output signal from among a first output signal ix, which has an in-phase component i responsive to a high frequency excitation signal for exciting ferromagnetic resonance of the magnetic sensor and which has an in-phase component x of a detected magnetic field; a second output signal iy, which has an in-phase component i responsive to the high frequency excitation signal and which has a quadrature component y of the detected magnetic field; a third output signal qx, which has a quadrature component q responsive to the high frequency excitation signal and which has the in-phase component x of the detected magnetic field; and a fourth output signal qy, which has a quadrature component q responsive to the high frequency excitation signal and which has the quadrature component y of the detected magnetic field; and time derivatives of the first output signal ix, the second output signal iy, the third output signal qx, and the fourth output signal qy; The present invention includes an element for identifying an output signal responsive to a reference signal from the output signals of the magnetic sensor using a learning model that is deep-learned using as input data a first data set consisting of a time series of the output signal in a first signal section T1 that can respond to a reference signal and a second data set consisting of a time series of the output signal in a second signal section T2 that cannot respond to the reference signal. El .

[0008] According to the bioinformation measurement system of the present invention, a learning model is established through deep learning, and magnetic measurement means are used to improve the accuracy of identifying manifestations such as neural firing in living organisms. [Brief explanation of the drawings]

[0009] [Figure 1] 1 is an explanatory diagram illustrating the configuration of a biological information measuring system according to an embodiment of the present invention. [Figure 2] FIG. 1 is an explanatory diagram illustrating a configuration of a magnetic sensor according to an embodiment of the present invention. [Figure 3] FIG. 4 is an explanatory diagram of a magnetic field distribution generated near a signal line layer. [Figure 4] FIG. 2 is a block diagram of a signal processing device constituting the biological information measurement system. [Figure 5] FIG. 10 is an explanatory diagram of the detection results of magnetic signals representing nerve excitation. [Figure 6] FIG. 1 is an explanatory diagram of a magnetic signal waveform resulting from nerve excitation in response to multiple stimuli. [Figure 7] 10 is an explanatory diagram relating to the arithmetic averaging of time-series data of output signals from a magnetic sensor placed near a living body when a reference signal is given to the living body. [Figure 8] 4 is an explanatory diagram relating to the time waveforms of time-series data (ix, iy, qx, qy) of the output signal of the magnetic sensor in each of the first signal section T1 and the second signal section T2. ​​FIG. [Figure 9] A conceptual diagram of the construction of a learning model using CNN and the estimation method using the learning model. [Figure 10] FIG. 10 is an explanatory diagram regarding the accuracy rate of signal detection in response to reference signals for 12 data sets (J-0-0 to S-0-5). [Figure 11] FIG. 10 is an explanatory diagram of the accuracy rate of signal detection in response to a reference signal when all 12 data sets are used as training data in the learning phase. [Figure 12]FIG. 10 is an explanatory diagram regarding the accuracy rate of signal detection in response to a reference signal when training data is used in the learning phase, with the length of the signal section, that is, the number of signal samples in the input data, as a parameter. [Figure 13] FIG. 10 is an explanatory diagram regarding the accuracy rate of signal detection in response to a reference signal when input data with the number of data channels as a parameter is used as training data in the learning phase. [Figure 14] FIG. 1 is an explanatory diagram showing the accuracy rate of signal detection in response to a reference signal when 15 data sets consisting of combinations of each signal series of time series data (ix, iy, qx, qy) of the output signal of magnetic sensor 1 listed in Table 1 are used as training data. [Figure 15] 1A and 1B are conceptual explanatory diagrams of a conventional TCN and a TCN according to an embodiment of the present invention. [Figure 16] FIG. 10 is an explanatory diagram showing the results of a comparison of the accuracy rates of signal detection in response to a reference signal of a learning model using a conventional TCN and the TCN of this embodiment. [Figure 17] FIG. 10 is an explanatory diagram showing the results of a comparison between the CNN of the first embodiment and the TCN of this embodiment in terms of the accuracy rate of signal detection in response to a reference signal when 15 data sets consisting of combinations of each signal series of time-series data (ix, iy, qx, qy) of the output signal of the magnetic sensor 1 listed in Table 1 are used as training data. [Figure 18] FIG. 10 is an explanatory diagram regarding the accuracy rate of signal detection in response to a reference signal when an ensemble method is used. DETAILED DESCRIPTION OF THE INVENTION

[0010] (First embodiment) 1 includes a reference signal generator 20, an electrode 202, a magnetic sensor 1, a signal amplifier 24, a mixer 26, a filter 27, a signal processing circuit 28, and a data analysis device 29. The reference signal generator 20, the signal amplifier 24, the mixer 26, and the filter 27 constitute a signal processing circuit 200, which will be described later. Here, the magnetic sensor 1 may be a magnetic sensor array in which single magnetic sensors are arranged in an array.

[0011] The reference signal generator 20 is a signal generated at a specific location on the animal's body, for example, the ulnar nerve N in a human. v Electrical or magnetic stimulation is applied through an electrode 202 placed at a first designated location through which the ulnar nerve N passes. v When the nerve cell is excited (the membrane potential of the nerve cell reaches an action potential due to stimulation), and the action potential is transmitted along the axon, v A magnetic field is generated in the vicinity of a nerve. The strength of this field decreases rapidly with distance from the nerve. Therefore, the distribution of the magnetic field varies depending on the anatomical structure, location, and functional state of the nerve.

[0012] The magnetic sensor 1 detects the ulnar nerve N. v The magnetic sensor 1 is disposed near a second designated location through which a magnetic field passes, and detects the magnetic signal. The signal amplifier 24 amplifies the magnetic signal. The mixer 26 multiplies the reference signal from the reference signal generator 20 by the magnetic signal detected by the magnetic sensor 1. The filter 27 removes high-frequency noise components from the output signal of the mixer 26. The signal processing circuit 28 generates output signals based on four signal sequences: a first output signal ix having an in-phase component i responsive to a high-frequency excitation signal for exciting ferromagnetic resonance in the magnetic sensor 1 and including an in-phase component x of the detected magnetic field; a second output signal iy having an in-phase component i responsive to the high-frequency excitation signal and including an orthogonal component y of the detected magnetic field; a third output signal qx having an orthogonal component q responsive to the high-frequency excitation signal and including the in-phase component x of the detected magnetic field; and a fourth output signal qy having an orthogonal component q responsive to the high-frequency excitation signal and including the orthogonal component y of the detected magnetic field. The data analysis device 29 uses a deep learning model that uses as input data a first data set consisting of a time series of the output signal in a first signal section T1 that can respond to a reference signal and a second data set consisting of a time series of the output signal in a second signal section T2 that cannot respond to a reference signal, to identify output signals that respond to the reference signal from the output signals of the magnetic sensor and generate an image representing the anatomical structure, location, and functional state of the nerve or muscle.

[0013] 2 includes a dielectric substrate 100, a signal line layer 110 laminated in a substantially rectangular shape extending in the front-to-rear direction on the upper surface of the dielectric substrate 100, a pair of substantially rectangular ground conductor layers 121 and 122 extending in the front-to-rear direction on the upper surface of the dielectric substrate 100 and spaced apart from the signal line layer 110 on the left and right sides, respectively, on the upper surface of the dielectric substrate 100, and a ground conductor layer 124 formed on the lower surface of the dielectric substrate 100. An input / output terminal is electrically connected to one end of the transmission line 120, a soft magnetic thin film 140 partially laminated on the other end of the transmission line 120, and a thin-film magnet 150. The ground conductor layer 124 may be omitted.

[0014] The dielectric substrate 100 may be made of, for example, semiconductor single crystal silicon, germanium, compound semiconductor GaAs, GaN, SiC, ZnSe, CdS, ZnO, InP, SiGe, and / or quartz, sapphire, glass, or ceramic alumina, silicon nitride, aluminum nitride, zirconia, silicon carbide, titania, yttria, or a composite material thereof.

[0015] The signal line layer 110 and the ground conductor layers 121, 122, and 124 may be made of Al, Cu, Au, Pt, Ag, Ti, or a laminated thin film of these materials.

[0016] The soft magnetic thin film 140 may be a Co-Fe-Si-B or Co-Nb-Zr amorphous thin film, a Fe-Si or Fe-Zr-N microcrystalline thin film, a multilayer thin film formed by stacking these thin films with a thin insulating layer such as SiO sandwiched between them, a Ni-Fe or Fe-Si-Al crystalline thin film, or a bulk material such as an Fe-Si alloy, an Fe-Co-Ni alloy, or a sendust alloy, or a thin film made of a Co-Fe-Ni-Si-B, Fe-Co-Ni-Zr, Fe-Ni-B, Co-Fe-Zr, or Co-Zr amorphous alloy ribbon, or a thin film made of soft magnetic ferrite.

[0017] The thin film magnet 150 may be a sputtered thin film of a Pt-Fe system, SmCo system, Nd-Fe-B system, Sm-Fe-N system, or Nd-Fe-N system, or a metal-based bulk magnet such as Mn-Al-Co, Co-Pt, Fe-Pt, or Fe-Al-Ni, an oxide-based ferrite magnet, or a rare earth-based SmCo or Nd-Fe-B magnet.

[0018] The soft magnetic thin film 140 is preferably disposed near the signal line layer 130. FIG. 3 shows the magnetic field distribution generated near the signal line layer 130. When a high-frequency current (angular frequency ω) flows through the signal line layer 130, a magnetic field B(ω) is generated around the signal line layer 130, as shown in FIG. 3. The strength of the magnetic field B(ω) rapidly decreases with increasing distance from the signal line layer 130. The electric field E(ω) is strongly distributed between the signal line layer 130 and the ground conductor layers 121 and 122. The characteristic impedance of the transmission line 120 is defined by the ratio of the electric field E(ω) to the magnetic field B(ω) and is strongly dependent on the permittivity ε and permeability μ of the medium near the signal line layer 110 where the electric field E(ω) and magnetic field B(ω) are distributed. Therefore, by disposing the high-permeability soft magnetic thin film 140 near the signal line layer 110, changes in the characteristic impedance that are sensitive to external magnetic fields can be obtained. By subjecting this change in characteristic impedance to signal processing, which will be described later, a highly sensitive magnetic sensor or magnetic sensor array can be realized.

[0019] The magnetic sensor has a structure in which the end portion of a transmission line 120, which has a signal line layer 130 and ground conductor layers 121 and 122, is short-circuited, and a thin-film magnet 150 and a soft magnetic thin film 140 are laminated on the end portion. The thin-film magnet 150 is magnetized in advance in the longitudinal direction of the soft magnetic thin film 140, and the magnetic field of the thin-film magnet 150 is induced in the soft magnetic thin film 140, which has high magnetic permeability. In other words, the thin-film magnet 150 and the soft magnetic thin film 140 form a closed magnetic circuit structure. Therefore, a magnetic bias is efficiently applied to the soft magnetic thin film 140.

[0020] The strength of the magnetic bias is desirably equal to the magnitude of the anisotropic magnetic field of the soft magnetic thin film 140. The magnitude of this anisotropic magnetic field can be adjusted by the composition, film formation conditions, heat treatment conditions, etc. of the soft magnetic thin film 140. Furthermore, the remanent magnetization, which indicates the strength of the thin film magnet 150, can also be adjusted by the composition, film formation conditions, heat treatment conditions, etc.

[0021] 4 is a block diagram of a signal processing circuit of the biological information measurement system of the present invention. Signal processing circuit 200 includes high-frequency signal generator 210, amplifiers 212, 214, 216, 218, 241, 242, circulator 220, distributors 222, 224, 226, 251, 252, mixers 231, 232, 261, 262, 263, 264, reference signal generator 20 that generates a reference signal, and low-pass filters 271, 272, 273, 274. From the output terminals of the low-pass filters 271, 272, 273, and 274, output signals are obtained that are four signal series consisting of a time series of a first output signal ix having an in-phase component i responsive to a high-frequency excitation signal for exciting ferromagnetic resonance of the magnetic sensor and an in-phase component x of the detected magnetic field, a second output signal iy having an in-phase component i responsive to the high-frequency excitation signal and an orthogonal component y of the detected magnetic field, a third output signal qx having an orthogonal component q responsive to the high-frequency excitation signal and an in-phase component x of the detected magnetic field, and a fourth output signal having an orthogonal component q responsive to the high-frequency excitation signal and an orthogonal component q of the detected magnetic field.

[0022] Reference signal generator 20 is configured as reference signal generator 20 of Fig. 1 and is configured to generate a reference signal equivalent to a reference signal applied to the human body via electrode 202. Third amplifier 216 configures signal amplifier 24 of Fig. 1. First mixer 231 and second mixer 232 configure mixer 26 of Fig. 1. First low-pass filter 271, second low-pass filter 272, third low-pass filter 273, and fourth low-pass filter 274 configure filter 27 of Fig. 1.

[0023] The high-frequency signal generated by high-frequency signal generator 210 is amplified by first amplifier 212 and distributed in phase to two directions, circulator 220 and fourth amplifier 218, via first distributor 222. The high-frequency signal supplied to circulator 220 is amplified by second amplifier 214 and supplied to magnetic sensor 1 or to the input / output terminals of magnetic sensors 1 constituting a magnetic sensor array via ports p0 to p1 of circulator 220. The high-frequency modulated signal, obtained by modulating the high-frequency signal with a magnetic signal (emitted by the object to be detected) by magnetic sensor 1, passes from port p1 to port p2 of circulator 220, is amplified by third amplifier 216, and is distributed in phase to two directions, first mixer 231 and second mixer 232, via second distributor 224.

[0024] On the other hand, the high-frequency signal supplied from first divider 222 to first mixer 231 and second mixer 232 is amplified by fourth amplifier 218 and distributed via third divider 226 to two directions, first mixer 231 and second mixer 232, in in-phase and quadrature. The output signal of first mixer 231 is amplified by fifth amplifier 241 and distributed via fourth divider 251 to third mixer 261 and fourth mixer 262, respectively, as a quadrature component. The output signal of second mixer 232 is amplified by sixth amplifier 242 and distributed via fifth divider 252 to fifth mixer 263 and sixth mixer 264, respectively, as a quadrature component.

[0025] In third mixer 261, the output signal of fourth divider 251 is mixed with the reference signal generated by reference signal generator 20, and after unnecessary noise is removed through first low-pass filter 271, the output signal is output as a signal (first output signal) that is an in-phase component of the high-frequency signal and has a quadrature component of the reference signal. In fourth mixer 262, the output signal of fourth divider 251 is mixed with the reference signal generated by reference signal generator 20, and is output as a signal (second output signal) that is an in-phase component of the high-frequency signal and has the in-phase component of the reference signal through second low-pass filter 272. In fifth mixer 263, the output signal of fifth divider 252 is mixed with the reference signal generated by reference signal generator 20, and is output as a signal (third output signal) that is a quadrature component of the high-frequency signal and has the in-phase component of the magnetic signal through third low-pass filter 273. In the sixth mixer 264, the output signal of the fifth distributor 252 is mixed with the reference signal generated by the reference signal generator 20, and is output via the fourth low-pass filter 274 as a signal (fourth output signal) that is an orthogonal component of the high-frequency signal and also has an orthogonal component of the magnetic signal.

[0026] FIG. 5 shows the nerve N detected by the magnetic sensor 1 at a second designated location in response to a reference signal provided by the reference signal generator 20 to a first designated location on the upper limb. v The magnetic signal representing the excitation or muscle activity of the stimulation period T1 is shown. ’ In this case, the specified intensity B of the frequency range from a few Hz to 3 kHz p The reference signal S1 is applied to the first specified location. For example, the intensity of the reference signal S1 is adjusted in the range of 1 to 25 mA. o is 1 kHz and the number of stimulations n is 100, the stimulation period T1 ’ is 0.1 seconds. v The magnetic signal S2 caused by the excitation or muscle activity of the MRI sensor gradually increases slightly after the start of application of the reference signal S1, and reaches a constant magnitude A p After that, when the application of the reference signal 211 is stopped, the stimulation stop period T2 ’ At this point, the magnetic signal S2 gradually decreases to approximately zero.

[0027] Figure 6 shows the magnetic signal waveforms resulting from nerve excitation or muscle activity in response to multiple stimuli. ’ The magnetic signal waveform M{m1, m2, m3, . . . , m p-1 ,m p} signal amplitude A{A1,A2,A3,...,A p-1 ,A p} can be obtained, and statistical quantities such as the mean and standard deviation can be calculated. s Set a single stimulus m p Magnetic signal amplitude A p The magnitude of the signal can also be used to determine whether nerve excitation or muscle activity is present. ’ The presence or absence of nerve excitation or muscle activity may be determined from the average magnitude of the magnetic signal waveform M during the evaluation period T3. ’ For example, during the stimulation period T1 ’ (0.1 seconds (stimulation frequency f0 = 1 kHz, number of stimuli n = 100)), stimulation stop period T2 ’ If a cycle of 1.1 seconds (1 second) is repeated 10 times, the total time is 11 seconds.

[0028] Figure 7 shows the reference signal (stimulation period (or stimulation time) T1 ’ The graph shows the arithmetic average (number of arithmetic additions z=99) of the time series data of the output signal of the magnetic sensor 1 placed near a living body (for example, the limbs of a vertebrate animal such as a human) when a reference signal (for example, 0.1 seconds) is applied to the living body. From this result, it is not possible to confirm a clear response signal of the magnetic sensor 1 to the reference signal.

[0029] The time series data of the output signal of the magnetic sensor 1 is composed of four signal series: a first output signal ix, which is an in-phase component of the excitation frequency of the magnetic sensor 1 and an in-phase component of the detected magnetic field; a second output signal iy, which is an in-phase component of the excitation frequency and an orthogonal component of the detected magnetic field; a third output signal qx, which is an orthogonal component of the excitation frequency and an in-phase component of the detected magnetic field; and a fourth output signal qy, which is an orthogonal component of the excitation frequency and an orthogonal component of the detected magnetic field.

[0030] 8 shows the time waveforms of the time series data (ix, iy, qx, qy) of the output signal of the magnetic sensor 1 in each of the first signal section T1 and the second signal section T2. ​​The "first signal section T1" is defined as a period during which it is estimated that the magnetic sensor 1 can detect a response signal resulting from the occurrence of a neural firing phenomenon in response to a reference signal in a living organism. The "second signal section T2" is defined as a period during which it is estimated that the magnetic sensor 1 cannot detect a signal resulting from the occurrence of a neural firing phenomenon in response to a reference signal in a living organism. For example, each of the first signal section T1 and the second signal section T2 is a period during which it is estimated that the response signal resulting from the occurrence of a neural firing phenomenon in response to a reference signal in a living organism can be detected. ’ = 0.4 seconds, which is four times 0.1 seconds. The first signal section T1 is a period of 0.4 seconds from the start of stimulation, and the second signal section T2 is a period of 0.4 seconds from the end of the first signal section T1. The number of data included in the first signal section T1 or the second signal section T2 is the number of signal samples. Furthermore, the number of repetitions M when providing multiple reference signals is the number of data channels.

[0031] As conceptually shown in the upper part of Figure 9, during the training phase, a learning model is constructed using deep learning, specifically a convolutional neural network (CNN), with the time-series data (i-x1, i-y1, q-x1, q-y1) of the output signal of magnetic sensor 1 during the first signal interval T1 and the time-series data (i-x2, i-y2, q-x2, q-y2) of the output signal of magnetic sensor 1 during the second signal interval T2 as input data. The CNN consists of an input layer for inputting the time-series data from magnetic sensor 1, intermediate layers (multiple (e.g., four) convolutional layers, a pooling layer, and multiple (e.g., three) fully connected layers), and an output layer for outputting the identification results of the response signal of magnetic sensor 1 in response to neural stimulation. The intermediate layer network is connected via a nonlinear activation function (e.g., a ReLU function). To prevent overfitting, nodes are randomly dropped out only during training. The number of nodes in the input layer is 320,000 when, for example, magnetic sensor 1 has four signal sequences (ix, iy, qx, qy), 4,000 signal samples, and 20 data channels. The number of nodes in the middle layer is adjusted by hyperparameters such as the filter size, pooling interval, stride, and total number of connections in the convolutional layer. In the convolutional layer, a convolutional filter (e.g., stride 4) is applied independently to the time series data of each of the four signal sequences, and the output is processed via an activation function (e.g., ReLU). In the pooling layer, maximum values ​​are extracted from the time series data (e.g., stride 3). In the fully connected layer, the outputs of the previous stage (four time series data) are fully connected, integrating the information on magnetic sensor 1's time series data (ix, iy, qx, qy). In the output layer, the softmax function is used to convert the output of the middle layer into the probability of the first signal interval T1 and the probability of the second signal interval T2, and the higher probability is output as the final classification result.

[0032] As conceptually shown in the lower part of Figure 9, in the inference phase, the time series data (ix, iy, qx, qy) of the new output signal from the magnetic sensor 1 is used as input data, and the learning model constructed in the learning phase is used to estimate the presence or absence of a response signal resulting from the manifestation of the phenomenon of neural firing in a living organism.

[0033] Figure 10 shows the accuracy rate (number of correct answers / number of tests) of signal detection in response to reference signals for 12 data sets (J-0-0 to S-0-5). The accuracy rate when the time series data (ix, iy, qx, qy) of the output signal from magnetic sensor 1 itself was used as input data was 48.4±7.0%, and the accuracy rate when the ratio of the standard deviation to the mean value of each signal sequence in the time series data (ix, iy, qx, qy) was used as input data was 48.9±8.5%. On the other hand, the accuracy rate when the time fluctuation rate of each signal sequence in the time series data (ix, iy, qx, qy) was used as input data was 64.5±10.5%, which is a relatively high value.

[0034] Figure 11 shows the accuracy rate of signal detection in response to the reference signal when all 12 data sets were used as training data during the learning phase. The accuracy rate when all data sets were used was 83.4±6.0%, which is 18.9% higher than when only individual data sets were used as training data. Here, the time fluctuation rate of each signal sequence of the time series data (ix, iy, qx, qy) of the output signal of magnetic sensor 1 was used as input data.

[0035] Figure 12 shows the accuracy rate of signal detection responding to a reference signal when training data is used in the learning phase, with the lengths of the first signal interval T1 and the second signal interval T2 (i.e., the number of signal samples in the input data) as parameters. The signal interval reaches a maximum between 150 and 160 ms, with an accuracy rate of 91%. This result indicates that the most useful information for recognizing the presence or absence of a response signal is contained in a signal interval approximately 1.5 times the duration of the reference signal (100 ms). The CNN hyperparameters (e.g., number of hidden layers, number of nodes, filter size, dropout rate, learning rate, etc.) were also adjusted.

[0036] Figure 13 shows the accuracy rate of signal detection in response to a reference signal when input data with the number of data channels as a parameter is used as training data during the learning phase. Compared to when the number of data channels M is 1, the accuracy rate is significantly improved to 90.5±2.1% when M=20. Here, the length of the signal section of each input data is 150 ms.

[0037] Table 1 shows 15 data sets each consisting of a combination of signal sequences of the time-series data (ix, iy, qx, qy) of the output signal of the magnetic sensor 1.

[0038] [Table 1]

[0039] Figure 14 shows the accuracy rate of signal detection in response to the reference signal when the 15 data sets shown in Table 1 are used as training data. The results show that in the case of a data set including the fourth output signal qy of the signal sequence, the accuracy rate is 90.0±3.0%, which is approximately 36% higher than the data set not including the fourth output signal qy.

[0040] (Second embodiment) Next, we will explain how to construct a learning model using a Temporal Convolutional Network (TCN) and estimate the presence or absence of a response signal derived from the manifestation of neural firing in a living organism.

[0041] As conceptually shown in FIG. 15, conventional TCN employs dilated convolutions, which are convolutions performed at fixed intervals, whereas the TCN of this embodiment performs pooling after regular convolutions.

[0042] The difference between the CNN of the first embodiment and the TCN of the second embodiment is that a residual block layer is incorporated into the hidden layer, which consists of a convolutional layer, a weight regularization layer, a convolutional layer, and a weight normalization layer. The TCN of this embodiment consists of an input layer, hidden layers (multiple (e.g., four) residual block layers, a pooling layer, and multiple (e.g., three) fully connected layers), and an output layer.

[0043] 16 shows the comparison results of the accuracy rate of signal detection responding to the reference signal of the learning model between the conventional TCN and the TCN of this embodiment. The accuracy rate of the TCN of this embodiment is 92.7±3.0%, which is about 41.7% higher than the accuracy rate of the conventional TCN.

[0044] Fig. 17 shows the results of a comparison between the CNN of the first embodiment and the TCN of this embodiment regarding the accuracy rate of signal detection responding to a reference signal when the 15 data sets shown in Table 1 are used as training data. Similar to the results of Fig. 14, the TCN of this embodiment also shows a high accuracy rate of 91.0±5.0% for a data set including the fourth output signal qy of the signal sequence, and a low accuracy rate of 52.0±11.0% for a data set not including the fourth output signal qy.

[0045] (Third embodiment) We also explain how the TCN ensemble learning model can be used to estimate the presence or absence of response signals resulting from the manifestation of neural firing in living organisms.

[0046] The judgment results of the learning model by TCN in the second embodiment are weighted using an ensemble method to make a final classification judgment. Specifically, the judgment results of the learning model for N input data sets are evaluated as probabilities for cases where a signal is possible and cases where a signal is not possible, and the harmonic mean and geometric mean of the evaluation results are calculated to make a final judgment. The number of events containing the reference signal required for judgment is the sum (7) of the number of input data sets N (e.g., 3) and the number of events included in one input data set (e.g., 5) minus 1 (4).

[0047] Figure 18 shows the accuracy rate of signal detection in response to a reference signal when using an ensemble method. It can be seen that by applying either the harmonic mean or geometric mean ensemble method, the stability with respect to the number of events used for judgment is improved compared to when the ensemble method is not applied.

[0048] In the above embodiment, the data set is configured by the time series of the first output signal ix, the second output signal iy, the third output signal qx, and the fourth output signal qy. However, in another embodiment, the data set is configured by the time series of the first output signal ix, the second output signal iy, the third output signal qx, and the fourth output signal qy, and the time derivatives di-x / dt, d 2 ix / dt 2 , ‥d n ix / dt n , di-y / dt, d 2 iy / dt 2 , ‥d n iy / dt n , dq-x / dt, d 2 qx / dt 2 , ‥d n qx / dt n , dq-y / dt, d 2 qy / dt 2 , ‥d n qy / dt n The data set may be configured by the time series of at least one of the output signals (ix, di-x / dt, qx, dq-x / dt), (ix, d 2 iy / dt 2 , qx, d 2 qy / dt 2 ), (ix, di-x / dt, iy, di-y / dt), etc., may be employed. [Explanation of symbols]

[0049] 1. Magnetic sensor, 100. Dielectric substrate, 110. Signal line layer, 120. Transmission line, 121, 122, 124. Ground conductor layer, 140. Soft magnetic thin film, 20. Reference signal generator, 24. Signal amplifier, 26. Mixer, 27. Filter, 28. Signal processing circuit, 29. Data analysis device, 200. Signal processing circuit, 202. Electrode.

Claims

1. A system for measuring biological information, an element for generating a data set comprising a time series of at least one output signal selected from the group consisting of a first output signal i-x, an in-phase component i responsive to a high frequency excitation signal for exciting ferromagnetic resonance of the magnetic sensor, and having an in-phase component x of a detected magnetic field; a second output signal i-y, an in-phase component i responsive to the high frequency excitation signal, and having a quadrature component y of the detected magnetic field; a third output signal q-x, an quadrature component q responsive to the high frequency excitation signal, and having the in-phase component x of the detected magnetic field; and a fourth output signal q-y, an quadrature component q responsive to the high frequency excitation signal, and having the quadrature component y of the detected magnetic field; and time derivatives of the first output signal i-x, the second output signal i-y, the third output signal q-x, and the fourth output signal q-y; A first signal section T that can respond to a reference signal 1 a first data set consisting of a time series of the output signal in a second signal period T 2 and an element for identifying an output signal responsive to a reference signal from the output signals of the magnetic sensor using a learning model machine-learned using the second data set consisting of a time series of the output signals in Biological information measurement system.

2. 2. The biological information measuring system according to claim 1, The element for generating the data set generates a data set consisting of a time series of output signals of the time differential values ​​of the fourth output signal qy. Biological information measurement system.

3. A method for measuring biological information using a biological information measurement system, comprising: generating, by a signal processing circuit constituting the bioinformation measurement system, a data set consisting of a time series of at least one output signal selected from the group consisting of a first output signal i-x, which is an in-phase component i responsive to a high-frequency excitation signal for exciting ferromagnetic resonance of a magnetic sensor and has an in-phase component x of a detected magnetic field, a second output signal i-y, which is an in-phase component i responsive to the high-frequency excitation signal and has an orthogonal component y of the detected magnetic field, a third output signal q-x, which is an orthogonal component q responsive to the high-frequency excitation signal and has the in-phase component x of the detected magnetic field, and a fourth output signal q-y, which is an orthogonal component q responsive to the high-frequency excitation signal and has the orthogonal component y of the detected magnetic field, and time differential values ​​of each of the first output signal i-x, the second output signal i-y, the third output signal q-x, and the fourth output signal q-y; A first signal section T that can respond to a reference signal 1 a first data set consisting of a time series of the output signal in a second signal period T 2 and using a learning model machine-learned using the second data set consisting of a time series of the output signals in the above-mentioned data set as input data, identifying an output signal responsive to a reference signal from the output signals of the magnetic sensor by a data analysis device constituting the biological information measurement system. Biological information measurement method.

Images