Medical information processing device, method, and program

Abstract

To realize incentive compatibility in allocation of clinical decisions.SOLUTION: A medical information processing device is provided, comprising an update unit. The update unit updates an individual model, associated with a target subject, for computing an effect evaluation value of a clinical decision on the basis of first data corresponding to an effect of the clinical decision on the target subject and data limited to a partial range of second data corresponding to an effect of the clinical decision on another subject different from the target subject.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] Embodiments disclosed in the present specification and drawings relate to a medical information processing apparatus, method, and program.

Background Art

[0002] The act of assigning a specific medical decision to a patient from a plurality of medical decisions is performed in clinical research and daily medical practice. Methods for adaptively changing assignments include adaptive designs, Bandit algorithms, and the like. To maximize the benefit of all patients as a whole, it is necessary to deliberately select (explore) medical decisions that are unlikely to be optimal at an appropriate frequency, but from the perspective of each individual patient, there is usually no incentive to make such a selection. It is conceivable to provide patients with financial incentives to induce them to perform exploration, but bias may occur in the patients who choose to perform exploration. As described above, in the assignment of medical decisions, it is difficult to achieve incentive compatibility where the well-being of individual patients aligns with the well-being of all patients as a whole.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problem to be Solved by the Invention

[0004] One of the problems to be solved by the embodiments disclosed in the present specification and drawings is to achieve incentive compatibility in the assignment of medical decisions. However, the problems to be solved by the embodiments disclosed in the present specification and drawings are not limited to the above-mentioned problems. Problems corresponding to the respective effects of the respective configurations shown in the embodiments described later can also be positioned as other problems.

Means for Solving the Problem

[0005] The medical information processing apparatus according to the embodiment includes an update unit. The update unit updates an individual model, which is a model related to the target subject and calculates an effect evaluation value of clinical decision-making, based on first data corresponding to the effect of clinical decision-making for the target subject, and data limited to a partial range of second data corresponding to the effect of clinical decision-making for another subject different from the target subject. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] [Figure 1] FIG. 1 is a diagram illustrating a configuration example of the medical information processing apparatus according to the present embodiment. [Figure 2] FIG. 2 is a diagram illustrating the flow of medical information processing performed by a processing circuit in accordance with a medical information processing program. [Figure 3] FIG. 3 is a diagram schematically illustrating the medical information processing shown in FIG. 2. [Figure 4] FIG. 4 is a diagram illustrating an example of a use range according to Example 1. [Figure 5] FIG. 5 is a diagram illustrating an example of a use range according to Example 2. [Figure 6] FIG. 6 is a diagram illustrating an example of a use range according to Example 3. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0007] Hereinafter, embodiments of a medical information processing apparatus, method, and program will be described in detail with reference to the drawings.

[0008] In the medical field, efforts are being made to realize Clinical Decision Support (CDS) using data and Artificial Intelligence (AI) technology. As for AI technology, supervised learning represented by deep learning is widely used, and especially in the field of diagnostic imaging, some systems have emerged that can detect lesions with accuracy exceeding that of humans.

[0009] However, the following issues may arise in realizing medical decision support through supervised learning: 1. A large amount of data is required to improve accuracy, and the cost of data collection is high. 2. Even if accuracy is high, it may not lead to appropriate medical decision-making.

[0010] Problem 2 is technically equivalent to the problem that "supervised learning can reveal correlations, but it cannot reveal causal relationships." For example, even if there was an AI model that could predict with 100% accuracy that a heart failure patient would die within five years, this AI model would be useless to a doctor who wanted to know how to extend the patient's life expectancy beyond five years.

[0011] Clarifying the causal relationships of medical procedures and identifying the optimal medical procedures is not currently achievable with supervised learning; this is the role of randomized controlled trials (RCTs) (or medical statistics). In other words, true "support for medical decision-making to optimize treatment" does not exist as an extension of current AI technology, and randomized controlled trials will not disappear.

[0012] However, randomized controlled trials also have many problems, as listed below: 1. They take a long time to complete, making it difficult to keep up with the latest trends, and results tend to become outdated. 2. They are inefficient because they are extremely costly and carry the risk of hypothesis testing failure. 3. Because eligibility criteria for studies tend to be set strictly, results may not apply to the general patient population (low external validity). 4. Only the average intervention effect can be compared, and it is not possible to identify effects that differ from patient to patient. For example, cases such as personalized medicine cannot be considered. 5. Excessive reliance on statistical significance makes publication bias and p-value hacking more likely.

[0013] In particular, with the rapid changes in treatment methods and the increasing importance of personalized medicine, the limitations of randomized controlled trials are becoming apparent.

[0014] Therefore, there is a need to efficiently optimize clinical decision-making without requiring large amounts of prior data or conducting randomized controlled trials. Furthermore, there is a need to develop clinical decision-making support models for this purpose.

[0015] Figure 1 shows an example configuration of the medical information processing device 1 according to this embodiment. As shown in Figure 1, the medical information processing device 1 is an information processing terminal such as a computer having a processing circuit 11, a storage device 12, an input device 13, a communication device 14, and a display device 15. The processing circuit 11, storage device 12, input device 13, communication device 14, and display device 15 are connected to each other via a bus so that signals can be input and output from each other.

[0016] The processing circuit 11 includes processors such as a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit). By executing a medical information processing program, the processing circuit 11 realizes functions such as allocation 111, observation 112, storage 113, decision 114, acquisition 115, update 116, and display control 117. Note that each of the functions 111 to 117 is not limited to being realized by a single processing circuit. A processing circuit may be configured by combining multiple independent processors, and each processor may realize each of the functions 111 to 117 by executing a program. Furthermore, each of the functions 111 to 117 may be a modularized program that constitutes the medical information processing program. These programs are stored in the storage device 12.

[0017] The storage device 12 includes ROM (Read Only Memory) and RAM (Random Memory) for storing various types of information. These include Access Memory (CHD), HDD (Hard Disk Drive), SSD (Solid State Drive), and integrated circuit memory devices. In addition to the above-mentioned memory devices, the storage device 12 may also be a drive device that reads and writes various information to portable recording media such as CDs (Compact Discs), DVDs (Digital Versatile Discs), and flash memory, or to semiconductor memory elements. Furthermore, the storage device 12 may be located in another computer connected via a network.

[0018] The input device 13 receives various input operations from the operator, converts the received input operations into electrical signals, and outputs them to the processing circuit 11. Specifically, the input device 13 is connected to input devices such as a mouse, keyboard, trackball, switch, button, joystick, touchpad, and touch panel display. The input device 13 outputs electrical signals to the processing circuit 11 corresponding to the input operations to the input device. The input device 13 may also be an input device provided on another computer connected via a network or the like.

[0019] The communication device 14 is an interface for sending and receiving various types of information with other computers. Information communication via the communication device 14 is carried out in accordance with standards suitable for medical information communication, such as DICOM (Digital Imaging and Communications in Medicine).

[0020] The display device 15 displays various information via the display control function 117 of the processing circuit 11. As the display device 15, for example, a liquid crystal display (LCD), a CRT (Cathode Ray Tube) display, an organic electroluminescent display (OELD), a plasma display, or any other display can be used as appropriate. Alternatively, a projector may be used as the display device 15.

[0021] In this embodiment, the processing circuit 11, through the implementation of the assignment function 111, determines which medical decisions to assign to subjects such as patients based on the effectiveness evaluation value of the medical decisions. The effectiveness evaluation value of the medical decisions is calculated using an individual model for calculating the effectiveness evaluation value of the medical decisions. The medical procedure corresponding to the assigned medical decision is performed on the subject. An effect resulting from the medical procedure occurs in the subject. The processing circuit 11, through the implementation of the observation function 112, observes the effect that occurred in the subject. The effect is expressed as a numerical value. This will be called the effect observation value. The processing circuit 11, through the implementation of the storage function 113, stores the effect observation value in a database.

[0022] The processing circuit 11 updates an individual model that calculates an evaluation value of the effectiveness of a medical judgment for a target subject, based on first observational data corresponding to the effect of a medical judgment on the target subject and observational data limited to a certain range of second observational data corresponding to the effect of a medical judgment on a subject other than the target subject, by implementing the update function 116. The target subject is a subject that is the subject of training for the individual model. The target subject and other subjects may be set to an individual subject, a group of subjects including one or more subjects, a hospital, or a group of hospitals including one or more hospitals. The first observational data is a sequence of data that includes at least multiple effect observation values ​​corresponding to multiple observation times of the effect of a medical judgment on the target subject. The second observational data is a sequence of data that includes multiple effect observation values ​​corresponding to multiple observation times of the effect of a medical judgment on the other subjects. Medical procedures and other features may be associated with the effect observation value at each observation time. The first observational data and the second observational data are stored in a database. The processing circuit 11 determines the said range by implementing the decision function 114. The processing circuit 11 acquires data limited to a specific range of the second set of observation data by implementing the acquisition function 115. The processing circuit 11 also displays various data on the display device 15 by implementing the display control function 117.

[0023] The subjects in this embodiment are typically a single patient with a disease. However, the subjects in this embodiment are not limited to this and may be a patient group consisting of two or more patients. Furthermore, the subjects in this embodiment do not necessarily have to be individuals with a disease; they may be healthy individuals. The models may be used for each physician, each medical department, each hospital, each region, or for each subject.

[0024] The model according to this embodiment is a function for calculating an effectiveness evaluation value for medical treatment decisions. This function may be a function determined manually, a function determined by experimental or deterministic methods, a function trained by machine learning, or a function determined by any other arbitrary method. The effectiveness evaluation value is an index value that evaluates the effectiveness of the medical treatment decisions of the allocation candidates. For example, index values ​​such as the confidence interval of the effect, the expected value of the effect, the maximum value of the expected value, and the difference from the maximum value of the expected value may be used. The individual model according to this embodiment means a model that is specialized for the target subject.

[0025] The types of medical decisions according to this embodiment include decisions regarding whether or not to perform a procedure, decisions regarding the content of the procedure, decisions regarding the quantity of the procedure, and decisions regarding the timing of the procedure. Examples of decisions regarding whether or not to perform a procedure include decisions regarding whether or not to perform surgery or whether or not to perform blood tests. Examples of decisions regarding the content include decisions regarding the selection of therapeutic drugs to be used or the selection of disease names to be diagnosed. Examples of decisions regarding the quantity include decisions regarding the selection of drug dosages or the selection of rehabilitation duration. Examples of decisions regarding the timing include decisions regarding the selection of the timing of hospital visits or the selection of the timing of surgery.

[0026] Those who make medical decisions include not only doctors, nurses, and other healthcare professionals, but also the subjects themselves, their families, and anyone related to the subjects. Medical decisions do not necessarily have to be highly medically sophisticated, nor do they have to be decisions made for the improvement of health. In other words, decisions that have a negative impact on medical care or health are also within the scope of medical decisions. For example, the decision of whether or not a healthy person should smoke at a certain time is also subject to medical decision-making.

[0027] A certain effect occurs as a result of a medical judgment. This effect is also called a reward or outcome. In this embodiment, the effects are expected to be, for example, clinical outcomes, patient-reported outcomes, and economic outcomes. Examples of clinical outcomes include morbidity (including whether or not the patient is ill), 5-year survival rate (including whether or not the patient is alive), complication rate (including whether or not complications occur), readmission rate (including whether or not the patient is readmitted), laboratory values ​​(or degree of improvement in laboratory values), and independence in daily living. Examples of patient-reported outcomes include subjective symptoms, subjective health status, satisfaction with treatment, and subjective well-being. Examples of economic outcomes include medical expenses, medical resources invested, and length of hospital stay.

[0028] The effect is expressed numerically and can be ranked for learning purposes. A numerical value may be assigned to something that is not originally numerical. The effect may be observed immediately after a medical decision. For example, the effect of a medical decision on whether or not to send an exercise-encouraging message to a healthy person's smartphone could be measured by whether or not the person exercises within 5 minutes of receiving the message.

[0029] The effectiveness may also take into account the cost required for making a medical decision. For example, in the smartphone message example above, sending a notification offers the advantage of making the effect more likely to be achieved than not sending one, but on the other hand, notifications also incur disadvantages such as communication costs and potentially hindering user behavior. To consider these advantages and disadvantages, for example, if the first message always incurs a cost of "5" and the exercise is rewarded with "100", then the effect can be calculated as reward "100" - cost "5" = effect "95". This makes it possible to obtain an effect that reflects cost-effectiveness.

[0030] The features according to this embodiment include the subject's attributes and / or state. Attributes are information that does not change based on the most recent medical assessment, such as the subject's gender and age. State is information that changes based on the most recent medical assessment, such as the subject's current blood pressure, blood pressure value, and blood glucose level.

[0031] The following describes an example of the operation of the medical information processing device 1 according to this embodiment.

[0032] Figure 2 is a diagram illustrating the flow of medical information processing performed by the processing circuit 11 according to the medical information processing program. Figure 3 is a schematic diagram showing the medical information processing shown in Figure 2.

[0033] As shown in Figures 2 and 3, the processing circuit 11, by realizing the assignment function 111, assigns medical decisions to the target subject using an individual model for that subject (step S1). Specifically, the assignment is performed through the following process.

[0034] The processing circuit 11 has individual models that calculate effectiveness evaluation values ​​for medical decisions, specifically tailored to the target subject. Each individual model has multiple structural models, each corresponding to a set of medical decisions. Each structural model is a model that calculates effectiveness evaluation values ​​for the corresponding medical decision. Basically, the processing circuit 11 calculates effectiveness evaluation values ​​by applying observational data of the target subject observed before the processing target time (processing target step) to the individual models. The effectiveness evaluation values ​​may be calculated based only on observational data observed at the time immediately preceding the processing target time, or they may be calculated based on observational data observed at multiple times prior to the processing target time.

[0035] The processing circuit 11 assigns an appropriate medical judgment from among multiple medical judgments based on multiple effect evaluation values, according to an adaptive design or bandit algorithm. The specific algorithm used for the bandit algorithm is not particularly limited. In this embodiment, the bandit algorithm refers to an algorithm that solves the problem of sequentially selecting an appropriate medical judgment from multiple medical judgments (options) in order to maximize the sum of effects. The bandit algorithm in this embodiment includes not only a narrow bandit algorithm in which the effect does not depend on the features, but also a contextualized bandit algorithm in which the effect depends on the features, and reinforcement learning that solves sequential decision-making problems in which the state changes according to previous medical judgments. As a specific algorithm for the bandit algorithm in this embodiment, for example, Epsilon Greedy, Thompson Sampling, Linear Thompson Sampling, PSRL (Posterior Sampling for Reinforcement Learning), BDQN (Bayesian Deep Q-Networks), etc. can be used.

[0036] The number of structural models included in the individual model in this embodiment is set according to the number of treatment decisions to be considered. The number of structural models is not particularly limited; any number of models, two or more, is acceptable. However, in order to specifically explain the following embodiment, we will assume that there are two structural models: one corresponding to treatment decision A and one corresponding to treatment decision B. In this case, the processing circuit 11 determines one appropriate treatment decision from the two treatment decisions A and B to be assigned.

[0037] In this embodiment, as an example, the parameters of individual models are updated using Thompson sampling, a type of bandit algorithm. Thompson sampling is a method that models the expected value parameter of an effect within a Bayesian statistical framework and applies a policy based on the probability matching method. Here, Bayesian statistics is a statistical theory that interprets that the probability can change each time new information is obtained, and the probability (or probability distribution) before and after the information is obtained is updated based on Bayes' theorem. The probability matching method is a method of selecting each option at each time step based on "the probability that the option has the maximum expected value," and is a method (probability policy) that introduces randomness into the options to be adopted. In the probability matching method, the method of formulating "the probability that the option has the maximum expected value" is arbitrary, but Thompson sampling is a method of calculating that probability from the posterior probability using a Bayesian approach.

[0038] In Thompson sampling, the update of the probability distribution is calculated assuming a conjugate prior distribution. The probability distribution of the effects on subjects resulting from a medical decision is assumed to follow a Bernoulli distribution. In this case, the beta distribution, which is the conjugate prior distribution of the Bernoulli distribution, is used as the posterior distribution of the expected value of the effect of the medical decision. The Bernoulli distribution is a discrete probability distribution that takes the value "1" at probability p and "0" at probability 1-p. For example, the posterior distribution of the expected value of the effect of medical decision A is given by parameter α A ,β A Using Beta(α A ,β A) is represented by Here, the parameter α A means the number of times the effect observation value "1" is observed as the effect of clinical decision A, and the parameter α B means the number of times the effect observation value "0" is observed. Similarly, the posterior distribution of the expected value of the effect of clinical decision B is defined by the parameter α B , β B using Beta(α B , β B ).

[0039] The effect evaluation value Y A is calculated according to the following equation (1) based on the Bernoulli distribution followed by the effect of clinical decision A and the parameter p that defines said Bernoulli distribution A . Similarly, the effect evaluation value Y B is calculated according to the following equation (2) based on the Bernoulli distribution followed by the effect of clinical decision B and the parameter p that defines said Bernoulli distribution B . Equation (1) is the mathematical expression of the structural model related to clinical decision A, and equation (2) is the mathematical expression of the structural model related to clinical decision B.

[0040] [Mathematical formula]

[0041] In Thompson sampling, the processing circuit 11 sequentially determines the current option (clinical decision A or clinical decision B) based on the series of clinical decision selections and effect observation values observed before the processing target time. More specifically, for each of the options A and B, the processing circuit 11 calculates the expected value μ of their effect A , μ B is randomly generated from the posterior distributions Beta(α A , β A ), Beta(α B , β B ), and selects the clinical decision corresponding to the maximum expected value among the expected values μ A , μ B , and assigns it to the subject.

[0042] Medical professionals perform medical procedures on the target subject corresponding to the assigned medical judgment, and an effect occurs on the target subject. When step S1 is performed, the processing circuit 11 observes the effect occurring on the target subject by realizing the observation function 112 (step S2). The effect is observed numerically as an effect observation value. For example, if the assigned medical judgment is "perform surgery," the effect observation values ​​obtained may include the 5-year survival rate and the presence or absence of complications. The effect observation values ​​may be obtained by any method. For example, they may be input by the operator via the input device 13, or measured values ​​from testing equipment may be input. Alternatively, they may be received from another computer via the communication device 14.

[0043] When step S2 is performed, the processing circuit 11 stores observation data, including the effect observation value, in the database DB1 (step S3). The database DB1 is a computer having a storage device for storing observation data of the subject. The database DB1 is connected to the medical information processing device 1 via the communication device 14 so that data can be sent and received. The observation data shall include at least the subject's identifier and the effect observation value. The observation data may further include any feature quantities such as the subject's gender.

[0044] When step S3 is performed, the processing circuit 11 determines whether or not to update the individual model by implementing the update function 116 (step S4). The timing of updating the individual model is not particularly limited and may be based on the number of observation data stored in the database DB1, or on the elapsed time from a reference point. As for the timing based on the number of observation data stored in the database DB1, for example, it may be set to the timing when an arbitrary reference number of observation data, such as 1, 3, or 10, has been stored in the database DB1. As for the timing based on the elapsed time from a reference point, for example, it may be set to the timing when an arbitrary set time, such as one week, two weeks, or one month, has elapsed since the last individual model update or the start of using the individual model.

[0045] If it is determined in step S4 that the individual model will not be updated (step S4: NO), the processing circuit 11 returns to step S1 and repeats steps S1 to S4 until it is determined that the individual model will be updated.

[0046] If it is determined in step S4 to update the individual model (step S4: YES), the processing circuit 11, by implementing the decision function 114, determines a portion of the observation data of other subjects to be used for the update (hereinafter referred to as the "usage range") (step S5). In step S5, the processing circuit 11 first determines whether or not there is a usage range, that is, whether or not to use the observation data of other subjects. If it is determined that there is a usage range, the processing circuit 11 determines the specific range of the usage range. More specifically, the processing circuit 11 determines the usage range based on the behavior and / or features of the target subject. The presence or absence of a usage range and the method of determining it can be determined by various methods. The presence or absence of a usage range and the method of determining its range will be described later in several examples. Other subjects are arbitrarily selected from subjects different from the target subject. Other subjects may be one subject or multiple subjects.

[0047] If the usage range is determined in step S5, the processing circuit 11 determines whether or not there is a usage range by implementing the acquisition function 115 (step S6). If it is determined in step S6 that there is no usage range (step S6: NO), the process proceeds to step S8.

[0048] If it is determined in step S6 that there is a range of use (step S6: YES), the processing circuit 11 acquires observational data of other entities regarding the range of use by realizing the acquisition function 115 (step S7). The observational data of other entities is stored in the database DB2 in the same way as the target entity. Database DB2 is a computer that has a storage device for storing observational data of other entities. Database DB2 is connected to the medical information processing device 1 via the communication device 14 so that data can be sent and received. Database DB2 and database DB1 may be the same computer or different computers. There are two forms for acquiring observational data of other entities regarding the range of use, and either form may be adopted. Form 1: The processing circuit 11 first acquires observational data of the entire range concerning other entities from database DB2, and then extracts observational data regarding the range of use from the acquired observational data of the entire range. Form 2: The processing circuit 11 acquires observational data of the range of use concerning other entities from database DB2.

[0049] If it is determined in step S6 that there is no usage range (step S6: NO) or if step S7 is performed, the processing circuit 11 updates the parameters of the individual model by implementing the update function 116 (step S8). In step S8, the processing circuit 11 updates the parameters of the individual model based on the observed data of the target entity and the observed data of other entities included in the usage range determined based on the behavior and / or features of the target entity. For example, the observed effect value R for the target entity. A If obtained, the observed effect value R A Based on observational data including the parameter α A ,β A The parameter α is updated. A ,β A These represent the number of times "1" and "0" have been observed, respectively, and are updated according to equation (3) below. That is, the processing circuit 11 updates the effect observation value R A If it is "1", then the parameter α A Add "1" to this, and the observed effect value R A If the parameter β is not "1",A Add "1" to it.

[0050]

number

[0051] In this embodiment, it is assumed that the individual model will differ for each subject. Therefore, parameter α A ,β A The value of α differs for each subject. The parameter of the individual model possessed by subject X is α A X ,β A Y Therefore, α A X ,β A Y R is the observed effect value resulting from selecting treatment decision A for patient X. A X Updated based on the effect observation R observed in other subjects Y. A Y Whether it is updated based on this is appropriately controlled.

[0052] If it is determined in step S6 that there is no usable range (step S6: NO), the processing circuit 11 updates the parameters of the individual model based only on the observed data of the target entity (step S8). Specifically, the observed effect value R for the target entity. A X Based on this, the parameter α of the individual model for subject X A X ,β A Y It will be updated.

[0053] If it is determined in step S6 that there is a range of use (step S6: YES), the processing circuit 11 updates the parameters of the individual model based on the observed data of the target entity and the observed data limited to the range of use of other entities (step S8). Specifically, the observed effect value R for the target entity. A X and the observed effect value R for other entities A YBased on this, the parameter α of the individual model for subject X A X ,β A Y The parameters of the individual model are updated. However, it is not necessary for the parameters of the individual model to be updated based on both the observational data of the target entity and the observational data limited to the usage range of other entities. For example, the parameters of the individual model may be updated based on the observational data of the target entity, and then updated based only on the observational data limited to the usage range of other entities, or vice versa.

[0054] This concludes the medical information processing according to this embodiment.

[0055] Next, several embodiments of this embodiment will be described. In the following embodiments, patient X is the subject, and patient Y is the other subject. Furthermore, the individual model in the following embodiments is a model for calculating the effect evaluation value of daily exercise for heart failure patients. Specifically, the individual model has two types of structural models: structural model MX1 which calculates the effect evaluation value for low exercise levels, and structural model MX2 which calculates the effect evaluation value for moderate exercise levels. Structural model MX1 has a larger expected value of effect and a smaller variance value compared to structural model MX2. This means that the frequency of assigning a medical judgment regarding low exercise levels tends to be higher than the frequency of assigning a medical judgment regarding moderate exercise levels. A medical judgment regarding moderate exercise levels is less likely to be optimal than a medical judgment instructing low exercise levels. Patient X would normally hesitate to have the individual model select a medical judgment regarding moderate exercise levels. However, if patient X selects a medical judgment regarding moderate exercise levels, the estimation accuracy of the individual model for patient X can be improved by the effects obtained from that medical judgment. Selecting a medical decision that is not frequently chosen and may result in short-term losses, but which can help improve the estimation accuracy of a model, is called exploratory behavior. In this case, selecting a medical decision regarding moderate exercise levels constitutes exploratory behavior.

[0056] If the parameters of patient X's individual model can be updated based on observational data obtained when patient Y chooses a medical decision regarding moderate exercise, then patient X will have a reduced incentive to choose that decision. On the other hand, patient Y will feel it is unfair if the observational data regarding the medical decision about moderate exercise, which they obtained at their own risk, is used for patient X. The processing circuit 11 according to this embodiment determines the range of patient Y's observational data used to update patient X's individual model in order to achieve incentive compatibility, where the happiness of individual patients and the happiness of patients as a whole coincide.

[0057] (Example 1) The processing circuit 11 in Example 1, through the implementation of the decision function 114, determines the range of use of observation data of other entities based on the timing of use of individual models relating to the target entity.

[0058] Figure 4 is a diagram illustrating the scope of use according to Example 1. Following steps S1 to S4 in Figures 2 and 3, the allocation of medical decisions by the allocation function 111, the observation of effects by the observation function 112, and the accumulation of observation data by the accumulation function 113 are repeatedly performed for each patient X and patient Y.

[0059] As shown in Figure 4, patient Y has been using individual models of patient Y (not shown) since before patient X, and patient Y's observation data DY1 to DY11 are stored in database DB2. For example, observation data DY1, DY2, DY4, DY5, DY7, DY8, DY9, DY10, and DY11 include observed effects resulting from low-momentum exercise. Observation data DY3 and DY6 include observed effects resulting from moderate-momentum exercise.

[0060] As shown in Figure 4, it is assumed that the first effect observation value for patient X was observed at time T41. That is, time T41 is the start time of using the individual model for patient X. Then, at time T42, it is determined in step S4 to update the individual model for patient X. The processing circuit 11, through the implementation of the decision function 114, determines that the period TD4 prior to the start time T41 of using the individual model for patient X, the target subject, will be included in the usage range of the observation data of patient Y, the other subject. The observation data of patient Y for the period after the start time T41 is not included in the usage range.

[0061] Once the usage range TD4 is determined, the processing circuit 11 updates the individual model of patient X based on the observed data of patient X and the observed data of patient Y related to the usage range TD4 by implementing the update function 116. Specifically, to update the structural model MX1 for low momentum, the observed data DX5, DX6, DX8, DX9, DX11 of patient X and the observed data DY1, DV2, DY4, DY5, DY7, DY8, DY9, DY10, DY11 of patient Y related to low momentum are used. To update the structural model MX1 for moderate momentum, the observed data DX7, DX10 of patient X and the observed data DY3, DY6 of patient Y related to moderate momentum are used.

[0062] In other words, the processing circuit 11 detects that at time T41, patient X has performed an action involving low or moderate energy expenditure. This detection is performed, for example, by acquiring observational data corresponding to the action and by accumulating that observational data in the database DB1. Upon detecting the action, the processing circuit 11 can use observational data of other patients Y prior to time T41. In other words, upon detecting the action, it becomes impossible to use observational data of other patients Y after time T41. That is, according to Embodiment 1, the processing circuit 11 updates the structural model MX based on the observational data of patient X after time T41 and the observational data of patient Y prior to time T41.

[0063] Thus, at the stage when patient X begins making treatment decisions using the individual model, the parameters of the individual model can be updated using patient Y's observational data, limited to observational data prior to the start time T41. This allows patient X to update the structural model MX2 related to treatment decisions that are unlikely to be optimal, based on a large amount of observational data, without having to select a treatment decision that is unlikely to be optimal, thus satisfying patient X's sense of well-being. For patient Y, since the scope of use is limited to observational data prior to the start time T41, the sense of unfairness is not significant. Therefore, it becomes possible to achieve incentive compatibility, where the happiness of individual patients and the happiness of patients as a whole coincide.

[0064] The scope of use in Example 1 is not limited to the period before the patient X, the subject, begins using the individual model. For example, the processing circuit 11 may determine the period during which the individual model of patient X is not in use to be within the scope of the patient Y's observation data. The period of inactivity means the period during which the individual model is not used for any reason after the start time of use of the individual model. If patient X begins using the individual model after the inactivity period has elapsed, the individual model of patient X may be updated using the patient Y's observation data accumulated during the inactivity period.

[0065] It is conceivable that the end time for using an individual model is predetermined, such as near the end of treatment for patient X. In this case, the processing circuit 11 may determine the range of use for patient Y's observed data to be a period from the end time for using patient X's individual model to a time determined statically or dynamically. Here, as an example, a statically determined period is a period predetermined by the operator, and a dynamically determined period is a period determined according to the content of the individual model update. During this usage period, the processing circuit 11 updates the parameters of patient X's individual model based on patient X's observed data and patient Y's observed data accumulated during that usage period. The reason why the use of patient Y's observed data during such a period is permitted is that, near the end of treatment for patient X's individual model, it is assumed that patient X has a reduced incentive to take exploratory actions, regardless of whether patient Y's observed data is available or not. Therefore, it is expected that patient Y will not feel a great sense of unfairness. On the other hand, patient X can improve the accuracy of the individual model without taking exploratory actions themselves. Thus, it becomes possible to achieve incentive compatibility, where the happiness of individual patients and the happiness of patients as a whole coincide.

[0066] (Example 2) The processing circuit 11 in Example 2, through the implementation of the decision function 114, determines the scope of use of observation data from other entities based on the medical judgment selected by the target entity. In the following description, components having substantially the same function as those in Example 1 are denoted by the same reference numerals and will be described redundantly only when necessary.

[0067] Figure 5 is a diagram illustrating the scope of use in Example 2. As shown in Figure 5, in Example 2 as well, for each of the target patient X and the other target patient Y, the allocation of medical decisions by the allocation function 111, the observation of effects by the observation function 112, and the accumulation of observation data by the accumulation function 113 are repeatedly performed according to steps S1 to S4 in Figures 2 and 3.

[0068] As shown in Figure 5, the processing circuit 11 according to Embodiment 2, by realizing the decision function 114, determines the period prior to the processing target time when the patient X, the subject, selects an exploratory action as the usage range. An exploratory action refers to an action corresponding to a medical judgment other than the medical judgment that the individual model has determined to be unlikely to be optimal or that it has determined to be optimal. Here, "optimal" in this embodiment means "the expected value of the effect evaluation value is maximized." However, "optimal" in this embodiment is not limited to this meaning and includes the meaning that "a certain indicator is optimal for that subject patient." This indicator may be of any type as long as it satisfies the optimal condition.

[0069] For example, in Figure 5, it is assumed that performing moderate-intensity exercise corresponds to exploratory behavior, and that patient X selected exploratory behavior at time T51. In this case, the processing circuit 11 determines the period DY51 prior to time T51 as the usage range. The processing circuit 11 updates the parameters of structural models MX1 and MX2 based on the observed data DX7 of patient X corresponding to time T51 and the observed data DY1 to DY7 of patient Y accumulated during the period DY51.

[0070] If patient X selects the exploration action again at time T52, the processing circuit 11 determines the period DY52, from time T51 when patient X previously selected the exploration action to time T52 when patient X selected the exploration action this time, as the usage range. The processing circuit 11 updates the parameters of structural models MX1 and MX2 based on the observed data DX10 of patient X corresponding to time T52 and the observed data DY8~DY10 of patient Y accumulated during the period DY52.

[0071] In other words, the processing circuit 11 detects that patient X has taken an exploratory action at time T51 or T52. This detection is performed, for example, by acquiring observational data corresponding to the exploratory action and by accumulating that observational data in the database DB1. Upon detecting the exploratory action, the processing circuit 11 can use observational data of other patients Y prior to time T51 or T52. As a result, the processing circuit 11 updates the structural model MX based on the entirety of patient X's observational data DX and patient Y's observational data DY prior to the time when the exploratory action was detected.

[0072] Thus, the processing circuit 11 in Embodiment 2, triggered by patient X, the target subject, choosing to engage in exploratory behavior, uses observational data of patient Y, another subject, acquired before the time the exploratory behavior was chosen, to update the parameters of the personal model for patient X. In other words, by making patient Y's observational data available through the selection of exploratory behavior, it becomes possible to provide patient X with an incentive to choose exploratory behavior. The same applies to patient Y, the other subject. Therefore, according to Embodiment 2, it is possible to achieve incentive compatibility, where the happiness of individual patients and the happiness of patients as a whole coincide.

[0073] Furthermore, the processing circuit 11 may determine the type of observed data for patient Y within the usage range and / or the time range of the usage range based on a comparison between the effectiveness evaluation value of the medical decision selected by patient X, the subject of the study, and the effectiveness evaluation value of the medical decision selected by the individual model for patient X. Specifically, the processing circuit 11 should widen the time range of the usage range the smaller the effectiveness evaluation value of the medical decision selected by patient X is compared to the effectiveness evaluation value of the medical decision selected by the individual model. In this case, it can be said that patient X has chosen a high-risk exploratory behavior, and the processing circuit 11 should reward patient X, who has chosen that behavior, with more observed data.

[0074] As another example, the processing circuit 11 may increase the types of patient Y observation data included in the usage range used for updating the smaller the effectiveness evaluation value of the medical decision selected by patient X is compared to the effectiveness evaluation value of the medical decision selected by the individual model. For example, if the patient X selects moderate-intensity exercise, which is an exploratory behavior, the processing circuit 11 uses both the observation data related to moderate-intensity exercise and the observation data related to low-intensity exercise for updating. On the other hand, if the patient X selects low-intensity exercise, which is not an exploratory behavior, the processing circuit 11 uses only the observation data related to low-intensity exercise for updating. By appropriately controlling the types of observation data and / or the time width of the usage range in this way, it becomes possible to provide appropriate incentives to patient X.

[0075] (Example 3) The processing circuit 11 in Example 3 determines the range of use of observed data from other entities based on the similarity between the feature quantities of the target entity and the feature quantities of other entities, by implementing the decision function 114. In the following description, components having substantially the same function as those in Example 1 are denoted by the same reference numerals and are described redundantly only when necessary.

[0076] Figure 6 is a diagram illustrating the scope of use in Example 3. As shown in Figure 6, in Example 3 as well, for each of the target patient X and the other target patient Y, the allocation of medical decisions by the allocation function 111, the observation of effects by the observation function 112, and the accumulation of observation data by the accumulation function 113 are repeatedly performed according to steps S1 to S4 in Figures 2 and 3.

[0077] The individual model in Example 3 follows a contextual bandit algorithm that calculates effect evaluation values ​​considering the patient's features. As described above, the features include attributes, such as the subject's gender and age, which are information that does not change based on the previous medical judgment, and / or states, such as the subject's blood pressure and blood glucose levels, which are information that changes based on the previous medical judgment. The processing circuit 11 includes subjects in the usage range for which low-similarity features have been observed, and excludes subjects in the usage range for which high-similarity features have been observed. Furthermore, for subjects in which low-similarity features have been observed, the processing circuit 11 includes the period during which the low-similarity features were observed in the usage range. In this way, the processing circuit 11 in Example 3 controls not only the temporal range but also the human range as the usage range.

[0078] As an example, the feature quantities related to Figure 6 take two values: state P or state Q. The observed data DX1-DX11 and DY1-DY11 will include the effect observation value and state P or state Q. In the observed data DX1-DX4 and DY1-DY4 before time T61, patient X's state is state P, patient Y's state is state Q, and the similarity between patient X's state and patient Y's state is "0". In the observed data DX5-DX7 and DY5-DY7 from time T61 to time T62, both patient X and patient Y's state is state Q, and the similarity is "1". In the observed data DX8-DX11 and DY8-DY11 from time T62 to time T63, patient X's state is state Q, patient Y's state is state P, and the similarity is "0".

[0079] At processing time T63, the processing circuit 11 determines the usage range by implementing the determination function 114. Specifically, the processing circuit 11 compares the state of patient X with the states of other entities such as patient Y and patient Z (not shown), which have been accumulated prior to processing time T63, for each observed data, and searches for other entities that have a state different from that of patient X. If there are other entities that have a state different from that of patient X, the processing circuit 11 includes those other entities in the usage range and excludes other entities that have the same state as patient X from the usage range. If there are multiple other entities that have a state different from that of patient X, all entities may be included in the usage range, or any one of them may be included in the usage range. In Figure 6, since the states of patient X and patient Y are different during the period TD61 prior to time T61 and the period TD3 from time T62 to time T63, patient Y is assumed to be included in the usage range. Next, the processing circuit 11 includes the observed data DY1-DY4 and DY8-DY11 from periods TD61 and TD63, where features with low similarity were observed, within its usage range. Then, the processing circuit 11 updates the parameters of the individual model for patient X based on the observed data DY1-DY4 and DY8-DY11 for patient Y during periods TD61 and TD63.

[0080] In other words, the processing circuit 11 determines at each time step whether patient X has different features from patient Y, and detects that patient X has different features from patient Y. This detection is performed, for example, by acquiring observational data corresponding to the features in question, and by accumulating such observational data in the database DB1. The processing circuit 11 can use the observational data of patient Y that corresponds to features different from those of patient X. Based on this, the processing circuit 11 updates the structural model MX based on the entirety of the observational data DX of patient X and the observational data DY of patient Y that has features different from those of patient X.

[0081] For example, if the status is gender, observational data from patients of the same gender will not be used to update the individual model, but observational data from patients of different genders will be used to update the individual model. Another example is if the status is age, observational data from patients of similar age will not be used to update the individual model, but observational data from patients of significantly different ages will be used to update the individual model. This is because updating the individual model for patient X using observational data from patients with significantly different features will have little impact on the medical decisions made by patient X, and therefore will not reduce patient X's incentive to choose exploratory behavior.

[0082] Furthermore, when focusing on changeable states, the human scope of the scope of use may be appropriately changed at the time each patient's state changes. For example, when patient X's blood pressure is low, processing circuit 11 updates the individual model using observational data from patient Y, who has high blood pressure, but does not update the individual model using observational data from patient Z, who has low blood pressure. When patient X's blood pressure rises sharply, processing circuit 11 may update the model using observational data from patient Z, who has low blood pressure. As another example, when patient X's blood pressure is low, processing circuit 11 updates the individual model using observational data from patient Y, who has high blood pressure, but does not update the individual model using observational data from patient Z, who has low blood pressure. However, if patient Y's blood pressure subsequently decreases and patient Z's blood pressure increases, the model may be updated using observational data from patient Y, who has low blood pressure. By adaptively changing the human scope according to each patient's state in this way, it is possible to maintain an incentive for patient X to always choose exploratory behavior, regardless of changes in the state of patient X or other entities.

[0083] This concludes the description of the examples.

[0084] (others) This embodiment is subject to various modifications. For example, in the above embodiment, the allocation function 111, observation function 112, storage function 113, decision function 114, acquisition function 115, update function 116, and display control function 117 are mounted on a single computer. However, this embodiment is not limited to this. These functions 111 to 117 may be distributed and mounted on multiple computers. In other words, the medical information processing device 1 may be a computer system composed of multiple computers on which functions 111 to 117 are distributed. Furthermore, the databases DB1 and DB2, where observation data is stored, are assumed to be external computers to the medical information processing device 1. However, this embodiment is not limited to this, and the databases DB1 and DB2 may be included in the medical information processing device 1.

[0085] As another example, the processing circuit 11 may, by implementing the decision function 114, determine the range of observational data from other entities to be passed on for updating the individual entity, depending on the state of the individual entity model. The state of the individual entity is specifically represented by the number of observational data used to train each structural model, the performance index values ​​of each structural model, etc. The performance index values ​​are parameters of the mathematical expression of each structural model, and for example, the parameters α, β, p, etc. mentioned above can be used.

[0086] As an example, let's consider the case where the state of an individual model is the number of observational data (hereinafter referred to as training data) used to train each structural model. If the number of training data is small, it means that selecting a medical judgment corresponding to that structural model is an exploratory behavior. The processing circuit 11 compares the number of training data for each structural model against a threshold, for example. Suppose that structural model MX1 has a large number of data relative to the threshold, and structural model MX2 has a small number of data relative to the threshold. If the number is large relative to the threshold, the processing circuit 11 determines the range of use for observational data of other subjects Y related to the medical judgment corresponding to structural model MX1, and trains structural model MX1 using the observational data of subjects Y included in the determined range of use. If the number is small relative to the threshold, the processing circuit 11 does not determine the range of use for observational data of other subjects Y related to the medical judgment corresponding to structural model MX2, and does not use the observational data of subjects Y for training structural model MX2. Although we have stated that the number of training data is compared against a threshold to determine the magnitude of the number of training data, the magnitude can also be determined by comparing the number of training data between structural models.

[0087] (Summary) According to some of the embodiments described above, the medical information processing device 1 according to this embodiment has a processing circuit 11. The processing circuit 11 updates an individual model that calculates an evaluation value of the effectiveness of a medical judgment, which is a model relating to a target subject, based on observational data corresponding to the effect of a medical judgment on a target subject and observational data limited to a certain range of observational data corresponding to the effect of a medical judgment on a subject other than the target subject, by implementing an update function 116.

[0088] According to the above configuration, the observational data of other subjects used to update the individual model of the target subject is limited to a certain range (usage range). Therefore, it becomes possible to enhance the well-being of the target subject while suppressing feelings of unfairness among other subjects regarding the updating of the individual model of the target subject. This makes it possible to achieve incentive compatibility, where the happiness of individual patients and the happiness of patients as a whole coincide. Consequently, it becomes possible to improve the accuracy of individual models, thus realizing precision medicine. By realizing precision medicine, it becomes possible to construct models with a small amount of observational data compared to general-purpose models. Furthermore, since it is not necessary to consider the exclusion of inter-individual characteristics, it is expected that models can be constructed without conducting randomized controlled trials.

[0089] According to at least one embodiment described above, it is possible to achieve incentive compatibility in the allocation of medical decisions.

[0090] In the above description, the term "processor" refers to circuits such as CPUs, GPUs, or Application Specific Integrated Circuits (ASICs), programmable logic devices (e.g., Simple Programmable Logic Devices (SPLDs), Complex Programmable Logic Devices (CPLDs), and Field Programmable Gate Arrays (FPGAs)). The processor achieves its function by reading and executing a program stored in a memory circuit. Alternatively, instead of storing the program in a memory circuit, the processor may be configured to directly incorporate the program into its circuitry. In this case, the processor achieves its function by reading and executing the program incorporated into the circuitry. Furthermore, instead of executing a program, the processor may achieve the function corresponding to the program through a combination of logic circuits. In this embodiment, each processor is not limited to being configured as a single circuit; multiple independent circuits may be combined to form a single processor and achieve its function. Moreover, the multiple components shown in Figure 1 may be integrated into a single processor to achieve its function.

[0091] While several embodiments have been described, these embodiments are presented as examples only and are not intended to limit the scope of the invention. These embodiments can be implemented in a variety of other forms, and various omissions, substitutions, modifications, and combinations of embodiments are possible without departing from the spirit of the invention. These embodiments and their variations are included in the scope and spirit of the invention, as well as in the claims and their equivalents. [Explanation of Symbols]

[0092] 1. Medical Information Processing Device 11 Processing Circuit 12 Storage device 13 Input devices 14. Communication equipment 15 Display equipment 111 Assignment function 112 Observation Functions 113 Storage function 114 Decision Function 115 Acquisition function 116 Update function 117 Display Control Function

Claims

1. Based on first data corresponding to the effect of the medical judgment on the subject, and data limited to a portion of second data corresponding to the effect of the medical judgment on a subject other than the subject, an update unit updates a model relating to the subject, which is an individual model that calculates the effect evaluation value of the medical judgment. A medical information processing device equipped with the following features.

2. The medical information processing apparatus according to claim 1, further comprising a determination unit that determines the partial range based on the timing of use of the individual model relating to the subject.

3. The medical information processing apparatus according to claim 2, wherein the determination unit determines a period prior to the start time of use of the individual model, during the period of suspension of use, or a period statically or dynamically determined from the end time of use, within the specified range.

4. The medical information processing apparatus according to claim 1, further comprising a determination unit that determines the partial range based on the medical judgment selected by the subject entity.

5. The medical information processing apparatus according to claim 4, wherein, if the decision unit determines at the target time that the target entity has selected a medical judgment other than the one in which the model is deemed to be optimal or a medical judgment other than the one in which the model is deemed optimal, the period prior to the target time is determined to be within the specified range.

6. The medical information processing apparatus according to claim 5, wherein the determination unit determines the type of the second data in the partial range and / or the time width of the partial range to be used for updating the model, based on a comparison of the effectiveness evaluation value of the medical judgment selected by the subject entity and the effectiveness evaluation value of the medical judgment selected by the model.

7. The medical information processing apparatus according to claim 1, further comprising a determination unit that determines the partial range based on the similarity between the feature quantities of the target entity and the feature quantities of the other entity.

8. The medical information processing apparatus according to claim 7, wherein the determination unit includes in the partial range any other entities for which a feature quantity with low similarity has been observed, and excludes from the partial range any entities for which a feature quantity with high similarity has been observed.

9. The medical information processing apparatus according to claim 7, wherein the determination unit determines the period during which the feature quantities with low similarity were observed within the observation period of the second data of the other entity to be a partial range.

10. The system further includes an acquisition unit for obtaining the second data from a database, The update unit extracts data limited to the aforementioned partial range from the second data. The medical information processing device according to claim 1.

11. The medical information processing apparatus according to claim 1, further comprising an acquisition unit for acquiring data limited to a certain range of the second data from a database.

12. The first data is a series of data including multiple effect observation values ​​corresponding to multiple observation times of the effect of the medical judgment on the subject, The second data is a series of data including multiple effect observation values ​​corresponding to multiple observation times of the effect of the medical judgment on the other subject. The medical information processing device according to claim 1.

13. The medical information processing device according to claim 1, wherein the subject entity and the other entity are an individual subject, a group of subjects including one or more subjects, a hospital, or a group of hospitals including one or more hospitals.

14. The medical information processing apparatus according to claim 1, wherein the update unit updates the individual model based on the first data and the second data within a partial range determined based on the actions and / or features of the target entity.

15. The medical information processing apparatus according to claim 1, further comprising a determination unit that determines the partial range of the second data to be passed for updating the individual model, according to the state of the individual model relating to the subject.

16. A computer, We obtain the first data corresponding to the effect of the medical judgment on the target subject. By obtaining data limited to a portion of the second set of data corresponding to the effect of medical judgments on entities other than the aforementioned target entity, Based on the first data and the data limited to a certain range, the model relating to the subject is updated, and the individual model that calculates the effectiveness evaluation value of the medical judgment is updated. A medical information processing method that includes the following features.

17. On the computer, The first function is to acquire data corresponding to the effect of the medical judgment on the target subject, A function to acquire data limited to a portion of the second set of data corresponding to the effect of medical judgments on entities other than the aforementioned target entity, Based on the first data and the data limited to a certain range, the system has a function to update the model relating to the subject, which is an individual model for calculating the effectiveness evaluation value of the medical judgment, A medical information processing program that makes this possible.

Images