Seizure predicting program, storage medium, seizure predicting device, and seizure predicting method

Abstract

The present invention addresses the problem of providing a novel technique that is very useful for predicting seizures. A program according to the present invention causes a computer to function as an acquiring unit and a predicting unit, wherein: the acquiring unit acquires log information including at least one of behavior information indicating a behavior of a subject, emotion information input on the basis of a subjective opinion of the subject, and seizure information relating to a record of seizures that have occurred in the subject; and the predicting unit predicts seizures associated with the subject's brain activity or mental activity on the basis of the log information.

Description

Seizure prediction program, storage medium, seizure prediction device, and seizure prediction method

[0001] The present invention relates to a seizure prediction program, a storage medium, a seizure prediction device, and a seizure prediction method.

[0002] Epileptic seizures almost always occur without provocation and are difficult to detect or predict, which significantly reduces the quality of life (QOL) of patients. To address these issues, research has been conducted on epileptic seizures and similar seizures associated with brain activity or mental activity (hereinafter simply referred to as "seizures" in this specification).

[0003] For example, Patent Document 1 describes an epileptic seizure prediction technology that detects signs of an epileptic seizure from data generated from electrocardiogram signals. The technology described in Patent Document 1 detects signs of an epileptic seizure based on whether information based on heart rate exceeds a control limit that should be met when the patient is in an interictal period.

[0004] International Publication No. 2020 / 066430

[0005] While there are known technologies for detecting epileptic seizures and detecting signs of seizures based on biological information such as heart rate, such technologies require the acquisition of biological information. This requires the use of expensive devices for measuring the biological information and the placement of electrodes on the subject's body. Furthermore, there is a problem in that signs of an epileptic seizure appear in biological information just before a seizure occurs, and even if a sign is detected, there are limited countermeasures that can be taken.

[0006] In light of the above, an object of the present invention is to provide a novel technology for predicting seizures that is highly useful.

[0007] In order to solve the above problems, the present invention provides a seizure prediction program that causes a computer to function as an acquisition unit and a prediction unit, wherein the acquisition unit acquires log information including at least one of behavioral information indicating the behavior of a subject, emotional information input based on the subject's subjective opinion, and seizure information related to records of seizures that have occurred in the subject, and the prediction unit predicts seizures associated with the subject's brain activity or mental activity based on the log information.

[0008] With this configuration, seizures can be predicted based on log information without using biological information, making it possible to predict seizures based on information that can be obtained without placing the burden of constantly wearing a device on the subject or interfering with the subject's body.

[0009] In a preferred embodiment of the present invention, the acquisition unit acquires at least two of the behavioral information, the emotional information, and the seizure information as the log information, and the prediction unit predicts the seizure of the subject based on the multiple types of log information acquired by the acquisition unit.

[0010] This configuration allows for more accurate prediction of seizures by taking into account multiple combinations of the subject's behavior, emotions, and seizure records.

[0011] In a preferred embodiment of the present invention, the behavioral information includes information indicating the behavior and the timing of the behavior, the emotional information includes information indicating the emotion felt by the subject and the timing of the emotion, the seizure information includes information indicating the timing of a seizure that occurred in the subject, and the prediction unit predicts a seizure based on the timing indicated by the log information.

[0012] With this configuration, future seizures can be predicted from records of the subject's past behavior, emotions, or seizures.

[0013] In a preferred embodiment of the present invention, the prediction unit predicts that a seizure will not occur within a specified time or period after the time indicated by the log information.

[0014] This configuration makes it possible to predict that a seizure will not occur during a specified period, greatly improving the quality of life of epilepsy patients.

[0015] In a preferred embodiment of the present invention, the prediction unit predicts that a seizure will not occur on the specified day based on the behavioral information or emotional information from the day before the specified day.

[0016] With this configuration, the subject can know that a seizure is unlikely to occur on the specified date. By predicting a specified date as the specified period, it is expected that the subject will be able to plan their activities more easily. Specifically, on days when the possibility of a seizure occurring is low, the subject will be able to go out and do other activities with peace of mind, which is expected to have a significant stress reduction effect.

[0017] In a preferred embodiment of the present invention, the prediction unit makes the prediction by estimating whether or not a seizure will occur on the specified day using a seizure interval that indicates the elapsed period since the previous seizure, based on the timing of the seizure information recorded in the past immediately prior to the specified day.

[0018] In a preferred embodiment of the present invention, the prediction unit makes the prediction by estimating whether a seizure will occur within a specified time or period after the time indicated by the log information.

[0019] This configuration makes it easier for the subject to take measures according to the prediction results, such as taking medication or avoiding dangerous objects or actions during periods when seizures are likely to occur.

[0020] In a preferred form of the present invention, for the purpose of making the prediction, the prediction unit inputs the log information into a trained model generated by machine learning, and obtains as an output of the trained model an estimated result of whether or not the subject will have a seizure within a specified time or period after the time indicated by the log information, and the trained model is generated from training data including the subject's log information and information indicating whether or not the subject has had a seizure within a specified time or period after the time indicated by the log information.

[0021] With this configuration, it is possible to predict whether or not a seizure will occur within a specified time or period using a trained model that has been trained based on past performance.

[0022] In a preferred form of the present invention, the trained model is generated by machine learning using a dataset including training data on multiple subjects with the same seizure classification as the subject and training data on the subject.

[0023] This configuration makes it possible to realize more appropriate predictions that reflect the tendencies of the subject. Furthermore, while the amount of training data tends to be small when only the subject's data is used, combining the data of multiple subjects and the subject's data makes it possible to secure a large amount of training data while also reflecting the subject's characteristics.

[0024] In a preferred embodiment of the present invention, the prediction unit predicts a seizure of the subject based on the log information as well as a seizure classification of the subject.

[0025] Such a configuration is expected to achieve more accurate predictions.

[0026] In a preferred embodiment of the present invention, the acquisition unit acquires the behavioral information as the log information, and the behavioral information includes sleep information related to the sleep of the subject.

[0027] It is known that sleep is related to seizures, and this configuration is expected to result in more accurate predictions.

[0028] In a preferred embodiment of the present invention, the acquisition unit acquires the behavioral information as the log information, and the behavioral information includes usage information acquired from background data in a terminal used by the subject.

[0029] With this configuration, the subject can acquire behavioral information without having to input it himself, further reducing the burden on the subject.

[0030] In a preferred embodiment of the present invention, the acquisition unit acquires the behavioral information as the log information, and the behavioral information includes movement information related to a movement speed of the subject.

[0031] In a preferred form of the present invention, the acquisition unit further acquires evaluation information based on the results of an examination or ability test of the subject, and the prediction unit predicts a seizure of the subject based on the evaluation information in addition to the log information.

[0032] Such a configuration is expected to achieve more accurate predictions.

[0033] In order to solve the above problems, the present invention is a seizure prediction device comprising an acquisition unit and a prediction unit, wherein the acquisition unit acquires log information including at least one of behavioral information indicating the behavior of a subject, emotional information input based on the subject's subjective opinion, and seizure information relating to records of seizures that have occurred in the subject, and the prediction unit predicts seizures associated with the subject's brain activity or mental activity based on the log information.

[0034] In order to solve the above-mentioned problems, the present invention provides a seizure prediction method in which a computer including an acquisition unit and a prediction unit acquires, by the acquisition unit, log information including at least one of behavioral information indicating the behavior of a subject, emotional information input based on the subject's subjective opinion, and seizure information relating to a record of seizures that have occurred in the subject, and predicts, by the prediction unit, a seizure associated with the subject's brain activity or mental activity based on the log information.

[0035] In order to solve the above problem, the present invention provides a seizure prediction method that acquires log information including at least one of behavioral information indicating a subject's behavior and the timing of those behaviors, emotional information indicating the subject's subjective emotions and the timing of those emotions, and seizure information relating to a record of seizures that have occurred in the subject, and estimates whether a seizure associated with brain activity or mental activity will occur on a specified day based on the log information from the day before the specified day.

[0036] According to the present invention, a novel technique for easily predicting seizures can be provided.

[0037] The present invention relates to a method for predicting a seizure, a method for predicting a seizure using a terminal device, and a method for predicting a seizure using a computer.

[0038] The present invention relates to a technology for predicting seizures associated with brain activity or mental activity in a subject. In particular, a seizure in this invention refers to a sudden symptom that repeatedly occurs in the same patient associated with brain activity or mental activity. In this embodiment, epileptic seizures and seizures similar thereto (exhibiting similar symptoms), whose symptoms appear in the state of consciousness or motor state, are predicted. Such seizures are classified into, for example, focal seizures, generalized seizures, and psychogenic non-epileptic seizures (PNES). Hereinafter, in this specification, such seizures associated with brain activity or mental activity will be simply referred to as "seizures."

[0039] In the present invention, the term "subject" refers to a subject for whom a seizure is predicted. The subject of the present invention is particularly a person who is likely to have a seizure. Specifically, the subject hereinafter refers to an epilepsy patient or a patient with psychogenic nonepileptic seizures who is a subject for seizure prediction. Note that the present invention is intended to present the results of seizure prediction, and is not intended for therapeutic or diagnostic purposes.

[0040] The present invention will now be described in more detail with reference to the accompanying drawings, in which preferred embodiments are shown, but which may be embodied in many different forms and are not limited to the embodiments set forth herein.

[0041] For example, in this embodiment, the configuration, operation, etc. of a seizure prediction device will be described, but similar effects can also be achieved by a system, method, method executed by a device, computer program that causes a computer device to execute a method, etc. that has similar functions. The program may be provided as a non-transitory computer-readable recording medium, or may be provided so as to be downloadable from an external server.

[0042] <1. System Configuration> Figure 1 is a block diagram showing the configuration of a seizure prediction system including a seizure prediction device of this embodiment. As shown in Figure 1, the seizure prediction system 0 of this embodiment includes a seizure prediction device 1 and a user terminal 2. The seizure prediction device 1 and the user terminal 2 are configured to be able to communicate via a network NW. In this embodiment, the network NW is an IP (Internet Protocol) network, but there are no limitations on the type of communication protocol, type of network, etc.

[0043] One or more information processing devices 10 (computer devices), such as a general-purpose server or personal computer, can be used as the seizure prediction device 1. Furthermore, a terminal device 9 (computer device), such as a personal computer, smartphone, or tablet terminal, can be used as the user terminal 2. In this embodiment, the seizure prediction device 1 is an information processing device 10 in which a computer program (seizure prediction program) that executes a seizure prediction method is installed.

[0044] Fig. 2A is a hardware configuration diagram of the information processing device 10. As shown in Fig. 2, the information processing device 10 has a control unit 101, a storage unit 102, and a communication unit 103, which are used to perform the functions of each unit and each process.

[0045] The control unit 101 has a processor such as a CPU capable of executing an instruction set and executes an OS and programs. The storage unit 102 has a volatile memory such as RAM capable of storing an instruction set, and a non-volatile recording medium such as an HDD or SSD capable of recording the OS, a seizure prediction program, a DBMS (database server), etc. The communication unit 103 has an interface for physically connecting to a network and controls communication with the network NW to input and output information.

[0046] Fig. 2(b) is a hardware configuration diagram of the terminal device 9. As shown in Fig. 2, the terminal device 9 has a control unit 901, a storage unit 902, a communication unit 903, an input unit 904, and an output unit 905, which are used to perform the functions of each unit and each process.

[0047] The control unit 901 has a processor such as a CPU capable of executing an instruction set and executes an OS, programs, etc. The storage unit 902 has a volatile memory such as RAM capable of storing an instruction set, and a non-volatile recording medium such as an HDD or SSD capable of recording the OS, programs, etc. The communication unit 903 has an interface for physically connecting to a network and controls communication with the network NW to input and output information. The input unit 904 has an operation input device capable of input processing such as a touch panel or keyboard, an audio input device capable of audio input such as a microphone, etc. The output unit 905 has a display device capable of display processing such as a display, and an audio output device such as a speaker.

[0048] 2. Functional Configuration The seizure prediction device 1 has a functional configuration including an acquisition unit 11, a prediction unit 12, and a storage unit 13. This is a specific hardware implementation of software-based information processing. The user terminal 2 also has an input means for the user to input information, and an output means including means for outputting to the user and connecting to the network NW. The user terminal 2 is used by the subject or by another person who can observe the subject. In this embodiment, the user terminal 2 is primarily a device such as a smartphone owned or carried by the subject.

[0049] The acquisition unit 11 acquires at least one of behavioral information indicating the behavior of the subject, emotional information input based on the subject's subjective opinion, and seizure information related to records of seizures that have occurred in the subject, via the user terminal 2. In this embodiment, the acquisition unit 11 acquires behavioral information, emotional information, and seizure information via the user terminal 2. Specifically, it is assumed that the subject inputs this information by selection or free input via the user terminal 2, and the user terminal 2 transmits the information together with the subject ID via the network NW to the seizure prediction device 1. Hereinafter, records related to the subject that can be used to predict seizures, including seizure information, behavioral information, and emotional information, will be collectively referred to as log information.

[0050] The log information is linked to information about each subject by a subject ID that identifies the subject, and includes time information indicating the time corresponding to each type of record. The time corresponding to each type of record is, for example, date and time information indicating the date and time of the behavior, emotion, seizure, etc. indicated by each record. The acquisition unit 11 acquires log information within a specific time range that is a predetermined time before the time at which prediction is to be made, and passes it to the prediction unit 12, which will be described later.

[0051] The time period information may be information that specifies the day it belongs to without including the time of day. In this embodiment, a "day" starts from the time the "subject (target person)" wakes up, and a "day" here refers to the period from when the subject wakes up until when they wake up the next day.

[0052] Here, in the present invention, "behavior" refers to observable movements or reactions exhibited by a subject. Therefore, biological information that cannot be observed without using measuring equipment, such as heart rate, brain waves, and breathing, is not included in the behavioral information of the present invention. Among the log information, behavioral information includes, for example, sleep information about the subject's sleep, movement information about the subject's movements, usage information about the usage history of the user terminal 2 used by the subject, exercise information about the exercise performed by the subject, diet information about the meals the subject ate, medication information about the subject's medication history, free time and work time or the ratio thereof, etc. In addition, information indicating any behavior may be included in the behavioral information of the present invention.

[0053] Examples of behavioral information in this embodiment will be described in more detail. Each type of behavioral information includes a unique behavior log ID, time information indicating the date and time of the behavior, and behavior content indicating the type of behavior represented by the behavioral information. Furthermore, the target ID is used to link the information of the target who performed the behavior.

[0054] The sleep information included in the behavioral information includes the bedtime (date and time) and wake-up time (date and time) as the aforementioned time period information. Sleep duration can be calculated from the bedtime and wake-up time, but the sleep duration may also be input separately. Sleep may also be divided into types, such as "real sleep" (mainly at night) and "nap" (such as afternoon naps), and input separately. Here, the bedtime and wake-up time refer to the actual start and end times of sleep, but the planned start time (planned bedtime) and planned end time of sleep (planned wake-up time) may also be acquired. The difference between the planned start time and actual start time of sleep, and the difference between the planned end time and actual end time of sleep, may also be used as sleep information.

[0055] In this embodiment, the boundary between sleep days is 5:00 AM. Specifically, sleep that began between 5:00 AM the previous day and 4:59:59 AM on the current day is treated as sleep on the "previous day." If a person awakens during the night, the sleep information before and after the awakening is not separated, and the sleep up to the point at which the person feels awake is treated as a single piece of sleep information. However, the time of the awakening during the night may also be recorded as sleep information, and the time during which the awakening occurred may be excluded from the sleep time.

[0056] The sleep information also includes information about sleep quality. Specifically, information indicating how one feels when waking up can be used as information about sleep quality. Here, the "feelings when waking up" refers to the subjective feelings about sleep felt by the subject. For example, it is assumed that the subject inputs how one feels when waking up by selecting from multiple options such as "well rested," "fairly rested," "average," "not very rested," and "not at all rested."

[0057] The information relating to sleep quality may also be information representing other indicators such as brain waves during sleep (e.g., the time or percentage at which alpha waves, beta waves, theta waves, and delta waves are detected during sleep), waking up during sleep, body movements during sleep, etc. However, the information relating to sleep quality in this embodiment is preferably information that can be obtained without measuring brain waves, body movements, etc.

[0058] The movement information included in the behavioral information includes a movement speed. For example, the movement speed of the user terminal 2 can be identified using a GPS (Global Positioning System) receiver or a sensor included in the user terminal 2, and this can be used as the movement information. For example, the movement information can include an average movement speed, a maximum movement speed, a minimum movement speed, etc. for a predetermined period, for example, a day (from waking up to going to bed).

[0059] Among the behavioral information, the usage information relating to the usage history of the user terminal 2 used by the subject can be background data acquired by the user terminal 2. For example, the usage time and usage duration (total usage duration in a predetermined period) for each app (or for each app category), the message sending history and number of replies for a messaging app, etc. can be used as usage information. Here, app categories refer to the classification of each app according to its function or purpose, such as tools, SNS, games, music, etc.

[0060] Other examples of usage information include screen-off time, screen-on time, the number of times photos are saved, the number of times videos are saved, and the number of calendar events registered.

[0061] Furthermore, the exercise information included in the behavioral information includes information such as the type, intensity, and duration of exercise performed by the subject. Specifically, for example, the exercise information may include the number of times exercised in a predetermined period, the total time, the number of steps, the distance walked and run, the stride length, the number of floors climbed, the time spent supporting both feet, stability, speed, active energy (energy consumed during exercise), energy consumed at rest, the number of falls, etc. In addition to the above, any information that can be input by the user or that can be obtained by a wearable device may be used.

[0062] Furthermore, the dietary information included in the behavioral information includes information on the content and amount of meals eaten by the subject. For example, the number of meals eaten and calorie intake can also be used as exercise information or dietary information. For example, whether or not a person ate breakfast, lunch, snack, dinner, or midnight snack may be received.

[0063] The medication information included in the behavioral information includes information such as the timing at which the subject took medication (time of day, whether morning, noon, or night, or whether before or after a meal, etc.), the type and amount of medication, etc. Furthermore, for example, the number of times a medication was taken per day may also be used as medication information. The medication information is input by the subject or an observer via the user terminal 2, and is acquired by the acquisition unit 11.

[0064] The above-described behavioral information is merely an example, and various types of behavioral information may include any other information. In this embodiment, all of the above information is used to predict a seizure, but only some of the information may be used for prediction.

[0065] In addition, in the present invention, "emotion" refers to a subjective feeling of a subject. Therefore, anything that can be inferred from biological information such as brain activity information or heart rate is not included in the "emotion" of the present invention.

[0066] The emotional information in the log information is information indicating the type and intensity of an emotion expressed in any manner, such as the four types of joy, anger, sadness, and pleasure, the eight basic emotions of joy, anticipation, anger, disgust, sadness, surprise, fear, and trust (Plutchik's Wheel of Emotions), or indicators of pleasure, displeasure, and arousal (Russell's Circumplex of Emotions). For example, the type and intensity of an emotion selected from a plurality of pre-set options may be used as the emotional information. Alternatively, only the type of emotion may be used as the emotional information, without using the intensity.

[0067] The emotional information may be input only once a day, or may be input at any time (multiple times) when the subject feels an emotion. In this embodiment, the acquisition unit 11 acquires emotional information, once a day, about the overall emotion of that day, which is subjectively selected by the user from nine categories: "happy," "fun," "calm," "tranquil," "bored," "anxious," "tired," "angry," and "sad."

[0068] The seizure information is one type of log information in the present invention, and is information relating to past seizures that have occurred in the subject. The seizure information includes the date on which the seizure occurred, the time the seizure started and ended, the type of seizure, etc. For example, the type of seizure may be the presence or absence of consciousness and / or the presence or absence of convulsions. The seizure information is input by the subject or an observer via the user terminal 2, and is acquired by the acquisition unit 11.

[0069] In addition to the log information, the acquisition unit 11 of this embodiment further acquires environmental information related to the environment surrounding the subject and biometric information of the subject acquired by a wearable device communicatively connected to the user terminal 2. Like the log information, the environmental information and biometric information are linked to the subject information of each subject by a subject ID that identifies the subject, and both include information indicating the time corresponding to various records.

[0070] Possible environmental information includes illuminance, temperature, atmospheric pressure, weather, humidity, etc. The environmental information may be acquired by a sensor included in the user terminal 2, or may be acquired from an external database that provides weather data based on location information of the user terminal 2 obtained by GPS or the like.

[0071] Furthermore, location information indicating the area of ​​stay and activity range information indicating the range of activity may be used as environmental information. Both location information and activity range information can be acquired based on the position information of the user terminal 2 obtained by GPS or the like. Location information is assumed to be classified by area, such as downtown, outskirts, residential areas, etc., and to indicate which area classification the location indicated by the location information falls into. Furthermore, the activity range information may be, for example, the difference between the activity range on weekdays and on holidays.

[0072] The biological information can be any information that can be acquired by a sensor equipped in the wearable device. Examples of biological information include heart rate, activity level, blood pressure, etc. As described above, the present invention predicts seizures primarily based on log information, and biological information can be used as a supplementary measure.

[0073] The acquisition unit 11 also acquires subject information that is pre-stored in the storage unit 13. The subject information includes information such as a subject ID that uniquely identifies the subject, the subject's name, seizure classification, attributes, diagnostic results for the subject, and psychological evaluation results.

[0074] Seizure classification refers to the classification of seizures exhibited by a subject. Seizures associated with brain or mental activity are classified according to the brain activity at the time of the seizure. For example, epileptic seizures cause excessive electrical excitation in the brain, and are classified into "focal seizures (also called partial seizures)" and "generalized seizures" depending on the location and spread of this excitation.

[0075] Furthermore, seizures that do not show abnormal epileptic waves in the EEG and that do not technically constitute "epileptic" seizures but that suddenly cause mental and physical states similar to those of epileptic seizures are called psychogenic non-epileptic seizures (PNES). In the present invention, "seizures" include psychogenic non-epileptic seizures, and "psychogenic non-epileptic seizures" are treated as one type of seizure classification.

[0076] Seizures associated with brain activity or mental activity are classified according to the patient, and in this embodiment, one of the three categories described above, "focal seizures," "generalized seizures," and "psychogenic non-epileptic seizures," that occurs in a subject is registered as the seizure classification in the subject information. Note that seizure classifications are not limited to the above three categories, and classifications that further subdivide focal seizures and generalized seizures may be used, or an "unclassifiable" seizure category may be added.

[0077] Attributes refer to the subject's attributes, such as age or generation, sex, family structure (whether or not the subject is married or has family members living together, etc.), occupation, years of employment, weekly working hours, drinking and smoking habits, recent (e.g., one month) health condition, whether or not the subject is receiving care, housework and exercise habits, etc. Diagnostic results refer to the diagnosis given to the subject and test data obtained through hospital examinations. Psychological evaluation results refer to the results of interviews and psychological evaluation tests administered to the subject, and are an example of evaluation information in the present invention. For example, evaluation information includes evaluation results based on the subject's responses regarding depressive symptoms, quality of life, stigma, etc., and ability evaluation results based on the results of ability tests regarding the subject's verbal IQ, memory, etc.

[0078] The prediction unit 12 predicts a seizure in the subject based on the log information acquired by the acquisition unit 11. The prediction unit 12 of this embodiment combines the various types of information acquired by the acquisition unit 11 described above and uses them to predict a seizure. The combination of information to be used can be determined arbitrarily, but it is particularly preferable to use a combination of at least two of behavioral information, emotional information, and seizure information to predict a seizure, such as a combination of behavioral information and emotional information, a combination of behavioral information and seizure information, or a combination of emotional information and seizure information.

[0079] As described above, the log information, environmental information, and biological information each include date and time information indicating a date and time as information indicating the time period. The prediction unit 12 predicts a seizure at a predetermined time based on the date and time information. More specifically, the prediction unit 12 predicts a seizure that will occur later than the time indicated by the information used for the prediction (log information).

[0080] Specifically, the prediction unit 12 makes a prediction using information from a predetermined time before the specified time or period for which the prediction is to be made. For example, the specified period is assumed to be a specific day, and log information, environmental information, and biological information of the subject on the day before the specified date that is the target of prediction are used to predict whether or not the subject will have a seizure on the specified date (specified period). The period of information used for the prediction and the range of time for which the prediction is to be made may be changed as desired. For example, information from one week immediately preceding the specified date may be used to predict whether or not a seizure will occur on the specified date. Furthermore, information from one week may be used to predict whether or not a seizure will occur in the following week.

[0081] Here, subjects who repeatedly experience seizures often have constant anxiety about seizures, which limits their daily lives. Given this background, providing information that "seizures will not occur during this period" is particularly important for improving the subject's QOL. Therefore, the prediction unit 12 of this embodiment outputs a prediction result that "seizures will not occur" (are unlikely to occur) when the possibility of a seizure occurring during a predetermined period, for example, on a specified date, is low.

[0082] In this embodiment, the prediction unit 12 predicts seizures using a trained model generated using machine learning technology and teacher data. That is, the prediction unit 12 inputs information acquired by the acquisition unit 11 into the trained model, receives an estimated result of whether or not a seizure has occurred at a specified time as the output of the trained model, and outputs a prediction result based on the estimated result. In this embodiment, the memory unit 13 stores the trained model in advance, and the prediction unit 12 inputs various information into the trained model stored in the memory unit 13 and receives the presence or absence of a seizure as the output. The procedure for generating the trained model will be described later.

[0083] The storage unit 13 stores the above-mentioned subject information, log information, environmental information, biological information, etc. The subject information is pre-registered for each subject before receiving information such as log information. The storage unit 13 then receives log information, environmental information, biological information, etc., along with the subject ID from the registered subject as needed, and stores the information in association with the subject information. As a result, various pieces of information are registered in association with the subject and the date and time, and can be used to predict seizures.

[0084] In this embodiment, the storage unit 13 receives and registers various types of information as needed, and the acquisition unit 11 acquires information on the timing required for seizure prediction from the information and passes it on to the prediction unit 12. However, the procedure for acquiring information can be changed as desired. For example, various types of information may be stored in the user terminal 2 without storing the information in the storage unit 13, and the user terminal 2 may transmit the information required for a specified timing to the seizure prediction device 1 each time it is needed. In this embodiment, the storage unit 13 also stores a trained model, which will be described later.

[0085] The above configuration is merely an example, and various functions may be realized by, for example, placing some of the means of each device in another device, and having multiple devices work together. Also, the functions of the seizure prediction device 1 and the user terminal 2 may be provided in a single device, and the input reception of various information such as log information and the prediction of seizures may be performed in a single device.

[0086] 3. Learning Process Next, we will explain how to generate a trained model that takes subject information (particularly attributes, diagnostic results, and psychological evaluation results), log information, environmental information, biological information, etc. as input and outputs an estimation result as to whether a seizure will occur.

[0087] In this embodiment, machine learning is performed on a different model for each seizure classification, and a trained model is generated for each seizure classification. That is, the seizure prediction device 1 generates a trained model trained only on training data related to focal seizures, a trained model trained only on training data related to generalized seizures, and a trained model trained only on training data related to psychogenic non-epileptic seizures, and stores these in the storage unit 13.

[0088] Regarding the type of model, any known classification model can be used. For example, models such as logistic regression, k-nearest neighbor (KNN), decision tree, support vector machine (SVM), and artificial neural network (ANN) can be used. Note that all of these models are widely used in classification problems, and those skilled in the art can understand their mechanisms, so a description thereof will be omitted.

[0089] In this embodiment, a seizure is predicted using a prediction model that outputs a prediction result of whether or not a seizure will occur. Specifically, the prediction model of this embodiment uses, as explanatory variables, subject information as well as log information, environmental information, and biological information within a certain range prior to the specified period (or specified time) for which prediction is to be made, and estimates whether or not a seizure will occur within the specified period, and outputs the result. The prediction unit 12 then receives the estimation result and outputs the prediction result based on it.

[0090] In the learning process, a set of information used as input (explanatory variables) for the model and seizure information that serves as training data for the output is acquired as learning data. Machine learning techniques are widely used, and those skilled in the art can understand the learning procedure, so a detailed explanation will be omitted.

[0091] In this embodiment, a model is generated that outputs a prediction result of whether or not a seizure will occur on a specified date using subject information, a seizure interval determined based on seizure information over a specified period, and behavioral information, emotional information, environmental information, and biological information from the day before the specified date as explanatory variables. Therefore, the training data includes a set of subject information (particularly attributes, diagnostic results, and psychological evaluation results), seizure information for a specific subject on a certain day, seizure information for the most recent specified period (or the period since the previous seizure), behavioral information, emotional information, environmental information, and biological information from the previous day. Here, as described above, a "day" in this embodiment refers to the time from when the "subject (target)" wakes up, and a "day" here refers to the period from when the subject wakes up to when they wake up the next day.

[0092] Here, data from multiple different subjects is used as training data. Although a trained model dedicated to a single subject may be generated, in this embodiment, in order to ensure a sufficient amount of training data, sets of subject information, log information, environmental information, and biological information, and seizure information for multiple subjects assigned the same seizure classification are trained.

[0093] It is known that seizure tendency varies from person to person, and the trained model generated by the above procedure may be adjusted for each subject. For example, a trained model may be generated by performing machine learning using a dataset that includes not only the training data described above but also training data obtained from the subject as a test subject. Furthermore, for example, a trained model may be generated by prioritizing training data from time periods when the subject is prone to seizures.

[0094] Alternatively, a trained model may be first generated using training data on multiple subjects other than the subject, and then the model may be further trained using additional training data in which the subject is a subject, thereby adjusting the model to suit the subject.

[0095] For the learning data on the subject, it is assumed that the subject's seizure information, behavioral information, and / or emotional information over a certain period of time prior to the specified date is used as the explanatory variables, and the presence or absence of a seizure on the specified date is used as the objective variable. As described above, the explanatory variables can be not only seizure information, behavioral information, or emotional information, but also any combination of seizure information, behavioral information, emotional information, environmental information, biological information, and subject information.

[0096] The period of time for which the information used as the explanatory variables is to be a certain period or longer is preferable to reflect the subject's tendency toward seizures. In particular, it is preferable that the information (seizure information, behavioral information, or emotional information) used for adjustment to suit the subject be a period longer than the period of the information (seizure information, behavioral information, or emotional information) used in the aforementioned learning involving multiple subjects.

[0097] More specifically, for example, in the aforementioned learning involving multiple subjects, learning is performed using information from the previous day as the explanatory variable and the presence or absence of a seizure on a specified date as the objective variable, while in the adjustment tailored to the subject, additional learning can be performed using information from one month prior to the specified date as the explanatory variable and the presence or absence of a seizure on the specified date as the objective variable. Furthermore, in the adjustment tailored to the subject, all available past data can be used without any particular time period limitation. This allows sufficient information about the subject's seizures to be learned, enabling highly accurate predictions that better reflect the subject's seizure tendencies.

[0098] 4. Prediction Processing Next, the seizure prediction processing by the prediction unit 12 will be described with reference to FIG. 3. In this embodiment, seizures are predicted using the trained models that correspond to the three seizure classifications of focal seizures, generalized seizures, and psychogenic non-epileptic seizures and that have been further adjusted for each subject using the method described above. Note that predictions can be made using any method based on the correlation between seizure information, behavioral information, or emotional information and the presence or absence of a seizure, and the prediction method is not limited to this.

[0099] 3 is a flowchart showing the processing steps for predicting a seizure by the prediction unit 12. First, in step S1, the acquisition unit 11 acquires information to be used for predicting a seizure. The acquired information includes subject information, log information about the subject, environmental information, and biological information, which are the same types of information used to generate the trained model. The timing of the log information, environmental information, and biological information may also be acquired within the ranges used as explanatory variables in generating the trained model.

[0100] In this embodiment, the user terminal 2 transmits log information, environmental information, and biological information for each subject as needed and stores them in the storage unit 13. The acquisition unit 11 acquires information for the required period from the storage unit 13 according to the period to be predicted. In this embodiment, the acquisition unit 11 acquires log information, environmental information, and biological information, as well as subject information, for the same period as the period used for the aforementioned adjustment to the subject. However, there is no limit to the period for acquiring the log information, environmental information, and biological information, and the acquisition period may be shorter than the period used for adjusting the model to the subject. Note that when adjustment for each subject is not performed, for example, to estimate the presence or absence of a seizure on a specified day using information from the previous day, the acquisition unit 11 acquires log information, environmental information, and biological information belonging to the day before the specified day, as well as subject information.

[0101] Next, in step S2, the prediction unit 12 receives the information acquired in step S1 from the acquisition unit 11 and inputs it into the trained model. Here, as described above, a trained model is generated for each subject and stored in the memory unit 13. Therefore, in step S2, the prediction unit 12 references the subject ID and inputs subject information (particularly attributes, diagnosis results, and psychological evaluation results), log information, environmental information, and biological information into the corresponding trained model. Note that if adjustment for each subject is not performed, the prediction unit 12 references the seizure classification in the subject's subject information and inputs the subject information (particularly attributes, diagnosis results, and psychological evaluation results), log information, environmental information, and biological information into the trained model corresponding to that seizure classification.

[0102] Then, in step S3, the prediction unit 12 receives the probability that a seizure will occur in the subject on the specified date and the probability that a seizure will not occur in the subject on the specified date as the output of the trained model. The prediction unit 12 outputs the higher probability as the prediction result. The user terminal 2 receives the output from the prediction unit 12 and outputs the result to the subject by displaying it on a screen, speaking, or the like. The output format may be determined arbitrarily, but for example, the user terminal 2 may display the prediction result as a symbol, such as a "sunny" symbol if the prediction result is that a seizure will occur, or a "rainy" symbol if the prediction result is that a seizure will not occur.

[0103] As described above, the seizure prediction device according to this embodiment can predict a seizure based on easily available information, such as behavioral information or emotional information. The seizure prediction method according to the present invention is a method in which a computer executes the above-described procedures, or a method in which any subject executes the above-described procedures. The seizure prediction method according to the present invention is intended solely to provide a subject with information to help guide their behavior by predicting a seizure, and is not intended for diagnosis or treatment.

[0104] Furthermore, unlike conventional technologies that detect seizures and their symptoms based on biological information, the seizure prediction device according to this embodiment can predict the presence or absence of a seizure not just before a seizure occurs, but for example, within the day, etc. Below, we will explain examples in which a seizure was actually predicted using the seizure prediction device, seizure prediction program, and seizure prediction method of the present invention.

[0105] (1) Generation of trained models First, trained models were generated for each seizure classification using the log information, environmental information, and biological information of multiple subjects using the method described in <3. Training process>. The subjects were 12 epilepsy patients aged 18 years or older who required hospitalization for detailed examination to diagnose epilepsy and determine treatment options. The data collection period was 667 days, of which seizures associated with brain activity or mental activity were observed on 82 days.

[0106] The explanatory variables used were 368 items, including behavioral information such as sleep duration, wake-up time, average daily movement speed, and maximum daily movement speed, emotional information input by the subject at any time, most recent seizure information indicating the date and time of the previous seizure, biological information measured by a wristwatch-type wearable device, and attribute information indicating the subject's attributes. Note that the elapsed time since the previous seizure may be used instead of or in addition to the most recent seizure information. Specific examples of explanatory variables are as described in the above-mentioned embodiment. In this example, a trained model is generated by performing training for each seizure classification, and no adjustment is made to suit each subject.

[0107] The analysis algorithm used was LightGBM (Light Gradient Boosting Machine). LightGBM is a supervised learning method that classifies explanatory variables according to the target variable. (Reference URL: https: / / lightgbm.readthedocs.io / en / latest / Python-Intro.html)

[0108] (2) Prediction Results Using the model created using the above procedure, seizure predictions were performed using test data. The accuracy rate, precision rate, recall rate, specificity, and F-measure were calculated from the prediction results, and the accuracy rate was 87.22%, precision rate 54.09%, recall rate 78.57%, specificity 88.51%, and F-measure 62.74%.

[0109] These results showed that it is possible to predict with high accuracy whether or not a seizure occurred on a specified date.

[0110] (3) Evaluation of learning results: The contribution of each explanatory variable was evaluated using the SHAP (Shapley Additive Explanations) method. SHAP uses the trained model and explanatory variables to calculate the SHAP value, which indicates the contribution of the explanatory variables to the estimation results. (Reference URL: https: / / shap.readthedocs.io / en / latest / )

[0111] When the SHAP value was calculated for the trained model according to this example, the following explanatory variables were considered important in the estimation: Seizure interval (the period since the previous seizure); attribute information; sleep information; movement information; psychological evaluation results; and emotional information.

[0112] These results suggest that it is possible to predict whether or not a seizure will occur based on behavioral information such as sleep and movement information, emotional information, and seizure information.

[0113] Furthermore, the inventors performed an experiment to confirm the accuracy by generating a trained model under different detailed conditions and adding training data. Hereinafter, differences between Example 2 and Example 1 will be described. Explanation of parts common to Example 1 will be omitted.

[0114] (1) Generation of trained models First, trained models were generated for each seizure classification using the log information, environmental information, and biological information of multiple subjects using the method described in <3. Training process>. The subjects were 23 epilepsy patients aged 18 years or older who required hospitalization for detailed examination to diagnose epilepsy and determine treatment options. The data collection period was 1,590 days, of which seizures associated with brain activity or mental activity were observed on 126 days.

[0115] As explanatory variables, 270 items were used, including behavioral information, emotional information, seizure information, biological information, and attribute information. Specifically, from the 368 items used in Example 1, items that were substantially overlapping or meaningless on their own were excluded or combined, and items indicating the presence or absence of abnormalities in MRI tests and medication history were added.

[0116] Furthermore, in Example 2, adjustments were made for missing data and imbalanced data during learning. Specifically, median imputation was performed for missing data. Furthermore, based on the frequency of seizures, the learning data was imbalanced data in that there were more days without seizures than days with seizures. Therefore, in Example 2, adjustments for imbalanced data were made by oversampling.

[0117] Various methods are known for dealing with imbalanced data, such as undersampling, weighting, and learning as an anomaly detection problem. While oversampling was performed in this example, other methods may be used to create training data. The analysis algorithm is the same as in Example 1.

[0118] (2) Prediction Results Using the model created using the above procedure, seizure predictions were performed using test data. The accuracy rate, precision rate, recall rate, specificity, and F-measure were calculated from the prediction results, and the accuracy rate was 90.07%, precision rate 65.72%, recall rate 68.57%, specificity 93.71%, and F-measure 66.79%.

[0119] Compared to the results of Example 1, the precision rate was particularly improved. Here, precision rate and recall rate are in a trade-off relationship. In Example 2, precision rate and recall rate were at a high level in a well-balanced manner, and the F-value, which indicates the harmonic mean of precision rate and recall rate, was improved compared to Example 1. From this, it can be said that the accuracy was improved by adjusting the explanatory variables, complementing missing data, oversampling, and other adjustments.

[0120] 0: Seizure prediction system 1: Seizure prediction device 2: User terminal 11: Acquisition unit 12: Prediction unit 13: Storage unit 10: Information processing device 101: Control unit 102: Storage unit 103: Communication unit 9: Terminal device 901: Control unit 902: Storage unit 903: Communication unit 904: Input unit 905: Output unit

Claims

1. A seizure prediction program that causes a computer to function as an acquisition unit and a prediction unit, wherein the acquisition unit acquires log information including at least one of behavioral information indicating a subject's behavior and the timing of those actions, emotional information indicating the emotions felt by the subject and the timing of those emotions, which is input based on the subject's subjective opinion, and seizure information indicating seizures that have occurred in the subject and the timing of those seizures, and the prediction unit inputs the log information into a trained model generated by machine learning and obtains, as an output from the trained model, an estimated result of whether or not the subject will have a seizure within a specified time or period after the time indicated in the log information, thereby predicting a seizure associated with the subject's brain activity or mental activity, and the trained model is generated from training data including the subject's log information and information indicating whether or not the subject has had a seizure within a specified time or period after the time indicated in the log information.

2. The seizure prediction program of claim 1, wherein the acquisition unit acquires at least two of the behavioral information, the emotional information, and the seizure information as the log information, and the prediction unit inputs the multiple types of log information acquired by the acquisition unit into the trained model and predicts the seizure of the subject based on the output of the trained model.

3. The seizure prediction program according to claim 1, wherein the prediction unit estimates that a seizure will not occur within a specified time or period after the time indicated by the log information.

4. The seizure prediction program of claim 3, wherein the prediction unit inputs the behavioral information or emotional information for the day before the specified date into the trained model, and, based on the output of the trained model, predicts that a seizure will not occur on the specified date.

5. The seizure prediction program of claim 1, wherein the prediction unit inputs a seizure interval indicating the period elapsed since the previous seizure into the trained model based on the timing of the seizure information recorded in the past immediately prior to the specified day, and makes the prediction by using the output of the trained model to estimate whether a seizure will occur on the specified day.

6. The seizure prediction program of claim 1, wherein the trained model is generated by machine learning using a dataset including the training data for multiple subjects with the same seizure classification as the subject and the training data for the subject.

7. The seizure prediction program according to claim 1, wherein the prediction unit predicts the subject's seizure based on the seizure classification of the subject in addition to the log information.

8. The seizure prediction program of claim 1, wherein the trained model is generated from training data that includes, in addition to the log information and information indicating whether the seizure has occurred, environmental information regarding the environment surrounding the subject, and the prediction unit predicts a seizure in the subject by inputting, in addition to the log information, the environmental information into the trained model and obtaining the estimation result as an output of the trained model.

9. The seizure prediction program of claim 1, wherein the trained model is generated from training data that includes biometric information of the subject in addition to the log information and information indicating whether the seizure has occurred, and the prediction unit predicts a seizure in the subject by inputting the biometric information in addition to the log information into the trained model and obtaining the estimation result as an output of the trained model.

10. The seizure prediction program of claim 1, wherein the acquisition unit acquires the behavioral information as the log information, and the behavioral information includes sleep information regarding the subject's sleep.

11. The seizure prediction program of claim 1, wherein the acquisition unit acquires the behavioral information as the log information, and the behavioral information includes usage information acquired from background data on a terminal used by the subject.

12. The seizure prediction program of claim 1, wherein the acquisition unit acquires the behavioral information as the log information, and the behavioral information includes movement information regarding the subject's movement speed.

13. The seizure prediction program of claim 1, wherein the acquisition unit further acquires evaluation information based on the results of an examination or ability test of the subject, and the prediction unit inputs the evaluation information, in addition to the log information, into the trained model, and predicts the subject's seizures based on the output of the trained model.

14. A storage medium storing the seizure prediction program according to any one of claims 1 to 13.

15. A seizure prediction device comprising an acquisition unit and a prediction unit, wherein the acquisition unit acquires log information including at least one of behavioral information indicating a subject's behavior and the timing of those actions, emotional information indicating the emotions felt by the subject and the timing of those emotions, which is input based on the subject's subjective opinion, and seizure information indicating seizures that have occurred in the subject and the timing of those seizures, and the prediction unit inputs the log information into a trained model generated by machine learning and obtains, as an output of the trained model, an estimated result of whether the subject will have a seizure within a specified time or period after the time indicated in the log information, thereby predicting a seizure associated with the subject's brain activity or mental activity, and the trained model is generated from training data including the subject's log information and information indicating whether the subject has had a seizure within a specified time or period after the time indicated in the log information.

16. A seizure prediction method comprising: a computer having an acquisition unit and a prediction unit, wherein the acquisition unit acquires log information including at least one of behavioral information indicating a subject's behavior and the timing of those actions; emotional information indicating the emotions felt by the subject, which is input based on the subject's subjective opinion, and the timing of those emotions; and seizure information indicating seizures that have occurred in the subject and the timing of those seizures; and the prediction unit inputs the log information into a trained model generated by machine learning and obtains, as an output of the trained model, an estimated result of whether the subject will have a seizure within a specified time or period after the time indicated in the log information, thereby predicting a seizure associated with the subject's brain activity or mental activity; and the trained model is generated from training data including the subject's log information and information indicating whether the subject has had a seizure within a specified time or period after the time indicated in the log information.

Images