Behavioral restriction estimation system

Abstract

To provide an action restriction estimation system capable of estimating feasibility of a variety of restrictions on a patient in the future.SOLUTION: An action restriction estimation system includes: an MRI 1 photographing a plurality of tomograms of the brain of a patient P; tomogram storage means 2 storing the plurality of tomograms of the patient P obtained by the MRI 1; patient brain image transmission means 3 capable of configuring a patient three-dimensional brain image on the basis of the plurality of stored tomograms and transmitting them via an Internet line; a neural network 4 which receives the transmitted patient's three-dimensional brain images and outputs a prediction result about a possibility for the patient P to experience action restriction in the future; and action restriction notification means 5 which notifies the possibility for the patient P to experience the action restriction in the future in accordance with the outputted prediction result.SELECTED DRAWING: Figure 1

Description

【Technical Field】 【0001】 The present invention relates to a behavior restriction estimation system for estimating the tendency of future patient behavior restrictions (the possibility of implementing various restraints on patients) by using tomographic images of the brain obtained by MRI (Magnetic Resonance Imaging). 【Background Art】 【0002】 Conventionally, physical restraints (behavior restrictions focusing on protecting the life of the patient and preventing serious physical injuries) have been imposed on patients suspected of having dementia or the like. Examples of physical restraints include the following items depending on the purpose. (1) To prevent wandering, tie the trunk and limbs to a wheelchair or a bed with a rope or the like. (2) To prevent falling, tie the trunk and limbs to a bed with a rope or the like. (3) To prevent the patient from getting off by themselves, surround the bed with a fence (side rail). (4) To prevent the removal of tubes for intravenous drip, enteral nutrition, etc., tie the limbs with a rope or the like. (5) To prevent the removal of tubes for intravenous drip, enteral nutrition, etc., or to prevent scratching the skin, put on a mitt-type glove or the like that restricts finger function. (6) To prevent slipping off a wheelchair or a bed or standing up on a wheelchair or a bed, attach a Y-shaped restraint belt, a waist belt, or a wheelchair table. (7) Use a chair that hinders the standing up of a person who can stand up. (8) Put on a nursing gown (connected clothes) to restrict undressing and diaper changing. (9) To prevent nuisance behavior to others, tie the trunk and limbs to a bed or the like with a rope or the like. (10) To calm the behavior, make the patient take an excessive amount of psychotropic drugs. (11) Isolate the patient in a room or the like that cannot be opened at will by the patient's own will. 【0003】 However, the Ministry of Health, Labour and Welfare ordinance states that "except in urgent and unavoidable circumstances, physical restraint or other acts that restrict the behavior of residents shall not be performed in order to protect the life or physical safety of the resident or other residents." The "Guidelines for Zero Physical Restraint" compiled by the Ministry of Health, Labour and Welfare's Council for the Promotion of Zero Physical Restraints outlines the following three principles for urgent and unavoidable circumstances. (Principle 1) Urgency: There is a very high possibility that the life or physical safety of the user or other users will be endangered. (Principle 2) Non-substitutability: There must be no alternative care method other than physical restraint or other behavioral restrictions. (Principle 3) Temporariness: Physical restraint and other restrictions on behavior must be temporary. In response to this, local governments and hospitals are taking steps to eliminate physical restraints, but the number of people using physical restraints has continued to increase from just over 5,000 in fiscal year 2003 to just under 11,000 in fiscal year 2014, and has remained high since then (approximately 11,000 in fiscal year 2019). 【0004】 Incidentally, Patent Document 1 (Japanese Patent Publication No. 5641629) describes a method of using machine learning techniques to infer an individual's characteristics (IQ, attention, memory, social skills, illness, personality, values, and aptitudes) based on brain images scanned with an MRI or the like (see paragraphs 0045-0047 in particular). Furthermore, Patent Document 2 (JP 2018-531648) describes a machine learning system that classifies input 3D radiation volume of the human brain as anomalies, including the identification of bleeding, evidence of possible Alzheimer's disease, or signs of seizure, and also provides a confidence level for the classification (see paragraph 0048 in particular). Furthermore, Patent Document 3 (Japanese Patent Publication No. 6483890) describes a method for predicting whether or not a subject will develop Alzheimer's disease within a predetermined period of time, according to a machine learning-based prediction algorithm (see, in particular, claim 1). However, the personal characteristics prediction system described in Patent Document 1 requires storing information regarding the correlation between characteristic values ​​of each part of the cerebrum and abilities or aptitudes (see Claim 1). Therefore, as described in paragraph 0042 and Figure 10, it is necessary to predetermine which characteristic values ​​of which parts of the brain have a high correlation with a particular ability or aptitude. Furthermore, the machine learning system described in Patent Document 2 classifies the identification of bleeding, evidence of possible Alzheimer's disease, or signs of seizures, and the diagnostic support device described in Patent Document 3 predicts whether or not a subject will develop Alzheimer's disease within a predetermined period. Both are systems that support diagnoses made by doctors and other medical professionals. [Prior art documents] [Patent Documents] 【0005】 [Patent Document 1] Japanese Patent Publication No. 5641629 (Japanese Unexamined Patent Publication No. 2015-84970) [Patent Document 2] Special Publication No. 2018-531648 (Patent No. 6450053) [Patent Document 3] Japanese Patent Publication No. 6483890 (Japanese Unexamined Patent Publication No. 2019-187966) [Overview of the Initiative] [Problems that the invention aims to solve] 【0006】 As mentioned above, the number of people physically restrained doubled in the 10 years from fiscal year 2003 and has remained high since then, indicating that efforts toward eliminating physical restraint are proving difficult. To overcome this situation, it is necessary to predict the likelihood of each patient experiencing various behavioral restrictions and to take measures in advance to prevent such restrictions from occurring. However, the systems described in Patent Documents 1-3 do not predict future behavioral restrictions and therefore cannot be used as a measure to achieve zero physical restraint. In view of the above situation, the problem to be solved by the present invention is to provide a behavior restriction estimation system that can estimate a tendency for future patient behavior restrictions to be implemented. 【Means for Solving the Problem】 【0007】 The behavior restriction estimation system of the invention according to claim 1 for solving the above problem is a tomographic image storage means for storing a plurality of tomographic images of a patient's brain obtained by MRI, Suspected dementia a patient brain image transmission means for transmitting a patient brain image based on the plurality of tomographic images stored in the tomographic image storage means, a neural network that receives the patient brain image transmitted from the patient brain image transmission means and outputs a prediction result regarding the possibility that the patient will be subject to behavior restrictions in the future, 3D and a behavior restriction notification means for notifying regarding the possibility that the patient will be subject to behavior restrictions according to the prediction result, A learning dataset storage means stores a learning dataset that links a 3D brain image of each subject, constructed using the same method as the patient brain image transmission means, with the behavioral restriction history data of each subject, based on multiple subject tomographic images of each subject's brain obtained by MRI from a large number of subjects suspected of having dementia. wherein the neural network is optimized by being obtained and trained in advance with a large number of learning data sets. 3D when From the time of reception in the future any point in time The learning dataset stored in the aforementioned learning dataset storage means a large number of learning data sets is given are obtained and optimized by being trained. Occasionally, The aforementioned behavioral restriction history data includes, for each subject, at least the date on which the need for behavioral restriction was determined, and information regarding whether or not the behavioral restriction was necessary on that day. The aforementioned restrictions on movement include at least five-point restraint or strong restraint on the bed with the installation of bed rails, weak restraint on the bed other than the strong restraint on the bed, and restraint on the wheelchair. The aforementioned behavioral restriction notification means provides notification regarding the possibility that the patient may be subjected to strong restraint in the bed, weak restraint in the bed, and restraint in the wheelchair. It is characterized by this. 【0008】 The invention according to claim 2 for solving the above problem is, in the behavior restriction estimation system of the invention according to claim 1, It is equipped with a time input means for specifying any time in the future from the time of reception. When the neural network receives the patient's 3D brain image transmitted from the patient brain image transmission means and the specified time specified by the time input means, it outputs a prediction result regarding the likelihood that the patient will experience behavioral restrictions during the specified time. The aforementioned learning dataset storage means continuously stores learning datasets that link the subject's 3D brain image with each subject's behavioral restriction history data. It is characterized by this. 【Effect of the Invention】 【0010】 Claim​​​​​ 1 The behavioral restriction estimation system of the invention obtained by MRI Suspected dementia Patients based on multiple tomographic images of the patient's brain 3D Receiving brain images when , From the time of reception future any point in time In the case of patients Physical restraint (small) Even without Severe restraint on a bed with five-point restraint or bed rail installation, mild restraint on a bed other than severe restraint on a bed, and on a wheelchair A neural network that outputs prediction results regarding the likelihood of being subjected to (including constraints), and, depending on the prediction results, at least Patient Severe restraint in bed, light restraint in bed, and restraint in wheelchair It is equipped with a means of notifying about the possibility of being subjected to behavioral restrictions, A neural network is pre-programmed with a large number of Suspected dementia Based on multiple tomographic images of each subject's brain obtained by MRI from the subject Furthermore, it is configured by the same method as the patient brain image transmission means. Subject 3D Brain imaging and behavioral restriction history data for each subject. linked Since a large number of training datasets are provided in advance and optimized through training, From the time of reception future any point in time This allows us to estimate the feasibility of implementing various restraints on patients in a given situation, and the likelihood of the patient being subjected to certain restraints. The bundle To prevent it from happening Fine texture Measures can be implemented in advance. In addition to the effects achieved by the invention of claim 1, the behavioral restraint estimation system of claim 2 can output a prediction result indicating whether or not patient P is likely to be physically restrained at any future time specified. [Brief explanation of the drawing] 【0012】 [Figure 1] Conceptual diagram of the behavioral restriction estimation system according to the embodiment. [Figure 2] A block diagram showing the configuration of neural network 4. [Figure 3] The learning curve for neural network 4 during training. [Figure 4] A table showing a comparison between predicted results regarding behavioral restrictions and actual implementation information regarding those restrictions. [Modes for carrying out the invention] 【0013】 Embodiments of the present invention will be described below with reference to examples. [Examples] 【0014】 Figure 1 is a conceptual diagram of a behavioral restriction estimation system according to an embodiment. As shown in Figure 1, the behavioral restriction estimation system according to the embodiment includes an MRI 1 that takes multiple tomographic images of the brain of a patient P suspected of having dementia, a tomographic image storage means 2 that stores the multiple tomographic images of patient P obtained by the MRI 1, a patient brain image transmission means 3 that can construct a three-dimensional brain image of patient P (hereinafter referred to as "patient 3D brain image") based on the stored multiple tomographic images and transmit the patient 3D brain image via an internet connection, a neural network 4 that receives the transmitted patient 3D brain image and outputs a prediction result regarding the possibility that patient P may be subjected to behavioral restrictions in the future, and a behavioral restriction notification means 5 that notifies patient P about the possibility of being subjected to behavioral restrictions in the future according to the outputted prediction result. Furthermore, the tomographic image storage means 2, the patient brain image transmission means 3, and the behavioral restriction notification means 5 are installed on the hospital PC 6. The behavioral restriction notification means 5 displays information predicted by the neural network 4 regarding the likelihood of patient P receiving behavioral restrictions in the future. Therefore, physician D can accurately formulate a treatment plan for patient P by referring to this information. Furthermore, the behavioral restriction notification means 5 notifies patient P that there is a possibility of being physically restrained in the future if the output of the neural network 4 is "there is a possibility of being physically restrained in the future," and notifies patient P that there is no possibility of being physically restrained in the future if the output is "there is no possibility of being physically restrained in the future." Methods of notification include displaying on the screen using text such as "Yes" or "No," symbols such as "×" or "○," or using colors such as "making part or all of the screen red" or "making part or all of the screen blue," and audio notification is also possible. 【0015】 Prior to taking multiple tomographic images of patient P's brain, multiple tomographic images of the brains of numerous subjects suspected of having dementia were taken using MRI 1 over a period of several years to about 10 years. Using the same method as the patient brain image transmission means 3, a three-dimensional brain image (hereinafter referred to as "subject three-dimensional brain image") was constructed based on the multiple subject tomographic images, and behavioral restriction history data was created for each subject. A learning dataset linking each subject's subject three-dimensional brain image and behavioral restriction history data is stored in the learning dataset storage means 7. The neural network 4 is then provided with the training datasets of each subject that have been previously stored in the training dataset storage means 7. By training the neural network 4, it is optimized and the AI ​​parameters are set. In other words, without the need for cumbersome processing such as storing information on the correlation between characteristic values ​​of each part of the cerebrum and abilities or aptitudes, as in the personal characteristic prediction system of Patent Document 1, data generated by tasks performed in normal examinations and medical treatments, such as constructing a 3D brain image of the subject and creating data on the subject's behavioral restriction history, can be linked together to form a learning dataset. By simply feeding the neural network 4 with the learning datasets of many subjects and allowing it to learn, information on the likelihood that patient P will experience behavioral restrictions in the future can be obtained by inputting a 3D brain image of a new patient P into the neural network 4. The neural network 4 and the training dataset storage means 7 are located on an analysis server 8 that is connected to the hospital PC 6 via an internet connection or the like. 【0016】 As shown in Figure 2, the neural network 4 is a convolutional neural network constructed with an input layer 4I, an output layer 4O, a sample data storage means 4L, a sample data input means 4T that inputs sample data consisting of numerous training datasets from the training dataset storage means 7 to the sample data storage means 4L, and multiple convolutional layers, pooling layers, BN layers (Batch-normalization layers), dropout layers, fully connected layers, and others (not shown). Generally, a unit is considered to be a combination of 2-3 convolutional layers and 1 pooling layer, and several units are often connected in series, with 2-4 fully connected layers added to them. Although BN layers and dropout layers are not essential, their inclusion can be expected to suppress overfitting of the entire neural network. In this example, as the neural network 4, we used an improved version of ResNet152 (abbreviated as "152-layer residual Network"), which has high predictive performance in image recognition tasks, adapted for use with 3D data. 【0017】 The input layer 4I is supplied with a patient's three-dimensional brain image from the patient brain image transmission means 3 described above. The output layer 4O outputs a prediction result regarding the likelihood that patient P may be subject to behavioral restrictions in the future, and this prediction result is provided to the behavioral restriction notification means 5 described above. The neural network 4 is optimized by training it with a large amount of sample data (typically hundreds to thousands of 3D brain images and behavioral restriction history data from subjects) stored in the sample data storage means 4L. By optimizing the neural network 4 with a sufficient number of sample data, a prediction accuracy of over 80% can be achieved. 【0018】 In this embodiment, patient 3D brain images and subject 3D brain images are generated using data acquired using the same imaging method as VSRAD®, image processing and statistical analysis software used to read the degree of atrophy near the parahippocampal gyrus, which is characteristic of early Alzheimer's disease, from MRI images. Specifically, these are 256×256×110 dot 3D images covering the entire brain, generated according to a medical data standard called DICOM, and consist of 2 bytes of grayscale data. In this embodiment, for use as a training dataset, the 256×256×110 dot 3D images are reduced and converted to 128×128×110 dot 3D images, and data augmentation such as translation and rotation in the vertical, horizontal, and depth directions is performed to supplement the less data, and the grayscale information is normalized from 0 to 1. Furthermore, regarding information recorded along with 3D brain images that could identify a patient or subject, in accordance with the Act on the Protection of Personal Information and the Guidelines for Anonymization Technology in the Utilization of Medical Information, a patient ID is used instead of the patient's name, and a subject ID is used instead of the subject's name. Corresponding to this patient ID or subject ID, at least the date of image acquisition and the gender and age of the patient or subject are recorded. 【0019】 The subject's behavioral restriction history data used in the training dataset includes a subject ID to link with the subject's 3D brain image. Corresponding to each subject ID, it records at least the date on which the need for physical restraint was determined, and information regarding whether or not physical restraint was necessary on that day. In cases where physical restraint was deemed necessary, the items (1) to (9) above, which are examples of physical restraint, are divided into three categories (implementation of strong restraint on the bed with five-point restraint or bed rail installation, implementation of weak restraint on the bed other than strong restraint, and implementation of restraint on the wheelchair using safety belts or waist belts, etc.), and information regarding which type of physical restraint was necessary is recorded. 【0020】 In this embodiment, 366 training datasets were prepared by linking the subject's 3D brain image with the subject's behavioral restriction history data. Of these, 268 were used as sample data stored in the sample data storage means 4L, 68 were used as test data necessary for optimizing the neural network 4, and 30 were used as evaluation data to verify the prediction results by the behavioral restriction estimation system. Figure 3 is a graph showing the learning curve (evolution of estimation error) during the training performed to optimize neural network 4. The vertical axis represents the value of "val_loss," an index for evaluating estimation error, and the horizontal axis represents the number of training iterations. Note that the vertical axis is displayed on a logarithmic scale because the value changes significantly in the early stages of training. Training was performed 2000 times, and the training was terminated when the estimation error stopped decreasing due to the effects of training (when the accuracy was maximized). In this embodiment, the estimation error did not decrease from the 177th to the 277th iteration. Therefore, in Figure 3, the maximum value on the horizontal axis (number of iterations) is 277. In this embodiment, the AI ​​parameters at the 177th iteration, which resulted in the smallest estimation error, were used. 【0021】 Figure 4 is a table showing a comparison between the predicted results regarding behavioral restrictions and the actual behavioral restriction implementation information, obtained by inputting 3D brain images of subjects from 30 evaluation data points into the neural network 4 optimized in this manner. Here, the outputted prediction results regarding behavioral restrictions are predictions (presence or absence of possibility) regarding the likelihood of the subject receiving behavioral restrictions (physical restraint) in the future, linked to the input subject's 3D brain image. A prediction result of "1" means that physical restraint will be applied, and a prediction result of "0" means that physical restraint will not be applied. In addition, the actual behavioral restriction implementation information is information regarding whether physical restraint was required or not, recorded in the subject's behavioral restriction history data. A "1" means that physical restraint is recorded as required, and a "0" means that physical restraint is recorded as not required. As can be seen in Figure 4, there were 19 prediction results of "1," of which 17 were correct, resulting in an accuracy rate of 89.5% when predicting physical restraint. On the other hand, there were 11 prediction results of "0," of which 9 were correct, resulting in an accuracy rate of 81.8% when predicting no physical restraint. The breakdown of whether or not physical restraints were used shows a bias towards those with physical restraints (63.3%) and those without (36.7%), but since the accuracy rate for both exceeds 80%, it can be said that the neural network 4 is sufficiently capable of predicting behavioral restrictions. In other words, by simply analyzing MRI images of the brain of patient P suspected of having dementia, it is possible to predict with a certain degree of accuracy whether or not behavioral restrictions will be imposed on patient P in the future. Therefore, the behavioral restriction estimation system of the present invention can be evaluated as a system that is sufficiently useful when physician D considers other conditions of patient P and takes measures to prevent patient P from being subjected to behavioral restrictions in the future, or when deciding whether or not to impose behavioral restrictions on patient P now. 【0022】 The following are examples of modifications of the embodiments. (1) In the embodiment, the patient brain image transmission means 3 constructs a patient 3D brain image based on multiple tomographic images of patient P and transmits the patient 3D brain image to the neural network 4. However, if the neural network 4 is pre-provided with a large number of training datasets consisting of multiple tomographic images of each subject and behavioral restriction history data, and is optimized by training, the patient brain image transmission means 3 only needs to send the multiple tomographic images of patient P directly to the neural network. If the neural network 4 is pre-provided with a large number of training datasets consisting of multiple 2D brain images of each subject constructed based on multiple tomographic images of each subject and behavioral restriction history data, and is optimized by training, the patient brain image transmission means 3 only needs to send multiple 2D brain images of patient P constructed based on multiple tomographic images to the neural network. Therefore, in the claims, the terms "patient brain image" and "subject brain image" are used instead of "patient 3D brain image" and "subject 3D brain image," respectively. 【0023】 (2) In the embodiment, the prediction result regarding behavioral restrictions output from the neural network 4 was whether or not there was a possibility of physical restraint. However, as shown in Figure 1, the magnitude of the possibility of physical restraint (restraint occurrence probability) can also be used. Furthermore, while the embodiment only outputted a prediction of whether or not patient P is likely to be subjected to physical restraint in the future, the behavioral restraint history data includes, in correspondence with each subject ID, at least the date on which the need for physical restraint was determined and information regarding the necessity or non-necessity of physical restraint on that day. Therefore, if patient 3D brain images and behavioral restraint history data are continuously created for each subject over a long period and stored in the learning dataset storage means 7, it is also possible to improve the system to output a prediction of whether or not patient P is likely to be subjected to physical restraint at any future time specified. 【0024】 (3) In the embodiment, because the sample data stored in the sample data storage means 4L was small, even though information was recorded on which of the three categories of physical restraints the subject needed as behavioral restriction history data, it was not possible to output a prediction result regarding the likelihood of receiving physical restraint. However, if more than 1000 sample data and more than 200 test data can be prepared, it will be possible to output a prediction result regarding the likelihood of the patient receiving any of the following types of physical restraint in the future: strong restraint in bed, weak restraint in bed, or restraint in a wheelchair, while achieving a sufficient accuracy rate. Furthermore, if we can prepare thousands of sample data points and over 500 test data points, and if, in addition to the three categories adopted in the example, we select at least one of the following as examples of physical restraint: the implementation of items (10) and (11) above, communication restrictions, visitation restrictions, restrictions on movement within the hospital, and restrictions on going out, and record information on which of the selected behavioral restrictions is necessary, we can output predictive results regarding the likelihood of receiving a wider variety of behavioral restrictions while achieving a sufficient accuracy rate. Furthermore, it is expected that accuracy will improve not only by increasing the number of sample and test data points and the number of training iterations, but also by applying Transformer (Attention), a new artificial intelligence (AI) technology. [Explanation of symbols] 【0025】 1. MRI 2. Tomographic image storage means 3. Patient brain image transmission means 4. Neural Network 4I. Input Layer 4L. Sample Data Storage Method 4O Output layer 4T Sample data input means 5 Action restriction notification means 6. In-hospital PC 7. Training dataset storage method 8. Analysis server D: Doctor P: Patient suspected of having dementia

Claims

[Claim 1] A tomographic image memory means for storing multiple tomographic images of the brain of a patient suspected of having dementia, obtained by MRI, A patient brain image transmission means that transmits a patient's three-dimensional brain image based on the plurality of tomographic images stored in the tomographic image storage means, A learning dataset storage means stores a learning dataset that links a three-dimensional brain image of each subject, constructed using the same method as the patient brain image transmission means, with the behavioral restriction history data of each subject, based on multiple subject tomographic images of each subject's brain obtained by MRI from a large number of subjects suspected of having dementia. A neural network that, upon receiving the patient's three-dimensional brain image transmitted from the patient brain image transmission means, outputs a prediction result regarding the possibility that the patient will experience behavioral restrictions at some point in the future from the time of reception, The system includes a behavioral restriction notification means that provides notification regarding the possibility of the patient being subject to behavioral restrictions based on the prediction results, The neural network is optimized by being trained using a large number of training datasets that have been previously stored in the training dataset storage means. The aforementioned behavioral restriction history data includes, for each subject, at least the date on which the need for behavioral restriction was determined, and information regarding whether or not the behavioral restriction was necessary on that day. The aforementioned restrictions on movement include at least five-point restraint or strong restraint on the bed with the installation of bed rails, weak restraint on the bed other than the strong restraint on the bed, and restraint on the wheelchair. The aforementioned behavioral restriction notification means provides notification regarding the possibility that the patient may be subjected to strong restraint in the bed, weak restraint in the bed, and restraint in the wheelchair. A behavioral restriction estimation system characterized by the following: [Claim 2] comprising a time input means for specifying any future time, When the neural network receives the patient's three-dimensional brain image transmitted from the patient brain image transmission means and the specified time specified by the time input means, it outputs a prediction result regarding the likelihood that the patient will experience behavioral restrictions during the specified time. The learning dataset storage means continuously stores learning datasets that link the subject's 3D brain image with each subject's behavioral restriction history data. The behavioral restriction estimation system described in feature 1.

Images