JP2011203996A - Inference device for executing inference by bayesian network, and program for attaining the inference device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To execute appropriate inferences, even when input information is insufficient.SOLUTION: A Bayesian network model 21A includes one or more knowledge variables (21b) with expert knowledge regarding correlation between one or more result variables (21d) and one or more factor variables (21a) described therein, and one or more detail variables (21c) with details of the result variables (21d) described therein. Furthermore, each of the variables is constituted, such that it has dependency onto the one or more knowledge variables (21b) from each of the factor variables (21a), dependency onto the one or more detail variables (21c) from the one or more knowledge variables (21b), and dependence on each of the result variables (21d) from the one or more detail variables (21c). An inference unit 12 sets, as evidence, the values of either or both of the result variables (21d) and the detail variables (21c), in response to one or more items of input information provided from an input unit 30 to calculate probabilities of each of the factor variables (21a) by means of inference depending on a Bayesian network.

Description

  The present invention relates to an inference apparatus that executes inference by a Bayesian network, and a program that realizes the inference apparatus.

  As a technique for performing statistical estimation, there is a technique called a Bayesian network. A Bayesian network assumes in advance in a statistical model called a “Bayesian network model” whether or not arbitrary information is generated through a process based on the causal relationship of a plurality of variables (events). This is a method for estimating an information source (mainly a probability of a factor variable that is a factor of a result variable) that is a source for generating the information from arbitrary information (mainly a result variable related to an observation result).

  There is an inference device that performs inference by the Bayesian network. This inference device is, for example, a diagnostic device for a medical field (especially emergency medicine) that estimates a disease, severity, urgency, etc. from a patient's symptoms, and a failure diagnostic device that diagnoses a failure of a machine tool, an information device, etc. It is used in a recommendation system for evaluating the priority of a recommendation object in an Internet service.

  This inference apparatus uses a Bayesian network model (hereinafter sometimes referred to as “model information”) constructed by hierarchizing a plurality of variables (nodes). The inference apparatus sets (sets) the value of the input information as evidence, and infers the probability of the factor variable based on the model information. At that time, the inference apparatus sets the variable corresponding to the superordinate concept or the subordinate concept as evidence according to the value of the input information, thereby reducing the probability of the factor variable regardless of the abstraction level and granularity of the input information. Can be inferred.

  “Evidence” means the value of an observable variable that is set to execute an inference. However, the inference device outputs a factor variable that is an output (for example, in the medical field, the probability of each disease when inferring a disease such as “subarachnoid hemorrhage” or “cerebral infarction”) or a variable that could not be observed (for example, In the medical field, the age information when the age is unknown and the pain information when the pain level is unknown) are inferred without setting a value.

  In addition, “granularity” in the inference process means the roughness of information. For example, the medical information “severe headache in the back of the head” is finer than the medical information “having a headache”. In the medical field, usually, ambulance crews can input finer-scale medical information than ordinary persons, and doctors and nurses can input more detailed medical information.

  As model information used by the inference apparatus, the dependency relation of each variable is described as a probability table by a plurality of hierarchized variables (nodes), and each variable is binary ("Y (Yes)" or Some have a configuration of “N (No)”) (see, for example, Patent Document 1).

  For example, as model information used in the medical field, “symptoms (symptoms)” such as headache and fever and “diseases” such as influenza and pneumonia are used as variables, and the dependency relationship between “symptoms” and “diseases” (onset) Probability) is described as a probability table, and each variable has a binary value (“Y” or “N”).

  The model information uses various variables related to the result variable in order to improve the estimation accuracy of the probability of the factor variable. For example, model information used in the medical field includes “headache” in order to improve the accuracy of estimating the probability of a disease when a disease such as “influenza” is a factor variable and a symptom such as “headache” is a result variable. In addition to the presence or absence of symptoms such as “headache” (“Y” or “N”), variables related to various symptoms (for example, “variables related to the degree of pain”, “variables related to the onset status”, "Variables related to pain sites" etc.) are used.

The “variable relating to the degree of pain” means, for example, a variable relating to the degree of pain (subordinate concept or attribute concept of pain) such as “severe (severe pain)” or “mild (slight pain)”.
The “variable related to the onset status” means a variable related to the onset status (subordinate concept or attribute concept) such as “acute (acute onset)” or “gradual progress”.
Furthermore, the “variable relating to the pain region” means, for example, a variable relating to the pain region such as “occipital region” or “temporal region” (subordinate concept of the region).
The following are the components of symptom variables such as “severe (severe pain)”, “slight (slight pain)”, “acute (acute onset)”, “gradual”, “occipital”, “temporal”, etc. This variable is called “element variable”.

  Specifically, the model information used in the medical field is, for example, as shown in FIG. 12, by applying FIG. 24 and FIG. It is configured to take values such as “headache / severe / acute / occipital”, “headache / severe / acute / temporal”, “headache / severe / gradual / occipital”,…, “no headache”, etc. . FIG. 12 is a diagram illustrating a configuration of a Bayesian network model used conventionally. FIG. 12A shows the configuration of variables used for model information, and FIG. 12B shows the configuration of a variable probability table. The inference apparatus estimates the probability of a disease such as “influenza” as a factor variable based on the model information.

JP 2007-293601 A (paragraphs 64 to 80, FIGS. 19 to 25)

  However, as described below, the conventional reasoning apparatus has a problem that it cannot execute a suitable reasoning when input information is insufficient.

  (1) The conventional inference apparatus can perform inference when input information is input without a shortage, but when input information is short, for example, when some element variables are unknown, Since there is no element variable value to set as evidence, the value of the element variable cannot be set. Therefore, the conventional reasoning apparatus cannot execute a suitable reasoning.

  (2) For example, “subarachnoid hemorrhage” often becomes “headache / severe, acute, occipital region” as a symptom of headache. Such knowledge is difficult to reflect in the model information when element variables such as the degree of pain, the onset status, and the pain site are unknown. For example, as illustrated in FIG. 13, the user of the inference apparatus includes an unknown element variable as the value of the “headache detail” variable in consideration of the case where some element variables are unknown. Suppose that the model information is added with values such as “headache / severe / acute / unknown part”, “headache / severe / onset unknown / back of head”, “headache / severe / onset unknown / unknown part”. FIG. 13 is a diagram showing a configuration of a Bayesian network model used conventionally. FIG. 13A shows the configuration of variables used for the model information, and FIG. 13B shows the configuration of the variable probability table.

  Even in this case, the user of the inference apparatus can use the probability table of the result variables (eg, variables related to symptoms such as “headache”) and the probability of the cause variables (eg, variables related to diseases such as “subarachnoid hemorrhage”). Since it is not easy to determine how much probability value should be set in the table, it is necessary to adjust the probability value setting while repeating trials. Therefore, the user of the inference apparatus has expertise regarding the correlation between the outcome variable and the factor variable (here, “subarachnoid hemorrhage” becomes “headache / severe, acute, occipital” as a symptom of headache. The knowledge that there are many cases) could not be immediately reflected in the reasoning. Therefore, it is difficult for the user of the inference apparatus to easily create model information that reflects such expertise and to easily adjust the probability value set in the model information. Therefore, the conventional reasoning apparatus cannot execute a suitable reasoning.

  (3) Temporarily, the user of the reasoning apparatus considers the case where some element variables are unknown as shown in FIG. 14, for example, variables related to the pain level (headache level) and onset status (headache onset). ) And variable values related to pain sites (headache sites) are simply configured as model information to which an “unknown” value is added. FIG. 14 is a diagram illustrating a configuration of a Bayesian network model used conventionally. FIG. 14A shows the configuration of variables used for model information, and FIG. 14B shows the configuration of a variable probability table.

  Even in this case, as in the case of (2) described above, it is easy for the user of the inference device to set the probability value to be set in the probability variable table of the result variable and the probability table of the factor variable. Therefore, it is necessary to adjust the probability value setting while repeating trials. Therefore, it has been difficult for a user of the inference apparatus to easily create model information that reflects his / her expertise and to easily adjust the probability value set in the model information. Therefore, the conventional reasoning apparatus cannot execute a suitable reasoning.

  The present invention has been made to solve the above-described problems, and provides an inference device that executes suitable inference even when input information is insufficient, and a program that realizes the inference device. Main purpose.

  In order to achieve the above object, a first invention is an inference apparatus for performing inference by a Bayesian network, wherein a model holding unit in which a Bayesian network model is held in advance, and the Bayesian network model held in the model holding unit. An inference unit that infers the probability of a factor variable that is a factor of a result variable related to an observation result, an input unit that provides the inference unit with one or more input information that is a condition for inference, and the inference An output unit that outputs a probability of a factor variable calculated as an execution result of inference by the unit, wherein the Bayesian network model is between one or more of the result variables and one or more of the factor variables, One or more knowledge variables describing the expertise of the correlation between the result variable and the factor variable, and the result variable One or more detailed variables described in detail, each variable further comprising a dependency from each said factor variable to one or more said knowledge variables, one or more from one or more said knowledge variables The inference unit is configured to have a dependency relationship to the detailed variable and a dependency relationship from one or more of the detailed variables to each of the result variables, and the inference unit is configured to input the input based on the Bayesian network. In accordance with one or more pieces of input information provided from the section, the value of one or both of the result variable and the detailed variable is set as evidence, and the probability of each factor variable is calculated. The configuration.

  The inference apparatus includes a dependency from each factor variable to one or more knowledge variables, a dependency from one or more knowledge variables to one or more detail variables, and one or more detail variables to each A Bayesian network model (model information) configured to have dependency on the result variable is used. Then, the inference unit sets the value of one or both of the result variable and the detailed variable as evidence according to one or more input information provided from the input unit. Thereby, this inference apparatus performs inference by a Bayesian network according to the input information, and calculates the probability of each factor variable.

  The knowledge variable describes the expertise related to the correlation between the result variable and the factor variable. This knowledge variable functions to absorb a deviation between one or both of the result variable and the detail variable and the factor variable in inference. Therefore, this inference apparatus can execute a suitable inference even when input information is insufficient by using knowledge variables. Moreover, the probability value is set separately from the detailed variable in the knowledge variable. Therefore, the user of the inference apparatus can easily create model information that reflects expert knowledge, and can easily adjust the probability value set in the model information.

In addition, the second invention relates to a correlation between the result variable and the factor variable between one or more result variables related to the observation result and one or more factor variables that cause the result variable. One or more knowledge variables describing expertise, and one or more detail variables describing details of the result variable, each variable including one or more of the factor variables from each of the factor variables Dependency on knowledge variable, one or more knowledge variables to one or more detail variables, and one or more detail variables to each result variable A program that causes a computer having a model holding unit in which a configured Bayesian network model is held in advance to function as an inference device that executes inference by a Bayesian network, the computer comprising: Based on one or more input information provided from the input unit based on the Bayesian network model held by the model holding unit, the value of one or both of the result variable and the detailed variable is set. By setting the evidence and calculating the probability of each of the factor variables, an inference unit for inferring the probability of the factor variable and one or more input information as a condition for inference are input to the inference unit. The input unit to be provided and the output unit that outputs the probability of the factor variable calculated as the execution result of the inference by the inference unit are configured.
This program can realize the inference apparatus according to the first invention.

According to the first aspect of the present invention, it is possible to provide an inference device that performs appropriate inference even when input information is insufficient.
According to the second invention, a program for realizing the inference apparatus according to the first invention can be provided.

It is a figure which shows the structure of the inference apparatus which concerns on embodiment. It is a figure (1) for demonstrating the variable regarding a headache. It is a figure (2) for demonstrating the variable regarding a headache. It is a figure (1) which shows the hierarchical structure of the Bayesian network model used by embodiment. It is a figure (2) which shows the hierarchical structure of the Bayesian network model used by embodiment. It is a figure which shows the structure of the disease variable used in embodiment. It is a figure which shows the structure of the symptom knowledge variable used by embodiment. It is a figure which shows the structure of the symptom detailed variable used by embodiment. It is a flowchart which shows operation | movement of the inference apparatus which concerns on embodiment. It is a figure which shows an example of the input screen of the reasoning apparatus which concerns on embodiment. It is a figure for demonstrating operation | movement of the inference apparatus which concerns on embodiment. It is a figure which shows an example of the output screen of the reasoning apparatus which concerns on embodiment. It is a figure for demonstrating operation | movement of the inference apparatus which concerns on embodiment. It is a figure which shows an example of the output screen of the reasoning apparatus which concerns on embodiment. It is a figure (1) which shows the structure of the Bayesian network model used conventionally. It is a figure (2) which shows the structure of the Bayesian network model used conventionally. It is a figure (3) which shows the structure of the Bayesian network model used conventionally.

  Hereinafter, an embodiment of the present invention (hereinafter referred to as “the present embodiment”) will be described in detail with reference to the drawings. Each figure is only schematically shown so that the present invention can be fully understood. Therefore, the present invention is not limited to the illustrated example. Moreover, in each figure, the same code | symbol is attached | subjected about the common component and the same component, and those overlapping description is abbreviate | omitted.

<Configuration of inference device>
Hereinafter, the configuration of the inference apparatus according to the present embodiment will be described with reference to FIG. FIG. 1 is a diagram illustrating a configuration of an inference device according to an embodiment. Here, a case where the inference apparatus 1 according to the embodiment is used as a diagnostic apparatus for a medical field (particularly emergency medical care) that estimates a disease, a severity, an urgency level, and the like from a patient's symptoms will be described as an example.

  As shown in FIG. 1, an inference apparatus 1 according to the embodiment is configured by a computer such as a personal computer or a server, and a control program (hereinafter referred to as “inference program”) 20A for causing the computer to function as an inference apparatus. And a Bayesian network model (hereinafter also referred to as “model information”) 21A and the like are stored in the storage unit 20 in a readable manner in advance.

  Note that the inference program 20A and the model information 21A are, for example, a recording medium configured in the form of a memory member such as a USB memory, an HDD device, an optical medium such as a CD-ROM or DVD, and a magnetic medium such as a flexible disk. 60 is stored in advance, and is installed in the computer from the recording medium 60.

  The inference program 20A can be configured by, for example, commercially available general-purpose Bayesian network processing software.

  The model information 21A can be created by using a model creation function of general-purpose Bayesian network processing software that is commercially available. Here, since the inference device 1 is used as a diagnostic device for the medical field, the model information 21A is model information having a configuration suitable for the medical field (specifically, a disease model representing a correlation between symptoms and diseases) Information).

The inference apparatus 1 includes a control unit 10, a storage unit 20, an input unit 30, an output unit 40, and a communication unit 50 configured by a communication interface. Hereinafter, the input unit 30 and the output unit 40 may be integrated and referred to as “input / output unit 3”.
The control unit 10 is configured by a CPU.
The storage unit 20 includes a ROM, a RAM, an HDD, and the like.
The input unit 30 includes a keyboard, a mouse, a reader / writer for the recording medium 60, and the like, and provides input information as a condition for executing inference to the inference unit 12 described later.
The output unit 40 includes a display (including a touch panel), a printer, and the like, and outputs a probability of a result variable calculated as an inference execution result by the inference unit 12 described later.
The communication unit 50 is configured by a communication interface.

  The control unit 10 functions as a main control unit 11, an inference unit 12, a display control unit 13, and a communication control unit 14. The CPU implements each of these functions by executing the inference program 20 </ b> A stored in the storage unit 20.

The main control unit 11 is a functional unit that controls the operation of each functional unit.
The inference unit 12 is a functional unit that executes an inference based on model information 21 </ b> A held in advance in a model holding unit 21 described later on the storage unit 20. The inference unit 12 executes processes such as acquisition of model information 21A from the storage unit 20, acquisition of variable values based on the model information 21A, setting or resetting of variable values as evidence, and inference.

The display control unit 13 is a functional unit that controls the display operation of the display as the output unit 40. In addition, the display control part 13 displays on the display the information regarding the screen for inputting each symptom and its condition, and a plurality of buttons on the screen. Each button is associated with each symptom or condition. FIG. 7 is a diagram showing an example of the input screen, and shows that buttons for inputting each symptom and its status are arranged.
The communication control unit 14 is a functional unit that executes communication with an external device (not shown) by the communication unit 50.

  In addition, a model holding unit 21 is secured in the storage unit 20. The model holding unit 21 is a storage area in which model information 21A as a Bayesian network model is held in advance. The model information 21A is in a format that can be processed by the inference unit 12. The model information 21A includes a factor variable 21a, a knowledge variable 21b, a detailed variable 21c, and a result variable 21d.

The factor variable 21a means an information source that is a source for generating arbitrary information (that is, a variable that causes the result variable 21d).
The knowledge variable 21b means a variable in which expert knowledge related to the correlation between the result variable 21d and the factor variable 21a is described.
The detailed variable 21c means a variable in which details of the result variable 21d are described.
The result variable 21d means a variable related to the observation result.

  Here, as described above, the model information 21A is created as model information (disease model information) having a configuration suitable for the medical field. Therefore, for example, a variable related to headache is used in the model information 21A. Hereinafter, the variables relating to headache will be described with reference to FIGS. 2 and 3 are diagrams for explaining variables related to headache, respectively.

<Variables related to headache>
As variables related to the headache, for example, as shown in FIG. 2, there are a pain level, an onset condition, a pain site, and the like. The “pain degree” variable takes a value representing the degree of pain (here, headache) such as “severe”, “mild”, and “no headache”. The “onset status” variable takes a value representing the onset progress status, such as “acute”, “gradual progress”, and “no headache”. The “pain site” variable takes a value representing a site of pain such as “occipital region”, “temporal region”, and “no headache”.

  These variables have a typical pattern depending on the disease. For example, at the time of subarachnoid hemorrhage, the pattern shown in FIG. 3 is obtained. That is, these variables are the pattern at the time of subarachnoid hemorrhage, the value of the “pain degree” variable is “severe”, the value of the “onset condition” variable is “acute”, and the value of the “pain site” variable is “ It is easy to become “back of head”

  In the model information 21A (see FIG. 1), a pattern of symptoms (result variable 21d) for such a disease (factor variable 21a) is described as a knowledge variable 21b. The knowledge variable 21b is set so that the probability value becomes higher as the pattern having a deeper dependency.

<Details of model information>
The details of the model information 21A will be described below with reference to FIGS. 4A and 4B and FIGS. 5A to 5D. 4A and 4B are diagrams each showing a hierarchical structure of a Bayesian network model used in the embodiment. FIG. 5A is a diagram illustrating a configuration of factor variables used in the embodiment. FIG. 5B is a diagram illustrating a configuration of knowledge variables used in the embodiment. FIG. 5C is a diagram illustrating a configuration of detailed variables used in the embodiment. FIG. 5D is a diagram illustrating a configuration of result variables used in the embodiment.

  The model information (disease model information) 21A for the medical field is usually created in the direction from the disease to the symptom because the symptom appears according to the disease. FIG. 4A shows the hierarchical structure of the model information 21A. In the example shown in FIG. 4A, the model information 21A includes the disease variable 211 at the top, the symptom variable 212 at the bottom, and the symptom detailed variable 213 and the symptom variable 213 between the disease variable 211 and the symptom variable 212. The knowledge variable 214 is arranged.

  The disease variable 211 is a variable related to diseases such as influenza and pneumonia. The disease variable 211 corresponds to the factor variable 21a shown in FIG. The probability table of the disease variable 211 is configured so that the possible value is a binary value (“Y” or “N”) or a plurality of predetermined variables. Details of the disease variable 211 will be described later with reference to FIG. 5A.

  The symptom variable 212 is a variable related to symptoms such as headache and fever. The symptom variable 212 corresponds to the result variable 21d shown in FIG. The probability table of the symptom variable 212 is configured such that the condition variable is the value of the symptom detailed variable 213 and the possible value is a binary value (“Y” or “N”). The probability table of the symptom variable 212 indicates that the probability of the value “N” is 0 when the probability of the value “Y” is 1, and the probability of the value “N” when the probability of the value “Y” is 0. Is set to be 1. For example, in the example shown in FIG. 4A, the probability table of the symptom variable 212 uses the value of one corresponding symptom detail variable 213 as a condition variable, and when the condition variable is “no headache”, the probability of the value “Y” is It is set to be 0 and the probability of the value “N” is 1. Otherwise, the probability of the value “Y” is 1 and the probability of the value “N” is 0. Set as follows.

  The symptom detailed variable 213 is a variable related to each detailed information of each symptom. The detailed symptom variable 213 corresponds to the detailed variable 21c shown in FIG. The symptom detail variable 213 is used to enter details of the observed symptom. In the probability table of the symptom detailed variable 213, the condition variable is the value of the symptom knowledge variable 214, and a possible value is a value of a combination of one or more variables (for example, “headache / severe / acute / occipital”). ). Details of the symptom detail variable 213 will be described later with reference to FIG. 5C.

  The symptom knowledge variable 214 is a variable related to medical and clinical knowledge of each symptom. The symptom knowledge variable 214 corresponds to the knowledge variable 21b shown in FIG. The symptom knowledge variable 214 is used to describe medical and clinical knowledge. In the probability table of the symptom knowledge variable 214, the condition variable is the value of the disease variable 211, and a possible value is a value of a combination of one or more variables (for example, “headache / severe / acute / occipital”). It is comprised so that it may become. However, the symptom knowledge variable 214 includes a value that does not specify some detailed information. Details of the symptom knowledge variable 214 will be described later with reference to FIG. 5B.

  The model information 21A separates the symptom detailed variable 213 and the symptom knowledge variable 214 because the observed symptom does not necessarily match medical / clinical knowledge. For example, when the patient himself / herself becomes an observer and the patient reports the degree of pain such as headache or chest pain to an ambulance crew or a doctor / nurse, the patient reports lightly with a heavy pain, There are times when you report pain exaggeratedly. In addition, when a patient declares a pain site, the patient sometimes reports abdominal pain as abdominal pain or chest pain. The symptom knowledge variable 214 can absorb these deviations.

  The disease variable 211 and the symptom knowledge variable 214 have a dependency relationship and are associated with each other. Further, the symptom knowledge variable 214 and the symptom detailed variable 213 have a dependency relationship and are associated with each other. Furthermore, the symptom detailed variable 213 and the symptom variable 212 are in a dependency relationship and are associated with each other.

  For example, the disease variable 211 of “subarachnoid hemorrhage” is dependent on the symptom knowledge variable 214 of “headache”. For example, when the value of the symptom knowledge variable 214 of “headache” is “severe / acute / occipital”, the disease variable 211 has a high probability under the condition of “subarachnoid hemorrhage” (for example, 1.00 Become).

  The hierarchical structure shown in FIG. 4A appears to be reversed at first glance because the inference direction is shown from the disease to the symptom direction. However, Bayesian networks can perform reverse reasoning. That is, the inference apparatus 1 can execute inference in the direction opposite to the inference direction illustrated in FIG. 4A. Therefore, the hierarchical structure shown in FIG. 4A has a correct configuration.

  FIG. 4B shows a state in which model information 21A created using the model creation function of general-purpose Bayesian network processing software is displayed on the display. In FIG. 4B, the square frame part is an area where the disease variable 211 is displayed, and the ellipse part is an area where any one of the symptom variable 212, the symptom detailed variable 213, and the symptom knowledge variable 214 is displayed. It has become.

(Details of disease variables)
Hereinafter, the details of the disease variable 211 will be described with reference to FIG. 5A. FIG. 5A shows a specific configuration of the disease variable 211. The disease variable 211 has a configuration shown in FIG. 5A (a) or FIG. 5A (b).

  In the example shown in FIG. 5A (a), the disease variable 211 is provided for each disease such as subarachnoid hemorrhage, cerebral infarction, acute myocardial infarction (not shown), for example. The disease variable 211 is configured to take a value of “Y” or “N” corresponding to each disease. In the example shown in FIG. 5A (a), there are a plurality of dependencies between the disease variable 211 and the symptom knowledge variable 214. Therefore, a plurality of symptom knowledge variables 214 correspond to an arbitrary disease variable 211, and a plurality of disease variables 211 correspond to an arbitrary symptom knowledge variable 214.

  In the example shown in FIG. 5A (b), the disease variable 211 is provided by integrating each disease into one. The disease variable 211 has a configuration in which the disease variable 211 takes a value uniquely set corresponding to each disease. In the example shown in FIG. 5A (b), there is a one-to-N dependency between the disease variable 211 and the symptom knowledge variable 214. Therefore, a plurality of symptom knowledge variables 214 correspond to one disease variable 211.

(Details of symptom knowledge variable)
The details of the symptom knowledge variable 214 will be described below with reference to FIG. 5B. FIG. 5B shows a specific configuration of the symptom knowledge variable 214.

  As shown in FIG. 5B, the probability table of the symptom knowledge variable 214 uses the value of the corresponding disease variable 211 as a conditional variable, and “headache / severe / acute / occipital” or “headache / severe / acute / temporal”. , “Headache / severe / acute”, “headache / severe / gradual / occipital”, ..., “headache”, “no headache”. The probability table of the symptom knowledge variable 214 is a value in which the probability of the value “Y” approximates to 1 (preferably 0.7 or more, and 1 when the condition variable matches a typical pattern of each disease. (Value less than .0)).

(Details of symptom detail variables)
Hereinafter, the details of the symptom detailed variable 213 will be described with reference to FIG. 5C. FIG. 5C shows a specific configuration of the symptom detail variable 213.

  As shown in FIG. 5C, the probability table of the symptom detailed variable 213 uses the value of the corresponding symptom knowledge variable 214 as a conditional variable, and “headache / severe / acute / occipital” or “headache / severe / acute / temporal”. ”,“ Headache / severe / acute / unknown part ”,“ headache / severe / progressive / occipital ”,...,“ Headache / detail unknown ”,“ no headache ”. The probability table of the symptom detail variable 213 is a value in which the probability of the value “Y” is close to 1 (preferably 0.7 or more, and when the condition variable matches a typical pattern of each disease, and In other cases, the value “N” is a value close to 1 (preferably a value of 0.7 or more and less than 1.0). Set to.

  Here, the value approximated to 1 is because the symptom detailed variable 213 is assumed to input a value as evidence based on the patient's chief complaint or observation, and includes an error.

  The dependence relationship of each variable defined by the model information 21A indicates that a certain disease presents a specific medically and clinically specific symptom, and further, a specific symptom can be decomposed into detailed symptoms from several viewpoints. It shows that general symptoms can be derived from the symptoms.

  When the inference device 1 estimates a disease from a symptom, the symptom information is input as input information based on the main complaint or observation of the patient, and the inference unit 12 determines the symptom variable 212 and the symptom details according to the symptom information. A value is set in one or both of the variables 213. Then, in the inference apparatus 1, the inference unit 12 sets the value of the set variable as evidence of the symptom knowledge variable 214 (however, depending on the setting of the dependency, the value of the set variable is set to another symptom detailed variable. 213 or disease variable 211 evidence). Thereby, the inference apparatus 1 performs inference by the Bayesian network, and calculates the probability of the value of each disease variable 211.

<Operation of the inference device>
Hereinafter, the operation of the inference apparatus 1 will be described with reference to FIG. FIG. 6 is a flowchart showing the operation of the inference apparatus according to the embodiment. The inference device 1 operates based on the time measured by a timer (not shown). The operation of the inference device 1 is defined by an inference program 20 </ b> A stored in advance in a readable manner in the storage unit 20, and is realized by the control unit 10. Hereinafter, since these points are conventional means in information processing, detailed description thereof will be omitted.

(Inference processing when the patient's symptoms cannot be grasped in detail)
Here, first, an inference process in the case where the patient's symptoms cannot be grasped in detail will be described with reference to FIG. Note that “when the details cannot be grasped” means, for example, that an emergency crew member or doctor / nurse is unable to observe the patient directly, or that the patient falls into a consciousness disorder, etc. When the symptoms cannot be heard directly from the patient himself, there is a case where there is no time for inputting detailed information because the urgency level is determined to be high according to the finding determination.

  First, as shown in FIG. 6, in the inference apparatus 1, the main control unit 11 outputs a command for model acquisition processing to the inference unit 12 as a pre-process (S101). When an instruction for model acquisition processing is input, the inference unit 12 acquires model information 21A from the model holding unit 21 (S102). Thereby, the inference apparatus 1 enters a state of waiting for input of variables.

  Thereafter, the ambulance crew or the doctor / nurse inputs the patient's symptoms to the inference device 1 from the input screen 41A (see FIG. 7) displayed on the display of the input / output unit 3. FIG. 7 is a diagram illustrating an example of an input screen of the inference apparatus according to the embodiment. As shown in FIG. 7, items such as a symptom, a headache level, and a headache onset status are displayed on the input screen 41 </ b> A, and a plurality of selection buttons are displayed corresponding to each item.

  Here, it is assumed that the ambulance crew was unable to hear the symptoms from the patient himself. The ambulance crew member or doctor / nurse designates a selection button corresponding to the symptoms that the patient knows from the buttons displayed on the input screen 41A, and designates the decision button. Here, explanation will be made assuming that the ambulance crew designates the selection button of “headache”, “vomiting”, and “loss of consciousness” and designates the decision button.

  As shown in FIG. 6, when the selection button and the decision button are designated, the input / output unit 3 outputs information corresponding to the designated selection button to the main control unit 11 as input information (S103). As a result, the input / output unit 3 outputs the information “headache”, “vomiting”, and “loss of consciousness” to the main control unit 11 as input information.

When the input information is input, the main control unit 11 outputs an instruction (variable set instruction) for setting a value of a variable corresponding to the input information to the inference unit 12 (S104).
When the variable setting instruction is input, the inference unit 12 sets the input information as the symptom variable 212 and sets the value corresponding to the input information as the evidence of the symptom variable 212 (S105). Here, the inference unit 12 corresponds to the input information “headache”, “vomiting”, and “loss of consciousness” as the symptom variable 212 as a “headache” variable, “vomiting” variable, and “loss of consciousness” variable. Is set to “Y”. That is, the inference unit 12 sets “<headache> = <Y>”, “<vomiting> = <” as a combination of “variable” and “value” of the symptom variable 212 (“<variable> = <value>”). Y> ”and“ <disappointment> = <Y> ”are set. Each variable before the evidence is set is in a state shown in FIGS. 4A and 5A to 5C, for example.

Next, the main control unit 11 outputs an inference process execution instruction (inference instruction) to the inference unit 12 (S106).
When the inference command is input, the inference unit 12 performs inference using the input variable value (S107). The inference unit 12 calculates the probability value of each disease variable 211 such as “subarachnoid hemorrhage”, “cerebral infarction”, and “myocardial infarction” based on this inference. Here, for example, the inference unit 12 sets the value of the symptom variable 212 as evidence of the symptom detailed variable 213 based on the model information 21A illustrated in FIG. 8, and sets the value of the symptom detailed variable 213 obtained thereby as the symptom knowledge. The probability of the value “Y” of each disease variable 211 such as “subarachnoid hemorrhage”, “cerebral infarction”, “myocardial infarction (not shown)” is calculated by setting as evidence of the variable 214. FIG. 8 is a diagram for explaining the operation of the inference apparatus according to the embodiment.

  As a result, for example, the inference unit 12 calculates the probability of the value “Y” of the disease variable 211 of “subarachnoid hemorrhage” as 0.40 (that is, the probability of the value “N” is 0.60), The probability of the value “Y” of the disease variable 211 of “infarction” is calculated as 0.55 (that is, the probability of the value “N” is 0.45), and the value of the disease variable 211 of “myocardial infarction (not shown)” The probability of “Y” is calculated as 0.05 (that is, the probability of value “N” is 0.95), and the probability of the value “Y” of other disease variables 211 is calculated as 0.00. In the example shown in FIG. 8, the probability table of the symptom variable 211 has the probability of the set evidence value set to 1.00, and the probability table of the other variables is the result of the inference. Probability is set. Model information 21 </ b> A actually prototyped includes about 30 disease variables 211 and about 50 symptom variables 212.

Next, as shown in FIG. 6, the main control unit 11 outputs a command (variable acquisition command) for acquiring the value of each variable related to each disease to the inference unit 12 (S108).
When the variable acquisition command is input, the inference unit 12 outputs each probability value of each disease variable 211 to the main control unit 11 as a response result to the variable acquisition command (S109). Here, for example, the inference unit 12 sets the probability value of each disease variable 211 as “<subarachnoid hemorrhage” variable: value “Y” = 0.40: value “N” = 0.60>;<“cerebral infarction” Variable: value “Y” = 0.55: value “N” = 0.45>;<“myocardial infarction” variable: value “Y” = 0.05: probability of value “N” = 0.95> The value is output to the main control unit 11.

Finally, the main control unit 11 outputs a result display command to the input / output unit 3 via the display control unit 14 (see FIG. 1) (S110).
When the result display command is input, the input / output unit 3 displays the probability value of the value “Y” of each disease variable 211 on the display 41 as an output screen 41 </ b> B (see FIG. 9) representing the probability of each disease. (S111). FIG. 9 is a diagram illustrating an example of an output screen of the inference apparatus according to the embodiment. FIG. 9 shows an inference result when a patient's symptom cannot be grasped in detail (that is, an inference result when an inference process is performed in a state where information related to the patient's symptom as input information is insufficient). In the example shown in FIG. 9, the output screen 41 </ b> B has an input information display screen 41 </ b> Ba representing symptoms inputted as input information provided on the left side of the screen, and an inference result display that displays a list of the probability of each disease as a horizontal bar graph. The screen 41Bb is provided on the right side of the screen.

  Since the input information display screen 41Ba and the inference result display screen 41Bb are displayed on the same screen, the user of the inference apparatus 1 (here, an emergency worker or a doctor / nurse), for example, an inference result (here, When there are too many candidates for (disease), this can be detected, and the input information and the inference results are compared to determine what information is needed to narrow down the candidates. can do. As a result, the ambulance crew or doctor / nurse can quickly input additional information to narrow down candidates and obtain an appropriate treatment method for the patient. Note that the display method of the inference result display screen 41Bb is not limited to the bar graph, and may be another method (for example, a display method by percentage). Further, although not shown, the output screen 41B is provided with an end button for ending the inference process, a symptom input button for restarting the inference process, and the like.

  Thereafter, when the emergency member or doctor / nurse terminates the inference process, for example, the emergency member or doctor / nurse selects an end button (not shown). In addition, the emergency crew or the doctor / nurse selects a symptom input button (not shown) when resuming the inference process. The input / output unit 3 detects which button is selected by the emergency member or the doctor / nurse (S112).

  When the button detected in S112 is an end button (not shown), the input / output unit 3 outputs information (end information) indicating the end of the inference process to the main control unit 11. In response to this, the main control unit 11 ends the inference process.

  On the other hand, when the button detected in S112 is a symptom input button (not shown), the input / output unit 3 displays the input screen 41A (FIG. 7) on the display 41 again (S113). Thereafter, the inference apparatus 1 repeats the processes after S103.

(Inference processing when the patient's symptoms are understood in detail)
Next, with reference to FIG. 6, the inference process when the patient's symptoms can be grasped in detail will be described. Here, it is assumed that the ambulance crew was able to hear the symptoms in detail from the patient himself.

  As a result, the ambulance crew can select “headache”, “vomiting” from the buttons displayed on the input screen 41A based on the report from the patient and the measurement results such as the respiratory rate, pulse rate, and blood pressure. ”And“ disappearance ”selection buttons, for example, a“ severe ”selection button as a headache degree, an“ acute ”selection button as a headache onset condition, and a“ mild ”selection button as a degree of vomiting (see FIG. (Not shown), “respiratory normal” selection button (not shown), “tachycardia” selection button (not shown), “systolic blood pressure 160” selection button (not shown), and “diastolic blood pressure” Assume that a selection button (not shown) of “100” is added and designated.

  In S103, the input / output unit 3 outputs information corresponding to the designated selection button as input information to the main control unit 11. As a result, the input / output unit 3 has “headache”, “vomiting”, “loss of consciousness”, “(headache degree) severe”, “(headache onset situation) acute”, “(vomiting degree) mild”, “normal breathing” ”,“ Tachycardia ”,“ systolic blood pressure 160 ”, and“ diastolic blood pressure 100 ”are output to the main control unit 11 as input information.

In S104, when the input information is input, the main control unit 11 outputs an instruction (variable set instruction) for setting a value of a variable corresponding to the input information to the inference unit 12.
In S <b> 105, the inference unit 12 sets the input information as the symptom variable 212 and sets a value corresponding to the input information as evidence of the symptom variable 212.

  At that time, first, the inference unit 12 responds to the input information “headache”, “vomiting”, and “loss of consciousness” as the symptom variable 212 as the “headache” variable, “vomiting”. Each value of the variable and the “loss of consciousness” variable is set to “Y”. That is, the inference unit 12 sets “<headache> = <Y>”, “<vomiting> = <” as a combination of “variable” and “value” of the symptom variable 212 (“<variable> = <value>”). Y> ”and“ <disappointment> = <Y> ”are set.

  The inference unit 12 sets the value of the “abnormal breathing” variable to “N” as the symptom variable 212 in response to the input information “normal breathing”, and further, “tachycardia”, “systolic blood pressure”. Corresponding to the input information of “160” and “diastolic blood pressure 100”, as the symptom variable 212, the values of the “abnormal pulse” variable and the “abnormal blood pressure” variable are set to “Y”. That is, the inference unit 12 sets “<breathing abnormality> = <N>”, “<pulse abnormality>” as a combination of “variable” and “value” of the symptom variable 212 (“<variable> = <value>”). = <Y> ”and“ <blood pressure abnormality> = <Y> ”are set.

  Next, the inference unit 12 corresponds to input information of “(Headache degree) Severe”, “(Headache onset situation) Acute”, “Tachycardia”, “systolic blood pressure 160”, and “diastolic blood pressure 100”. As the symptom detail variable 213, the value of the “headache degree (pain degree)” variable is set to “severe”, the value of the “headache onset status (onset status)” variable is set to “acute”, and “pulse abnormality (pulse status)” The value of the “variable” variable is set to “tachycardia” and the value of the “blood pressure abnormality (blood pressure situation)” variable is set to “high blood pressure”. That is, the inference unit 12 sets “<headache degree (pain degree)> = <severe>” as a combination of “variable” and “value” of the symptom detailed variable 213 (“<variable> = <value>”). “<Headache onset status (onset status)> = <Acute>”, “<Pulse abnormality (pulse status)> = <Tachycardia>” and “<Blood pressure abnormality (blood pressure status)> = <High blood pressure>” are set .

Thereafter, in S106, the main control unit 11 outputs an inference process execution instruction (inference instruction) to the inference unit 12.
Then, in S107, the inference unit 12 performs inference using the input variable value. The inference unit 12 calculates the probability value of each disease variable 211 such as “subarachnoid hemorrhage”, “cerebral infarction”, and “myocardial infarction” based on this inference. Here, since the detailed information of the symptom is input as the input information, the inference unit 12 can select either the value of the symptom variable 212 or the value of the symptom detailed variable 213 based on the model information 21A illustrated in FIG. Probability of the value “Y” of each disease variable 211 such as “subarachnoid hemorrhage”, “cerebral infarction”, “myocardial infarction (not shown)” by setting either or both as evidence of the symptom knowledge variable 214 Calculate FIG. 10 is a diagram for explaining the operation of the inference apparatus according to the embodiment.

  As a result, the inference unit 12 calculates, for example, the probability of the value “Y” of the disease variable 211 of “subarachnoid hemorrhage” as 0.95 (that is, the probability of the value “N” is 0.05), and the “brain Other disease variables including “myocardial infarction (not shown)” calculated by setting the probability of the value “Y” of the disease variable 211 of “infarct” to 0.05 (that is, the probability of the value “N” being 0.95) The probability of the value “Y” of 211 is calculated as 0.00. In the example shown in FIG. 10, the probability table of the symptom variable 211 and the probability table of the symptom detail variable 213 are set to 1.00 for the probability of the set evidence value, and the probability table of each other variable is The probability as a result of inference is set.

Thereafter, in S108, the main control unit 11 outputs a command (variable acquisition command) for acquiring the value of each variable related to each disease to the inference unit 12.
Then, in S109, the inference unit 12 outputs each probability value of each disease variable 211 to the main control unit 11 as a response result to the variable acquisition command. Here, for example, the inference unit 12 sets the probability value of each disease variable 211 as “<subarachnoid hemorrhage” variable: value “Y” = 0.95: value “N” = 0.05>;<“cerebral infarction” ”Variable: value“ Y ”= 0.05: value“ N ”= 0.95> and the like are output to the main control unit 11.

Finally, in S110, the main control unit 11 outputs a result display command to the input / output unit 3 via the display control unit 14 (see FIG. 1).
Then, in S111, the input / output unit 3 displays the probability value of the value “Y” of each disease variable 211 on the display 41 as an output screen 41B (see FIG. 11) showing the probability of each disease. FIG. 11 is a diagram illustrating an example of an output screen of the inference apparatus according to the embodiment. FIG. 11 shows an inference result when a patient's symptom can be grasped in detail (that is, an inference result when an inference process is performed in a state where information related to the patient's symptom as input information is substantial). As shown in FIG. 11, the inference result when the patient's symptom can be grasped in detail is much higher than the inference result when the patient's symptom shown in FIG. 9 cannot be grasped in detail. It is high. Thus, the inference device 1 can execute more accurate inference as the amount of information as input information increases.

  As described above, the model information 21A includes a knowledge variable (symptom knowledge variable 214) and a detailed variable (symptom detailed variable 213) between a factor variable (disease variable 211) and a result variable (symptom variable 212). Each variable has a dependency from a factor variable (disease variable 211) to a knowledge variable (symptom knowledge variable 214), a dependency from a knowledge variable (symptom knowledge variable 214) to a detailed variable (symptom detailed variable 213), and details. A variable (symptom detail variable 213) is configured to have a dependency relationship from a result variable (symptom variable 212).

  In this model information 21A, knowledge about the correlation between the factor variable (disease variable 211) and the result variable (symptom variable 212) is reflected by the probability table of the knowledge variable (symptom knowledge variable 214).

  The inference device 1 uses the model information 21A to infer the probability of each factor variable (disease variable 211). At this time, the knowledge variable (symptom knowledge variable 214) causes a deviation between one or both of the result variable (symptom variable 212) and the detailed variable (symptom detailed variable 213) and the factor variable (disease variable 211). Functions to absorb. Therefore, this inference apparatus can execute a suitable inference even when input information is insufficient by using knowledge variables. Moreover, a probability value is set in the knowledge variable (symptom knowledge variable 214) separately from the detailed variable (symptom detailed variable 213). Therefore, the user of the inference apparatus 1 can easily create the model information 21A reflecting the specialized knowledge, and can easily adjust the probability value set in the model information 21A.

  In addition, the inference device 1 knows the value of one or both of the result variable (symptom variable 212) and the detailed variable (symptom detailed variable 213) when there is a lot of input information (in this case, information on symptoms). Set as evidence of variable (symptom knowledge variable 213). On the other hand, when the input information is small, the inference apparatus 1 sets the value of the result variable (symptom variable 212) as evidence of the knowledge variable (symptom knowledge variable 213). Thereby, even if there is little input information, the inference apparatus 1 can perform the inference process similar to the case where there is much input information, and can thereby obtain an accurate inference result.

As described above, according to the inference device 1 according to the present embodiment, even when the input information is small, the inference processing can be executed in the same manner as when the input information is large, thereby obtaining an accurate inference result. Can do.
Moreover, according to the inference apparatus 1, it is possible to easily create the model information 21A reflecting the specialized knowledge, and to easily adjust the probability value set in the model information 21A.

  Moreover, according to the inference apparatus 1, by using a knowledge variable (symptom knowledge variable 213) for inference, the discrepancy between the subjective recognition of the observer and the specialized knowledge is eliminated, and an accurate inference result is obtained. Can do. For example, when the patient becomes an observer and the patient reports the degree of pain such as headache or chest pain to an ambulance crew or a doctor / nurse, the patient may endure severe pain and lightly report Sometimes pains are reported exaggeratedly. In addition, when a patient declares a pain site, the patient may complain of abdominal pain and chest pain in the epigastric region. Since the inference apparatus 1 functions so that the knowledge variable (symptom knowledge variable 213) absorbs these deviations, an accurate inference result can be obtained.

  The present invention is not limited to the above-described embodiment, and various changes and modifications can be made without departing from the gist of the present invention.

  For example, in the embodiment, the case where the inference device 1 is used in a diagnostic device for the medical field that estimates a disease from a patient's symptom has been described as an example. However, the reasoning device 1 may, for example, use a disease variable as a failure factor variable (such as a “power supply plug not inserted” variable or a “disk drive failure” variable), and a symptom variable as a phenomenon result variable (such as a “power-on not possible” variable) "Non-renewable" variable), symptom knowledge variable to phenomenon knowledge variable ("Power lamp lighting" variable, "Disk drive rotation" variable, etc.), symptom detailed variable to phenomenon detailed variable ("Power lamp lighting color") Variable, “disk drive rotation sound” variable, etc., and depending on the input information, set either the phenomenon variable or the phenomenon detail variable as evidence of the phenomenon knowledge variable, It is also possible to use it as a failure diagnosis device by using a configuration that outputs a probability.

1 Reasoning device (computer)
DESCRIPTION OF SYMBOLS 3 Input / output part 10 Control part 11 Main control part 12 Inference part 13 Display control part 14 Communication control part 20 Storage part 20A Inference program (control program)
21 Model holding unit 21A Bayesian network model (model information)
21a Factor variable 21b Knowledge variable 21c Detailed variable 21d Result variable 30 Input unit 40 Output unit 50 Communication unit 60 Recording medium 211 Disease variable (factor variable)
212 Syndrome Knowledge Variable (Knowledge Variable)
213 Detailed symptoms variables (detailed variables)
214 Symptom variables (result variables)

Claims (6)

In an inference device that performs inference by a Bayesian network,
A model holding unit in which a Bayesian network model is held in advance;
Based on the Bayesian network model held by the model holding unit, an inference unit that infers the probability of a factor variable that is a factor of a result variable related to an observation result;
An input unit that provides the inference unit with one or more pieces of input information as a condition for inference;
An output unit that outputs a probability of a factor variable calculated as an execution result of inference by the inference unit;
The Bayesian network model includes one or more expert knowledge related to a correlation between the result variable and the factor variable between the one or more result variables and the one or more factor variables. A knowledge variable and one or more detail variables describing details of the result variable, each variable having a dependency from each of the factor variables to one or more knowledge variables, one or more A dependency relationship from the knowledge variable to one or more detailed variables and a dependency relationship from the one or more detailed variables to each of the result variables,
Based on the Bayesian network, the inference unit uses the value of one or both of the result variable and the detailed variable as evidence according to one or more pieces of the input information provided from the input unit. An inference device characterized in that the probability of each of the factor variables is set and calculated. The inference device according to claim 1,
The factor variable probability table is configured to take a binary value or a plurality of predetermined variables,
The result variable probability table is configured such that a conditional variable is a value of the detailed variable and a possible value is a binary value.
The detailed variable probability table is configured such that a conditional variable is a value of the knowledge variable, and a possible value is a value of a combination of one or more predetermined variables.
The knowledge variable probability table is configured such that a conditional variable is a value of the factor variable, and a possible value is a predetermined combination of one or more variables. . In the inference device according to claim 1 or 2,
The output unit is configured as a display,
The display displays one or more result variables in a selectable state as an input screen for the input information,
The input unit provides the inference unit with the selected result variable as the input information to the inference unit. The inference device according to any one of claims 1 to 3,
The output unit is configured as a display,
The display device displays an input information display screen for displaying the input information used for inference and an inference result display screen for displaying an inference result by the inference unit as output screens. The inference device according to claim 4,
The display displays a list of probabilities of the factor variables calculated by the inference unit as the inference result display screen. Expert knowledge about the correlation between the result variable and the factor variable is described between one or more result variables related to the observation result and one or more factor variables that cause the result variable 1 Including one or more knowledge variables and one or more detail variables describing details of the result variable, each variable having a dependency from each of the factor variables to one or more of the knowledge variables; A Bayesian network model configured to have a dependency from one or more of the knowledge variables to one or more of the detailed variables, and a dependency of one or more of the detailed variables to each of the result variables. A program that causes a computer having a model holding unit held in advance to function as an inference device that executes inference by a Bayesian network,
The computer,
Based on one or more input information provided from the input unit based on the Bayesian network model held by the model holding unit, the value of one or both of the result variable and the detailed variable is set. An inference part for inferring the probability of the factor variable by setting as evidence and calculating the probability of each factor variable;
The input unit providing the inference unit with one or more pieces of the input information which are conditions for inferring;
A program that functions as an output unit that outputs a probability of a factor variable calculated as an execution result of inference by the inference unit.

Images