Brain activity detection device and brain activity detection method

Abstract

To provide a brain activity detection apparatus and a brain activity detection method which can effectively detect a seizure due to a brain activity of a subject in a daily life.SOLUTION: A brain activity detection apparatus detects a brain wave signal of a single channel representing a frontal lobe brain wave via a pair of electrodes arranged in a forehead of a subject, converts the signal into data of the segment image represented by a two-dimensional coordinate system with the time and amplitude as a coordinate axis, and then estimates a corresponding seizure symptom by referring to a seizure symptom estimation model constructed by deep learning of brain wave feature data in each segment image.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to a brain activ Motion check ity detection apparatus and a brain activity detection method, and particularly aims to provide a brain activity detection apparatus and a brain activity detection method for detecting epileptic seizures in daily life.

Background Art

[0002] Epilepsy is a neurological disease that causes sudden seizures due to abnormal excitation of nerve cells in the brain. According to the World Health Organization (WHO), it is estimated that there are approximately 50 million people with epilepsy worldwide. Seizure symptoms vary widely, including tonic-clonic seizures, absence seizures, myoclonic seizures, and atonic seizures.

[0003] With appropriate diagnosis and treatment, many patients can live a life free of seizures. However, approximately 30% of patients are unable to control their seizures with medication, which currently impairs their social life. In addition, after a seizure occurs, there is a risk of collision with objects or drowning while bathing, so caregivers need to ensure the patient's safety.

[0004] However, since seizures can occur anytime and anywhere, caregivers must constantly accompany the patient, which may restrict their daily life. Therefore, seizure detection systems are needed to improve the quality of life of patients and caregivers.

[0005] As such a seizure detection system, a seizure detection method using multi-channel EEG (electroencephalogram) has been conventionally proposed (see Non-Patent Document 1). In addition, an analysis method for EEG recording based on multi-channel assumption (see Patent Document 1) and a method for identifying pathological brain activity patterns of a subject using nonlinear classification (see Patent Document 2) have also been proposed.

Prior Art Documents

Patent Documents

[0006] [Patent Document 1] Japanese Patent Publication No. 2017-80483 [Patent Document 2] Special Publication No. 2020-512860 [Non-patent literature]

[0007] [Non-Patent Document 1] Zabihi M, Kiranyaz S, Ince T, Gabbouj M (2013) Patient-specific epileptic seizure detection in long-term eeg recording in paediatric patients within tractable seizures. [Overview of the Initiative] [Problems that the invention aims to solve]

[0008] The seizure detection method using multichannel EEG described in Non-Patent Document 1 can detect seizures with higher accuracy than single-channel EEG by employing a method to extract statistical features and frequency domain features.

[0009] However, the electroencephalogram (EEG) measuring device itself is relatively large and heavy, and electrodes must be placed all over the head, which limits the locations where EEG measurements can be taken, usually within hospitals.

[0010] Similarly, the electroencephalogram measurement methods using multichannel EEG described in Patent Documents 2 and 3 also had the problem that it was extremely difficult to continuously measure the electroencephalogram signals of the subject in their daily life.

[0011] This invention was made in consideration of the above points, and aims to propose a brain activity detection device and a brain activity detection method that can effectively detect seizures caused by the brain activity of a subject in their daily life. [Means for solving the problem]

[0012] To solve these problems, the present invention includes: a signal detection unit that detects a single-channel electroencephalogram (EEG) signal representing a frontal lobe EEG via a pair of electrodes placed on the frontal lobe of a subject; an image conversion unit that extracts partial time series from the EEG signal detected by the signal detection unit by sequentially dividing it into predetermined time widths so as to partially overlap in the time direction, and converts each of these partial time series into a segment image represented in a two-dimensional coordinate system with time and amplitude as coordinate axes; an EEG feature extraction unit that sequentially analyzes the segment images converted by the image conversion unit and extracts EEG feature data from each of the segment images; and a seizure symptom estimation unit that estimates the corresponding seizure symptom from the EEG feature data sequentially extracted by the EEG feature extraction unit, while referring to a seizure symptom estimation model constructed by deep learning using EEG feature patterns classified according to seizure symptoms caused by abnormal human EEG as training data.

[0013] As a result, the brain activity detection device can effectively detect seizures caused by brain activity by placing electrodes only in the frontal lobe of the subject, so as not to interfere with daily life, and using a single-channel frontal lobe electroencephalogram.

[0014] Furthermore, in this invention, the image conversion unit extracts partial time series from the electroencephalogram signal detected by the signal detection unit by sequentially dividing it into sliding windows of predetermined time intervals with a predetermined overlap ratio using the sliding window method, and converts each of these partial time series into a segment image by setting it in a two-dimensional coordinate system in which the amplitude is increased by a predetermined multiple.

[0015] As a result, the brain activity detection device, when converting EEG signals into segment images corresponding to partial time series, clearly distinguishes between normal and abnormal EEGs, and processes the data to make it easier to identify EEG patterns from the images. This significantly improves the accuracy of extracting EEG feature data from each analyzed segment image in the EEG feature extraction unit. In particular, EEGs are superimposed with noise from head movements, facial muscle tension, blinking, etc., and evaluating them within a single sliding window is insufficient for classifying normal and abnormal EEGs; this insufficiency can be resolved.

[0016] Furthermore, in the present invention, when the electroencephalogram feature extraction unit extracts electroencephalogram feature data from each segment image, it uses a predetermined seizure detection algorithm to sequentially determine the timing of changes between the seizure period, which represents the period including the seizure and predetermined time before and after the seizure, and the interseizure period, which represents the period from the end of the seizure period to the next seizure period.

[0017] As a result, the brain activity detection device can distinguish between the interictal period, which accounts for most of the subject's daily life, and the seizure period, which includes the actual seizure and the predetermined time before and after the seizure.

[0018] Furthermore, in this invention, the seizure symptom estimation unit, for each segment image extracted by the electroencephalogram feature extraction unit, undersamples the electroencephalogram feature data corresponding to the interictal period to match the number of electroencephalogram feature data corresponding to the seizure period, and then creates n sub-datasets, and each sub dataset Using a corresponding classifier, the system estimates the relevant seizure symptoms based on the final output determined by a majority vote of the outputs of all the classifiers.

[0019] As a result, in the brain activity detection device, since the ictal period is very short compared to the interictal period which accounts for most of the subject's daily life, it becomes possible to balance the proportion of data obtained in the interictal period and the ictal period. This can prevent overfitting in learning using a classifier, and also prevent a decrease in classification accuracy that is the result of such learning.

[0020] Furthermore, in the present invention, the seizure symptom estimation unit applies a Residual Network as a classifier, and applies the Backpropagation method as a method for updating teacher data of the seizure symptom estimation model.

[0021] As a result, in the brain activity detection device, by separating input feature quantities and residual feature quantities in advance using a residual network, even when the residual feature quantities cannot be calculated correctly, the feature quantities supplied from an upstream layer can be transmitted to a downstream layer as they are. This makes it possible to avoid phenomena that impede the progress of learning, such as gradient vanishing that occurs in each layer upstream of the intermediate layer. As a result, degradation of feature quantities caused by repeated operations accompanying deepening of the network can be suppressed.

[0022] Furthermore, in the present invention, the method comprises: a first step of detecting a single-channel electroencephalogram signal representing frontal lobe brain waves via a set of electrodes arranged on the subject's frontal head; a second step of extracting partial time series from the electroencephalogram signal detected in the first step while sequentially dividing the signal into predetermined time widths so as to partially overlap in the time direction, and converting each partial time series into data into a segment image represented by a two-dimensional coordinate system with time and amplitude as coordinate axes; a third step of sequentially analyzing the segment images converted in the second step, and extracting electroencephalogram feature data in each segment image respectively; and a fourth step of estimating a corresponding seizure symptom from the electroencephalogram feature data sequentially extracted in the third step, while referring to a seizure symptom estimation model constructed by deep learning using electroencephalogram feature patterns classified for each seizure symptom caused by abnormal human brain waves as teacher data.

[0023] As a result, in the brain activity detection method, electrodes are arranged only on the frontal region of a subject so as not to interfere with daily life, and ictal events caused by brain activity can be effectively detected using single-channel frontal lobe electroencephalogram. Effects of the Invention

[0024] According to the present invention, a brain activity detection apparatus and a brain activity detection method capable of detecting ictal events caused by brain activity even in daily life can be implemented. Brief Description of the Drawings

[0025] [Figure 1] It is a conceptual diagram provided for explaining the brain activity detection system according to the present embodiment. [Figure 2] It is a block diagram showing the internal configuration of the brain activity detection apparatus according to the present embodiment. [Figure 3] It is a schematic diagram showing the arrangement state of single-channel electrode positions (Fp1-F7, Fp2-F8). [Figure 4] It is a schematic diagram provided for explaining an ictal period and an interictal period. [Figure 5] It is a chart showing types of abnormal electroencephalograms that appear during epileptic seizures. [Figure 6] It is a schematic diagram provided for explaining an image conversion method using the sliding window method. [Figure 7] It is a conceptual diagram provided for explaining a seizure detection algorithm. [Figure 8] It is a conceptual diagram provided for explaining the architecture of a seizure symptom estimation model. [Figure 9] It is a chart showing information of epileptic patients serving as subjects. [Figure 10] It is a chart showing electroencephalogram feature data corresponding to the ictal period with ID number chb01. [Figure 11] It is a chart showing electroencephalogram feature data corresponding to the ictal period with ID number chb05. [Figure 12] It is a chart showing experimental results obtained using the Fp1-F7 channel (Fig. 3). [Figure 13] This is a diagram showing the experimental results using the Fp2-F8 channel (Figure 3). [Figure 14] This chart shows the results of selecting the optimal channel in terms of sensitivity and specificity. [Figure 15] This is a diagram showing an example of the interpretation results of patient electroencephalogram (EEG) characteristic data. [Modes for carrying out the invention]

[0026] An embodiment of the present invention will be described in detail below with reference to the drawings.

[0027] (1) Brain activity detection system according to this embodiment Figure 1 shows the configuration of the brain activity detection system 1 according to the present invention. The brain activity detection system 1 includes a brain activity detection device 2 that detects frontal lobe electroencephalograms via a pair of electrodes 2A placed in the frontal lobe of the subject, a terminal device 3 that receives estimated seizure symptom data obtained as a result of the brain activity detection device 2, and an external cloud server 5 that transmits data from the terminal device 3 via a network 4.

[0028] A brain activity detection device 2 is attached to the subject's frontal lobe via electrodes 2A, and estimated seizure symptom data detected by the brain activity detection device 2 is received by a terminal device 3, which then manages the data centrally.

[0029] Terminal device 3 consists of, for example, an information wireless terminal such as a smartphone, and is configured to enable wireless connection with brain activity detection device 2 via wireless communication such as Wi-Fi, and to connect to an external cloud server 5 via network 4.

[0030] On the cloud server 5, after user authentication of the subject and device authentication of the corresponding brain activity detection device 2, estimated seizure symptom data obtained from the subject is collected in a database (not shown). Then, the data analysis platform 6 connected to the cloud server 5 analyzes the estimated data collected for each subject read from the database using a predetermined analysis method, and then manages the analysis results.

[0031] In practice, the data analysis platform 6 performs statistical processing and big data analysis on estimated data obtained from subjects to perform inference and learning aimed at improving quality of life (especially brain activity), and then stores and manages this data in the database of the cloud server 5.

[0032] In this way, with the brain activity detection system 1, the subject can use the terminal device 3 to receive feedback information from the cloud server 5 as needed, which includes analysis of estimated seizure symptom data.

[0033] (2) Brain activity detection device according to this embodiment Figure 2 shows the configuration of the brain activity detection device 2 according to the present invention. The brain activity detection device 2 consists of a pair of electrodes 2A placed in the frontal lobe of the subject, a signal detection unit 10, an image conversion unit 11, an electroencephalogram feature extraction unit 12, a seizure symptom estimation unit 13, and a transmission unit 14 that transmits estimated seizure symptom data to a terminal device 3 (Figure 1) via wireless communication.

[0034] The signal detection unit 10 detects a single-channel electroencephalogram (EEG) signal representing the frontal lobe EEG via a pair of electrodes 2A placed in the frontal lobe of the subject.

[0035] In this invention, we focused on the electroencephalogram (EEG) of the subject's frontal lobe. The prefrontal association cortex is a brain region that receives information from other association cortices and plans actions based on that input. It is also interconnected with the motor cortex, basal ganglia, and limbic system by fibers, and receives information related to movement, memory, arousal, and autonomic nervous system regulation.

[0036] Therefore, when seizure symptoms such as involuntary twitching or loss of consciousness occur, it is thought that abnormal nerve excitation propagates to the prefrontal cortex. Furthermore, because the prefrontal cortex is interconnected with other areas of the brain, it is thought that if a localized seizure occurs in another area, abnormal nerve excitation may propagate to the prefrontal cortex.

[0037] Furthermore, the subject's forehead has exposed skin, making it easier to place electrodes compared to other measurement points. From these perspectives, frontal lobe electroencephalography (EEG) is considered the most suitable method for detecting epileptic seizures in daily life.

[0038] In this embodiment, the CHB-MIT Scalp EEG database was used as the electroencephalogram (EEG) database included in the signal detection unit 10. This EEG database was compiled using the 10-20 method (a method of placing electrodes on the head during EEG measurement) recommended by the International Society of Electroencephalography, and EEG data was collected from pediatric patients with intractable seizures.

[0039] This electroencephalogram (EEG) database is recorded with a sampling frequency of 256 Hz and a resolution of 16 bits. The EEG data is saved as an EDF (European Data Format) file after power line noise has been removed. Figure 3 shows the electrode placement diagram according to the International 10-20 method. In the electrode placement diagram shown in Figure 3, the EEG signals from single-channel electrode positions (Fp1-F7, Fp2-F8), which are bipolar derivations in the frontal lobe, are used as channel positions.

[0040] Epilepsy has two periods: the "seizure period" during a seizure and the "interseizure period" between seizures. The seizure period refers to the time from the onset of the first symptoms until the seizure ends. During the seizure period, abnormal electroencephalograms (EEGs) characteristic of epilepsy appear in part or all of the brain. The EEG database contains annotation files on the start and end times of seizures identified by clinical researchers, and this information was used to identify the seizure period.

[0041] The interictal period is defined as the time between the end of one seizure and the onset of the next. This period accounts for 99% of the patient's daily life. Since some patients experience physical, sensory, or emotional changes before a seizure, the period immediately preceding a seizure must be excluded from the interictal period. Also, after a seizure, consciousness is often clouded, and it takes time for normal brain activity to return as consciousness gradually returns.

[0042] For these reasons, as shown in Figure 4, the interictal period is defined as a period of 30 minutes or more from the start and end of a seizure. Therefore, in this invention, the "seizure period" is defined as the period including the seizure and the predetermined time (30 minutes) before and after the seizure, and the "interictal period" is defined as the period from the end of a seizure period to the next seizure period.

[0043] The image conversion unit 11 extracts partial time series from the electroencephalogram signal detected by the signal detection unit 10, sequentially dividing it into predetermined time widths so as to partially overlap in the time direction, and converts each of these partial time series into a segment image represented in a two-dimensional coordinate system with time and amplitude as coordinate axes.

[0044] In this invention, we focused on the shape of abnormal electroencephalograms (EEGs) that appear during seizures. Nerve cells become excessively excited due to repeated paroxysmal depolarization at the epileptic focus, resulting in epileptic discharges such as spikes and sharp waves. Slow waves are also generated during hyperpolarization. Since the abnormal EEGs of epilepsy are composed of a combination of these spikes and slow waves, they can be clearly distinguished from the background activity of normal EEGs.

[0045] Figure 5 shows a table illustrating the types of abnormal electroencephalogram (EEG) patterns that appear during epileptic seizures. During generalized tonic-clonic seizures, a combination of spikes and slow waves (sp-w) appears, while during absence seizures, a 3 Hz sp-w appears. In juvenile myoclonic epilepsy, polyspike EEG patterns are frequently observed. Delta bursts and theta bursts are high-amplitude, rhythmic slow waves that occur when brain function is impaired, such as during loss of consciousness. Because each type of seizure produces characteristic abnormal EEG patterns, we implemented an image recognition algorithm to classify the shapes of normal EEGs from those of epileptic abnormal EEGs.

[0046] Specifically, as shown in Figure 6, the image conversion unit 11 extracts partial time series from the electroencephalogram signal detected by the signal detection unit 10 by sequentially dividing it into 2-second intervals with a 50% overlap using the sliding window method, and converts each of these partial time series into a segment image by setting it in a 2D coordinate system in which the amplitude is increased by 6 to 60 times.

[0047] Normal scalp electroencephalogram (EEG) amplitudes are approximately 10-100 μV, while abnormal EEGs often have amplitudes more than twice that of normal background activity. Therefore, to ensure sufficient area for displaying the EEG, segmented EEGs were acquired within a range of ±600 μV. During acquisition, the background color was set to black and the EEG image to white. The EEG images were resized to 256 dpi x 256 dpi, and the central portion was cropped to 244 dpi x 244 dpi. Furthermore, since a pre-trained model from the machine learning library (Pytorch torchvision library) was used, the mean and standard deviation of the RGB levels of the ImageNet training data were used for normalizing the input image.

[0048] As a result, the brain activity detection device 2, when converting electroencephalogram (EEG) signals into segment images corresponding to a partial time series, clearly distinguishes between normal and abnormal EEGs and processes the data to make it easier to identify EEG patterns from the images. This significantly improves the accuracy of extracting EEG feature data from each analyzed segment image in the EEG feature extraction unit 12, which will be described later.

[0049] In particular, brainwaves are superimposed with noise from head movements, facial muscle tension, blinking, etc., making it insufficient to classify normal and abnormal brainwaves using only a single sliding window. This method can overcome that inadequacy.

[0050] The electroencephalogram (EEG) feature extraction unit 12 (Figure 2) sequentially analyzes the segment images converted by the image conversion unit 11 and extracts EEG feature data from each segment image.

[0051] Specifically, when the electroencephalogram feature extraction unit 12 extracts electroencephalogram feature data from each segment image, it uses a predetermined seizure detection algorithm to sequentially determine the timing of changes between the seizure period, which represents the period including the seizure and predetermined time before and after the seizure, and the interseizure period, which represents the period from the end of the seizure period to the next seizure period.

[0052] Because electroencephalograms (EEGs) are superimposed with noise from head movements, facial muscle tension, blinking, etc., evaluating them within a single window (the sliding window shown in Figure 6) is insufficient for classifying normal and abnormal EEGs. Therefore, a seizure detection algorithm was adopted to broaden the judgment range.

[0053] As shown in Figure 7, the seizure detection algorithm defines the classes estimated by the ensemble model as "Class 0 (normal)" and "Class 1 (abnormal)," and consists of two phases (Phase I and Phase II). The role of Phase I is to determine the timing of the transition from the interictal period to the seizure period, and the role of Phase II is to determine the timing of the transition from the seizure period to the interictal period.

[0054] If Class 1 is counted five times in a row, our algorithm determines that the EEG is abnormal and proceeds to Phase II. Otherwise, it is considered normal. On the other hand, if Class 0 is counted N times in a row, it is considered normal and proceeds to Phase I. Otherwise, it is considered abnormal. The N count mentioned above is the number of times Class 1 is counted after proceeding to Phase II.

[0055] As a result, the brain activity detection device 2 can distinguish between the interictal period, which accounts for most of the subject's daily life, and the seizure period, which includes the actual seizure and the predetermined time before and after the seizure.

[0056] The seizure symptom estimation unit 13 (Figure 2) uses brainwave feature patterns classified according to seizure symptoms caused by abnormal human brainwaves as training data, and while referring to a seizure symptom estimation model constructed by deep learning, it estimates the corresponding seizure symptoms from brainwave feature data sequentially extracted by the brainwave feature extraction unit 12.

[0057] Figure 8 shows the architecture of the seizure symptom estimation model. Since the interictal period accounts for most of a patient's daily life, the amount of seizure-induced electroencephalogram (EEG) data that can be collected from patients is very small, and the dataset has an unbalanced distribution of data. If a dataset with an unbalanced distribution of data is used to train a classifier, the classification accuracy will be low, so it is necessary to balance the distribution of data.

[0058] The seizure symptom estimation unit 13 undersamples the electroencephalogram (EEG) feature data in each segment image extracted by the EEG feature extraction unit 12, matching the number of EEG feature data corresponding to the seizure period with the EEG feature data corresponding to the interictal period.

[0059] Next, the seizure symptom estimation unit 13 creates n subdatasets, and each sub dataset Using a corresponding classifier, the corresponding seizure symptoms are estimated based on the final output determined by a majority vote of the outputs of all the classifiers. Since the amount of data in each sub-dataset was small, the batch size was set to "8".

[0060] As a result, the brain activity detection device 2 can balance the proportion of data obtained during the interictal period and the seizure period, as the seizure period is very short compared to the interictal period, which accounts for the majority of the subject's daily life. This prevents overfitting in learning using the classifier and prevents a decrease in the classification accuracy resulting from that learning.

[0061] Furthermore, in Figure 8, the seizure symptom estimation unit 13 applies a Residual Network (ResNet) 18 as a classifier to detect abnormalities in epileptic electroencephalograms using image recognition technology, and also applies backpropagation as a method for updating the training data of the seizure symptom estimation model.

[0062] This residual network is a neural network that can achieve layer depth and improved accuracy by introducing shortcut connections. The machine learning library (PyTorch torchvision library) provides five versions of pre-trained models (18, 34, 50, 101, and 152 layers). Since training a model takes longer with deeper layers, we decided to use the 18-layer ResNet as the classifier. The weights of the pre-trained ImageNet were then applied using the torchvision model subpackage. I changed it to load from subpackage.

[0063] Furthermore, since ResNet is a model designed to classify input images into 1,000 categories, the output layer needs to be replaced for two-class classification. Therefore, a linear layer with two output nodes was created and used to replace the output layer of ResNet18.

[0064] In the seizure symptom estimation unit 13, the number of epochs (the number of training iterations for each data subset) was set to "50". During training, the loss was calculated for each epoch, and backpropagation was performed. In the seizure symptom estimation unit 13, during validation, the accuracy for each epoch was calculated using validation data, and the model's weight parameters were saved when the accuracy improved.

[0065] Since seizure detection is a binary classification task, cross-entropy loss was used as the loss function. Furthermore, Adam (Adaptive Momentum) was used as the optimizer to minimize the loss function. Estimation was used, and the learning rate was set to 0.005. In this invention, a ResNet trained on the ImageNet dataset was fine-tuned for each subject to create a classifier tailored to the subject's seizure symptoms.

[0066] As a result, the brain activity detection device 2, by pre-separating input features and residual features using a residual network, can transmit features supplied from the upstream layer directly to the downstream layer even if the residual features cannot be calculated correctly. This avoids phenomena that hinder learning progress, such as vanishing gradients, that occur in each layer upstream of the hidden layer. Consequently, the degradation of features due to iterative calculations associated with deepening the network can be suppressed.

[0067] (3) Experimental results using a brain activity detection device In order to confirm whether the brain activity detection device 2 described above could detect abnormal electroencephalograms with sufficient accuracy using a deep learning method with a seizure symptom estimation model, an experiment was conducted on 10 epilepsy patients (ages 3 to 19) registered in the CHB-MIT Scalp EEG database.

[0068] Figure 9 shows the information of the epilepsy patients who will be the subjects of the study. In Figure 9, the ID number, gender, age, and seizure count of the 10 patients are shown. The data for each seizure period, the average time of ictal EEG (s) corresponding to the seizure period, and the total time of interictal EEG (h) corresponding to the interictal period are displayed in a list.

[0069] Two types of experiments were conducted: one using the Fp1-F7 channels and another using the Fp2-F8 channels. To evaluate the deep learning method using the seizure symptom estimation model according to the present invention, leave-one-out cross-validation (LOOCV) was performed, with only one test data point used for cross-validation.

[0070] The experimental procedure is as follows: First, N subgroups were created according to the number of seizures N for each patient. Next, the N seizure data were distributed to each subgroup, and then the non-seizure data was distributed to each subgroup so that the total time of non-seizure data for each subgroup was equal.

[0071] For example, if the ID number is chb01, the dataset contains seven EDF files (chb01_03.edf, chb01_04.edf, chb01_15.edf, chb01_16.edf, chb01_18.edf, chb01_21.edf, chb01_26.edf) that include seizure counts. Therefore, seven subgroups are created, and the seizure data is distributed to each subgroup.

[0072] Furthermore, since the dataset includes 21 hours of non-seizure data (total time of EEG feature data corresponding to the interictal period), the 21 hours of EEG feature data was split into 3 hours of EEG feature data and distributed to each subgroup. One of the N subgroups was used as validation data, and the others were used as training data.

[0073] This cross-validation process was repeated N times, with each of the N subgroups used once as validation data. Sensitivity, specificity, and detection delay time (DLT) were then evaluated.

[0074] Here, TP is the number of true positives, FP is the number of false positives, TN is the number of true negatives, and FN is false When the number of negative results is used, the sensitivity is given by the following formula (1):

number

number

number

[0075] Furthermore, if Ts is the time of seizure onset and Td is the time when the seizure is first detected, the detection delay time (DLT) is given by the following equation (3):

number

[0076] To determine the seizure type of epilepsy, it is necessary to estimate the epileptic lesion using electroencephalography (EEG). In this invention, 18 channels of EEG were read for each patient to estimate the epileptic lesion. EEG interpretation was performed based on criteria such as phase inversion between channels, large amplitude (200 μV or more), appearance of spike waves, appearance of sharp waves, and other abnormal epileptic EEG patterns.

[0077] As an example, Figure 10 shows the electroencephalogram (EEG) characteristic data corresponding to the seizure period for ID number chb01. The seizure began at 2,996 seconds, and immediately after onset, high-amplitude EEG phase inversion (±200 [μV]) was observed between Fp2-F4 and F4-C4. Five seconds after the onset of the seizure, polyspikes appeared between Fp2-F8 and F8-T8. From these findings, it is considered that the epileptic lesion was located around F4 and F8.

[0078] Figure 11 shows the electroencephalogram (EEG) characteristic data corresponding to the seizure period for patient ID number chb05. The seizure began at 2,317 seconds, and polyspikes appeared immediately after onset between F7-T7, T7-P7, F8-T8, and T8-P8. Additionally, 4 seconds after the onset of the seizure, spikes and slow-wave complexes were observed throughout the entire head. Based on these findings, it is considered that the seizure lesion was located in T7 and T8. Therefore, EEG measurements were performed on all 10 patients.

[0079] The experimental results using the Fp1-F7 channels (Figure 3) are shown in Figure 12. The average sensitivity was 88.08% (%), the average specificity was 98.98% (%), and the average detection delay time (DLT) was 7.55 seconds. The experimental results using the Fp2-F8 channels (Figure 3) are shown in Figure 13. The average sensitivity was 86.63% (%), the average specificity was 97.76% (%), and the average detection delay time (DLT) was 8.16 seconds.

[0080] Figure 14 shows the results of selecting the optimal channel in terms of sensitivity and specificity for each patient. The mean sensitivity, mean specificity, and mean detection delay time (DLT) were 88.73% and 98.98%, respectively.

[0081] Figure 15 shows the interpretation of the electroencephalogram (EEG) characteristic data of 10 patients, i.e., the estimated seizure foci. F7, T7, and P7 represent the anterior, middle, and posterior temporal regions of the left hemisphere, respectively, while F8, T8, and P8 represent the anterior, middle, and posterior temporal regions of the right hemisphere (Figure 3). Based on the EEG measurements, it was estimated that all 10 patients had epileptic foci in the temporal region.

[0082] (4) Effects of this embodiment Several methods using multi-channel EEG (electroencephalography) have been proposed to detect seizures, and while these can be applied to automated interpretation of EEGs and the diagnosis of epilepsy, most of the measurement devices are relatively large, and electrodes must be placed all over the subject's head, making it extremely difficult to measure multi-channel EEG in everyday life.

[0083] Therefore, experimental results have confirmed that by detecting seizures using a single-channel frontal lobe electroencephalogram (EEG) in the brain activity detection device 2 as in the present invention, an average sensitivity of 88.73% can be obtained, which is equivalent to the sensitivity of conventional methods using multi-channel EEG (average sensitivity of 90.62%). Since abnormal epileptic EEG can be clearly distinguished from the background activity of normal EEG, it is possible to achieve highly sensitive seizure detection based on waveform patterns.

[0084] Furthermore, experimental results have confirmed that the detection delay time (DLT) according to the present invention is 7.39 seconds. Since epileptic seizures typically last 30 to 60 seconds, if caregivers can be notified within 8 seconds of the onset of a seizure, they will be able to provide prompt care to the subject after the seizure.

[0085] On the other hand, according to the present invention, the average specificity was 98.98%, which is slightly lower than the conventional method using multi-channel EEG (average sensitivity 99.32%). One reason for this is that the amount of brain activity information obtained from one channel of electroencephalography is small, and the electroencephalography is superimposed with noise from head movements, facial muscle tension, blinking, etc. To improve the average specificity, it is possible to resolve this by setting a larger window width (overlap ratio) using the sliding window method in the image conversion unit 11 (Figure 2) described above.

[0086] Furthermore, in order to estimate the location of the epileptic lesions in the patients, electroencephalogram (EEG) signals were examined using the brain activity detection device 2 according to the present invention. As a result, it was estimated that all 10 patients had epileptic lesions in the temporal lobe. Therefore, the brain activity detection device 2 using a single-channel frontal lobe EEG can also detect temporal lobe seizures in the patients.

[0087] Temporal lobe epilepsy has two types: "medial temporal lobe epilepsy," in which seizures are triggered by rigidity of the hippocampus in the limbic system, and "lateral temporal lobe epilepsy," in which seizures are triggered by epileptic discharges in the neocortex. Because the prefrontal cortex is interconnected with the temporal cortex and limbic system by fibers, it is thought that when epileptic discharges occur in the temporal region, the abnormal excitation propagates to the prefrontal cortex.

[0088] Some studies have shown that using a single-channel temporal lobe electroencephalogram (FT10-T8) is effective in detecting temporal lobe seizures. However, when measuring the temporal channel, it is often difficult to place electrodes at the measurement point due to the presence of the subject's hair.

[0089] Therefore, the method of detecting a single-channel frontal lobe electroencephalogram (EEG) from the subject's frontal lobe using the brain activity detection device 2 of the present invention is more effective for detecting temporal lobe seizures during daily life compared to other methods using single-channel EEG, in terms of the ease of placing electrodes on the subject's head.

[0090] With the above configuration, a brain activity detection device 2 can be realized that can effectively detect seizures caused by the subject's brain activity in daily life using a method that measures a single-channel frontal lobe electroencephalogram.

[0091] (5) Other embodiments As described above, in this embodiment, the seizure symptom estimation unit 13 in the brain activity detection device 2 is described as estimating the corresponding seizure symptoms using supervised deep learning with a seizure symptom estimation model. However, the present invention is not limited to this, and a neural network other than a residual network may be applied as the classifier in deep learning. Furthermore, a method other than backpropagation may be applied as the method for updating the training data of the seizure symptom estimation model.

[0092] In this embodiment, the image conversion unit 11 extracts partial time series from the electroencephalogram signal detected by the signal detection unit 10 by sequentially dividing it into 2-second intervals with a 50% overlap using the sliding window method, and converts each partial time series into a segment image by setting it to a 2D coordinate system in which the amplitude is increased by 6 to 60 times. However, the present invention is not limited to this, and as long as the electroencephalogram feature extraction unit 12 can effectively extract electroencephalogram feature data within the segment image, the overlap ratio may be 50% or more, the division time may be set to something other than 2 seconds, and the amplitude increase ratio does not have to be limited to 6 to 60 times. [Explanation of Symbols]

[0093] 1...Brain activity detection system, 2...Brain activity detection device, 2A...Electrodes, 3...Terminal device, 4...Network, 5...Cloud server, 6...Data analysis platform, 10...Signal detection unit, 11...Image conversion unit, 12...EEG feature extraction unit, 13...Seizure symptom estimation unit, 14...Transmission unit.

Claims

1. A signal detection unit that detects a single-channel electroencephalogram (EEG) signal representing the frontal lobe EEG via a pair of electrodes placed on the subject's frontal lobe, An image conversion unit extracts partial time series from the electroencephalogram signal detected by the signal detection unit, by sequentially dividing it into predetermined time widths so as to partially overlap in the time direction, and converts each of these partial time series into a segment image represented in a two-dimensional coordinate system with time and amplitude as coordinate axes. An electroencephalogram (EEG) feature extraction unit sequentially analyzes the segment images converted by the image conversion unit, extracts EEG feature data from each segment image, and, based on the extracted EEG feature data, uses a predetermined seizure detection algorithm to sequentially determine, in chronological order of the segment images, whether to distinguish between the seizure period, which represents the period from 30 minutes before the start of a seizure through the seizure and 30 minutes after the end of the seizure, and the interseizure period, which represents the period of 30 minutes or more from the start and end of a seizure. The EEG feature data in each segment image extracted by the EEG feature extraction unit is undersampled to match the number of EEG feature data corresponding to the seizure period, creating n sub-datasets. Using a classifier corresponding to each sub-dataset, the final output determined by a majority vote of the outputs of all the classifiers is used as input. The seizure symptom estimation unit estimates the corresponding seizure symptoms by referring to a seizure symptom estimation model constructed by deep learning, using EEG feature patterns classified according to each seizure symptom caused by abnormal human brain waves as training data. A brain activity detection device characterized by comprising the following features.

2. The image conversion unit, From the electroencephalogram (EEG) signal detected by the signal detection unit, the partial time series is extracted by sequentially dividing it into predetermined time intervals with a predetermined overlap using the sliding window method, and each of these partial time series is set in a two-dimensional coordinate system that is amplified by a predetermined number of times in the amplitude direction, and converted into the segment image. The brain activity detection device according to claim 1.

3. The seizure symptom estimation unit applies a residual network as the classifier and backpropagation as the method for updating the training data of the seizure symptom estimation model. The brain activity detection device according to claim 1.

4. The first step involves detecting a single-channel electroencephalogram (EEG) signal representing the frontal lobe electroencephalogram via a pair of electrodes placed on the subject's frontal lobe, The second step involves extracting partial time series from the electroencephalogram signal detected in the first step by sequentially dividing it into predetermined time widths so as to partially overlap in the time direction, and converting each of these partial time series into a segment image represented in a two-dimensional coordinate system with time and amplitude as coordinate axes. A third step involves sequentially analyzing the segment images converted in the second step, extracting electroencephalogram (EEG) feature data from each segment image, and, based on the extracted EEG feature data, using a predetermined seizure detection algorithm, sequentially determining, in chronological order of the segment images, whether to distinguish between the seizure period, which represents the period from 30 minutes before the start of a seizure through the seizure to 30 minutes after the end of the seizure, and the intercalation period, which represents the period of 30 minutes or more from the start and end of the seizure. In the third step, the electroencephalogram (EEG) feature data in each segment image extracted in the third step is undersampled to match the number of EEG feature data corresponding to the seizure period, creating n sub-datasets. Using a classifier corresponding to each sub-dataset, the final output determined by a majority vote of the outputs of all the classifiers is used as input. The fourth step involves using EEG feature patterns classified according to each seizure symptom caused by abnormal human EEG as training data, and estimating the corresponding seizure symptoms while referring to a seizure symptom estimation model constructed by deep learning. A method for detecting brain activity, characterized by comprising the following features.

5. In the second step described above, From the electroencephalogram (EEG) signal detected in the first step, the partial time series is extracted by sequentially dividing it into sliding windows of predetermined time intervals with a predetermined overlap using the sliding window method, and each of these partial time series is set in a two-dimensional coordinate system that is amplified by a predetermined factor in the amplitude direction and converted into the segment image. The brain activity detection method according to feature 4.

6. In the fourth step, a residual network is applied as the classifier, and backpropagation is applied as the method for updating the training data of the seizure symptom estimation model. The brain activity detection method according to feature 4.

Images