Analysis system, analysis device, analysis method, and analysis program for evaluating pain levels

Abstract

This system provides the ability to calculate and display pain estimates stably and with high accuracy. [Solution] The analysis system 100 for evaluating the level of pain felt by a subject comprises an electroencephalograph 10, a pain analysis server 1000 that performs pain level analysis, and an in-hospital terminal 20 that receives signals from the electroencephalograph, transmits them to the pain analysis server 1000, and has a display screen 22 for presenting the analysis results to the user. The pain analysis server 1000 evaluates the electromyography signals included in the electroencephalogram signals according to the electromyography level and calculates the pain level using a trained model generated by machine learning. The trained model learns based on a training dataset corresponding to the electromyography level and outputs a pain level discrimination result.

Description

Technical Field

[0001] The present invention relates to a technique for analyzing electroencephalogram signals obtained from a subject, analyzing pain levels, and standardizing them as indices.

Background Art

[0002] Pain is essentially subjective. However, in treatment, it is desirable to be objectively evaluated. There are many situations where patients suffer disadvantages due to underestimation of pain. Therefore, a method for objectively estimating pain using electroencephalograms has been proposed (for example, see Patent Document 1).

[0003] However, the intensity of pain is subjective and difficult to objectively evaluate. Also, a technique for effectively monitoring the temporal change of pain is not necessarily in an established state.

[0004] For example, on a computerized visual analog scale (COVAS), pain intensity was continuously evaluated in the range of 0 to 100 (0: "no pain"; 100: "intolerable pain"). Methods such as recording COVAS data simultaneously with the change in stimulus intensity may be taken.

[0005] In Patent Document 1, an analysis of the correlation between COVAS data and feature amounts of electroencephalogram data is carried out. Also, there are examples of devices and methods for detecting pain by analyzing electroencephalogram data collected from an electroencephalogram (for example, see Patent Document 2).

[0006] Furthermore, there are other reports on techniques for attempting to evaluate pain using electroencephalograms (for example, see Non-Patent Document 1 and Non-Patent Document 2).

[0007] However, electroencephalogram signals have a large fluctuation and do not necessarily always correspond to the subject.

[0008] However, objective evaluation of a person's pain level is attracting considerable attention for its potential use in hospital treatment and healthcare, particularly for the purpose of managing analgesics.

[0009] Compared to various modalities used in medical settings, using electroencephalography (EEG) to quantitatively assess pain levels is suitable for clinical use due to its safety and non-invasiveness. [Prior art documents] [Patent Documents]

[0010] [Patent Document 1] Japanese Patent Publication No. 2020-203121 Specification [Patent Document 2] Patent No. 5642536 specification [Non-patent literature]

[0011] [Non-Patent Document 1] Moritz M. Nickel et al., “Brain oscillations differentially encode noxious stimulus intensity and pain intensity”, NeuroImage 148 (2017) 141―147 [Non-Patent Document 2] Duo Chen et al., “Scalp EEG-Based Pain Detection Using Convolutional Neural Network”, IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 30, 2022, p.274-285 [Overview of the project] [Problems that the invention aims to solve]

[0012] As mentioned above, building reliable EEG signal-based models for pain estimation has been challenging due to several factors, including the complexity of EEG and individual variability. Currently, further technological development is needed to stabilize, improve accuracy, and increase speed in these models.

[0013] The present invention was made to solve these problems, and aims to provide a system that can acquire multiple electroencephalogram (EEG) data, convert the EEG data into digital signals in real time, and use machine learning techniques to stably and accurately calculate and display pain estimates from the converted digital data. [Means for solving the problem]

[0014] According to one aspect of this invention, an analysis system for evaluating the level of pain felt by a subject comprises a measuring device for detecting the subject's electroencephalogram (EEG) signal, an analysis device that analyzes the pain level based on the signal from the measuring device, and an information presentation device having a display unit that receives the signal from the measuring device, transmits it to the analysis device, and presents the analysis results from the analysis device to the user, wherein the analysis device includes an electromyogram (EMG) level determination unit that evaluates the EEG signal contained in the EEG signal according to the EMG level, and a discrimination processing unit that receives the signal from the measuring device and calculates the pain level using a trained model generated by machine learning, wherein the trained model learns based on a training dataset corresponding to the EMG level and outputs a pain level discrimination result. Preferably, the system further includes a model selection unit that switches between trained models, which calculate pain levels over time, according to the electromyography level determined by the electromyography level determination unit. Preferably, the measuring device is an electroencephalograph that detects electroencephalogram signals over time from multiple predetermined positions on the subject's head, the electromyography level determination unit evaluates the electromyography signals included in the electroencephalogram signals from the electroencephalograph as multiple electromyography level ranks over time, the trained model includes multiple ensemble trained models corresponding to each rank generated by ensemble learning based on a training dataset for each of the multiple ranks of electromyography levels, each ensemble trained model integrates the discrimination results of multiple submodels trained with different machine learning algorithms, and the discrimination processing unit further includes a model selection unit that switches the trained model that calculates the pain level over time from among the multiple trained models according to the rank of the electromyography level determined by the electromyography level determination unit. Preferably, the electroencephalogram (EEG) signal is divided into segments at predetermined time intervals, and the electromyography (EMG) level determination unit determines the EMG level according to the power of the signal in a predetermined frequency band within each segment of the EEG signal. Preferably, the analysis device outputs information indicating the electromyography level evaluated by the electromyography level determination unit over time, and the information display device outputs information indicating the electromyography level to the display unit. Preferably, the electroencephalograph is attached to multiple predetermined positions on the subject's head and outputs monitoring signals over time, which measure the impedance of multiple electrodes for measuring electroencephalogram signals. The analysis device further includes a signal quality monitoring unit that determines the signal quality according to the monitoring signals and excludes signals below a predetermined quality from the analysis. Preferably, the different machine learning algorithms include at least two of the following: deep learning algorithms, clonet algorithms, EEG inception network algorithms, EEGnet models, and EEG LSTM algorithms. According to another aspect of this invention, an analysis device for evaluating the level of pain felt by a subject based on signals from a measuring device for detecting electroencephalogram (EEG) signals from a plurality of predetermined positions on the subject's head, comprising: an electromyogram (EMG) level determination unit that evaluates the electromyogram (EMG) signals included in the EEG signals according to the EMG level; and a discrimination processing unit that receives signals from the measuring device and calculates the level of pain using a trained model generated by machine learning. The trained model learns based on a training dataset corresponding to electromyography levels and outputs a result that determines the level of pain. Preferably, the system further includes a model selection unit that switches between trained models, which calculate pain levels over time, according to the electromyography level determined by the electromyography level determination unit. Preferably, the measuring device is an electroencephalograph that detects electroencephalogram signals over time from multiple predetermined positions on the subject's head, the electromyography level determination unit evaluates the electromyography signals included in the electroencephalogram signals from the electroencephalograph as multiple electromyography level ranks over time, the trained model includes multiple ensemble trained models corresponding to each rank generated by ensemble learning based on a training dataset for each of the multiple ranks of electromyography levels, each ensemble trained model integrates the discrimination results of multiple submodels trained with different machine learning algorithms, and the discrimination processing unit further includes a model selection unit that switches the trained model that calculates the pain level over time from among the multiple trained models according to the rank of the electromyography level determined by the electromyography level determination unit. Preferably, the electroencephalogram (EEG) signal is divided into segments at predetermined time intervals, and the electromyography (EMG) level determination unit determines the EMG level according to the power of the signal in a predetermined frequency band within each segment of the EEG signal. Preferably, the electroencephalograph is attached to multiple predetermined positions on the subject's head and outputs monitoring signals over time, which measure the impedance of multiple electrodes for measuring electroencephalogram signals. The analysis device further includes a signal quality monitoring unit that determines the signal quality according to the monitoring signals and excludes signals below a predetermined quality from the analysis. Preferably, the different machine learning algorithms include at least two of the following: deep learning algorithms, clonet algorithms, EEG inception network algorithms, EEGnet models, and EEG LSTM algorithms. Preferably, the analysis device transmits information on the pain level and information indicating the rank of the electromyography level evaluated by the electromyography level determination unit, in a format that can be presented to the user by an information presentation device that receives signals from the electroencephalograph and transmits them to the analysis device. Preferably, the analysis device outputs information indicating the rank of the electromyography level evaluated by the electromyography level determination unit over time so that it is displayed on the display unit of the information presentation device. According to yet another aspect of this invention, an analysis method for evaluating the level of pain felt by a subject using a computer, wherein the computer includes a computing device and a memory device, and the method comprises the steps of: the computing device storing electroencephalogram (EEG) signal data from a measuring device for detecting the subject's electroencephalogram (EEG) signal in the memory device; the computing device evaluating the components of electromyography (EMG) signals included in the EEG signal data according to the EMG level; and the computing device receiving the EEG signal data and calculating the pain level using a trained model generated by machine learning, wherein the trained model learns based on a training dataset corresponding to the EMG level and outputs a pain level determination result, and presenting the pain level analysis result to the user. According to still another aspect of the present invention, there is an analysis program for causing a computer to evaluate the level of pain felt by a subject. The computer includes an arithmetic unit and a storage unit. The analysis program causes the computer to perform the steps of: storing, in the storage unit, electroencephalogram signal data from a measuring device for detecting the electroencephalogram signal of the subject by the arithmetic unit; evaluating, by the arithmetic unit, the component of the electromyogram signal included in the electroencephalogram signal data according to the electromyogram level; calculating, by the arithmetic unit, the level of pain by a learned model generated by machine learning upon receiving the electroencephalogram signal data; and presenting the analysis result of the pain level to the user. The learned model is learned based on a learning dataset according to the electromyogram level and outputs a discrimination result of the pain level.

Advantages of the Invention

[0015] According to the present invention, it is possible to realize a system capable of calculating and displaying a pain estimation value stably and with high accuracy.

Brief Description of the Drawings

[0016] [Figure 1] It is a conceptual diagram for explaining the overall configuration of the pain analysis system 100 of the present embodiment in operation. [Figure 2] It is a conceptual diagram for explaining the arrangement of electrodes attached to the subject 1. [Figure 3] It is a functional block diagram for explaining the configurations of the electroencephalograph 10 and the in-hospital terminal 20. [Figure 4] It is a functional block diagram for explaining the configuration of the pain level analysis server 50. [Figure 5] It is a block diagram for explaining the hardware configuration of the pain analysis server 1000. [Figure 6] It is a conceptual diagram showing the learning process and prediction process of the artificial intelligence model used for calculating the pain score PS in the pain analysis server 1000. [Figure 7] It is a diagram showing an example of thermal stimulation given to the subject. [Figure 8]This figure shows an example of selecting pain zones / painless zones based on CoVAS mean values. [Figure 9] This is a conceptual diagram showing the training dataset used for learning. [Figure 10] This diagram illustrates the model configuration of the Chrononet model, which is a learning model corresponding to the first machine learning algorithm. [Figure 11] This figure shows the Inception Network model, which is a learning model corresponding to the second machine learning algorithm. [Figure 12] This figure shows the EEG network model, which is a learning model corresponding to the third machine learning algorithm. [Figure 13] This diagram shows the EEGLSTM model, which is a learning model corresponding to the fourth machine learning algorithm. [Figure 14] This diagram illustrates a feature-based model, which is a learning model corresponding to the fifth machine learning algorithm. [Figure 15] This flowchart illustrates the process of predicting pain scores (PS) using a trained ensemble machine learning model. [Figure 16] This diagram illustrates the process of coordinating between the in-hospital terminal 20 and the pain level analysis server 50. [Figure 17] This figure shows an example of the content displayed on the display screen 22 of the in-hospital terminal 20 during real-time measurement. [Figure 18] This figure shows the evaluation results of the predictive performance of encapsulated individual models and ensemble models. [Modes for carrying out the invention]

[0017] The following describes embodiments of the present invention and the brain image analysis data utilization system of the present invention.

[0018] In this embodiment, the pain analysis device (or pain analysis server) functions as a device that receives and analyzes electroencephalogram (EEG) data from the subject. This pain analysis device is a device that evaluates the subject's objective pain score.

[0019] Furthermore, the following description assumes a configuration where the pain analysis server evaluates pain scores on the cloud, but the embodiment is not limited to this configuration. For example, it may operate as an on-premise system within a facility such as a hospital, as a pain analysis system (a pain analysis server operating on the facility's local network). Alternatively, it may be configured as a standalone computer system, as a pain analysis device.

[0020] In this specification, the system configuration described above will be collectively referred to as the "pain analysis system."

[0021] Figure 1 is a conceptual diagram illustrating the overall operating configuration of the pain analysis system 100 according to this embodiment.

[0022] In the following explanation, a hospital will be used as an example of a facility for measuring the brainwaves of subjects.

[0023] Referring to Figure 1, multiple electrodes 2 are attached to the head of subject 1, more specifically to the forehead, in the hospital to acquire electroencephalogram (EEG) data.

[0024] The signals from the multiple electrodes 2 attached to the body are amplified by the electroencephalograph 10, subjected to predetermined analog filtering, and converted into digital signals by an AD converter that samples the signals.

[0025] The electroencephalogram (EEG) signals, converted into digital signals, are processed at the in-hospital terminal 20 and then transmitted via network 1 to the pain level analysis server 50.

[0026] As will be described later, the display screen of the in-hospital terminal 20 shows the information being measured to the operator operating the system within the hospital. A typical "operator" is a physician, but this also includes anyone operating the device under the supervision of a physician.

[0027] The pain level analysis server 50 includes, as described later, a pain analysis server 1000 that calculates a "pain score PS (Pain Score)" representing the pain level from the electroencephalogram (EEG) signals of subject 1, and a database server 2000 that stores the EEG data transmitted to the pain level analysis server 50 in association with the attributes of subject 1 (for example, patient ID, name, gender, age, doctor's findings, etc.).

[0028] Furthermore, the data structure and corresponding security specifications for "Subject 1's Attributes" are not limited to those exemplified, and the data structure can be modified as appropriate depending on the system configuration and requirements (e.g., security specifications).

[0029] The pain analysis server 1000 sends back to the in-hospital terminal 20 not only the "pain score PS" but also information such as the intensity of the electromyography (EMG) signal components in the electroencephalogram (EEG) signal.

[0030] As explained below, the processing performed by the pain analysis server 1000 is achieved by using artificial intelligence based on ensemble learning to execute a model selection algorithm in order to improve the accuracy and speed of predicting pain scores (PS) based on electroencephalogram signals.

[0031] This configuration allows for the prediction of a highly reliable pain score (PS). Furthermore, in situations where signal interference from facial and skeletal muscles is present, the system periodically checks the level of the electromyography (EMG) component and selects a "weak learner" within the ensemble learning model that has been pre-trained to correspond to each level, in order to optimize the prediction of the pain score (PS).

[0032] Furthermore, the algorithm executed by the pain analysis server 1000 in this embodiment combines ensemble learning with model switching based on electromyography levels. This enables a robust, real-time pain estimation system based on electroencephalogram (EEG) signals.

[0033] Figure 2 is a conceptual diagram illustrating the arrangement of electrodes to be attached to subject 1.

[0034] Regarding electrode placement for electroencephalography (EEG), for example, a conventional technique is the "10% method (extended 10-20 method)," which was recommended as an international standard by the International Federation of Electroencephalography and Clinical Neurophysiology Societies (now the International Federation of Clinical Neurophysiology Societies) and defined in 1991 to accommodate multi-channel recording.

[0035] Figure 2(c) shows the electrode arrangement for the extended 10-20 method. The advantages of the extended 10-20 electrode arrangement are as follows:

[0036] • Because the placement is based on rules, the position is highly reproducible and easy to measure repeatedly.

[0037] It covers the cerebrum.

[0038] • The design ensures that the distance between each electrode is equal.

[0039] • They are arranged based on anatomical locations.

[0040] However, in this embodiment, for the sake of ease of measurement, the electrodes are mainly placed on the forehead of subject 1 for measurement purposes.

[0041] Figure 2(b) shows the names of the electrodes arranged in this embodiment, in comparison to the "10% method (extended 10-20 method)".

[0042] On the other hand, Figure 2(a) shows the channel names of the electrodes used for explanatory purposes in this embodiment. The channel names in Figure 2(a) correspond to the electrode arrangement shown in Figure 2(b). The reference electrode (REF) is, for example, attached near the left earlobe of subject 1. It is also grounded to electrode Z on the midline of the forehead.

[0043] For example, during measurement, the electroencephalogram (EEG) signal is sampled at 1 kHz, and the electrode impedance is maintained at less than 20 kΩ. When acquiring training data, as described later, continuous pain assessments (CoVAS signal) and stimulus intensity (e.g., temperature) are simultaneously recorded as additional channels at the same sampling frequency as the EEG acquisition.

[0044] Figure 3 is a functional block diagram illustrating the configuration of the electroencephalograph 10 and the in-hospital terminal 20.

[0045] The functions performed by the electroencephalograph 10 and the in-hospital terminal 20 shown in Figure 3 are examples only and are not limited to these. Furthermore, the functions performed by the electroencephalograph 10 and the in-hospital terminal 20 can be changed as appropriate depending on the system configuration.

[0046] Referring to Figure 3, the electroencephalograph 10 measures the temporal changes in the electrical potential occurring in the head of subject 1 using electrodes 2.1~2.N (N: natural number) arranged as described in Figure 2, acquires the data as an electroencephalogram (EEG), and outputs this EEG signal data. More precisely, the signals acquired by electrodes 2.1~2.N are changes in the electrical potential of the living body, and include not only EEG signals from the brain but also electromyographic signals generated by muscle activity, and are weak changes in electrical potential (tens of microvolts to tens of millivolts).

[0047] Therefore, in the sense of acquiring biological signals more generally, the electroencephalograph 10 is called a "biological signal measuring device." It should be noted that, in addition to electroencephalography, other devices capable of non-invasively measuring brain activity, such as magnetoencephalography, can also be used as such measuring devices.

[0048] The electroencephalograph 10 includes an amplifier 100, a bandpass filter (BPF) 110, an A / D converter 120, a digital filter processing unit 130, and an interface unit 140.

[0049] The electroencephalogram (EEG) signals acquired by measuring the potential of each electrode 2.1 to 2.N are amplified by amplifiers 100.1 to 100.N in individual transmission lines (channels), passed through an A / D converter 120, subjected to predetermined digital filtering by a filter processing unit 130, and output to the in-hospital terminal 20 via an interface 140. These electrodes include one or more reference electrodes and one or more measurement electrodes, and the EEG signals are acquired by measuring the potential difference between each measurement electrode and the reference electrode corresponding to that measurement electrode.

[0050] The A / D converter 130 converts the electroencephalogram (EEG) signal, which is an analog signal input from the amplifier 120, into a digital signal. Although not particularly limited, for example, the A / D converter 120 can be a ΔΣ type A / D converter that converts the analog signal into a digital signal with a predetermined number of bits or more.

[0051] The in-hospital terminal 20 controls the entire operation of the electroencephalograph 10. The in-hospital terminal 20 includes an interface unit 210 that receives signals from the electroencephalograph 10, a storage unit 220 for temporarily storing the input signals, a central processing unit (CPU) 230 for performing predetermined calculations on the data in the storage device 220, a clock generation unit 202 that supplies the operating clocks for each unit, and an output transmission unit 240 that transmits data from the in-hospital terminal 20 to the pain level analysis server 50 via the network 1.

[0052] The in-hospital terminal 20 includes a receiving unit that receives data from the pain level analysis server 50 and a display unit that presents information about the operation to the operator, but these are not shown in Figure 3.

[0053] The CPU 230 performs several functions, including a baseband digital signal processing unit 232 that performs data preprocessing and a contact state detection unit 230 that monitors the contact state of the electrodes with the subject 1.

[0054] The CPU 230 reads and executes programs from a non-volatile memory unit (not shown) and manages and controls the operation of each component that makes up the electroencephalograph 10.

[0055] The contact state detection unit 230 is an electronic circuit (contact state detection circuit) that detects the contact state between the electrode and the living body. Specifically, the contact state detection unit 230 measures the contact impedance between the electrode and the skin of the living body and outputs it as contact state information.

[0056] The in-hospital terminal 20 may be configured to determine whether the contact state is good or bad based on the magnitude of this contact impedance. For example, by performing this determination at predetermined timings during attachment and during measurement of the electroencephalogram signal, the in-hospital terminal 20 can evaluate the attachment state and signal quality of the electroencephalograph 10. The criteria for determining whether the contact state is bad may be changed as appropriate by the settings. For example, if there are 8 channels for signal input from the electrodes, the contact state may be determined to be bad if lead-off from the patient is detected in even one of these 8 channels.

[0057] Figure 4 is a functional block diagram illustrating the configuration of the pain level analysis server 50.

[0058] The pain level analysis server 50 includes the pain analysis server 1000 and the database server 2000.

[0059] The pain analysis server 1000 includes a communication interface 1002 for exchanging data with in-hospital terminals 20.1 to 20.N via network 1, a calculation processing unit 1100, a memory 1200 for temporarily storing data, and a storage unit 1300 for non-volatile data storage. The pain analysis server 1000 exchanges data with multiple in-hospital terminals 20.1 to 20.N. Hereafter, when referring to in-hospital terminals 20.1 to 20.N collectively, they will be called in-hospital terminal 20.

[0060] The memory unit 1300 stores not only the program for the pain level analysis server to operate, but also data for identifying trained models, as described later, as trained model data 1310.

[0061] Furthermore, the database server 2000 includes a communication interface 2310 for communicating with the pain analysis server 1000, a subject attribute database 2320, a subject electroencephalogram database 2330, and a hospital / user attribute database 2340. Hereafter, "database" will be abbreviated as DB.

[0062] Generally speaking, a "database" can refer to a database with a well-defined schema and transaction capabilities, such as a relational database or an object-relational database, or it can refer to an object database or a column-oriented database management system, where data in the same column is aggregated and stored in physically close areas.

[0063] Based on the program stored in the memory unit 1300 and deployed in the memory 1200, the arithmetic processing unit 1100 in the electroencephalogram (EEG) data receiving module 1110 receives EEG signal data and electrode contact information data from the in-hospital terminal 20. Furthermore, the EEG data receiving module 1110 divides the received EEG signal data into segments as described later.

[0064] The signal quality management module 1120 evaluates the signal quality of the electroencephalogram (EEG) signal data through the processing described later.

[0065] The electromyography rank determination module 1130 determines the level of the electromyography signal in the electroencephalogram (EEG) signal data according to the intensity level of data in a predetermined frequency band in the EEG signal data, and outputs an electromyography rank as described below.

[0066] The pain assessment module 1140, in determining the pain score PS from electroencephalogram (EEG) signal data, uses the model selection module 1150 to select an ensemble-trained model 1160.1 to 1160.M that calculates the pain score PS, based on the timing of the assessment and the rank of the electromyography level.

[0067] Figure 5 is a block diagram illustrating the hardware configuration of the pain analysis server 1000 shown in Figure 4.

[0068] In the following, it is assumed that the pain analysis server 1000 and the database server 2000 operate on servers running on the cloud. However, the pain analysis server 1000 and the database server 2000 may also operate on, for example, on-premises servers.

[0069] Although the following explanation uses the configuration of the pain analysis server 1000 as an example, the configurations of the database server 2000 and the in-hospital terminals 20 are basically the same.

[0070] Server 1000 may be configured so that its own CPU (Central Processing Unit) performs the calculations, or it may be configured so that part of the program's processing is executed on another server. In the following explanation, we will assume that the CPU within the server itself performs the calculations.

[0071] Referring to Figure 5, the server 1000 comprises a computer device 1010, a network communication unit 1012 for communicating with a network, and a recording medium (for example, a memory card) 1210 for recording external data and providing it to the computer device 1010.

[0072] For example, the recording medium 1210 can be a USB memory stick, memory card, or external storage device. The network communication unit 1012 can utilize, for example, wired LAN or wireless LAN communication functions. The network communication unit 1012 and the input / output interface 1090 constitute the communication interface 1002.

[0073] As shown in Figure 5, the computer body constituting this computer device 1010 includes, in addition to the disk drive 1030 and memory drive 1020, a CPU 1100 connected to the bus 1050, memory 1200 including ROM (Read Only Memory) 1200.1 and RAM (Random Access Memory) 1200.2, a non-volatile rewritable non-volatile storage device 1300, and an input / output interface 1090 for communication over a network and data exchange with the outside. For example, the non-volatile storage device 1300 can be an HDD (Hard Disk Drive) or an SSD (Solid State Drive). In the following description, it will be described as an SSD. An optical disc can be installed in the disk drive 1030. A memory card 1210 can be installed in the memory drive 1020.

[0074] In this explanation, the data and programs that store the information fundamental to the operation of the computer device 1010 are assumed to be stored in the SSD 1300.

[0075] In Figure 5, the medium on which information such as programs to be installed on the computer can be recorded may be, for example, a DVD-ROM (Digital Versatile Disc), a memory card, or a USB memory stick. To accommodate such cases, the computer is equipped with drive devices (memory drive 1020, disk drive 1030) capable of reading these media.

[0076] The main components of the computer device 1010 consist of computer hardware and software executed by the CPU 1100. Generally, such software is stored and distributed on a storage medium or via a network, retrieved via the disk drive 1030 or network communication unit 1012, and temporarily stored in the SSD 1300. It is then read from the SSD 1300 into the RAM 1200.2 in memory and executed by the CPU 1100. In the case of a network connection, the software may be loaded directly into RAM and executed without being stored in the SSD 1300.

[0077] The program for functioning as a computer device 1010 does not necessarily need to include an operating system (OS) that causes the computer body 1010 to execute functions such as an information processing device. The program only needs to contain the instruction portion that calls appropriate functions (modules) in a controlled manner and obtains the desired result. How the computer system 1010 operates is well known, so a detailed explanation is omitted.

[0078] Furthermore, the CPU 1100 may be a single-core processor or a multi-core processor. In other words, it may be a single-core processor or a multi-core processor. Also, the server 1000 may be configured with multiple servers to perform distributed processing.

[0079] Figure 6 is a conceptual diagram showing the learning and prediction processes of the artificial intelligence model used to calculate the pain score PS in the pain analysis server 1000.

[0080] Referring to Figure 6, the collection of training data is carried out, for example, at clinical research facilities or clinical trial hospitals.

[0081] As will be described later, subjects are given reference stimuli such as temperature stimuli, and electroencephalogram (EEG) data is acquired and stored in a database corresponding to the magnitude of each reference stimulus.

[0082] Based on the training data stored in the database, a machine learning model is generated that regresses the pain level onto the features of the electroencephalogram (EEG) data. Here, "ensemble learning" is used in this machine learning process, as will be described later.

[0083] The trained model generated by machine learning performs predictive processing on the input data on the pain analysis server 1000 in the cloud. Specifically, based on the electroencephalogram (EEG) data transmitted from the EEG machine 10 and in-hospital terminal 20 installed in hospitals and clinics, an estimated amount of pain (pain score PS) is calculated and sent back to the in-hospital terminal 20, where it is displayed on the terminal's display screen 22.

[0084] Furthermore, the program for predicting pain scores (PS) that operates on the cloud may be configured to be usable by third-party devices as an API (Application Programming Interface). (Conditions for acquiring training data)

[0085] When acquiring training data, the first measurement condition involves applying temperature stimulation as a pain stimulus to multiple subjects.

[0086] The setup conditions for thermal stimulation experiments are disclosed, for example, in the following literature.

[0087] Publicly available document 1: International Publication No. 2022 / 145429 Publicly known document 2: Patent No. 7401067

[0088] Figure 7 shows an example of thermal stimulation given to such subjects for generating training data.

[0089] The state in which thermal stimulation is applied is shown as "reg-ref" in Figure 7(a). In this state, electroencephalogram (EEG) data is recorded simultaneously with the gradual increase in thermal stimulation applied to the subject's left arm. Pain recording progresses until the maximum tolerable pain level of the thermal stimulation is reached, after which the thermal stimulation gradually decreases.

[0090] During the recording of electroencephalogram (EEG) data, subjects are asked to continuously quantify their pain perception on a scale of 0 to 100, using a Computerized Visual Analog Scale (CoVAS) device, with painless pain and the worst tolerable pain being used as baselines.

[0091] As shown in Figure 7(a), this process is repeated two or more times, creating the shapes of two peaks.

[0092] For example, the duration of the 6-minute stimulation is the same for all subjects, while the maximum stimulation intensity (i.e., peak temperature) is individually adjusted within the range of 48.9°C to 47°C based on the subject's tolerance.

[0093] The second measurement condition, as shown in Figure 7(b), is a state of painless stimulation. During this state, a constant, painless thermal stimulus is applied to the participant's left arm for approximately 5 minutes in this case. The stimulation temperature is customized for each subject within the range of 42°C to 40°C, and there is no pain or discomfort; only the sensation of heat is felt.

[0094] Under both measurement conditions, subjects were instructed to lie in a comfortable armchair, close their eyes to minimize any movement that might affect brainwave signals, and remain in a fixed position. (Signal preprocessing)

[0095] The following describes the signal preprocessing pipeline. In principle, the same preprocessing is performed on both the training data and the real-time measurement data.

[0096] In the training data, data from subjects with a data loss rate exceeding 5% were excluded from the analysis.

[0097] The electroencephalogram (EEG) signals are preprocessed using a pipeline designed as follows:

[0098] 1) The electroencephalogram (EEG) signals are divided into fixed-length windows, for example, 15 seconds long. The signals within a fixed-length window are called "segments." The duration of a segment may be longer or shorter than 15 seconds, as long as it is an appropriate length for feature extraction.

[0099] 2) Unwanted frequency components are removed using online bandpass filters with cutoff frequencies of 0.1 Hz and 110 Hz.

[0100] 3) Noise was removed using power line rejection notch filters with frequencies of 60Hz and 120Hz.

[0101] 4) Apply an average standard to remove noise that is common to all electrodes. (Calculation of electromyography signal score)

[0102] In the pain analysis server 1000, an electromyogram score (hereinafter referred to as "EMG score") is calculated to estimate the amount of muscle signals contained in the electroencephalogram data (of each processed segment) for signals that have undergone predetermined processing and have been transmitted from the in-hospital terminal 20. The EMG score is calculated by analyzing the intensity of the frequency component power in the higher frequency gamma band (i.e., 70Hz to 110Hz) as follows. Although not particularly limited, in this embodiment, the EMG score (rank of the electromyogram signal intensity level) is classified into the following four groups. Classification 1: Low EMG component (EMG signal intensity < 30 dB) Classification 2: Moderate EMG component (30 dB < EMG signal intensity < 38 dB) Classification 3: High EMG component (38 dB < EMG signal intensity < 47 dB) Classification 4: Excluded level signal (EMG signal intensity > 47 dB)

[0103] However, the classification of EMG scores is not limited to such classification and is only for illustration. Therefore, more ranks or fewer ranks of classification may be adopted. (Calculation of Signal Quality Index (SQI:Signal Quality Indicator))

[0104] Here, the signal quality index SQI is an index for evaluating the quality of EEG signals and is determined based on electrode impedance and noise artifacts. Although not particularly limited, for example, the signal quality index SQI can be defined as follows. SQI = 100×1 / (1 + Z / Zref) Here, Z is the electrode impedance and Zref is the reference impedance. For example, under normal experimental conditions, it is set to 10 kΩ. For example, when the electrode impedance is 5 kΩ, 10 kΩ, and 20 kΩ, the values of the signal quality index SQI are 67, 50, and 33, respectively. That is, the signal quality index SQI takes values in the range of 0 to 100, and the higher the value, the higher the signal quality. When the signal quality index SQI is 0, it theoretically means the optimal value. When it is around 50, it indicates that the impedance is close to the reference value and is in a standard state. When the signal quality index SQI is 0, it indicates that the impedance is very high and the signal quality is poor. Specifically, to monitor the signal acquisition status, electrode impedance is measured periodically and supplied to a pre-processing module, where a corresponding SQI score indicating the signal quality status during the PS prediction process is calculated. An SQI value greater than, for example, 67% is considered a good signal, while a value less than 33% is considered a completely bad signal. However, values ​​between 67% and 33% are treated as acceptable signals. This evaluation of the signal quality index (SQI) is performed for all channels (electrodes) at predetermined time intervals and timings. (Pain prediction algorithm) (Variation in individual responses)

[0105] The frequency components of the signal band tend to show significant variability among subjects' data, particularly in the high-frequency EEG band (gamma band).

[0106] Therefore, using the EMG score, we divide the datasets used for training into four categories, as summarized in Table 1 below. Datasets with an EMG score exceeding 47 dB are excluded from the training set. [Table 1]

[0107] Due to the large variance in the distribution of the training data, the artificial intelligence models of each algorithm in ensemble learning, as described later, are trained individually using all dataset groups (i.e., the entire dataset, a sub-dataset of low EMG scores, a sub-dataset of medium EMG scores, and a sub-dataset of high EMG scores).

[0108] In this way, the ensemble-learned machine learning models use different models for PS prediction based on EMG scores of real-time measured brainwave signals, which are not included in the training data. (Data preparation for training the PS prediction model)

[0109] For each selected dataset category, valid electroencephalogram (EEG) signals are extracted and divided into segments with 15-second intervals using a 2.5-second movement step.

[0110] Each segment is preprocessed and stored for model training. The mean value of the CoVAS signal for each data segment is used as the PS ground truth label.

[0111] The reference signal is used to extract segments without pain (PS=0). Additionally, segments of the EEG signal representing pain are extracted from the reg-ref signal, which has a CoVAS mean corresponding to a temperature > 44°C (mean pain threshold).

[0112] Figure 8 shows an example of selecting pain zones / painless zones based on CoVAS mean values.

[0113] The lower panel of Figure 8 shows an example of temperature levels and CoVAS signals for the reference conditions. The pain state segment of the electroencephalogram is obtained using CoVAS from the reg-ref signal corresponding to temperature. (Learning model algorithms and machine learning)

[0114] In this way, we train five different learning model algorithms (different regression models) as an example using a relatively large dataset (e.g., 304 subjects) with different thermal stimulation conditions and optimized preprocessing pipelines.

[0115] Figure 9 is a conceptual diagram showing the training dataset used for this type of learning.

[0116] As shown in Figure 9(a), machine learning is performed on the entire dataset and on sub-datasets classified based on EMG score levels, as shown in Figures 9(b) to (d).

[0117] Next, models pre-trained on the same EMG level data group are ensemble-trained and then integrated into a single, optimal machine learning model format.

[0118] The method of integration is not particularly limited, but for example, it could be the average of the PS scores of each model within a single ensemble.

[0119] During PS prediction, the model selection algorithm periodically and quantitatively calculates the EMG level of the received data and selects a corresponding ensemble model set, thereby improving the accuracy and speed of PS prediction.

[0120] As described later, evaluation of a 50-subject test set using a holdout method with data at different EMG score levels revealed an overall improvement in pain estimation performance when using the ensemble machine learning model system of this embodiment (specifically, MSE, MAE, r², and correlation). Furthermore, the PS prediction computation time was also reduced through encapsulation of the machine learning model format, etc. (Machine learning model)

[0121] In this embodiment of ensemble machine learning, electroencephalogram (EEG) signals are used as input, and a regression task is performed to predict the PS score as the output.

[0122] Each model is trained to generate a continuous PS score from 0 (no pain) to 100 (highest pain level). Through ensemble learning, the five different models are combined, and the weighted average prediction of all models is used as the final PS prediction. The model architecture is designed to extract features from different temporal and spatial dimensions, achieving effective data representation learning.

[0123] The following describes the machine learning algorithms included in each machine learning model. (Deep learning model)

[0124] Four end-to-end trainable deep learning models are tuned to fit the task. The input to this group of models is a processed electroencephalogram (EEG) signal, and the output is a PS (Psychostatic Stress) prediction.

[0125] It should be noted that the deep learning models described below are not limited to these configurations, as long as they can implement regression models for discriminating labeled time-series signals, particularly electroencephalogram (EEG) signals; the following explanation is merely illustrative. Furthermore, assembly learning models, as described later, may be constructed using fewer or more learning models and learning algorithms. (1: Chrononet model)

[0126] Figure 10 is a diagram illustrating the model configuration of the Chrononet model, which is a learning model corresponding to the first machine learning algorithm.

[0127] This model is disclosed, for example, in the following literature. Publicly available document 1: Subhrajit Roy et al., “ChronoNet: A Deep Recurrent Neural Network for Abnormal EEG Identification”, arXiv:1802.00308v2 [eess.SP] 18 May 2018

[0128] As shown in Figure 10, this model consists of three convolutional blocks, followed by four gate iteration units (GRUs), and a final block with a fully connected (FC) layer for PS prediction. Each convolutional block is constructed from three parallel 1D convolutional layers with different kernel sizes (temporal receptive fields), followed by a ReLU activation unit. The convolutional kernel sizes are designed to induce the temporal representation of the signal.

[0129] Here, ReLU is a function whose output is always 0 when the input value to the function is 0 or less, and whose output is the same as the input value when the input value is greater than 0. (2: EEG Inception Network Model)

[0130] Figure 11 shows the inception network model, which is a learning model corresponding to the second machine learning algorithm.

[0131] This model is disclosed, for example, in the following literature. Publication 2: Eduardo Santamaria-Vazquez et al., “EEG-Inception: A Novel Deep Convolutional Neural Network for Assistive ERP-Based Brain-Computer Interfaces”, IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 28, NO. 12, DECEMBER 2020, p.2773-2782

[0132] As shown in Figure 11(a), the inception network model is designed to induce data representations at various receptive field scales. The network model consists of three starting blocks (convolutional layers and depth-unit convolutional layers). As shown in Figures 11(b) and (c), each starting block consists of three parallel threads with time-space convolutional layers. The convolutional layers and depth-unit convolutional layers consist of batch normalization, ELU nonlinear activation, and dropout layers. Each starting block is followed by a feature-connected layer, then an average pooling layer for dimensionality reduction, and the last block is followed by a fully connected (FC) layer for PS prediction. (3: EEG Net Model)

[0133] Figure 12 shows the EEG network model, which is a learning model corresponding to the third machine learning algorithm.

[0134] This model is disclosed, for example, in the following literature. Publicly available document 3: Vernon J. Lawhern et al., "EEGNet: A Compact Convolutional Neural Network for EEG-based Brain-Computer Interfaces", arXiv:1611.08024v4 [cs.LG] 16 May 2018

[0135] This EEG network model consists of a 2D convolutional layer comprising three parallel time convolutional layers followed by a spatial convolutional layer, a feature concatenation layer, a batch normalization layer, and a subsequent separable convolutional layer. Each convolutional layer is composed of a batch normalization layer and an ELU (Exponential Linear Unit) nonlinear activation layer. Furthermore, two mean pooling layers are used for dimensionality reduction. Finally, a fully connected (FC) layer follows.

[0136] Here, ELU is a function whose output value is between "0.0" and "-α" (where α is basically 1.0, i.e., "-1.0") when the input value to the function is 0 or less, and whose output value is the same as the input value when the input value is greater than 0. (4: EEGLS™ model)

[0137] Figure 13 shows the EEGLSTM model, which is a learning model corresponding to the fourth machine learning algorithm.

[0138] Furthermore, the following configuration of this model is defined as the "EEGLSTM model".

[0139] As shown in Figure 13, the EEGLSTM model consists of three time convolutional layers 3000.1 to 3000.3, a feature concatenation layer 3010, a batch normalization layer 3020, followed by a processing layer 3030 consisting of a 2D convolutional layer, a batch normalization layer, and a PReLU (Parametric Rectified Linear Unit) activation layer. Subsequently, an average pooling layer 3040 is used to reduce the dimensionality of the intermediate feature map, and after processing by a dropout layer 3050 to reduce overfitting to the training data, a processing layer 3060 consisting of a 2D*2 convolutional layer, a batch normalization layer, and a PReLU (Parametric Rectified Linear Unit) activation layer is performed. Furthermore, an average pooling layer 3070 is used again to reduce the dimensionality of the intermediate feature map, and a dropout layer 3080 is performed to reduce overfitting. Finally, two spatially separable bidirectional LSTM (Long Short Term Memory) layers 3090 and 3100 are provided. The final fully connected (FC) layer 3110 predicts the pain score (PS) from the estimated features.

[0140] Here, PReLU is a function in which, if the input value to the function is less than 0, the output value is α times the input value (where α is a parameter determined by learning), and if the input value is 0 or greater, the output value is the same as the input value.

[0141] Furthermore, LSTM is an improved version of the RNN (Recurrent Neural Network) layer that takes time series into account, in order to solve the vanishing gradient problem inherent in RNNs. Also, a "2D*2 convolutional layer" performs the 2D convolution process twice. (5: Function-based model)

[0142] Figure 14 shows a function-based model, which is a learning model corresponding to the fifth machine learning algorithm. The configuration in Figure 14 defines a "function-based model".

[0143] For this model, the "feature extraction" technique can be used, for example, as disclosed in the following document by the applicant. Publicly available document 4: International Publication No. 2019 / 022242 Publicly available document 5: International Publication No. 2019 / 009420 Furthermore, as an LSTM model, a general-purpose LSTM layer capable of learning long-term dependencies in time-series data can be used.

[0144] Referring to Figure 14, in the function-based model, power spectral density features are extracted based on the frequency bands of the electroencephalogram (EEG) signal. The features are pooled channel by channel and used to train an LSTM-based predictive model. The predictive model network consists of LSTM layers that receive an input sequence and pass it to fully connected (FC) layers that output a continuous PS prediction of the hidden state for the last time step. (Model training)

[0145] Model Training: All deep networks (including LSTMs in feature-based models) are trained with the Adam optimizer using a mean squared error or mean absolute error loss function (with different penalty weights for pain / painless regions, in some cases).

[0146] Approximately 20% of the training dataset is used for validation, which is performed after each epoch to ensure training stability. The initial learning rate is modified to avoid plateaus in learning. An early stopping strategy is also implemented based on the loss value of the validation set to avoid overfitting to the training data. (Combined with the model's ensemble)

[0147] Candidate models for each EMG score dataset category are used in ensemble machine learning for PS prediction. In this configuration, the outputs from all models are weighted and averaged to obtain the final PS prediction.

[0148] Figure 15 is a flowchart illustrating the pain score PS prediction process using the ensemble machine learning model trained in this manner.

[0149] As mentioned above, in this embodiment, the electromyography (EMG) level is divided into three ranks (low, medium, and high). However, the number of such ranks is not limited to three. In Figure 15, it is assumed that there are M ranks, and an ensemble machine learning model is generated corresponding to each. Adding this to the initial model that uses data from all ranks as training data, (M+1) models are used, and the pain score (PS) prediction process is performed on the EEG data measured online, switching (selecting) the model according to the EMG score.

[0150] Referring to Figure 15, in the pain analysis server 1000, when the electroencephalogram data receiving module 1110 acquires the subject's electroencephalogram signal data (S100), it divides the electroencephalogram signal data into segments (S102).

[0151] The pain discrimination module 1140 sets the variable i, which represents the segment obtained as a time series, to 1 (S104).

[0152] First, the signal quality management module 1120 determines the signal quality (S106), and if the signal quality does not reach a predetermined level (deemed "bad" in S106), it is excluded from the prediction determination (S110).

[0153] On the other hand, if the signal quality management module 1120 determines that the signal quality has reached a predetermined level (good in S106), the electromyography rank determination module 1130 determines the rank of the electromyography (S108).

[0154] Depending on the determination result of the electromyography rank determination module 1130 and the stage of the prediction processing, the pain determination module 1140 selects a model to be used for prediction (S120).

[0155] In other words, the pain assessment module 1140 initially uses a default model (a model trained on the entire dataset), but continues to track the EMG score values ​​of the electroencephalogram signals (segments) being processed. Periodically, the average EMG score is calculated, and an ensemble machine learning model of the corresponding EMG score is selected accordingly.

[0156] This configuration allows for optimization of PS prediction under various ranges of EMG interference. In other words, PS prediction is improved with varying levels of muscle signal interference.

[0157] Here, initially, a default model is selected, and from the signal of the segment after the next predetermined period, a model of rank 1 to rank 3 (in this case, M=3 in Figure 15) is selected according to the average EMG score for the segment during that period (S122.0 to S122.M).

[0158] The pain score PS is output from the selected model (S124) and sent back to the in-hospital terminal 20. At the same time, the EMG score, or more specifically, the average EMG score used to calculate the pain score PS, is also sent back to the in-hospital terminal.

[0159] Next, it is determined whether the process has been instructed to end. If the process has not been instructed to end (N in S126), the pain determination module 1140 increments the variable i by 1 and returns the process to step S106. If the process has been instructed to end (Y in S126), the pain determination module 1140 terminates the process.

[0160] Figure 16 is a diagram illustrating the coordination process between the in-hospital terminal 20 and the pain level analysis server 50.

[0161] Figure 17 shows an example of the content displayed on the display screen 22 of the in-hospital terminal 20 during real-time measurement.

[0162] Referring to Figure 16, the patient's attribute information is registered on the in-hospital terminal 20 (S200), and the acquisition of electroencephalogram signals by the electroencephalograph 10 begins (S202).

[0163] In the electroencephalograph 10, the electroencephalogram signal is amplified, A / D converted, filtered, etc. (S204), and the electroencephalogram signal data is transmitted to the in-hospital terminal 20 and displayed on the display screen 22 of the in-hospital terminal 20 (S206).

[0164] Figure 17 shows an example of an electroencephalogram (EEG) signal waveform.

[0165] Meanwhile, the in-hospital terminal 20 detects the contact status of the electrodes (S206). If necessary, the measurement status is displayed in the message field.

[0166] Measurement signal data (EEG measurement signal data and electrode contact status information) is transmitted from the in-hospital terminal 20 to the pain analysis server 1000 (S210).

[0167] When the pain analysis server 1000 receives measurement signal data (S212), it stores the data in the database server 2000 (S214) and executes the "pain level analysis process" described in Figure 15 (S220).

[0168] The pain analysis server 1000 sends back the pain score (PS), electromyography rank information, and signal quality information to the in-hospital terminal 20 (S222).

[0169] The in-hospital terminal 20 displays the electromyography rank, signal quality indicators, and pain score (PS) on the display screen 22.

[0170] Figure 17 shows an example of such a display. In Figure 17, as an example, signal quality and electromyography signal rank are displayed using predetermined symbols indicating levels, and the pain score (PS) is displayed as a numerical value. However, the display method is not limited to this example.

[0171] On the in-hospital terminal 20, if the measurement is not yet complete, the process returns to step S202; if the process is complete, the measurement is terminated. [Evaluation Results]

[0172] The following describes the results of evaluating the pain score (PS) calculation using the algorithm described above, with an unknown test set (50 subjects). (Evaluation of the predictive performance of ensemble models) (Performance evaluation criteria)

[0173] Test set: To evaluate PS prediction performance across different models, a dataset of 50 subjects is used as the final evaluation test set.

[0174] This test set was randomly selected from samples of all EMG score groups. Table 2 shows the distribution of EMG scores for subjects in the test dataset. [Table 2]

[0175] The following is an overview of the evaluation results for the model's performance.

[0176] A detailed comparison of the model's performance will be analyzed using the following metrics. a) Overall mean absolute error (MAE):

[0177] It is calculated as the average of the absolute differences (errors) between the predicted value and the actual value.

number

[0178] Here, yi is the true value, and yi (hat) is the predicted value. MAE measures the error, with a lower value indicating better performance. b) Mean Squared Error (MSE) of the whole

[0179] It is calculated as the mean square of the difference (error) between the predicted value and the actual value.

number

[0180] A lower MSE value indicates higher model accuracy. c) Coefficient of determination (R²)

[0181] This provides a measure of how well the model's predictions reproduce the actual score. The best possible score is 1.0, and negative values ​​are possible. d) Pearson correlation coefficient (Corr)

[0182] The degree of linear association between the true variable and the predictor variable is quantified on a scale of (+1 to -1), ranging from a perfectly positive linear relationship to a perfectly negative linear relationship, respectively.

number

[0183] Figure 18 shows the evaluation results of the predictive performance of encapsulated individual models and ensemble models.

[0184] It can be seen that ensemble models tend to show better performance metrics compared to individual models.

[0185] Furthermore, a comparison of the average time spent on single-segment prediction and memory usage between the ensemble model and its corresponding individual models confirmed that the ensemble model is effective in terms of inference time and memory usage. (Evaluation of the model selection approach)

[0186] Table 3 summarizes the PS predicted performance scores, which are the average scores of all subjects in the test set. [Table 3]

[0187] By selecting a model, all evaluation metrics improve compared to using only the default model.

[0188] As described above, the brain image analysis device of this embodiment makes it possible to provide information that can assist in diagnosis by utilizing brain tomography image data accumulated as big data and performing analysis that characterizes multiple neurodegenerative diseases using at least two of the volume and image signal intensity of multiple brain structures.

[0189] The embodiments disclosed herein are illustrative of configurations for specifically carrying out the present invention and do not limit the technical scope of the present invention. The technical scope of the present invention is indicated by the claims rather than by the description of the embodiments, and modifications within the literal scope and equivalent meaning of the claims are intended. [Explanation of Symbols]

[0190] 2 electrodes, 10 electroencephalographs, 20 in-hospital terminals, 22 display screens, 50 pain level analysis servers, 100 pain analysis systems, 1000 pain analysis servers, 1100 arithmetic processing units, 1110 electroencephalogram data receiving modules, 1120 signal quality management modules, 1130 electromyography rank determination modules, 1140 pain determination modules, 1150 model selection modules, 1160.1~1160.M ensemble trained models, 1200 memory, 1300 storage units, 2000 data servers.

Claims

1. An analysis system for evaluating the level of pain felt by a subject, A measuring device for detecting the electroencephalogram signals of the subject, An analysis device that analyzes the pain level based on the signal from the aforementioned measuring device, The system includes an information display device that receives signals from the measuring device, transmits them to the analysis device, and has a display unit for presenting the analysis results from the analysis device to the user. The aforementioned analysis device is An electromyography level determination unit that evaluates the electromyography signals included in the electroencephalogram signal according to the electromyography level, The system includes a discrimination processing unit that receives a signal from the measuring device and calculates the pain level using a trained model generated by machine learning, The aforementioned trained model is an analysis system that learns based on a training dataset corresponding to the electromyography level and outputs a result for determining the pain level.

2. The analysis system according to claim 1, further comprising a model selection unit that switches between the learned models used to calculate the pain level over time, according to the electromyography level determined by the electromyography level determination unit.

3. The measuring device is an electroencephalograph that detects electroencephalogram signals over time from multiple predetermined positions on the subject's head. The electromyography level determination unit evaluates the electromyography signals included in the electroencephalograph signal as a rank of multiple electromyography levels over time, The trained model includes multiple ensemble trained models corresponding to each rank, generated by ensemble learning based on training datasets for each of the multiple ranks of the electromyography level. Each of the aforementioned ensemble trained models integrates the discrimination results of multiple submodels trained with different machine learning algorithms. The analysis system according to claim 1, wherein the discrimination processing unit further includes a model selection unit that switches the learned model used to calculate the pain level from among the plurality of learned models according to the rank of the electromyography level determined by the electromyography level determination unit, as time progresses.

4. The electroencephalogram signal is divided into segments of predetermined time intervals, The analysis system according to any one of claims 1 to 3, wherein the electromyography level determination unit determines the electromyography level according to the power of the signal in a predetermined frequency band in each segment of the electroencephalogram signal.

5. The analysis device outputs information indicating the electromyography level evaluated by the electromyography level determination unit over time. The analysis system according to claim 4, wherein the information display device outputs information indicating the electromyography level to the display unit.

6. The electroencephalograph is attached to multiple predetermined positions on the subject's head and outputs monitoring signals over time, which measure the impedance of multiple electrodes used to measure the electroencephalogram signals. The analysis system according to claim 3, further comprising a signal quality monitoring unit that determines the quality of the signal in accordance with the monitoring signal and excludes signals below a predetermined quality from the analysis target.

7. The analysis system according to claim 3, wherein the distinct machine learning algorithms include at least two of the following: a deep learning algorithm, a Kronnet algorithm, an EEG inception network algorithm, an EEG net model, and an EEG LSTM algorithm.

8. An analysis device for evaluating the level of pain felt by a subject, based on signals from a measuring device for detecting electroencephalogram signals from multiple predetermined locations on the subject's head, An electromyography level determination unit that evaluates the electromyography signals included in the electroencephalogram signal according to the electromyography level, The system includes a discrimination processing unit that receives a signal from the measuring device and calculates the pain level using a trained model generated by machine learning, The aforementioned trained model is an analysis device that learns based on a training dataset corresponding to the electromyography level and outputs a result for determining the level of pain.

9. The analysis apparatus according to claim 8, further comprising a model selection unit that switches among the learned models, which calculate the pain level over time, according to the electromyography level determined by the electromyography level determination unit.

10. The measuring device is an electroencephalograph that detects electroencephalogram signals over time from multiple predetermined positions on the subject's head. The electromyography level determination unit evaluates the electromyography signals included in the electroencephalograph signal as a rank of multiple electromyography levels over time, The trained model includes multiple ensemble trained models corresponding to each rank, generated by ensemble learning based on training datasets for each of the multiple ranks of the electromyography level. Each of the aforementioned ensemble trained models integrates the discrimination results of multiple submodels trained with different machine learning algorithms. The analysis apparatus according to claim 8, wherein the discrimination processing unit further includes a model selection unit that switches the learned model used to calculate the pain level from among the plurality of learned models according to the rank of the electromyography level determined by the electromyography level determination unit, as time progresses.

11. The electroencephalogram signal is divided into segments of predetermined time intervals, The analysis device according to claim 8 to 10, wherein the electromyography level determination unit determines the electromyography level according to the power of the signal in a predetermined frequency band in each segment of the electroencephalogram signal.

12. The electroencephalograph is attached to multiple predetermined positions on the subject's head and outputs monitoring signals over time, which measure the impedance of multiple electrodes used to measure the electroencephalogram signals. The analysis apparatus according to claim 10, further comprising a signal quality monitoring unit that determines the quality of the signal in accordance with the monitoring signal and excludes signals below a predetermined quality from the analysis target.

13. The analysis apparatus according to claim 10, wherein the distinct machine learning algorithms include at least two of the following: a deep learning algorithm, a Kronnet algorithm, an EEG inception network algorithm, an EEG net model, and an EEG LSTM algorithm.

14. The analysis device according to claim 10, wherein the analysis device transmits information on the pain level and information indicating the rank of the electromyography level evaluated by the electromyography level determination unit in a format for the user, via an information presentation device that receives a signal from the electroencephalograph and transmits it to the analysis device.

15. The analysis device according to claim 14, wherein the analysis device outputs information indicating the rank of the electromyography level evaluated by the electromyography level determination unit over time so as to be displayed on the display unit of the information presentation device.

16. An analysis method for evaluating the level of pain felt by a subject using a computer, The computer includes an arithmetic unit and a memory device, The calculation device stores electroencephalogram (EEG) signal data from a measuring device for detecting the subject's EEG signal in the storage device. The calculation device performs the steps of evaluating the electromyography signal components included in the electroencephalogram signal data according to the electromyography level, The computing device includes the step of receiving the electroencephalogram signal data and calculating the pain level using a trained model generated by machine learning, The aforementioned trained model learns based on a training dataset corresponding to the electromyography level and outputs a result for determining the pain level. An analysis method comprising the step of presenting the results of the analysis of the pain level to the user.

17. An analysis program that uses a computer to evaluate the level of pain felt by a subject, The computer includes an arithmetic unit and a memory device, The analysis program is provided to the computer, The calculation device stores electroencephalogram (EEG) signal data from a measuring device for detecting the subject's EEG signal in the storage device. The calculation device performs the steps of evaluating the electromyography signal components included in the electroencephalogram signal data according to the electromyography level, The calculation device receives the electroencephalogram signal data and calculates the pain level using a trained model generated by machine learning. The process involves the user being presented with the results of the pain level analysis, The aforementioned trained model is an analysis program that learns based on a training dataset corresponding to the electromyography level and outputs a result for determining the pain level.

Images