Recognition and extraction of EEG signals

Abstract

A method for identifying and extracting electroencephalogram (EEG) signals associated with pain is disclosed. The method includes: receiving EEG data for each trial from one or more trials; determining a current density for each signal; estimating the current density for a set of neural activity regions of interest based on the calculated current density; and calculating at least one spectral feature for each trial based on the estimated current density. Thus, for each neural activity region of interest, a mean and variance of the variation in EEG data between EEG data labeled as indicating a pain state and EEG data labeled as indicating a non-pain state can be calculated, and based on at least one region of interest, an EEG signal associated with pain can be identified, wherein the variance is below a predetermined threshold in the at least one region of interest.

Description

Technical Field

[0001] The present invention generally relates to a process for identifying electroencephalogram (EEG) patterns associated with pain. The present invention also relates to an interactive process for pain relief training. Background Art

[0002] Chronic pain, defined as persistent pain lasting more than three months, is a complex condition that affects the central nervous system and may even be caused by its dysfunction. In 1986, the World Health Organization declared that pain continues to affect many people worldwide. Furthermore, in Singapore, chronic pain remains a common problem, affecting an estimated 1 in 10 people.

[0003] The GSK (GlaxoSmithKline) Global Pain Index 2017 report states that pain affects workplace productivity and quality of life, costing the global economy an estimated $245 billion annually.

[0004] Due to the complex pathophysiology of chronic pain, several treatment options exist, including medication, minimally invasive interventions, and open surgery. However, the most common pain treatment options are primarily medication-based, including opioids, anticonvulsants, and anti-inflammatory agents. These treatments are expensive and have relatively low success rates. Furthermore, these treatments carry potential drawbacks, including serious side effects and complications, which can lead to user dependence.

[0005] In short, the two most prominent problems facing the management of chronic pain are the lack of objective methods to detect pain and the lack of inexpensive, noninvasive, and safe methods to treat the central nervous system when dealing with pain.

[0006] It would therefore be desirable to provide solutions that address one or more of the above-mentioned shortcomings, eg, those associated with existing pain treatment methods, or at least provide useful alternatives. Summary of the Invention

[0007] The present invention was developed in view of the medical burden of managing chronic pain and the negative impact of chronic pain on individuals (physically, emotionally and financially), other people and society, as well as the challenges in restoring fear-avoidance behaviors associated with pain. The embodiments of the brain-computer interface (BCI) described herein are capable of identifying unique pain-specific EEG patterns that match pain-specific conditions and display fear-avoidance behaviors. These BCIs can also be used as therapeutic tools to regulate pain. If an active pain management model is maintained over the long term, this can significantly affect the way pain is rehabilitated during the healthcare process. Specifically, this can enable customized treatment methods for pain-specific conditions and help to better understand and treat fear-avoidance behaviors. In addition, if the benefits of BCI training are sustained, healthcare costs can also be reduced in the long run. Therefore, a BCI-based pain detection and / or relief system is disclosed.

[0008] In a first aspect, the present invention provides a method for identifying and extracting electroencephalogram (EEG) signals associated with pain, the method comprising:

[0009] receiving EEG data for each trial from one or more trials, the EEG data comprising one or more signals, each of the one or more signals being associated with a corresponding coordinate vector, the EEG data being labeled to indicate a pain state and / or a no-pain state;

[0010] determining a current density for each of the one or more signals in a corresponding coordinate vector;

[0011] estimating the current density of a set of neural activity regions of interest based on the calculated current density;

[0012] calculating at least one spectral feature for each trial based on the estimated current density;

[0013] calculating, for each region of neural activity of interest, a mean and a variance of EEG data variation between EEG data labeled as indicating a pain state and EEG data labeled as indicating a no-pain state based on the at least one spectral feature; and

[0014] Pain-related EEG signals are identified based on at least one of the regions of interest, wherein a variance is below a predetermined threshold in the at least one of the regions of interest.

[0015] The method may further include extracting pain-related EEG signals from the one or more regions of interest.

[0016] Thus, the method may search for the best set of cortical locations / EEG coordinates on the scalp / region of interest where the current density activity estimated from the potentials of the EEG shows the most consistent change from no-pain EEG to pain EEG.

[0017] The method may further comprise constructing a classifier to predict whether the EEG recording is associated with pain or the absence of pain. The method may further comprise constructing a regression machine to predict whether the EEG recording is associated with a level of perceived pain. The machine may employ an algorithm that uses current density activity measurements for each EEG recording. Constructing the classifier / regressor may comprise constructing a machine that uses one or a combination of a support vector machine, a deep neural network, etc., with linear mechanisms being preferred. The output of the classifier / regressor is a scalar indicator (a scalar (rather than a binary output value) can typically be calculated based on a binary classifier), and the method further comprises using the scalar indicator to indicate the level of painful EEG activation - that is, to which of at least two pain-related EEG categories the EEG belongs.

[0018] Calculating the mean and variance of the EEG data variation may include calculating at least one of an inter-class scatter matrix and a total scatter matrix based on the spectral features to determine the covariance between at least two pain-related EEG classes.

[0019] The step of receiving EEG data may include receiving EEG data that does not exceed a motion threshold, such as motion artifacts from the EEG. The motion threshold may be set at a level at which motion artifacts on the recorded EEG are insufficient to reduce the accuracy of the above method. Exceeding the motion threshold may be determined by a motion sensor attached to the subject from which the EEG data is captured, or by the subject's clinician actively marking the EEG data in the presence of movement, such as by pressing a button to mark the EEG data. Motion artifacts may also be reduced or removed by having the subject sit, stand, or lie still during the test, or by preprocessing the EEG to remove motion artifacts.

[0020] The EEG categories associated with pain can simply be a pain EEG category and a no-pain EEG category, where the pain EEG category indicates that the subject was in pain at the time the EEG was recorded, and the no-pain EEG category indicates that the subject was not in pain at the time the EEG was recorded. In some cases, there may be more than two categories to represent different degrees of pain. Therefore, at least two EEG categories associated with pain are a pain EEG category and a no-pain EEG category.

[0021] The current density may be a radial current density.

[0022] Calculating the mean and variance of the EEG data variation may include calculating at least one of an inter-class scatter matrix and a total scatter matrix. Both the inter-class scatter matrix and the total scatter matrix may be calculated, and identifying the EEG signal may include satisfying the following equation:

[0023] argminL f(L),

[0024] Where f(L) is the multivariate Fisher score of the neural activity region of interest L, where f(L) is:

[0025] f(L)=trace{S b (S t +γι) -1},

[0026] Among them, S b is the inter-category scatter matrix, S t is the total scatter matrix, γ is the forward regularization parameter, and ι is the identity matrix.

[0027] The region of interest may be a cortical location / EEG coordinate on the scalp.

[0028] The at least two pain-related EEG categories may be a pain EEG category and a no-pain EEG category.

[0029] The method may further comprise:

[0030] constructing a binary classifier, the binary classifier receiving current density or current density activity from further EEG data and outputting a scalar indicator;

[0031] receiving further EEG data;

[0032] applying a binary classification machine to the further EEG data; and classifying the further EEG data into EEG data indicative of a pain state or EEG data indicative of a no-pain state based on a scalar indicator associated with the further EEG data.

[0033] In one embodiment, a method for identifying and extracting electroencephalogram (EEG) signals associated with pain, and creating a classifier therefrom, includes:

[0034] (Pain EEG data acquisition) An EEG device is attached to a chronic pain patient, and when a pain attack event occurs, the patient / operator should register the event into the EEG data stream, which is recorded in digital form; if the pain sensation decreases significantly after a period of time, he / she should register the decreased pain event into the EEG data stream; the patient's perceived pain score can also be registered; and the process is repeated to record more such pain events.

[0035] The person is placed in a static position (eg, sitting, standing, lying down, etc.), and if there is no pain sensation, pain-free EEG data is recorded for a period of time.

[0036] Non-pain EEG recordings and pain EEG recordings from other chronic pain patients were collected according to the same protocol.

[0037] The recorded data are processed using a computer to identify pain-related brain activation patterns in the EEG; in principle, a computer algorithm described in detail in a technical report uses an iterative optimization process to search for the best set of cortical locations / EEG coordinates on the scalp where the current density activity estimated from the EEG potential shows the most consistent change from no-pain EEG to pain EEG, where

[0038] Current density activity can refer to the characteristic spectral characteristics of the natural variation of current density over time, such as power in a frequency band;

[0039] Statistics such as Fisher's discriminant or KL (Kullback–Leibler) divergence can be used to measure the consistency of the change from no-pain EEG to pain EEG;

[0040] Current density can be estimated using spherical spline interpolation of the surface electric potential (EEG), for example, as proposed by Perrin et al. (1989);

[0041] A classification algorithm / regression algorithm is constructed using a computer to predict whether an EEG recording is associated with pain or no pain (in the case of classification) or with perceived pain level (in the case of regression, as long as the patient's perceived pain level is registered as described above); the algorithm uses the current density activity measurements described above for each EEG recording; the classification mechanism can use any method, such as support vector machines, deep neural networks, etc. or a combination thereof, and linear mechanisms are preferred; the output of the algorithm is then used as a scalar indicator (a scalar (rather than a binary output value) can typically be calculated based on a binary classification machine) to indicate the level of painful EEG activation.

[0042] Also disclosed herein is a computer method for quantifying activation of EEG signal activity associated with pain, the computer method comprising:

[0043] Performing the computer method described above on the EEG data to calculate the current density activity measurements indicated above,

[0044] Classification / regression algorithms are used to calculate the level of pain-related EEG activity.

[0045] The computer method may further include: receiving an initial input comprising a numerical pain level estimate; wherein receiving further EEG data includes: receiving consecutive EEG time periods from continuously acquired EEG data; the computer method further includes: re-estimating the numerical pain level in each consecutive EEG time period based on the above-mentioned quantification of the activation level of the pain-related EEG signal.

[0046] The above-described computer process or method provides a novel, adaptive, and participant-specific pain neural matrix detection and analysis procedure that explores and identifies discriminative and robust patterns in spontaneous EEG for use in pain modeling and decoding.

[0047] Also disclosed herein is a computer method for quantifying activation of EEG signal activity associated with pain, the computer method comprising:

[0048] Performing the computer method described above on the EEG data to calculate the current density activity measurements indicated above,

[0049] Use classification / regression algorithms to calculate the level of pain-related EEG activity;

[0050] The further EEG data is classified using a binary classification machine to determine whether the further EEG data represents a pain EEG class or a no-pain EEG class.

[0051] Also disclosed herein is a computer method for classifying pain-related EEG signals, the computer method comprising:

[0052] Collect EEG data from multiple trials, including:

[0053] at least one trial performed by a first subject susceptible to pain perception; and

[0054] At least one test performed with a healthy second subject,

[0055] Wherein, the EEG data of each trial includes one or more signals;

[0056] performing the computer method as described above on the EEG data and setting an initial coordinate vector for each signal; and

[0057] constructing a binary classifier to distinguish between at least two pain-related EEG states based on EEG measurements of at least one region of interest;

[0058] receiving further EEG data; and

[0059] Apply the binary classification machine to further EEG data;

[0060] The scalar indicator is received from the binary classifier and further EEG data is classified into a pain-indicating EEG class or a no-pain EEG class based on the scalar indicator.

[0061] A binary classification machine can be at least one of the following:

[0062] Support vector machines;

[0063] Multi-layer neural networks; and

[0064] Generalized Linear Discriminant Analyzer.

[0065] Calculating the mean and variance may include calculating at least one of an inter-class scatter matrix and a total scatter matrix based on the at least one spectral feature to determine the covariance between the at least two pain-related EEG classes. Calculating the inter-class scatter matrix and the total scatter matrix may include:

[0066] generating a spline interpolation matrix for at least two electrical characteristics of the EEG data;

[0067] EEG data were converted into current density estimates using a spline interpolation matrix;

[0068] Calculate the power in the band of the current density estimate; and

[0069] Using the band powers of the current density estimates, the between-class scatter matrix and the total scatter matrix were calculated.

[0070] Calculating the power in the band for the current density estimate may include:

[0071] Determine a current density time series for a current density estimate;

[0072] A vector of band powers of the current density time series is calculated using at least one of Fourier-based decomposition and bandpass filtering followed by energy calculation.

[0073] Classification of further EEG samples using a binary classification machine may include:

[0074] converting the further EEG data into further current density estimates using the spline interpolation matrix; calculating frequency band power for the further current density estimates; and

[0075] A binary classifier is applied to the frequency band powers, wherein a positive output of the binary classifier indicates a first state of a pain EEG state and a no-pain EEG state of a source of further EEG data, and wherein a negative output of the binary classifier indicates a second state of the pain EEG state and the no-pain EEG state of the source, the second state being different from the first state.

[0076] The computer method may further include: receiving an initial input comprising a numerical pain level estimate; receiving further EEG data may include: receiving consecutive EEG time periods from continuously recorded (i.e., acquired) EEG data; the computer method may further include: re-estimating (e.g., increasing or decreasing) the numerical pain level estimate based on whether each consecutive EEG time period is classified as representing a pain EEG category or a no-pain EEG category, for example, whether the above-mentioned quantification of the activation level of the pain-related EEG signal is performed in each consecutive EEG time period.

[0077] The above-described computer process or method provides a novel, adaptive, and participant-specific pain neural matrix detection and analysis procedure that explores and identifies discriminative and robust patterns in spontaneous EEG for use in pain modeling and decoding.

[0078] Also disclosed herein is a system (e.g., a portable electronic device such as a smartphone) for identifying and extracting pain-related EEG signals, the system comprising:

[0079] EEG signal source;

[0080] Memory; and

[0081] At least one processor, the memory storing instructions, which, when executed by the at least one processor, cause the at least one processor to:

[0082] receiving EEG data for each trial from one or more trials, the EEG data comprising one or more signals, each of the one or more signals being associated with a corresponding coordinate vector, the EEG data being labeled to indicate a pain state and / or a no-pain state;

[0083] determining a current density for each of the one or more signals in a corresponding coordinate vector;

[0084] estimating the current density of a set of neural activity regions of interest based on the calculated current density;

[0085] calculating at least one spectral feature for each trial based on the estimated current density;

[0086] calculating, for each region of neural activity of interest, a mean and a variance of EEG data variation between EEG data labeled as indicating a pain state and EEG data labeled as indicating a no-pain state based on the at least one spectral feature; and

[0087] Pain-related EEG signals are identified based on the one or more regions of interest, wherein a variance is below a predetermined threshold in the one or more regions of interest.

[0088] In some embodiments, the EEG signal source may be a memory device or a wire that stores EEG data. In other embodiments, the EEG signal source may be a subject, and the EEG signal is recorded directly from the subject.

[0089] The electric potential field characteristic may be radial current density.

[0090] At least one processor may calculate both the between-class scatter matrix and the total scatter matrix and may identify EEG signals associated with pain by satisfying the following equation:

[0091] argmin L f(L),

[0092] where f(L) is the multivariate Fisher score of the neural activity region of interest L, where f(L) is:

[0093] f(L)=trace{S b (S t +γι) -1},

[0094] Among them, S b is the inter-category scatter matrix, S t is the total scatter matrix, γ is the forward regularization parameter, and ι is the identity matrix.

[0095] The EEG signal source may be arranged to collect EEG data from a plurality of trials, the plurality of trials comprising:

[0096] at least one trial performed by a first subject susceptible to pain perception; and

[0097] At least one test performed with a healthy second subject,

[0098] Wherein, the EEG data of each trial includes one or more signals;

[0099] Wherein, at least one processor is configured to:

[0100] applying an initial coordinate vector to each of said signals;

[0101] constructing a binary classifier to distinguish between at least two pain-related EEG states based on EEG measurements from at least one region of interest for extracting pain-related EEG signals;

[0102] receiving further EEG data; and

[0103] The further EEG data is classified using a binary classification machine to output a scalar indicator, and whether the further EEG data represents a pain EEG class or a no-pain EEG class is determined based on the scalar indicator.

[0104] The system may further include an input device for receiving an initial input comprising a numerical pain level estimate, wherein the at least one processor is configured to:

[0105] receiving further EEG data by receiving consecutive EEG time segments from the continuously recorded EEG data; and

[0106] The numerical pain level estimate is increased or decreased based on whether each consecutive EEG time segment is classified as representing a pain EEG class or a no-pain EEG class.

[0107] The system may further include:

[0108] monitor;

[0109] input devices; and

[0110] one or more EEG devices, the one or more EEG devices being for each of the at least one region of interest and being located in the respective regions of interest, the one or more EEG devices being configured to record continuously recorded EEG data, receiving continuous EEG time segments from the continuously recorded EEG data,

[0111] Wherein, at least one processor is further configured to:

[0112] determining an attention score for a first EEG time segment from the consecutive EEG time segments;

[0113] Display on the monitor:

[0114] Numerical estimation of pain levels;

[0115] Attention scores; and

[0116] interactive activities;

[0117] During interaction with the interactive activity via the input device, continuously updating the numerical pain level estimate and attention score based on subsequent EEG time segments from the continuous EEG time segment; and

[0118] The behavior of the interactive activity was adjusted to reduce the number of subsequent EEG time segments that were classified into the pain EEG category.

[0119] The one or more EEG devices may be adapted to apply simulation at one or more of the at least one region of interest, wherein recording the continuously recorded EEG using the one or more EEG devices comprises recording the EEG resulting from stimulation applied by the one or more EEG devices at the at least one region of interest.

[0120] The above system can provide a novel brain-computer interface (BCI) pain relief system. The system can be based on or operate through pain perception detection and cognitive functions, which are calculated based on interactive audio-visual feedback for pain modulation, or calculated based on other neural stimulation mechanisms (e.g., vagus nerve stimulation).

[0121] Some embodiments of the system can identify pain signatures by correlating electroencephalogram (EEG) signals with pain episodes in a group of healthy participants and chronic pain patients. The system can then develop a training and processing mechanism or sequence designed to model and decode the pain neural matrix. A closed-loop sensing and neurofeedback mechanism is then established to deliver pain neuromodulation therapy that uses shared attention and monitoring and stimulation of pain neural matrix activity to help chronic pain patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0122] Embodiments of the present invention will now be described by way of non-limiting examples with reference to the accompanying drawings, in which:

[0123] Figure 1 is a flowchart of a computer method for identifying and extracting pain-related electroencephalogram (EEG) signals according to the present teachings;

[0124] Figure 2 Photographs of a prototype or experimental setup used to capture EEG data during performance of a specific set of activities;

[0125] Figure 3 for Figure 2 A schematic diagram of the arrangement in;

[0126] Figure 4 A flowchart illustrating the process of acquiring an EEG associated with pain-inducing activity;

[0127] Figure 5 includes Figures 5a to 5d , showing the various sites where the specified activities were performed to induce pain;

[0128] Figure 6 A series of still images taken from a video of actors performing everyday tasks that can induce pain;

[0129] Figure 7Results from experiments are presented in which subjects with chronic pain exhibited elevated activation levels during movement in the alpha, medium beta, and high beta ranges compared to healthy subjects;

[0130] Figure 8 shows experimental results obtained by observing videos of actors performing everyday movements designed to induce pain;

[0131] Figure 9 An exemplary pain EEG time period is shown - marking both heavier and lighter pain events in the EEG signal;

[0132] Figure 10 Event counts for various pain events are shown, with the top graph showing the total number of heavier and lighter pain events for each subject, and the bottom graph showing pain events rejected due to noisy EEG;

[0133] Figure 11 A photograph of a training system arrangement for BCI-based pain neuromodulation therapy using shared attention and monitoring and feedback of pain neural matrix activity according to the present teachings;

[0134] Figure 12 for Figure 11 System diagram of the system in

[0135] Figure 13 for Figure 11 and Figure 12 Flowchart of the system usage;

[0136] Figure 14 A user interface for gamified presentation of a health score and an attention score according to the present teachings;

[0137] Figure 15 is a flow chart illustrating a non-drug based approach to chronic pain management according to the present teachings;

[0138] Figure 16 Shown is a participant in a gamified therapy program wearing an EEG cap and playing a space game on a laptop with a joystick;

[0139] Figure 17 showing a user interface displaying a first question for estimating a subject's pain level;

[0140] Figure 18 showing a user interface displaying a second question for estimating the subject's pain level; and

[0141] Figures 19 to 24A display is provided for displaying a user interface in the game, the user interface displaying a concentration score, a health or pain level estimation score, and interactive activities. DETAILED DESCRIPTION

[0142] While the present disclosure encompasses various embodiments of the computer methods and systems described above, two categories of embodiments are broadly described, namely:

[0143] (i) Adaptive and potentially participant-specific pain neural matrix detection and analysis procedures. This can be achieved by exploring and identifying discriminative and robust patterns in spontaneous EEG for use in pain modeling and decoding.

[0144] (ii) BCI-based pain neural matrix activity and attention monitoring and feedback training system based on pain perception detection and cognitive functions, which are calculated based on interactive audio-visual feedback or based on brain stimulation mechanisms to modulate pain.

[0145] For adaptive and potential participant or subject-specific pain neural matrix detection and analysis procedures, the advantages of well-labeled sensory and related specific cognitive and pain matrix functions can be exploited in patients with chronic low back pain or chronic lower limb pain compared to healthy controls. This helps to identify discriminative and robust patterns in spontaneous EEG and in induced sensory EEG for use in pain prediction. Therefore, EEG data can be extracted only from those regions or locations that help to distinguish between EEG signals of the pain EEG category and EEG signals of the no-pain category, that is, signals indicating that the subject was in pain when the EEG was recorded, and EEG signals of the no-pain category, that is, signals indicating that the subject was not in pain when the EEG was recorded.

[0146] In this example, modeling through pain neural matrix detection and analysis uses an experimental design involving acquisition of EEG signals from patients with chronic low back pain or chronic lower limb pain when compared to healthy controls, i.e., healthy or pain-free subjects, under two pain-inducing conditions: a movement-inducing condition and a fear-inducing condition.

[0147] A BCI-based pain neural matrix activity and attention monitoring and feedback training system was discussed, featuring a game interface with audiovisual feedback. This game interface informs users of their current brain activation and attention levels. The game also guides users in learning to modulate EEG features and develop skills in managing attention to alleviate perceived fear-related pain.

[0148] The BCI system captures EEG signals and decodes the underlying brain states associated with cognitive and fear-related pain perception. This decoded brain state is then presented to the participant in a visual or other form to guide the participant in learning to regulate the brain state to achieve better pain management. For example, over several sessions, a participant or subject can learn to focus on visual feedback while suppressing brain functional activity associated with fear-related pain perception.

[0149] Therefore, certain embodiments of the computer methods and systems described herein are intended to provide:

[0150] a. A shared monitoring and feedback mechanism for attention and pain neural matrix activity.

[0151] b. An interactive interface that helps patients or subjects learn to relieve perceived pain through a game based on audio-visual feedback that informs the patient or subject of their pain level and mental attention level in real time.

[0152] c. An adaptive neural stimulation / feedback mechanism that relieves pain by learning and optimizing stimulation / feedback parameters that are correlated with pain neural matrix activation estimates obtained from brain signals and empirical treatment effects.

[0153] In this regard, the neural matrix is a network of neurons that is believed to transmit pain-related signals that lead to the sensation of pain.

[0154] refer to Figure 1 , which shows a computer process or method 100 for identifying and extracting EEG signals related to pain. The method 100 is based on EEG pattern recognition of pain / no pain episodes. The computer method 100 generally includes:

[0155] 102: Receive EEG data, i.e., receive EEG data from one or more trials, recording one or more signals for each trial. The EEG data is labeled to indicate a pain state and / or a no-pain state. For example, a 32-channel EEG device used in the trial will output 32 EEG signals, one for each channel. Each EEG signal is associated with a corresponding coordinate vector. This vector can be the position of the EEG device recording the trial, or it can be an initial position vector that is selectively or arbitrarily set to enable calculation of an optimal position of the EEG sensor on the scalp.

[0156] 104: Determine the current density for each of the one or more signals in the corresponding coordinate vector - For each signal, calculate the current density in the corresponding coordinate vector assigned to the signal. Typically, the current density will be radial current density, but other potential field characteristics such as voltage potential distribution may be used.

[0157] 106: Estimate Current Density - Estimate current density for a set of neural activity regions of interest. The set of regions may be arbitrarily selected or may be predetermined to enable refinement of the regions of interest to those regions that are more likely to generate relevant signals for distinguishing between painful EEG states.

[0158] 108: Calculate at least one spectrum feature of the electric potential field feature, where the at least one spectrum feature is typically a frequency band power.

[0159] 110: For each region of neural activity of interest, calculate the mean and variance of the EEG data variation between EEG data labeled as indicating a pain state and EEG data labeled as indicating a no-pain state. This typically involves calculating a scatter matrix; generally, a between-class scatter matrix and a total scatter matrix are calculated. However, in some cases, only one of these matrices may be required, and thus only that one matrix is calculated. These matrices help understand the covariance between two or more pain-related EEG classes.

[0160] 112: Identify an EEG signal related to pain based on at least one of the regions of interest, wherein, in at least one of the regions of interest, a variance is lower than a predetermined threshold.

[0161] To obtain EEG patterns associated with pain episodes, design and conduct an experimental protocol with an estimated total duration of approximately 1.5 hours. The tasks are as follows:

[0162] i. Subjects or patients (referred to as "participants" in this study) completed a battery of self-report questionnaires [(i.e., demographics, Numeric Pain Rating Scale (NRS), healthcare use, Brief Pain Interference (BPI), Pain Catastrophic Scale (PCS), Tampa Kinesiophobia Scale, Short Form-36 (SF-36), and Patient Health Questionnaire-9 (PHQ-9)).

[0163] ii. Participants perform a series of 15 body movements.

[0164] iii. Participants watched a series of 15 videos involving individuals engaging in everyday activities.

[0165] A total of 11 healthy participants and 11 patients (a mix of patients with low back pain and patients with lower limb pain) participated in the experiment. Figure 2 The EEG arrangement 200 shown captures EEG. The EEG arrangement comprises:

[0166] 202: an EEG cap worn on the subject's scalp, wherein the EEG sensors or electrodes are located around the scalp;

[0167] 204: EEG amplifier;

[0168] 206: Backpack, which enables carrying the components of arrangement 200, in particular when performing the following reference Figure 4 the duration of the mission described;

[0169] 208: Electrocardiogram (ECG) electrodes for monitoring the subject's heart condition;

[0170] 210: One or more galvanic skin response (GSR) sensors for measuring skin conductivity and located on the forearm of a subject when in use;

[0171] 212: Buzzer. The buzzer enables self-registration or self-labeling of EEG data. The buzzer can be operated to simply label the data in a binary manner (e.g., pain / no pain) or to label the data in a progressive manner, such as clicking button A when the subject experiences a certain amount of pain, clicking button A again if the pain increases, or clicking button B if the pain decreases.

[0172] The overall system or arrangement 200 includes an EEG amplifier 204, secured in a backpack 206, coupled to a galvanic skin response (GSR) sensor module (scintillator) 210. The EEG headset or cap 202 captures electrical neurophysiological activity in the brain associated with cognitive / emotional states of perception, attention, and calmness.

[0173] This wearable BCI system was worn by the participant during the experiment, with the EEG cap 202 mounted on the participant's head and the GSR sensor 210 clipped onto the participant's forearm.

[0174] like Figure 3 As shown in schematic diagram 300 of arrangement 200 , EEG cap 202 includes 40 channels.

[0175] The EEG cap 202, ECG electrodes 208, and GSR electrodes are all fed to an amplifier 204. Since the readings from the GSR electrodes are relatively weak, the readings from the GSR electrodes are pre-amplified by a preamplifier 302 before being fed to the amplifier 204. The output of the amplifier 204 is fed to a computer 304 (e.g., a laptop), which also receives signals from the buzzer 212 (e.g., Figure 2 Thus, the computer 304 may label the EEG data, ECG data, and GSR data based on the button status or buzzer status (e.g., button A pressed / not pressed, and button B pressed / not pressed) when the EEG data, ECG data, and GSR data are received.

[0176] During the experiment, EEG signals, pain events, and pain intensity were recorded using the arrangement 200. For the first part of the Phase 1 recording, participants (healthy subjects and subjects believed to be patients, such as people experiencing chronic pain) were asked to perform 15 different movements to induce pain. These movements were:

[0177] 1. Lift the box from the floor

[0178] 2. Pick up and move cartons

[0179] 3. Pick up the dumbbells from the cart

[0180] 4. Pick up the box under the table

[0181] 5. Mobile Dumbbells

[0182] 6. Mopping the floor

[0183] 7. Carrying the suitcase upstairs

[0184] 8. Go downstairs

[0185] 9. Walking on a treadmill with a backpack in front of or behind your body

[0186] 10. Fast walking on the treadmill

[0187] 11. Walking on a treadmill with your backpack on your side

[0188] 12. Pulley for pulling stools

[0189] 13. Pull the pulley

[0190] 14.Put the boxes on the shelf

[0191] 15. Pushing a metal cart

[0192] Before starting each of the 15 different motor tasks, participants watched a video clip depicting the movements they needed to perform. This gave them a clearer understanding of the task they needed to perform and ensured a greater likelihood of performing the movements in a consistent manner.

[0193] Afterwards, participants remained still while a 30-second baseline EEG recording was performed. This resting period was used for neural matrix detection and analysis.

[0194] Each exercise was accompanied by a therapist. During the exercise task, the participant was advised to inform the therapist when they experienced an onset of pain during the pain-inducing exercise. The therapist was then able to mark the EEG segment (e.g., by clicking a button to electronically apply a time signature to the relevant EEG segment to mark the relevant EEG segment as a segment of the pain EEG category) and instruct the participant to stop exercising for 20 seconds before continuing. This 20-second rest period enables the system to capture relevant clean EEG segments that are not corrupted by motion artifacts generated when the participant performs the task. A description of the pain was also recorded, including: pain score, pain type, and area where the pain occurred.

[0195] By providing a therapist, delays that would otherwise occur if the subject had to free both hands and position the buzzer 212 to mark the data are significantly alleviated.

[0196] like Figure 4 As shown, the process 400 employed by the arrangements 200, 300 involves identifying a starting point for recording an EEG at 402. EEG, ECG, and GSR recordings then begin at 404, and an introductory video is played to the subject at 406, although this is not required. After the introductory video is played at 406, the subject remains still for 30 seconds at 408 so that the system can obtain a baseline EEG recording.

[0197] Then, an exercise video is played to the subject to show the exercise being performed by the subject at 410. This enables the participant to have a clearer understanding of the exercise task that needs to be performed and ensures a greater likelihood of performing the action in a consistent manner.

[0198] Then at 412, the subject, in a supervised environment, i.e., in the company of a therapist, performs the motion associated with the video at a predetermined number of times (between 1 and any suitable number of times), or performs the motion associated with the video at a predetermined time period. During the execution of the motion, if the pain experienced by the subject increases at decision box 414, the subject is asked to stop and rest for 20 seconds at 416, and the therapist will mark the EEG segment, for example, by clicking a button to electronically apply a time signature to the relevant EEG segment to mark the relevant EEG segment as a segment of the pain EEG category. A description of the pain is also recorded, including the pain score, the type of pain, and the area where the pain occurs. If, after resting at 416, the subject is unable to continue to perform the motion at decision box 418, the program ends at 420. Otherwise, at 412, the subject continues to perform the motion again. If, during the execution of the motion, the subject experiences lighter pain at decision box 422, steps 416 and 418 as described above are repeated.

[0199] If the subject's pain level remains consistent, the subject repeats or maintains the exercise until the desired number of repetitions and / or the predetermined duration is reached at 424. If the subject does not complete all 15 exercise tasks at decision block 426 (in other embodiments, this number may be different), the system proceeds to the next exercise video at 427. This process repeats, with the subject again providing a baseline recording at 408 and proceeding through steps 410 through 426 until all relevant exercises have been performed.

[0200] Figure 5a A table with a laptop and a rolling, backless office chair were shown, in which the subject sat and watched an introductory video before the motor task (in Part 1); this was followed by a series of 15 videos of individuals engaging in daily activities (in Part 2). Figures 5b to 5d The various sites where the movement is performed are shown in FIG. Figure 5b It is a treadmill on which the subjects can be asked to walk with the bag in different postures; Figure 5c Various fitness equipment are shown, and the range and intensity of movement on the fitness equipment can be viewed and tested differently; Figure 5d A series of pick-up and place activities are shown, with predefined foot positions on the floor to ensure that certain activities require extension of the arms and / or back.

[0201] After all exercises were performed, a further baseline recording was taken at 428. The video was played to the subject again at 430 and an EEG recording was taken. The series of videos, currently totaling 15, features individual actors or actresses performing tasks that mimic people engaging in everyday activities that can induce pain. These tasks include:

[0202] 1.Man lifting a heavy box of documents

[0203] 2. Men carrying heavy bags

[0204] 3. Man carrying a heavy bag of groceries

[0205] 4. Man lifts heavy boxes onto a cart

[0206] 5. Man carrying crate upstairs

[0207] 6. Man removes heavy documents

[0208] 7. A man pushes a cart of heavy boxes

[0209] 8. Man walking down stairs with sprained ankle

[0210] 9. Men flipping mattresses

[0211] 10. Woman lifting a heavy box of documents

[0212] 11. Women carrying heavy bags

[0213] 12.Woman lifting heavy boxes onto a trolley

[0214] 13. Women put the sheets in awkward corners

[0215] 14. Woman unloading washing machine

[0216] 15. Woman unloading groceries from shopping cart

[0217] Example still images taken from these videos are shown below. Figure 6 Similarly, the EEG recording during video playback can be used to determine the specific activities that led to the subject's anticipated pain experience. If the entire video has not been played at 432, a baseline EEG recording is performed again at 428, and the next video is played at 433. Otherwise, the method ends at 434.

[0218] When labeling data and monitoring patients in the presence of a therapist, the above process facilitates the construction of a well-labeled dataset of EEG segments that can be easily categorized into pain EEG and non-pain EEG categories. Furthermore, one or more EEG acquisition steps 408, 428 of method 100 involve acquiring EEG data from multiple trials, including one or more trials from subjects susceptible to pain (i.e., patients or subjects with chronic pain) and one or more trials from healthy subjects. For example, the step of storing the EEG data received in step 102 in a database or memory can be avoided.

[0219] A total of 11 healthy subjects (i.e., subjects without chronic pain) and 11 patients (a mix of subjects with low back pain and lower limb pain) participated in the experiment. EEG waveforms were screened (using visualization and manual labeling), and corrupted / noisy time periods were rejected. EEG analysis used a 10-second time window or time period starting from each event marker (immobilization, pain, less pain). The data finally collected and used for this analysis included a total of 19 participants from three categories, including: back pain (6 participants), lower limb pain (2 participants), and healthy controls (11 participants).

[0220] Compared with healthy controls, pain patients reported significantly higher pain scores (p = .001) and pain interference (p = .01). Pain patients also reported lower physical functioning (p = .02) and higher role limitations due to physical health status (p = .04). Among the results obtained through the above procedures, no significant differences were found between pain patients and healthy controls on measures of fear avoidance of physical activity and depressive symptoms.

[0221] For the second task, patients were asked to perform 15 body movements, and the EEG brain signals of healthy controls and pain patients were compared at rest. Pain patients showed higher activation levels in alpha, medium beta, and high beta frequencies. This was particularly seen in the frontal and prefrontal regions, particularly in the right hemisphere. Figure 7 Shown in.

[0222] For the third part, the subjects were asked to watch videos of actors or actresses performing everyday tasks that could induce pain. Each participant watched a series of 15 videos of individuals engaging in everyday activities. The results seemed to show that patients who reported pain (responders) had significant EEG pattern differences in the medium and high beta frequencies compared to those who did not (non-responders).

[0223] To extract local EEG signals, current source density (CSD) estimates were calculated using the spherical spline algorithm [KT06].

[0224]

[0225] Table 1: Comparison of category accuracy between pain EEG categories and no-pain EEG

[0226] In this regard, after data acquisition, a pain neural matrix modeling and decoding mechanism is implemented based on electric potential field features. In this embodiment, the feature is current density, or more specifically, radial current density, which represents the second-order derivative of the electric potential field on the scalp surface. Current density is used to implement a modeling and decoding mechanism referred to herein as the constrained discriminant current density algorithm (cDCD).

[0227] The radial current density measures the amount of current flowing radially at a specific coordinate. Consider the spherical coordinates E i , I∈1,...,N EEG montage stitching of N electrodes. The spherical spline surface Laplace algorithm for current density estimation (CSD) aims to obtain the true current density at any coordinate (I) as follows:

[0228]

[0229] Here, c must be estimated from each EEG sample z (scalp potential measurement) i An array that satisfies:

[0230] Gc+Tc0=z (2)

[0231] as well as

[0232] Tc=0 (3)

[0233] Where G is the matrix of the g function of the angle between the N electrodes, and T is the unit vector [1, 1..., 1].

[0234] Function cos represents point E and electrode E i The angular distance between . The function h is given by:

[0235]

[0236] m is a constant greater than 1, and p n (x) is a Legendre polynomial of order n.

[0237] The above-mentioned CSD can estimate the amount of current density at a given spherical coordinate. However, it does not answer which coordinate, i.e., which brain voxel (note that CSD is generally only sensitive to low-depth sources) may show neural activity associated with a specific mental task / condition.

[0238] The CSD-based modeling framework described below can be used to identify regions of neural activity that are involved in pain perception and cognitive processing.

[0239] For simplicity, assume that the {ci(t)} value has been calculated for each EEG time sample. A set of neural activity regions of interest can then be evaluated: L = {l1, I, ...} - the coordinate vectors in method 100 can be set to the corresponding regions in the neural activity regions of interest. The CSD value at the neural activity region of interest can then be calculated using the above equation to obtain {cj(t)}.

[0240] Here, we consider the oscillatory activity in these CSD estimates using M EEG samples (sampled at {t}) during the kth trial (instance) of the mental task / condition. The mean power of cj(t) over the trial is then calculated and expressed as d k (l j )express.

[0241] Now consider a classification / detection scenario where we need to discern contrasting neural activity between a mental task category ω0 and another category ω1. This may require understanding the covariance of the characteristics associated with the two tasks. The mean and variance of the CSD power values d(t) within each category can be determined, and their discriminative power can be expressed using the multivariate Fisher score:

[0242] f(L)=trace{S b (S t +γI) -1} (5)

[0243] Among them S b and S t are the between-class scatter matrix and the total scatter matrix, respectively, andγ is the positive regularization parameter.

[0244] Therefore, we model the optimal CSD coordinates as:

[0245] argmin L f(L) (6)

[0246] According to the above, for a given L, calculating f includes:

[0247] (Non-linear step) Calculate h for each electrode and each CSD coordinate in L i (L);

[0248] (Linear step) Use precomputed {c i} to calculate the CSDc for each CSD coordinate and each EEG time sample j estimated value of;

[0249] (Non-linear step) Calculate the mean power of the CSD in each trial: d k (l j );

[0250] (Non-linear step) Calculate the scatter matrix S b and S t ;

[0251] • (Non-linear step) Calculate Fisher score.

[0252] Therefore, in this optimization problem, the nonlinear step is dominant. To reduce the computational load and improve robustness, the solution for L is restricted to a plausible range or range of interest. This is achieved by setting upper and lower limits on the region to be optimized in spherical coordinates. Therefore, the modeling and optimization problem is rewritten as:

[0253] argmin L f(L), subject to ∑ z u(I i ,b)=0 (7)

[0254] Where b is the parameter that defines the region of approximate coordinates in spherical space, and u is a step function: if the position vector I i If the value is within the region, the function value is 0; otherwise, the function value is 1 or any non-zero value.

[0255] In a practical application of the cDCD method, method 100 involves investigating or monitoring the following rhythmic activity: Theta theta:

[48] ; Alpha alpha:

[812] ; Low beta beta:

[1216] ; Medium beta beta:

[1624] ; High beta beta: [24 32]; Delta delta:

[14] . In each trial, the power spectrum (dk) of the average CSD power value at discrete frequency bins (starting from a 256-point FFT with a sampling rate of 250 Hz) is calculated using Welch's method, and then the sum of the power from all frequency bins within a specific frequency band is used.

[0256] A total of 32 trials (cases) were conducted with pain-free and pain-free periods. Six-fold cross-validation classification tests were performed using either frequency band-power or CSD-frequency band-power features derived from scalp potentials. A linear support vector machine (SVM) was used as the classifier using the Matlab Statistics Toolbox.

[0257] In the current cDCD implementation, a simplified version is considered in which the L coordinates are the same as the coordinates of the electrodes. Therefore, the optimization is reduced to the portion of the spherical coordinate system within the corresponding electrode region. Although the result is not as ideal as solving the objective equation (7), restricting to the electrode coordinates enables the evaluation of solutions (although they may not be optimal) and allows comparison with the baseline.

[0258] Refer to Table 1 to compare the accuracy at each fold and the average accuracy. Compared with the average, the proposed method significantly improves the detection accuracy from 68% to 82%.

[0259] It is expected that as the optimization approaches the target optimization provided by equation (7), this suboptimal cDCD will yield increasingly accurate solutions for the EEG classes if used as an initial solution for the optimization process.

[0260] As mentioned above, the second category is the BCI-based pain neural matrix activity and attention monitoring and feedback training system.

[0261] and Figure 2 A similar arrangement as shown can be used for this process. Specifically, the overall BCI-based pain neuromodulation treatment method uses a shared attention and pain neural matrix activity monitoring and feedback training system 1100, such as Figure 11 As shown, the system 1100 includes a computer 1102 (which can be a desktop, laptop, or tablet) connected to an EEG amplifier 1104 and an EEG cap 1106. The EEG cap 1106 is meant to be worn by the participant.

[0262] The operator helps the participant put on the EEG cap 1106 and applies gel to the electrodes. The operator then launches the BCI client software application on the computer to check the impedance level of each electrode to ensure good connectivity. Scalp EEG signals can then be acquired. Additionally, a USB joystick allows the user to play an EEG attention / pain feedback game running on computer 1102.

[0263] Figure 12 A system diagram 1200 is shown of the system 1100. In this embodiment, a cap 1202 includes a 40 channel arrangement that feeds an amplifier 1204. The amplifier 1204 sends the amplified signal to a computer 1206, and the user manually interacts with the computer 1206 via a joystick 1208.

[0264] After the system is set up on the user, the method flow 1300 begins at 1302. Real-time EEG recording is then started at 1304 to correlate attention and pain or well-being scores with the EEG obtained during interaction with the computer 1206.

[0265] Initially, if the session is the first for this experiment at 1306, an introductory video is presented to the user at 1308. The subject then manually enters their pain and / or distress level at 1310 to establish a baseline from which to gradually develop a wellness score. The user then plays the game for a predetermined period (currently 3 minutes) at 1312 and again specifies their pain and / or distress level at 1314. If, after the gaming period ends, the subject has completed 8 rounds (or some predetermined number) of the game at decision block 1316, the method ends at 1318. If the subject has not completed 8 rounds, steps 1310 through 1316 are repeated. In some cases, steps 1310 and 1314 may not need to be performed between each round of the game, with steps 1310 and 1314 only being performed for one round.

[0266] For the second and subsequent sessions, the subject performs the same steps 1310 to 1316. However, at predetermined stages, the difficulty level of the game may be increased to cultivate the subject's cognitive pain management abilities. For example, the difficulty level may be increased every 3k+1 sessions: k∈N (a natural number), e.g., two difficulty levels are added (step 1322) in the 4th, 7th, 10th, 13th, and 16th sessions (step 1320).

[0267] Attentional strategies, which teach people to intentionally direct and maintain their attention away from pain, are existing psychological strategies often provided by clinicians as a way to alleviate pain and associated distress. Patients will attempt to redirect their attention by thinking about something else, or by changing its meaning, context, motivational relevance, or importance. Lay examples include statements such as "Try not to dwell on the pain" or "Think of something positive," which are perhaps the most common advice given to chronic pain patients. However, such strategies lack feedback to inform patients whether they are implementing the treatment correctly, and whether they should be doing more or less.

[0268] In this system, two modules are used, namely attention score and pain neural matrix (health) score, which are Figure 14 1400. These scores enable patients to implement personalized pain management strategies. In addition, especially during the game, patients receive visual feedback of their performance and therefore have a relatively objective measure of the success of their treatment.

[0269] Participants will be instructed to focus and maintain their attention to improve their attention scores. When the system detects that a participant may be in pain, the well-being score will decrease and the participant will be advised to practice pain management strategies taught by their clinician or therapist. These strategies may include cognitive behavioral therapy, breathing techniques, and other customized management strategies. This visual feedback mechanism provides a more direct and informative response to the subject. The subject will be informed whether such pain management strategies are effective, which will then enable the subject to mobilize or use these strategies more effectively in their daily activities.

[0270] This method is Figure 15 The invention provides an overview of a closed-loop sensing and neurofeedback / stimulation mechanism or system for pain neuromodulation training / therapy. To enable the method to be easily integrated into daily life, non-invasive EEG data is acquired through an EEG acquisition (ACQ) module and processed to identify attention scores and well-being or pain scores, as described below.

[0271] Various methods for detecting attention (i.e., concentration) have been proposed in the past. See, for example, U.S. Patent No. 8,862,581, entitled "Method and System for Detecting Concentration," the entire contents of which are incorporated herein by reference. The citation of such prior art should not be construed as common general knowledge.

[0272] A method for determining an attention score within an EEG time segment, for example, obtained from a time series of EEG time segments, includes detecting concentration by:

[0273] - Extracting temporal features from brain signals, i.e., one or more EEG time segments;

[0274] - Use a classifier to classify the extracted time features to give a score x1;

[0275] -Extract spectral spatial features from brain signals;

[0276] - selecting a spectrum-spatial feature containing distinguishing information between the focused state and the non-focused state from the extracted spectrum-spatial feature set;

[0277] - Use a classifier to classify the selected spectral spatial features to give a score x2;

[0278] - combining the scores x1 and x2 to give a single score; and

[0279] -Determine whether the subject is in a focused state based on the single score number.

[0280] A skilled artisan will be aware of methods other than those described above that are intended to fall within the current understanding of determining an attention score.

[0281] In addition to the attention score, a wellness score must also be determined, as discussed above. To accomplish this, one or more processors in the system implement a pain neural matrix modeling and decoding engine to determine the trainee's wellness score. This process is accomplished in two phases. Phase A involves neural matrix modeling through data acquisition and learning. Phase B involves decoding.

[0282] Data collection and learning phase A includes collecting non-pain EEG data and pain EEG data from the patient group and the healthy group, ie, receiving EEG data.

[0283] Automatic and / or manual EEG screening is then performed to select clean EEG episodes or time periods of at least N seconds long (as in our study, a preferred value of N=20) after each pain episode. Figure 4 Extracts from the trials discussed.

[0284] Then, an optimal set of scalp (cortical) locations is identified where the current density contains the most information related to the pain EEG compared to the non-pain EEG. This may involve implementing method 100 by setting a set of initial scalp locations, where the initial scalp locations are represented by L (or L 初始 ) represents the spherical coordinate vector corresponding to each position. A numerical optimization algorithm is then run to search for the optimal L value using this objective function:

[0285] argmin L f(L), obey

[0286] where f is a function used to evaluate the discriminant power of L, and the constraint term indicates that L must be located in the plausible region defined by b1 and b0.

[0287] The f function is given by the multivariate Fisher scoring:

[0288] f(L)=trace{S b (S t +γI) -1}

[0289] Among them S b and S t are the inter-class scatter matrix and the total scatter matrix, respectively, calculated as described below, and γ is an optional positive parameter used for regularization. These two matrices are calculated as follows:

[0290] Currently, the generated transformation matrices G and H are used for spherical spline interpolation of the surface potential (G) or current source density (H) at the location or region L of interest.

[0291] The EEG data is then converted into current density estimates using a spline interpolation matrix. Currently, the selected EEG waveform (potential) is converted into a current density estimate using matrices G and H to obtain the current density estimate as a matrix X of size nC × nT, where nC is the number of spatial points on the scalp when the estimate is made on the scalp, and nT is the number of time samples during each pain-free or pain-attack period.

[0292] Frequency band powers are then calculated for the current density estimates. Currently, a vector of frequency band powers is calculated for each current density time series x(t) using either a Fourier transform-based decomposition or an energy calculation after a bandpass filter array. Frequency bands can be set empirically (e.g., based on known brain signal frequencies) to maximize system performance, or using traditional EEG frequency bands including delta, theta, alpha, low beta, medium beta, and high beta.

[0293] The band powers of the current density estimates can then be used to calculate the inter-class scatter matrix and the total scatter matrix. For example, the total scatter matrix S can be calculated using all samples of the band power vectors t Similarly, the calculation of the mean frequency band power vectors for the pain EEG class and the no-pain EEG class enables the calculation of the between-class scatter matrix S b .

[0294] The optimal L is obtained. The optimal L is the position / region where the radial current density is considered to have the highest discrimination power between the signal of the pain EEG category and the signal of the no-pain EEG category.

[0295] The optimal L can then be used to calculate the current density time series x(t).

[0296] Then, build a binary classification machine ( Figure 1 The method further comprises the step 118 of classifying the subject to distinguish between at least two pain-related EEG states, the states being a pain EEG class (i.e., a state in which the EEG indicates that the subject is experiencing pain) and a no-pain EEG class (i.e., a state in which the EEG indicates that the subject is not experiencing pain). The binary classifier can be any suitable binary classifier, such as a support vector machine, a multilayer neural network, a generalized linear discriminant analyzer, and the like.

[0297] A classifier is selected and trained so that it can generate a score called the Pain Matrix Neural Activation Score (rPNAS). The rPNAS assigns positive values to samples in the pain EEG class and negative values to samples in the no-pain EEG class.

[0298] Therefore, through the process described above, the pain EEG model developed by learning from EEG data includes or is composed of: L, the coordinates of the current density estimation points; and matrices G and H, which are associated with the current density estimates (these matrices are discussed in "Spherical splines for scalp potential and current density mapping. Electroencephalography and clinical neurophysiology" published by Perrin, F., Pernier, J., Bertrand, O., Echallier, JF (1989), the entire content of which is incorporated by reference into this application), specified frequency bands and frequency band power calculation methods, and classification machine models.

[0299] Phase B involves decoding. Decoding relies on Figure 1 Step 120 receives further EEG data and Figure 1 The data is classified in step 122. In this embodiment, the decoding assumes that a continuous stream of EEG samples is received. The EEG sample sequence or EEG sample stream is converted into a new time series that indicates the raw pain neural matrix activation score (rPNAS) at each time point.

[0300] If the EEG montage stitching is the same as in phase A, the algorithm can reuse the same transformation matrices G and H for spherical spline interpolation of the surface potential (G) or current source density (H). Otherwise, the two matrices as described above must be recalculated using L and the new montage stitching.

[0301] In general, only clean EEG data should be processed. Therefore, for each new EEG sample, EEG sample rejection is performed. Afterwards, time-windowed EEG segments (e.g., EEG time segments) are extracted at fixed intervals (e.g., every N samples or t seconds), and if some EEG samples are corrupted, EEG segment rejection is performed. In this embodiment, any eye artifacts or other excessive artifacts will cause the entire segment to be rejected. Therefore, the data is clean.

[0302] The EEG samples are converted into current density estimates using matrices G and H to obtain the current density estimate as an nC × nT matrix X. As described above, nC is the number of spatial points on the scalp at which the estimate is made, and nT is the number of time samples during each pain-free or pain-attack period.

[0303] The vector of band powers for each current density time series x(t) is then calculated using the same method as in stage A. The band power vectors can then be converted into rPNAS using a classifier.

[0304] For pain treatment, the following process is used based on the classification scheme described above. Pain treatment can be performed through a game interface with audio-visual feedback and / or brain stimulation. Signal tracking and adaptive normalization will pre-process the acquired EEG signals.

[0305] Adaptive feedback and / or simulation modeling and optimization are performed to calculate the scores and parameters of the brain stimulation and audiovisual feedback game. Assume that a set of neural stimulation parameters is described by a vector V, and the goal of the adaptive stimulation mechanism is to manage rPNAS within a threshold (determined empirically). For closed-loop studies, rPNAS (rather than heart rate) is used as the target metric to be optimized (for argminf(L), to be minimized), and the response to a specific stimulation configuration is a dynamic process.

[0306] This study can be applied to a single person / patient / subject or to a group of people (including the general population).This is again done in two stages, whereby the EEG acquisition and neurostimulation equipment is attached to the person.

[0307] These phases are the training phase, Phase A, which involves the stimulation parameter vector V j Generate multiple random values. For each parameter vector, generate a stimulation pulse and collect EEG samples (i.e., responses) relative to the stimulation. Then, the above decoding method can be used to calculate the rPNAS time series.

[0308] When testing all random V j Afterwards (and optionally after several iterations for optimization), all rPNAS time series data are associated with the stimulation parameter vector V j Collected together.

[0309] Machine learning is used to construct a recursive regression model that relates the stimulus parameter vector V to the rPNAS response (either a time series or a specific time point). The specific regression model mechanism can be selected empirically to achieve the best prediction of the rPNAS response. The regression model can be, for example, a recursive neural network (such as long short-term memory), a linear transfer function, or any other suitable model.

[0310] The second phase, Phase B, is the adaptive phase. In this phase, the current rPNAS is measured using the decoding method described above. A baseline rPNAS is established, for example, by the operator, and a target rPNAS range is set by the consulting physician. This ensures that pain management is within a predetermined range, thereby enabling the subject's cognitive ability to manage pain to develop smoothly.

[0311] The algorithm selects the optimal parameter vector V according to the recursive regression model j If brain stimulation is used, the optimization performed by the algorithm may take into account rPNAS control as well as stimulation electrical power.

[0312] The resulting game feedback and / or stimulation is then delivered and rPNAS measurements are tracked. The measured or observed rPNAS measurements are used to update the recursive regression model.

[0313] The brain stimulation engine is used to determine the frequency, duration and pattern of non-invasive stimulation so as to maximize the efficacy of stimulation for the treatment of chronic pain. Brain stimulation can be transcranial direct current stimulation (tDCS), vagus nerve stimulation (VNS) or repetitive transcranial magnetic stimulation (rTMS). The audio-visual feedback game provides a game strategy to the trainee or subject to manage their pain. In the example shown, the interactive activity is displayed on a display in an interactive computer game environment. EEG control tasks are assigned to the trainee or subject to achieve a goal (such as shooting down as many targets as possible in a space hunter game) using the EEG activity controlled by a specific game.

[0314] Figure 11 The system shown in Figure 16 In use. Figure 17 and Figure 18 As shown, the display of computer 1102 displays a series of questions for assessing the subject's pain level. Figure 19 A user interface is shown on which attention and health or pain level assessment scores are displayed, as well as user interaction activities with the user interface through an input device such as a joystick. Figure 20Further features of the display are shown, including: a sound adjustment menu, the remaining time until session expiration or game end, missile attack (or other parameters required by the game), the current score of the subject playing the game, and a top-down view of the environment in which the subject is playing the game. Figures 21 to 24 Exemplary controls enabled by user attention and pain level assessments or wellness scores are provided.

[0315] It will be understood that many further modifications and permutations of various aspects of the described embodiments are possible. Accordingly, the described aspects are intended to embrace all such changes, modifications and variations that fall within the spirit and scope of the appended claims.

[0316] In this specification and the claims that follow, unless the context requires otherwise, the word "comprise" and variations such as "comprises" and "comprising" will be understood to mean the inclusion of a stated integer or step or group of integers or steps but not the exclusion of any other integer or step or group of integers or steps.

[0317] Reference in this specification to any previous publication (or information obtained therefrom) or to any known matter is not, and should not be taken as, an acknowledgment or admission or any form of indication that the previous publication (or information obtained therefrom) or known matter forms part of the common general knowledge in the field of business to which this specification relates.

Claims

1. A computer method for identifying and extracting pain-related electroencephalogram (EEG) signals, the computer method comprising: receiving EEG data for each trial from one or more trials, the EEG data comprising one or more signals, each of the one or more signals being associated with a corresponding coordinate vector, the EEG data being labeled to indicate a pain state and / or a no-pain state; determining a current density for each of the one or more signals in the corresponding coordinate vector; estimating the current density of a set of neural activity regions of interest based on the calculated current density; calculating at least one spectral feature for each trial based on the estimated current density; calculating, for each region of neural activity of interest, a mean and a variance of EEG data variation between EEG data labeled as indicating a pain state and EEG data labeled as indicating a no-pain state based on the at least one spectral feature; as well as Pain-related EEG signals are identified based on at least one of the neural activity regions of interest, wherein the variance is below a predetermined threshold in the at least one neural activity region of interest.

2. The computer method according to claim 1, wherein: The current density is a radial current density.

3. The computer method of claim 1 , wherein: Calculating the mean and variance of the EEG data variation includes calculating at least one of an inter-class scatter matrix and a total scatter matrix based on the spectral features to determine the covariance between at least two pain-related EEG classes.

4. The computer method of claim 1 , wherein: Calculating the mean and variance of the EEG data variation includes calculating at least one of a between-class scatter matrix and a total scatter matrix.

5. The computer method of claim 4, wherein: Calculating both the between-class scatter matrix and the total scatter matrix, and wherein identifying at least one region of neural activity of interest comprises satisfying the following equation: argmin L f(L), where f(L) is the multivariate Fisher score of the neural activity region of interest L, where f(L) is: f(L)=trace{S b (S t +y) -1 }, Among them, S b is the inter-class scatter matrix, S t is the total scatter matrix, γ is the forward regularization parameter, and ι is the identity matrix.

6. The computer method of claim 1 , wherein: The neural activity region of interest is a cortical location of the subject.

7. The computer method of claim 1 , wherein: The neural activity region of interest is defined by EEG coordinates on the subject's scalp.

8. The computer method of claim 3, wherein: The at least two pain-related EEG categories are a pain EEG category and a no-pain EEG category.

9. The computer method of claim 1 , further comprising: constructing a binary classifier that receives current density or current density activity from further EEG data and outputs a scalar indicator; receiving the further EEG data; applying the binary classifier to the further EEG data; and The further EEG data is classified as EEG data indicative of a pain state or EEG data indicative of a no-pain state based on a scalar indicator associated with the further EEG data.

10. A computer method for classifying EEG signals associated with pain, the computer method comprising: EEG data were collected from multiple trials, including: at least one trial performed by a first subject susceptible to pain perception; and At least one test performed with a healthy second subject, Wherein, the EEG data of each trial includes one or more signals; executing the computer method according to claim 1 on the EEG data and setting an initial coordinate vector for each signal; and constructing a binary classifier to distinguish between at least two pain-related EEG states based on EEG measurements of the at least one neural activity region of interest; receiving further EEG data; and applying the binary classifier to the further EEG data; A scalar indicator is received from the binary classifier and the further EEG data is classified into a pain-indicating EEG class or a no-pain EEG class based on the scalar indicator.

11. The computer method of claim 10, wherein: The binary classification machine is at least one of the following: Support vector machines; Multi-layer neural networks; and Generalized Linear Discriminant Analyzer.

12. The computer method of claim 10, wherein: Calculating the mean and the variance includes calculating at least one of a between-class scatter matrix and a total scatter matrix based on the at least one spectral feature to determine a covariance between at least two pain-related EEG classes.

13. The computer method of claim 12, wherein: Calculating the between-class scatter matrix and the total scatter matrix involves: generating a spline interpolation matrix for at least two electrical characteristics of the EEG data; converting the EEG data into current density estimates using the spline interpolation matrix; calculating a frequency band power for the current density estimate; and The between-class scatter matrix and the total scatter matrix are calculated using the band powers of the current density estimates.

14. The computer method of claim 13, wherein: Calculating the band power of the current density estimate includes: determining a current density time series of said current density estimate; A vector of frequency band powers of the current density time series is calculated using at least one of Fourier-based decomposition and bandpass filtering followed by energy calculation.

15. The computer method of claim 13, wherein: Classifying further EEG samples using the binary classifier includes: converting the further EEG data into further current density estimates using the spline interpolation matrix; calculating a frequency band power for the further current density estimate; and The binary classifier is applied to the frequency band power, wherein a positive output of the binary classifier indicates a first state of a pain EEG state and a no-pain EEG state of a source of the further EEG data, and wherein a negative output of the binary classifier indicates a second state of a pain EEG state and a no-pain EEG state of the source, the second state being different from the first state.

16. The computer method of claim 10, further comprising: receiving an initial input comprising a numerical pain level estimate; And wherein the receiving further EEG data includes: receiving consecutive EEG time segments from the continuously recorded EEG data; the computer method further includes: increasing or decreasing the numerical pain level estimate based on whether each consecutive EEG time segment is classified as indicating a pain EEG category or indicating a no pain EEG category.

17. A system for identifying and extracting pain-related EEG signals, the system comprising: EEG signal source; Memory; as well as at least one processor, the memory storing instructions that, when executed by the at least one processor, cause the at least one processor to: receiving EEG data for each trial from one or more trials, the EEG data comprising one or more signals, each of the one or more signals being associated with a corresponding coordinate vector, the EEG data being labeled to indicate a pain state and / or a no-pain state; determining a current density for each of the one or more signals in the corresponding coordinate vector; estimating the current density of a set of neural activity regions of interest based on the calculated current density; calculating at least one spectral feature for each trial based on the estimated current density; calculating, for each region of neural activity of interest, a mean and a variance of EEG data variation between EEG data labeled as indicating a pain state and EEG data labeled as indicating a no-pain state based on the at least one spectral feature; as well as Pain-related EEG signals are identified based on at least one of the neural activity regions of interest, wherein the variance is below a predetermined threshold in the at least one neural activity region of interest.

18. The system according to claim 17, wherein: The EEG signal source is arranged to collect EEG data from a plurality of trials, the plurality of trials comprising: at least one trial performed by a first subject susceptible to pain perception; and At least one test performed with a healthy second subject, Wherein, the EEG data of each trial includes one or more signals; Wherein, the at least one processor is configured to: applying an initial coordinate vector to each of said signals; constructing a binary classifier to distinguish between at least two pain-related EEG states based on EEG measurements from the at least one region of neural activity of interest for extracting pain-related EEG signals; receiving further EEG data; and The further EEG data is classified using the binary classifier to output a scalar indicator, and a determination is made based on the scalar indicator whether the further EEG data represents a pain EEG class or a no-pain EEG class.

19. The system of claim 18, further comprising an input device for receiving an initial input comprising a numerical pain level estimate, and wherein The at least one processor is configured to: receiving further EEG data by receiving consecutive EEG time segments from the continuously recorded EEG data; and The numerical pain level estimate is increased or decreased based on whether each consecutive EEG time segment is classified as representing a pain EEG category or a no-pain EEG category.

20. The system of claim 19, further comprising: monitor; Input devices; as well as one or more EEG devices for each of the at least one region of neural activity of interest, the one or more EEG devices being configured to record the continuously recorded EEG data, receiving the continuous EEG time segments from the continuously recorded EEG data, Wherein, the at least one processor is further configured to: determining an attention score for a first EEG time segment from the continuous EEG time segments; Displayed on said display: the numerical pain level estimate; the attention score; and interactive activities; During interaction of the interactive activity via the input device, continuously updating the numerical pain level estimate and the attention score based on subsequent EEG time segments from the continuous EEG time segments; and The behavior of the interactive activity is adjusted to reduce the number of subsequent EEG time segments that are classified into the pain EEG category.

Images