Diagnosis assistance device, diagnosis assistance system, and program

Abstract

Provided is a diagnosis assistance device (100) provided with: a data acquisition unit (201) that acquires medical data pertaining to an examined subject; a score calculation unit (202) that calculates a risk score indicating a possibility of having an injury / disease from the medical data, by using an AI model (300) created through a learning process using training medical data; and a determination unit (203) that determines a classification class to which the risk score belongs from among three or more classification classes, by using a diagnostic ability index of the AI model.

Description

Diagnostic support device, diagnostic support system, and program

[0001] The present invention relates to a diagnostic support device, a diagnostic support system, and a program.

[0002] In recent years, the practical application of AI (Artificial Intelligence) technology has progressed rapidly, and methods have been proposed for automatically analyzing electrocardiogram waveforms by treating them as time-series data or images and using machine learning or deep learning (AI learning) models. Patent Document 1 proposes a method for generating electrocardiogram analysis data that enables efficient deep learning of relationships between multiple types of waveform information and leads in an electrocardiogram. Patent Document 2 proposes a method for generating data representing a two-dimensional image having a waveform image region and a rhythm image region as electrocardiogram analysis data.

[0003] JP 2020-130772 A JP 2022-054202 A

[0004] The classification model obtained by learning represents the probability that a test subject has an injury or disease as a continuous risk score ranging from 0 to 1. The higher the risk score, the higher the probability that the test subject has an injury or disease. However, this only indicates a qualitative trend. Even if the same risk score is output, different AI model performance will result in different diagnostic capabilities. Therefore, it is difficult for medical professionals to accurately determine the probability that a test subject has an injury or disease after an AI test based solely on the risk score value. Some aspects of the present invention aim to provide a technology that enables medical professionals to more easily determine the probability that a test subject has an injury or disease after the test.

[0005] According to some embodiments, a diagnostic assistance device is provided, which includes a data acquisition unit that acquires medical data on a subject to be examined, a score calculation unit that calculates a risk score representing the possibility of the subject having an injury or illness for the medical data using an AI model created by learning using teacher medical data, and a determination unit that determines the classification class to which the risk score belongs from three or more classification classes using a diagnostic performance index of the AI ​​model.

[0006] Some embodiments allow medical personnel to more easily determine the probability that a test subject has a disease or injury following a test.

[0007] Other features and advantages of the present invention will become apparent from the following description taken in conjunction with the accompanying drawings, in which the same or similar elements are designated by the same reference numerals.

[0008] The accompanying drawings are incorporated into and constitute a part of the specification, illustrate embodiments of the present invention, and are used, together with the description, to explain the principles of the present invention.

[0023] Figure 1 is a block diagram illustrating an example of the hardware configuration of a diagnostic assistance device according to a first embodiment. Figure 2 is a block diagram illustrating an example of the functional configuration of a diagnostic assistance device according to a first embodiment. Figure 3 is a block diagram illustrating an example of an AI model according to a first embodiment. Figure 4 is a schematic diagram illustrating an example of a classification class according to a first embodiment. Figure 5 is a diagram illustrating an example of a method for determining a recommended action according to a first embodiment. Figure 6 is a flow diagram illustrating an example of a method executed by the diagnostic assistance device according to a first embodiment. Figure 7 is a flow diagram illustrating an example of a method executed by the diagnostic assistance device according to a first embodiment. Figure 8 is a schematic diagram displaying an example of a screen according to a first embodiment. Figure 9 is a block diagram illustrating an example of the configuration of a diagnostic assistance system according to a first embodiment. Figure 10 is a block diagram illustrating an example of the hardware configuration of a diagnostic assistance device according to a second embodiment. Figure 11 is a block diagram illustrating an example of a processing flow executed by a processor of a diagnostic assistance device according to a second embodiment. Figure 12 is a schematic diagram illustrating an example of a classification class output in a second embodiment. Figure 13 is a schematic diagram illustrating an example of an AI output display according to a second embodiment. Figure 14 is a schematic diagram illustrating another example of an AI output display according to a second embodiment. Figure 15 is a diagram illustrating an example of a method for determining a recommended action according to a second embodiment. 10 is a block diagram illustrating an example of the configuration of a diagnostic assistance system according to the second embodiment.

[0009] Hereinafter, the embodiments will be described in detail with reference to the accompanying drawings. Note that the following embodiments do not limit the scope of the invention as claimed, and not all combinations of features described in the embodiments are necessarily essential to the invention. Two or more of the features described in the embodiments may be combined in any desired manner. Furthermore, the same reference numerals are used to designate identical or similar components, and redundant descriptions will be omitted.

[0010] <First Embodiment> The first embodiment of the present invention relates to a diagnostic support device. A diagnostic support device is a device that supports diagnoses performed by medical professionals such as doctors, dentists, and veterinarians. The person to be diagnosed (hereinafter referred to as the test subject) may be a human or another animal (e.g., livestock or pet). A person diagnosed with an injury or illness may also be called a patient.

[0011] Diagnosis is the determination of whether a test subject has an illness or injury. For example, it may be determined whether a test subject has a specific illness or injury, or it may be determined whether a test subject has some illness or injury, even if the name of the illness or injury cannot be identified. In the following description, if a test subject has an illness or injury, it is referred to as the test subject being positive, and if a test subject does not have an illness or injury, it is referred to as the test subject being negative.

[0012] The diagnosis support device assists in diagnosis using medical data related to the subject. Medical data is data that can serve as the basis for diagnosis. For example, the medical data may include data collected from the subject using a testing device (e.g., electrocardiogram data, X-rays, blood test results, height, weight). The medical data may include findings on the subject by a medical professional. The medical data may include attributes of the subject (e.g., age, gender, lifestyle, medical history). The medical data may include family history or genetic information of the subject, and may include attributes of the subject's blood relatives (e.g., parents, brothers, sisters).

[0013] [Example of Hardware Configuration of Diagnosis Support Device] An example of the hardware configuration of the diagnosis support device 100 according to the first embodiment will be described with reference to FIG. 1 . The diagnosis support device 100 includes the components shown in FIG. 1 . The diagnosis support device 100 may include only one of each component shown in FIG. 1 , or may include multiple components. The diagnosis support device 100 may not include some of the components shown in FIG. 1 , or may include components not shown in FIG. 1 . For example, the diagnosis support device 100 may include an inspection device for collecting medical data from a subject. In this case, the diagnosis support device 100 may be called an inspection device having a diagnosis support function.

[0014] The processor 101 is a device that controls the overall operation of the diagnostic support device 100. The processor 101 is configured, for example, by a CPU (Central Processing Unit). The memory 102 is a device that stores programs and temporary data required for the operation of the diagnostic support device 100. The memory 102 is configured, for example, by a RAM (Random Access Memory) or a ROM (Read Only Memory). The operation of the diagnostic support device 100 may be performed, for example, by the processor 101 executing a program stored in the memory 102. Alternatively, some or all of the operation of the diagnostic support device 100 may be performed by a dedicated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array). Furthermore, the processor 101 may include a GPU (Graphical Processing Unit). A device having the processor 101 and the memory 102, such as the diagnostic support device 100, is sometimes called a computer.

[0015] The input device 103 is a device for acquiring input from a user of the diagnosis assistance device 100 (e.g., a medical professional; hereinafter, the user of the diagnosis assistance device 100 will be simply referred to as the user). The input device 103 is configured, for example, by a keyboard, a mouse, a touch panel, etc. The output device 104 is a device for outputting to the user. The output device 104 is configured, for example, by a display and a speaker. In the example of FIG. 1 , the input device 103 is described as constituting a part of the diagnosis assistance device 100, but the input device 103 may be a device separate from the diagnosis assistance device 100. In this case, the diagnosis assistance device 100 includes an input interface for connecting to the input device 103. Similarly, the output device 104 may be configured as a part of the diagnosis assistance device 100, or may be a device separate from the diagnosis assistance device 100.

[0016] The communication device 105 is a device that enables the diagnosis assistance device 100 to communicate with other devices. The other devices may be computers connected to a network (e.g., the Internet or a local area network). For example, the communication device 105 may be used to acquire medical data from the other devices.

[0017] The storage device 106 is a device that stores data used in the operation of the diagnostic assistance device 100. The storage device 106 is configured by a storage medium such as a hard disk drive (HDD), a solid state drive (SSD), or a DVD (digital versatile disc). At least a portion of the data stored in the storage device 106 may be stored in a device external to the diagnostic assistance device 100, such as a cloud environment, and may be read out during processing by the diagnostic assistance device 100.

[0018] [Example of Functional Configuration of Diagnostic Support Device] An example of the functional configuration of the diagnostic support device 100 will be described with reference to Fig. 2. The diagnostic support device 100 may not include some of the functions shown in Fig. 2, or may include functions not shown in Fig. 2. Each function of the diagnostic support device 100 may be realized by the processor 101 executing a program loaded into the memory 102. At least some of the functions of the diagnostic support device 100 may be realized by a dedicated circuit such as an ASIC or FPGA.

[0019] The data acquisition unit 201 acquires data used in the diagnosis assistance device 100. For example, the data acquisition unit 201 acquires medical data related to the subject. The data acquisition unit 201 may acquire medical data input by a user to the input device 103. The data acquisition unit 201 may acquire medical data from a device external to the diagnosis assistance device 100 (e.g., a file server or an examination device) using the communication device 105. If the diagnosis assistance device 100 includes an examination device, the data acquisition unit 201 may acquire medical data from the examination device. The medical data acquired by the data acquisition unit 201 may be stored in the memory 102 or the storage device 106 for subsequent processing.

[0020] The data acquisition unit 201 also acquires verification medical data. The verification medical data is medical data labeled with whether the provider of the verification medical data is positive or negative. The provider of the verification medical data is the person who created the medical data. For example, the medical data may include data about the provider that is similar to the medical data about the test subject described above.

[0021] The verification medical data is used to calculate a likelihood ratio of the AI ​​model 300 (described below) used by the diagnosis support device 100. The verification medical data has the same data structure as the medical data used to diagnose the test subject, except that it is labeled. The data acquisition unit 201 may acquire medical data from a device external to the diagnosis support device 100 (e.g., a file server) using the communication device 105. The data acquisition unit 201 acquires multiple pieces of verification medical data. The multiple pieces of verification medical data may include medical data acquired from the same provider at different times, or may include medical data acquired from different providers. The verification medical data acquired by the data acquisition unit 201 may be stored in the memory 102 or the storage device 106 for subsequent processing.

[0022] The score calculation unit 202 calculates a risk score that indicates the possibility that the person of the medical data has an injury or disease (i.e., is positive) using the AI ​​model 300. The person of the medical data may be a test subject or a provider of verification medical data.

[0023] As shown in Figure 3, the AI ​​model 300 outputs a risk score for the input medical data. In the following description, the risk score can take a value in the range of 0 to 1. Alternatively, the risk score may take a value in another range. The higher the risk score, the more likely the person in the medical data is positive, and the lower the risk score, the more likely the person in the medical data is negative.

[0024] The AI ​​model 300 is created by machine learning using training medical data. Machine learning is a technique in which a computer independently finds patterns potentially contained in data and uses those patterns to perform regression, classification, and the like. In this specification, machine learning includes deep learning. Machine learning may also be referred to as AI learning. The AI ​​model 300 may have any structure. For example, the AI ​​model 300 may be configured as a neural network, a convolutional neural network, or a transformer.

[0025] The teacher medical data is medical data labeled as positive or negative by the provider of the teacher medical data. The teacher medical data has the same data structure as the verification medical data described above. The teacher medical data is used to train the AI ​​model 300 through machine learning. The score calculation unit 202 calculates a risk score using the trained AI model 300. The machine learning of the AI ​​model 300 may be performed by the diagnosis support device 100 or by another device.

[0026] In one example, the medical data is electrocardiogram data, and the risk score represents the likelihood that the person of the electrocardiogram data has paroxysmal atrial fibrillation (PAF). In this case, the teacher medical data includes electrocardiogram data of a patient with paroxysmal atrial fibrillation (i.e., electrocardiogram data labeled as being positive for paroxysmal atrial fibrillation) and electrocardiogram data of a donor who does not have paroxysmal atrial fibrillation (i.e., electrocardiogram data labeled as being negative for paroxysmal atrial fibrillation).

[0027] Diseases such as PAF, which manifest spontaneously or suddenly, are known to be difficult to detect because symptoms often do not appear at the time of testing. For example, depending on the timing of acquisition of electrocardiogram data from a PAF patient, the electrocardiogram data may show sinus rhythm, making PAF difficult to detect. This type of PAF is sometimes called "hidden atrial fibrillation." Using the AI ​​model 300, it is possible to determine with a certain degree of accuracy whether a test subject is positive for PAF.

[0028] The class determination unit 203 determines, from three or more classification classes, the classification class to which the risk score calculated by the AI ​​model 300 belongs. The three or more classification classes divide the range that the risk score can take (i.e., the range from 0 to 1). Below, a case will be described in which the class determination unit 203 determines, from four classification classes, the classification class to which the risk score belongs. The same explanation applies when the class determination unit 203 determines, from three or five or more classification classes, the classification class to which the risk score belongs.

[0029] The four classification classes will be described with reference to Figure 4. The range that the risk score can take (i.e., the range from 0 to 1) is divided into four ranges by three boundary values ​​BL, BM, and BH. The boundary value BL (first boundary value) is greater than 0 and less than the boundary value BM. The boundary value BM (third boundary value) is greater than the boundary value BL and less than the boundary value BH. The boundary value BH (second boundary value) is greater than the boundary value BM and less than 1. The method for determining the boundary values ​​BL, BM, and BH will be described later.

[0030] The class determination unit 203 determines that a risk score belongs to the classification class "low" when the risk score is included in a range equal to or greater than 0 and less than the boundary value BL. The class determination unit 203 determines that a risk score belongs to the classification class "middle-low" when the risk score is included in a range equal to or greater than the boundary value BL and less than the boundary value BM. The class determination unit 203 determines that a risk score belongs to the classification class "middle-high" when the risk score is included in a range equal to or greater than the boundary value BM and less than the boundary value BH. The class determination unit 203 determines that a risk score belongs to the classification class "high" when the risk score is included in a range equal to or greater than the boundary value BH and less than 1. In the above example, when a risk score is equal to a boundary value, the risk score is determined to belong to a class determination having the boundary value as its lower limit. Alternatively, when a risk score is equal to a boundary value, the risk score may be determined to belong to a class determination having the boundary value as its upper limit. As shown in FIG. 4 , the range of possible risk scores is unequally divided by the four classification classes. Alternatively, the range of possible risk scores may be evenly divided.

[0031] The class presenting unit 204 presents the classification class determined by the class determining unit 203 to the user. A specific example of a method for presenting the classification class to the user will be described later. The action presenting unit 205 presents a recommended action to the user based on the classification class determined by the class determining unit 203. For example, when the risk score is determined to belong to the classification class "high," the action presenting unit 205 recommends the user to consider conducting a detailed examination on the test subject. When the risk score is determined to belong to the classification class "middle-high," the action presenting unit 205 recommends the user to consider conducting regular examinations on the test subject. When the risk score is determined to belong to the classification class "middle-low," the action presenting unit 205 recommends the user to consider conducting a simple examination that is less burdensome on the test subject. When the risk score is determined to belong to the classification class "low," the action presenting unit 205 recommends the user not to conduct additional examinations on the test subject. A specific example of a method for presenting recommended actions to the user will be described later.

[0032] The diagnostic performance index calculation unit 206 calculates the diagnostic performance index of the AI ​​model 300 using the verification medical data. The diagnostic performance index is an index that represents the diagnostic ability of the AI ​​model 300. Specifically, the diagnostic performance index calculation unit 206 calculates the sensitivity, specificity, positive likelihood ratio, and negative likelihood ratio of the AI ​​model 300 as the diagnostic performance index. Sensitivity is a diagnostic performance index that represents the ability of the AI ​​model 300 to determine a positive individual as positive. Specificity is a diagnostic performance index that represents the ability of the AI ​​model 300 to determine a negative individual as negative. The positive likelihood ratio is a diagnostic performance index that represents how likely a positive individual is to be positive compared to a negative individual. The positive likelihood ratio is sometimes expressed as "LR+". The negative likelihood ratio is a diagnostic performance index that represents how likely a positive individual is to be negative compared to a negative individual. The negative likelihood ratio is sometimes expressed as "LR-". The positive likelihood ratio and negative likelihood ratio of the AI ​​model 300 are calculated based on the sensitivity and specificity using the following formulas: Positive likelihood ratio = sensitivity / (1 - specificity), Negative likelihood ratio = (1 - sensitivity) / specificity.

[0033] The sensitivity, specificity, positive likelihood ratio, and negative likelihood ratio of the AI ​​model 300 vary depending on the cutoff value. Therefore, the diagnostic performance index calculation unit 206 calculates the sensitivity, specificity, positive likelihood ratio, and negative likelihood ratio of the AI ​​model 300 for each of multiple cutoff values ​​based on the risk score. The specificity, positive likelihood ratio, and negative likelihood ratio monotonically increase as the cutoff value increases. The sensitivity monotonically decreases as the cutoff value increases.

[0034] The boundary value determination unit 207 determines boundary values ​​of three or more classification classes for classifying the risk scores. For example, the boundary value determination unit 207 determines the three boundary values ​​BL, BM, and BH shown in Fig. 4. The boundary value determination unit 207 may determine the boundary values ​​BL, BM, and BH based on the diagnostic performance index (specifically, sensitivity, specificity, positive likelihood ratio, and negative likelihood ratio) of the AI ​​model 300 calculated by the diagnostic performance index calculation unit 206.

[0035] For example, the boundary value determination unit 207 sets the largest cutoff value among the multiple cutoff values ​​used to calculate the diagnostic performance index of the AI ​​model 300, at which the negative likelihood ratio is equal to or less than a first predetermined value A, as the boundary value BL. The first predetermined value A is a likelihood ratio that is clinically useful for determining a negative result. The boundary value determination unit 207 may also set another value based on the negative likelihood ratio (for example, the cutoff value among the multiple cutoff values ​​at which the negative likelihood ratio is closest to the above-mentioned first predetermined value A) as the boundary value BL.

[0036] The boundary value determination unit 207 sets the smallest cutoff value, among the multiple cutoff values ​​used to calculate the diagnostic performance index of the AI ​​model 300, at which the positive likelihood ratio is equal to or greater than the second predetermined value B as the boundary value BH. The second predetermined value B is a likelihood ratio that is clinically useful for determining a positive result. The boundary value determination unit 207 may also set another value based on the positive likelihood ratio (for example, the cutoff value, among the multiple cutoff values, at which the positive likelihood ratio is closest to the above-mentioned second predetermined value) as the boundary value BH.

[0037] The boundary value determination unit 207 determines the boundary value BM to be, for example, an optimal cutoff value that maximizes the sum of sensitivity and specificity among the multiple cutoff values ​​used to calculate the diagnostic performance index of the AI ​​model 300. The boundary value determination unit 207 may also determine another value between the boundary value BL and the boundary value BH (for example, the average value of the boundary value BL and the boundary value BH) as the boundary value BM. For example, the boundary value determination unit 207 may determine the boundary value BM based on a value that is based on sensitivity and specificity among the multiple cutoff values ​​used to calculate the diagnostic performance index of the AI ​​model 300 and is different from the likelihood ratio.

[0038] When the class determination unit 203 determines the classification class to which the risk score belongs from the three classification classes, any one of the three boundary values ​​BL, BM, and BH described above does not need to be used. For example, when the classification class "Low" and the classification class "Middle-Low" are combined, the boundary value BL is not used. When the classification class "Middle-Low" and the classification class "Middle-High" are combined, the boundary value BM is not used. When the classification class "Middle-High" and the classification class "High" are combined, the boundary value BH is not used.

[0039] As described above, the boundary value determination unit 207 determines multiple boundary values ​​based on the diagnostic performance index of the AI ​​model 300. The boundary values ​​divide the range that the risk score can take into three or more classification classes. Therefore, the range that the risk score can take is generally divided unevenly by the three or more classification classes.

[0040] The boundary value determined based on the diagnostic performance index of the AI ​​model 300 may vary depending on the performance of the AI ​​model 300. For example, suppose that the boundary value BH is determined using the same N pieces of validation medical data for an AI model with relatively high performance (hereinafter referred to as a high-performance AI model) and an AI model with relatively low performance (hereinafter referred to as a low-performance AI model). In this case, the boundary value BH determined for the high-performance AI model will be lower than the boundary value BH determined for the low-performance AI model. As a result, even if the high-performance AI model and the low-performance AI model output the same risk score, the high-performance AI model is more likely to belong to the classification class "high." In this way, in this embodiment, by determining the boundary value based on the diagnostic performance index of the AI ​​model 300, the risk score to which the classification class belongs can be accurately determined.

[0041] The reliability of the risk score output by the AI ​​model 300 varies depending on the performance of the AI ​​model 300. For example, suppose that both an AI model with relatively high performance and an AI model with relatively low performance output a risk score of "0.8." In this case, even if the same risk score is output, referring to the AI ​​model with relatively high performance increases the likelihood of correctly diagnosing the test subject. According to this embodiment, multiple boundary values ​​are determined based on the diagnostic performance index of the AI ​​model 300. Therefore, the values ​​of the multiple boundary values ​​may differ depending on the performance of the AI ​​model 300.

[0042] The probability value acquisition unit 208 acquires a prior probability value of an injury or disease based on the user's findings on the test subject. The prior probability value is a probability value that the test subject will test positive, determined by the user based on the test subject's findings before or after the diagnosis assistance device 100 presents the user with a risk score. The probability value acquisition unit 208 may acquire the prior probability value from the user using, for example, the input device 103. The probability value acquisition unit 208 may determine whether to acquire the prior probability value of an injury or disease, and acquire the prior probability value of an injury or disease from the user if it is determined that the prior probability value of an injury or disease should be acquired. The probability value acquisition unit 208 may determine whether to acquire the prior probability value of an injury or disease based on, for example, attribute information of the test subject (age, gender, medical history, test history), whether the test subject has a pacemaker, whether noise was mixed into the test subject's waveform during the test, and, if noise was mixed into the waveform, the type of noise.

[0043] The action presenting unit 205 may present a recommended action to the user based on the classification class determined by the class determining unit 203 and the prior probability value acquired by the probability value acquiring unit 208. For example, the action presenting unit 205 may determine an action by referring to a table 500 in FIG. 5. The table 500 is generated by the manufacturer of the diagnosis assistance device 100 and stored in the memory 102 or the storage device 106 of the diagnosis assistance device 100.

[0044] The rows of table 500 represent the classification classes determined by the class determination unit 203. The columns of table 500 represent whether the prior probability value acquired by the probability value acquisition unit 208 is higher or lower than a predetermined threshold (e.g., 50%), and may be configured with one column or two or more columns. "Detailed Examination" in table 500 indicates that the recommended action is to consider conducting a detailed examination for the test subject. "Regular Examination" in table 500 indicates that the recommended action is to consider conducting a regular examination for the test subject. "Simple Examination" in table 500 indicates that the recommended action is to consider conducting a simple examination that is less burdensome for the test subject. "No Examination" in table 500 indicates that the recommended action is not to conduct an additional examination for the test subject. The recommended actions may include other actions not shown in table 500.

[0045] [Method for Determining Boundary Values ​​of Classification Classes] An example of a method for determining boundary values ​​of classification classes will be described with reference to Fig. 6. The method of Fig. 6 is executed by the diagnosis assistance device 100. The method of Fig. 6 may be started in response to an instruction to execute the method of Fig. 6 from a user of the diagnosis assistance device 100.

[0046] In S601, the data acquisition unit 201 of the diagnosis assistance device 100 acquires verification medical data as described with reference to FIG.

[0047] In S602, the diagnostic performance index calculation unit 206 of the diagnosis assistance device 100 calculates diagnostic performance indexes (e.g., sensitivity, specificity, positive likelihood ratio, and negative likelihood ratio) of the AI ​​model 300 using the verification medical data acquired in S601, as described with reference to Fig. 2. The AI ​​model 300 may be stored in the memory 102 or the storage device 106 at the time of executing S602, or may be acquired from an external device at the time of executing S602.

[0048] In S603, the boundary value determination unit 207 of the diagnosis support device 100 determines a plurality of boundary values ​​(e.g., boundary values ​​BL, BM, and BH) of the classification classes based on the diagnostic performance index calculated in S602, as described with reference to Fig. 2. The boundary value determination unit 207 stores the determined plurality of boundary values ​​in the memory 102 or the storage device 106 for subsequent processing.

[0049] When multiple AI models 300 are available, the diagnostic support device 100 may determine multiple boundary values ​​by executing S602 and S603 for each AI model 300. In this case, the diagnostic support device 100 stores the corresponding multiple boundary values ​​in association with the AI ​​model 300.

[0050] [Method for assisting diagnosis] An example of a method for assisting diagnosis will be described with reference to Fig. 7. The method of Fig. 7 is executed by the diagnosis assistance device 100. The method of Fig. 7 may be started in response to an instruction to execute the method of Fig. 7 from a user of the diagnosis assistance device 100.

[0051] An example of a screen 800 for obtaining instructions from a user to execute the method of Fig. 7 will be described with reference to Fig. 8. The screen 800 is displayed on, for example, the output device 104 of the diagnosis assistance device 100. An input field 801 is a graphic object for obtaining designation of medical data from the user. The medical data may be designated by, for example, the address of a file storing the medical data.

[0052] The input field 802 is a graphic object for obtaining a prior probability value from the user. The prior probability value is specified as, for example, a value between 0 and 1. If the prior probability value is not used to determine the recommended action, the screen 800 does not need to include the input field 802.

[0053] The button 803 is a graphic object for obtaining an instruction from the user to execute the method of Fig. 7. The diagnosis assistance device 100 starts the method of Fig. 7 in response to the button 803 being pressed by the user.

[0054] In S701, the data acquisition unit 201 of the diagnosis support device 100 acquires medical data of the test subject. The address of the medical data of the test subject is specified in the input field 802. In S702, the score calculation unit 202 of the diagnosis support device 100 calculates a risk score for the medical data acquired in S701, as described with reference to FIG. 2 . The AI ​​model 300 may be stored in the memory 102 or the storage device 106 at the time of executing S702, or may be acquired from an external device at the time of executing S702. If the diagnosis support device 100 is capable of using multiple AI models 300, the diagnosis support device 100 may acquire, from the user, designation of the AI ​​model 300 to be used.

[0055] In S703, the class determination unit 203 of the diagnosis assistance device 100 determines the classification class to which the risk score calculated in S702 belongs, as described with reference to Fig. 2. In S703, the boundary value determined by the method of Fig. 6 and corresponding to the AI ​​model 300 used in S702 is used.

[0056] In S704, the class presenting unit 204 of the diagnosis assistance device 100 presents the classification class determined in S703 to the user. Furthermore, the action presenting unit 205 of the diagnosis assistance device 100 presents recommended actions to the user based on the classification class determined in S703 (and the prior probability value if entered in the input field 802).

[0057] Referring again to Figure 8, an example of a screen 810 used to present classification classes and recommended actions to a user will be described. Screen 810 includes bar-shaped regions 811 representing multiple classification classes (four in the example of Figure 8). Region 811 is divided into a portion 811L corresponding to the classification class "low," a portion 811ML corresponding to the classification class "middle-low," a portion 811MH corresponding to the classification class "middle-high," and a portion 811H corresponding to the classification class "high." These four portions are arranged horizontally in the order shown in Figure 8.

[0058] The screen 810 includes an indicator 812 that indicates the classification class determined in S703. In the example of FIG. 8, the indicator 812 has a triangular shape. The indicator 812 may have other shapes. In the example of FIG. 8, the indicator 812 is displayed in a position that points from above to the portion 811MH. This indicates to the user that the risk score was determined to belong to the classification class "middle-high" in S703.

[0059] The indicator 812 may present to the user the relative position of the risk score calculated for the test subject relative to the boundary values ​​of the four classification classes (the boundary values ​​BL, BM, and BH mentioned above). Specifically, the left end of the portion 811L represents the position where the risk score is 0. The boundary line between the portion 811L and the portion 811ML represents the position where the risk score is the boundary value BL. The boundary line between the portion 811ML and the portion 811MH represents the position where the risk score is the boundary value BM. The boundary line between the portion 811MH and the portion 811H represents the position where the risk score is the boundary value BH. The right end of the portion 811H represents the position where the risk score is 1.

[0060] As shown in FIG. 8 , assume that the risk score is determined to belong to the classification class "middle-high," and the risk score divides the lower limit (i.e., boundary value BM) and upper limit (i.e., boundary value BH) of the classification class "middle-high" internally at a:b. In this case, indicator 812 is positioned to indicate the position where the left and right ends of portion 811MH are internally divided at a:b. By positioning indicator 812 in this manner, the user can recognize how close the risk score is to other classification classes, making it easier for the user to make a diagnosis.

[0061] The screen 810 may include an indicator 813 that indicates the relative position of a predetermined positive likelihood ratio (1, 2, 3.25, 5, and 10 in the example of FIG. 8 ) of the AI ​​model 300 relative to the boundary values ​​of the four classification classes (the above-mentioned boundary values ​​BL, BM, and BH). In the example of FIG. 8 , the indicator 813 has a bar shape. The indicator 813 is added to the top of the region 811. "LR+" on the screen 810 indicates that the indicator 813 added to the top of the region 811 represents a positive likelihood ratio. There is a one-to-one correspondence between the positive likelihood ratio and the cutoff value. The indicator 813 is placed at a position that indicates the cutoff value corresponding to the predetermined positive likelihood ratio. As described above for the indicator 812, the position indicated by the cutoff value is the position that divides both ends of the classification class that includes this cutoff value internally.

[0062] The screen 810 may include an indicator 814 that indicates the relative position of a predetermined negative likelihood ratio (0, 0.1, 0.2, 0.36, 0.5, and 1 in the example of FIG. 8 ) of the AI ​​model 300 relative to the boundary values ​​of the four classification classes (the above-mentioned boundary values ​​BL, BM, and BH). In the example of FIG. 8 , the indicator 814 has a bar shape. The indicator 814 is added to the bottom of the area 811. "LR-" on the screen 810 indicates that the indicator 814 added to the bottom of the area 811 represents a negative likelihood ratio. There is a one-to-one correspondence between the negative likelihood ratio and the cutoff value. The indicator 814 is placed at a position that indicates the cutoff value corresponding to the predetermined negative likelihood ratio. As described above for the indicator 812, the position indicated by the cutoff value is the position that divides both ends of the classification class that includes this cutoff value internally.

[0063] On screen 810, of the four portions included in area 811, only the portion corresponding to the classification class to which the risk score is determined to belong is displayed, and the other three portions may not be displayed. For example, in the example of Fig. 8, portion 811MH and indicators 812 to 814 pointing to it may be displayed, while portions 811L, 811ML, and 811H and indicators 813 to 814 pointing to them may not be displayed.

[0064] The screen 810 includes a field 815 for displaying a recommended action. The action presenting unit 205 of the diagnosis assistance device 100 may display a sentence indicating the determined action in the field 815, as described with reference to FIG. 2 . If the diagnosis assistance device 100 does not present a recommended action to the user, the screen 810 does not need to include the field 815.

[0065] The information displayed on screen 810 may be presented to the user in other formats. For example, the information displayed on screen 810 may be presented to the user in a printed form, such as a report. Furthermore, the information displayed on screen 810 may be used by other programs in addition to or instead of being presented to the user.

[0066] In the above-described embodiment, the method of FIG. 6 and the method of FIG. 7 are executed by the same diagnostic assistance device 100. Alternatively, the method of FIG. 6 and the method of FIG. 7 may be executed by separate devices. For example, a device other than the diagnostic assistance device 100 may execute the method of FIG. 6, and the diagnostic assistance device 100 may execute the method of FIG. 7 using the result of the method of FIG. 6 (i.e., boundary values). Also, in the method of FIG. 7, instead of the diagnostic assistance device 100 calculating a risk score using the AI ​​model 300, the diagnostic assistance device 100 may calculate a risk score using an AI model 300 other than the diagnostic assistance device 100, and the diagnostic assistance device 100 may acquire this risk score. Furthermore, the method of FIG. 6 may be executed by multiple devices working together, and the method of FIG. 7 may be executed by multiple devices working together.

[0067] [Configuration Example of Diagnostic Support System] An example of the configuration of a diagnostic support system 900 according to the first embodiment will be described with reference to Fig. 9. The diagnostic support system 900 includes a user terminal 901 and a server 902. The user terminal 901 and the server 902 may each be configured by a computer including the same hardware components as the diagnostic support device 100 described with reference to Fig. 1.

[0068] The user terminal 901 is a device used by a user. The server 902 is a device that executes specific processing in response to a request from the user terminal 901. The server 902 may be located in an on-premise environment or in a cloud environment.

[0069] The user terminal 901 and the server 902 can communicate with each other through a network 903. The network 903 may be a local area network, the Internet, a cellular network, or any combination thereof.

[0070] 2 are distributed between a user terminal 901 and a server 902. For example, the user terminal 901 may include a data acquisition unit 201, a class determination unit 203, a class presentation unit 204, an action presentation unit 205, and a probability value acquisition unit 208. The server 902 may include the data acquisition unit 201, a score calculation unit 202, a diagnostic performance index calculation unit 206, and a boundary value determination unit 207.

[0071] The data acquisition unit 201 of the user terminal 901 may acquire medical data of the subject from the user of the user terminal 901. The data acquisition unit 201 of the server 902 may acquire the medical data acquired by the user terminal 901 from the user terminal 901. The above-described operations (e.g., the methods in FIGS. 6 and 7) by the diagnosis assistance device 100 are executed by the user terminal 901 and the server 902 in cooperation with each other.

[0072] Second Embodiment A second embodiment of the present invention relates to a diagnostic support device. A diagnostic support device is a device that supports diagnoses performed by medical professionals such as doctors, dentists, and veterinarians. The person to be diagnosed (hereinafter referred to as the test subject) may be a human or another animal (e.g., livestock or pet). A person diagnosed with an injury or illness may also be called a patient.

[0073] The purpose of the test is to determine whether the test subject has an illness or disease. For example, it may be determined whether the test subject has a specific illness or disease, or it may be determined whether the test subject has some illness or disease even if the name of the illness or disease cannot be identified. In the following description, the fact that the test subject has an illness or disease is referred to as "the test subject is positive," and the fact that the test subject does not have an illness or disease is referred to as "the test subject is negative."

[0074] The diagnostic support device assists in diagnosis using medical data related to the subject. Medical data is data that can serve as the basis for diagnosis. For example, the medical data may include data collected from the subject using a testing device (e.g., electrocardiogram data, X-rays, blood test results, height, weight). The medical data may also include findings on the subject by a medical professional. The medical data may also include attributes of the subject (e.g., age, gender, lifestyle, medical history). The medical data may also include attributes of the subject's blood relatives (e.g., parents, brothers, sisters).

[0075] [Example of Hardware Configuration of Diagnosis Support Device] An example of the hardware configuration of a diagnosis support device 1000 according to the second embodiment will be described with reference to FIG. 10 . The diagnosis support device 1000 includes the components shown in FIG. 10 . The diagnosis support device 1000 may include only one of each component shown in FIG. 10 , or may include multiple components. The diagnosis support device 1000 may not include some of the components shown in FIG. 10 , or may include components not shown in FIG. 10 . For example, the diagnosis support device 1000 may include an inspection device for collecting medical data from a subject. In this case, the diagnosis support device 1000 may be called an inspection device having a diagnosis support function.

[0076] The processor 1001 is a device that controls the overall operation of the diagnostic support device 1000. The processor 1001 is configured, for example, by a CPU (Central Processing Unit). The memory 1002 is a device that stores programs and trained AI models necessary for the operation of the diagnostic support device 1000. The memory 1002 is configured, for example, by a RAM (Random Access Memory) or a ROM (Read Only Memory). The operation of the diagnostic support device 1000 may be performed, for example, by the processor 1001 executing a program stored in the memory 1002. Alternatively, some or all of the operation of the diagnostic support device 1000 may be performed by a dedicated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array). Furthermore, the processor 1001 may include a GPU (Graphical Processing Unit). A device having the processor 1001 and the memory 1002, such as the diagnostic support device 1000, may also be called a computer.

[0077] The input device 1003 is a device for acquiring input from a user of the diagnostic assistance device 1000 (e.g., a medical professional; hereinafter, the user of the diagnostic assistance device 1000 will be simply referred to as the user). The input device 1003 is configured, for example, by extracting waveform data, images, etc. recorded from electrodes, sensors, cameras, etc., using a keyboard, touch panel, or mouse. The output device 1004 is a device for outputting data to the user. The output device 1004 is configured, for example, by a display or printer. In the example of FIG. 10 , the input device 1003 is described as constituting a part of the diagnostic assistance device 1000, but the input device 1003 may be a device separate from the diagnostic assistance device 1000. In this case, the diagnostic assistance device 1000 includes an input interface for connecting to the input device 1003. Similarly, the output device 1004 may be a part of the diagnostic assistance device 1000 or a device separate from the diagnostic assistance device 1000.

[0078] The communication device 1005 is a device that allows the diagnosis assistance device 1000 to communicate with other devices. The other devices may be computers connected to a network (e.g., the Internet or a local area network). For example, the communication device 1005 may be used to acquire medical data from the other devices.

[0079] The storage device 1006 is a device that stores data used for the operation of the diagnostic assistance device 1000. The storage device 1006 is configured by a storage medium such as a hard disk drive (HDD), a solid state drive (SSD), or a DVD (digital versatile disc).

[0080] [Example of a process flow in which functions of the diagnostic assistance device are executed by a processor] An example of a process flow in which functions of the diagnostic assistance device 1000 are executed by a processor will be described with reference to Fig. 11 . The diagnostic assistance device 1000 may not include some of the functions shown in Fig. 11 , or may include functions not shown in Fig. 11 . Each function of the diagnostic assistance device 1000 is realized by the processor 1001 executing a program or a trained AI model loaded into the memory 1002. At least some of the functions of the diagnostic assistance device 1000 may be realized by a dedicated circuit such as an ASIC or FPGA.

[0081] The data acquisition unit 1101 acquires data used in the diagnosis assistance device 1000. For example, the data acquisition unit 1101 acquires medical data related to the subject of examination. The data acquisition unit 1101 may acquire medical data input by a user to the input device 1003. The data acquisition unit 1101 may acquire medical data from a device external to the diagnosis assistance device 1000 (e.g., a file server or an examination device) using the communication device 1005. If the diagnosis assistance device 1000 includes an examination device, the data acquisition unit 1101 may acquire medical data from the examination device. The medical data acquired by the data acquisition unit 1101 may be stored in the memory 1002 or the storage device 1006 for subsequent processing.

[0082] The trained AI model unit 1102 is created by AI learning using training medical data. The trained AI model unit 1102 may have any structure. For example, the trained AI model unit 1102 may be configured as a neural network such as deep learning, or as a support vector machine such as machine learning.

[0083] The teacher medical data is medical data labeled with whether the provider of the teacher medical data is positive or not. The risk score calculation unit 1103 calculates a risk score that the person in the medical data has an injury or illness using the trained AI model unit 1102. The AI ​​learning of the trained AI model unit 1102 may be performed by the diagnosis support device 1000 or by another device.

[0084] The risk score calculation unit 1103 calculates a risk score that indicates the possibility that the person of the medical data has an injury or illness (i.e., is positive) using the trained AI model unit 1102. The person of the medical data may be a test subject or a provider of verification medical data.

[0085] As shown in Figure 11, the trained AI model unit 1102 calculates a risk score for the input medical data. In the following description, the risk score can take a value in the range of 0 to 1. Alternatively, the risk score may take a value in another range. The higher the risk score, the more likely the person involved in the medical data is positive, and the lower the risk score, the more likely the person involved in the medical data is negative.

[0086] In one example, the medical data is electrocardiogram data, and the risk score represents the likelihood that the person of the electrocardiogram data has paroxysmal atrial fibrillation (PAF). In this case, the teacher medical data includes electrocardiogram data of a patient with paroxysmal atrial fibrillation (i.e., electrocardiogram data labeled as being positive for paroxysmal atrial fibrillation) and electrocardiogram data of a donor who does not have paroxysmal atrial fibrillation (i.e., electrocardiogram data labeled as being negative for paroxysmal atrial fibrillation).

[0087] Diseases such as PAF, which manifest spontaneously or suddenly, are known to be difficult to detect because symptoms often do not appear at the time of testing. For example, depending on the timing of acquisition of electrocardiogram data from a patient with PAF, the electrocardiogram data may show sinus rhythm, making PAF difficult to detect. This type of PAF is sometimes called "hidden atrial fibrillation." The trained AI model unit 1102 quantitatively determines the risk of a test subject having PAF, enabling support for diagnosis.

[0088] The classification class determination unit 1104 determines, from three or more classification classes, the classification class to which the risk score calculated by the trained AI model unit 1102 belongs. The three or more classification classes divide the range that the risk score can take (i.e., the range from 0 to 1). Below, as an example, a case will be described in which the classification class determination unit 1104 determines, from four classification classes, the classification class to which the risk score belongs. The same explanation applies when the classification class determination unit 1104 determines, from three or five or more classification classes, the classification class to which the risk score belongs.

[0089] The four classification classes will be described with reference to Figure 12. The range that the risk score can take (i.e., the range from 0 to 1) is divided into four ranges by three boundary values ​​BL, BM, and BH. The boundary value BL is greater than 0 and less than the boundary value BM. The boundary value BM is greater than the boundary value BL and less than the boundary value BH. The boundary value BH is greater than the boundary value BM and less than 1. The method for determining the boundary values ​​BL, BM, and BH will be described later.

[0090] The classification class determination unit 1104 determines that the risk score belongs to the classification class "low" when the risk score is included in a range equal to or greater than 0 and less than the boundary value BL. The classification class determination unit 1104 determines that the risk score belongs to the classification class "middle-low" when the risk score is included in a range equal to or greater than the boundary value BL and less than the boundary value BM. The classification class determination unit 1104 determines that the risk score belongs to the classification class "middle-high" when the risk score is included in a range equal to or greater than the boundary value BM and less than the boundary value BH. The classification class determination unit 1104 determines that the risk score belongs to the classification class "high" when the risk score is included in a range equal to or greater than the boundary value BH and less than 1. In the above example, when the risk score is equal to the boundary value, the risk score is determined to belong to a class determination having the boundary value as its lower limit. Alternatively, when the risk score is equal to the boundary value, the risk score may be determined to belong to a class determination having the boundary value as its upper limit. As shown in FIG. 12, the range of possible risk scores is divided into four classification classes by predetermined values ​​of a plurality of diagnostic performance indices (described later).

[0091] The method for determining the boundary value is based on indicators (specifically, sensitivity, specificity, positive likelihood ratio, negative likelihood ratio) that represent the diagnostic ability of the AI ​​model corresponding to the output risk score of the trained AI model unit 1102.

[0092] Sensitivity is an index representing the ability to determine a subject with an injury or illness as positive when the risk score output by the trained AI model unit 1102 is equal to or greater than the cutoff value of a certain test using the cutoff value of a certain test. Specificity is an index representing the ability to determine a subject with an injury or illness as negative when the risk score output by the trained AI model unit 1102 is equal to or less than the cutoff value of a certain test using the cutoff value of a certain test using the cutoff value of a certain test using the cutoff value of a certain test using the cutoff value of a certain test using the cutoff value of a certain test using the cutoff value of a certain test using the risk score ...

[0093] The likelihood ratio is the ratio of the odds before the test to the odds after the test, and is an index that shows how much the probability of having an illness changes before and after the test when the test is positive or negative. The positive likelihood ratio of a test indicates how many times more likely a positive person is to be positive than a negative person, and the negative likelihood ratio indicates how many times more likely a positive person is to be negative than a negative person. It can be calculated from sensitivity and specificity using the following formula: Positive likelihood ratio (LR+) = sensitivity / (1 - specificity), Negative likelihood ratio (LR-) = (1 - sensitivity) / specificity.

[0094] From the calculated positive and negative likelihood ratios, the posterior probability of the test subject having the disease or injury after the test can be calculated using a nomogram, with reference to the prior probability value 1106 of the test subject having the disease or injury before the test. The prior probability value is a value that a clinician estimates the probability that the test subject has the disease or injury before the test is performed, based on a medical interview or other clinical information. It can be estimated based on literature from the test subject's medical history and physical examination findings, or it can be comprehensively determined based on the clinician's experience. In addition, the season, region, and characteristics of the facility can also affect the prior probability determination.

[0095] By referring to the prior probability value 1106 estimated by the clinician, the classification class determination unit 1104 of the AI ​​model and the AI ​​output display unit 1105 (described later) use a nomogram to quantitatively determine the posterior probability of the test subject after the AI ​​test based on the positive likelihood ratio or negative likelihood ratio of the test corresponding to the determined class. This helps doctors make accurate decisions to make a definitive diagnosis.

[0096] A method for determining the boundary values ​​of the four classification classes will be described using an example. For example, the boundary value BL is set to the maximum cutoff value (e.g., 0.1) at which the negative likelihood ratio is equal to or less than a first predetermined value A. When the negative likelihood ratio of a certain test is equal to or less than A, the possibility that the test subject has an injury or illness can be more strongly denied. The boundary value BL may also be set to another value based on the negative likelihood ratio (e.g., a cutoff value among multiple cutoff values ​​at which the negative likelihood ratio is closest to or even lower than the first predetermined value).

[0097] The boundary value BH is set to the smallest cutoff value (e.g., 0.8) at which the positive likelihood ratio is equal to or greater than the second predetermined value B. When the positive likelihood ratio of a certain test is equal to or greater than B, it is more likely that the test subject has an illness. The boundary value BH may also be determined using another value based on the positive likelihood ratio (e.g., a cutoff value at which the positive likelihood ratio is closest to or even higher than the second predetermined value).

[0098] The boundary value BM is determined as an optimal cutoff value C (e.g., 0.5) that maximizes the sum of sensitivity and specificity. The boundary value BM may also be determined as another value between the boundary values ​​BL and BH (e.g., the average value between the boundary values ​​BL and BH) or a third predetermined value of the positive and negative likelihood ratios.

[0099] When the classification class determination unit 1104 determines the classification class to which the risk score belongs from the four classification classes, any one of the three boundary values ​​BL, BM, and BH described above does not need to be used. For example, when the classification class "low" and the classification class "middle-low" are combined, the boundary value BL is not used. When the classification class "middle-low" and the classification class "middle-high" are combined, the boundary value BM is not used. When the classification class "middle-high" and the classification class "high" are combined, the boundary value BH is not used.

[0100] As described above, the boundary values ​​of the classification classes are determined by a plurality of predetermined values ​​based on the diagnostic performance index of the trained AI model unit 1102. These boundary values ​​divide the range of possible risk scores into three or more classification classes. Therefore, in general, the range of possible risk scores is divided by the predetermined values ​​of the plurality of diagnostic performance indexes for three or more classification classes.

[0101] The boundary value determined based on the diagnostic performance index of the trained AI model unit 1102 may differ depending on the performance of the trained AI model unit 1102. For example, even if the same risk score is output for an AI model with relatively high performance (hereinafter referred to as a high-performance AI model) and an AI model with relatively low diagnostic performance (hereinafter referred to as a low-performance model), the boundary values ​​determined for the multiple classification classes will differ depending on the performance of the models, and the probability change (test effect) before and after the test corresponding to the determined class will also differ. Thus, in this embodiment, by determining the boundary value based on the diagnostic performance index of the trained AI model unit 1102, the positive and negative likelihood ratios of the test corresponding to the determined class can be presented to users, including doctors, by the AI ​​output display unit 1105 (described below).

[0102] FIG. 13 shows an example of four classification classes output by the AI ​​output display unit 1105. An example of a screen 1300 used to present the classification classes and recommended actions to the user will be described. The screen 1300 includes bar-shaped regions 1301 representing the four classification classes. The region 1301 is divided into a portion 1301L corresponding to the classification class "low," a portion 1301ML corresponding to the classification class "middle-low," a portion 1301MH corresponding to the classification class "middle-high," and a portion 1301H corresponding to the classification class "high." These four portions are arranged horizontally in the order shown in FIG. 13. The result of the classification class determination unit 1104 is shown by displaying the portion 1301H with a weighted color or pattern.

[0103] FIG. 14 shows another example of four classification classes output by the AI ​​output display unit 1105. Screen 1400 includes an indicator that indicates the risk score output by the risk score calculation unit 1103 of FIG. 11. In the example of FIG. 14, indicator 1402 has a triangular shape, but may have other shapes. Portion 1401MH to which indicator 1402 belongs indicates the class determined by the classification class determination unit 1104 of FIG. 11. This indicates to the user that the risk score calculated by the trained AI model has been determined to belong to the classification class "middle-high."

[0104] Indicator 1402 may present to the user the relative position of the risk score calculated for the test subject relative to the boundary values ​​of the four classification classes (the boundary values ​​BL, BM, and BH described above). Specifically, the left end of portion 1401L represents the position where the risk score is 0. The boundary line between portion 1401L and portion 1401ML represents the position where the risk score is boundary value BL. The boundary line between portion 1401ML and portion 1401MH represents the position where the risk score is boundary value BM. The boundary line between portion 1401MH and portion 1401H represents the position where the risk score is boundary value BH. The right end of portion 1401H represents the position where the risk score is 1.

[0105] As shown in FIG. 14 , assume that the risk score calculated by the AI ​​model is determined to belong to the classification class "middle-high," and the risk score divides the lower limit (i.e., boundary value BM) and upper limit (i.e., boundary value BH) of the classification class "middle-high" into a:b. In this case, indicator 1402 is positioned to indicate the position where the left and right ends of portion 1401MH are divided into a:b. By positioning indicator 1402 in this manner, the user can recognize how close the risk score is to other classification classes and the corresponding positive and negative likelihood ratios. This allows the user to quantitatively grasp the change in probability before and after testing with greater accuracy, making it easier to make a definitive diagnosis.

[0106] The screen 1400 may include an indicator 503 that indicates the relative position of a predetermined positive likelihood ratio (LR+) (1, 2, 3.4, 5, and 10 in the example of FIG. 14 ) of the model, in which the risk score output from the trained AI model unit 1102, ranging from 0 to 1, is calculated as each cutoff value. In the example of FIG. 14 , the indicator 503 has a bar shape. The indicator 503 is added to the top of the area 1401. The indicator 503 is placed at a position indicating each cutoff value corresponding to the predetermined positive likelihood ratio. As described above for the indicator 1402, the position indicated by the cutoff value is the position that divides both ends of the classification class that includes this cutoff value internally.

[0107] The screen 1400 may include an indicator 504 that indicates the relative position of a predetermined negative likelihood ratio (LR-) (0, 0.1, 0.2, 0.36, 0.5, and 1 in the example of FIG. 14 ) of the model, in which the risk score output from the trained AI model unit 1102 ranging from 0 to 1 is calculated as each cutoff value. In the example of FIG. 14 , the indicator 504 has a bar shape. The indicator 504 is added to the bottom of the area 1401. The indicator 504 is placed at a position indicating each cutoff value corresponding to the predetermined negative likelihood ratio. As described above for the indicator 1402, the position indicated by the cutoff value is the position that divides both ends of the classification class that includes this cutoff value internally.

[0108] On screen 1400, of the four portions included in area 1401, only the portion corresponding to the classification class to which the risk score is determined to belong is displayed, and the other three portions may not be displayed. For example, in the example of Figure 14, portion 1401MH and indicators 1402 to 1404 pointing to it may be displayed, while portions 1401L, 1401ML, and 1401H and indicators 1403 to 1404 pointing to them may not be displayed. Figure 15 shows an example of such an output display.

[0109] Screen 1500 indicates the risk score calculated by risk score calculation unit 1103 in Fig. 11 using a triangular indicator 1502. A bar-shaped area 1501MH displays the discrimination class to which the risk score indicated by indicator 1502 belongs. The lower and upper limits of the risk score of class 1501MH are represented by boundary values ​​BM and BH. Area 1501MH indicating the discriminated class is divided into five equal parts by vertical bars, and each equal part may be further divided into two equal parts by indicator 1503.

[0110] The upper part of the area 1501MH indicating the determined class may include an indicator 1504 indicating the relative positions of three predetermined values ​​(3.4, 4.8, and 6.2 in the example of FIG. 15 ) including the lower limit, upper limit, and predetermined values ​​between them of the positive likelihood ratio (LR+) of the model determined from the cutoff values ​​corresponding to the lower limit BM and upper limit BH of the risk score of this class. In this way, the example of FIG. 15 displays a predetermined plurality of positive likelihood ratios and negative likelihood ratios distributed over the entire range of the classification class to which the risk score has been determined to belong.

[0111] The lower part of the area 1501MH indicating the class to be determined may include an indicator 1505 indicating the relative positions of three predetermined values ​​(0.36, 0.48, and 0.6 in the example of Figure 15) including the lower limit, upper limit, and predetermined values ​​between them of the negative likelihood ratio (LR-) of the model obtained from the cutoff values ​​corresponding to the lower limit BM and upper limit BH of the risk score of this class.

[0112] As described above, Figure 15, which is an example of the display of the AI ​​output display unit 1105, differs from the output display examples of Figures 13 and 14 in that it provides the user with a more accurate quantitative indication of the change in probability before and after the test based on an indicator 1502 that indicates the relative position of the risk score calculated from the AI ​​model, and the three positive and negative likelihood ratios presented within the determined class area, thereby supporting a definitive diagnosis.

[0113] The recommended action presentation unit 1107 in Fig. 11 presents recommended actions to the user based on the judgment class displayed by the AI ​​output display unit 1105 and the prior probability value 1106 of the test target obtained from the user. For example, the recommended action presentation unit 1107 may determine the action by referring to the table 1600 in Fig. 16. The table 1600 is generated by the manufacturer of the diagnostic support device 1000 and stored in the memory 1002 or the storage device 1006 of the diagnostic support device 1000. Along with the output display of the determined classification class, the presentation of recommended actions can also be displayed on the diagnostic screen.

[0114] The rows of table 1600 represent the classification classes determined by the classification class determination unit 1104. The columns of table 1600 represent whether the pre-test prior probability value 1106 in FIG. 11 is higher or lower than a predetermined threshold (e.g., 50%) and may be composed of two or more columns. The "Detailed Examination" in table 1600 recommends that the subject be considered for detailed examination. The "Regular Examination" in table 1600 recommends that the subject be considered for regular examination. The "Simple Examination" in table 1600 recommends that the subject be considered for simple examination, which is less burdensome. The "No Examination" in table 1600 recommends that the subject not be considered for additional examination. Other descriptions may be used for the recommended actions.

[0115] Based on the positive and negative likelihood ratios of the test corresponding to the determination class displayed by the AI ​​output display unit 1105 in Fig. 11 and the prior probability value 1106 of the test subject obtained from the user, the posterior probability of the test subject having an injury or illness after the test can be quantitatively determined by referring to a nomogram. Once the posterior probability is determined, a recommended action plan such as that shown in table 1600 in Fig. 16 can be presented to medical professionals, and based on the quantitative posterior probability, medical professionals can be supported in making an accurate definitive diagnosis.

[0116] The rows of table 1700 in FIG. 17 represent the posterior probability of the estimated test subject having an injury or disease after the test. As an example, the posterior probability ranging from 0 to 100% is divided into four classes using three predetermined thresholds (e.g., 20%, 50%, and 80%), and the recommended action for each class is shown in the columns of table 1700. If the posterior probability value is lower than the first predetermined threshold, the recommended action in table 1700 is "No additional test." If the posterior probability value is higher than the first predetermined threshold and lower than the second predetermined threshold, the recommended action in table 1700 is "Simple test." If the posterior probability value is higher than the second predetermined threshold and lower than the third predetermined threshold, the recommended action in table 1700 is "Regular test." If the posterior probability value is equal to or greater than the third predetermined threshold, the recommended action in table 1700 is "Detailed test." If the posterior probability is equal to the predetermined threshold, it is determined that the posterior probability belongs to a class whose lower limit is this threshold. Alternatively, if the posterior probability is equal to a predetermined threshold, the posterior probability may be determined to belong to a class whose upper limit is the threshold.

[0117] A proposed action plan (FIG. 16 or FIG. 17) is displayed on the screen based on the test results, the AI ​​output display, the determined classification class, and the user's estimated prior probability of the test subject. This information is then saved to a recording device or printed by the test result report output unit 1108. This ensures that the test results and recommended actions are properly documented and accessible to medical professionals for future reference.

[0118] [Configuration Example of Diagnostic Support System] An example of the configuration of a diagnostic support system 1800 according to the second embodiment will be described with reference to Fig. 18. The diagnostic support system 1800 includes a user terminal 1801 and a server 1802. The user terminal 1801 and the server 1802 may each be configured by a computer including the same hardware components as those of the diagnostic support device 1000 described with reference to Fig. 10.

[0119] The user terminal 1801 is a device used by a user. The server 1802 is a device that executes specific processing in response to a request from the user terminal 1801. The server 1802 may be located in an on-premise environment or in a cloud environment.

[0120] The user terminal 1801 and the server 1802 can communicate with each other through a network 1803. The network 1803 may be a local area network, the Internet, a cellular network, or any combination thereof.

[0121] 11 are distributed between a user terminal 1801 and a server 1802. For example, the user terminal 1801 may include a data acquisition unit 1101, a trained AI model unit 1102, a risk score calculation unit 1103, a classification class determination unit 1104, an AI output display unit 1105, a pre-test prior probability value 1106 of the test subject, a posterior probability calculation and recommended action presentation unit 1107, and a test result report output unit 1108. The server 1802 may include the data acquisition unit 1101, the trained AI model unit 1102, the risk score calculation unit 1103, the classification class determination unit 1104, and the AI ​​output display unit 1105.

[0122] The data acquisition unit 1101 of the user terminal 1801 may acquire medical data of the subject of examination from the user of the user terminal 1801. The data acquisition unit 1101 of the server 1802 may acquire, from the user terminal 1801, the medical data acquired by the user terminal 1801. The above-described operation of the diagnosis assistance device 1000 (the flow of the execution process in FIG. 11 ) is executed by the user terminal 1801 and the server 1802 in cooperation with each other.

[0123] Summary of the Embodiments (Item 1) A diagnostic support device includes: a data acquisition unit that acquires medical data on a subject; a score calculation unit that calculates a risk score representing the likelihood of the subject having a disease or injury for the medical data using an AI model created by learning using training medical data; and a determination unit that determines the classification class to which the risk score belongs from three or more classification classes using a diagnostic performance index of the AI ​​model. This item allows medical professionals to more easily determine the probability that the subject will have a disease or injury after testing. For example, the posterior probability is quantitatively calculated based on the subject's prior probability and the output of the AI ​​model, allowing medical professionals to more accurately determine the posterior probability of the disease or injury after testing. This quantitative information is important for medical professionals to appropriately evaluate the patient's condition and recommend appropriate detailed examinations, treatments, and follow-ups. (Item 2) The diagnostic support device described in Item 1, wherein two or more boundary values ​​of the three or more classification classes are determined based on different diagnostic performance indexes of the AI ​​model. This item allows appropriate determination of the boundary values ​​of the three or more classification classes. (Item 3) The diagnostic support device according to item 1 or 2, wherein the first boundary values ​​of the three or more classification classes are determined based on a negative likelihood ratio of the AI ​​model. According to this item, it is possible to accurately determine the classification class to which a test subject who is unlikely to have an injury or disease belongs. (Item 4) The diagnostic support device according to any one of items 1 to 3, wherein the second boundary values ​​of the three or more classification classes are determined based on a positive likelihood ratio of the AI ​​model. According to this item, it is possible to accurately determine the classification class to which a test subject who is likely to have an injury or disease belongs. (Item 5) The diagnostic support device according to any one of items 1 to 4, wherein the third boundary values ​​of the three or more classification classes are determined based on a value that is based on the sensitivity and specificity of the AI ​​model and is different from the likelihood ratio. According to this item, it is possible to accurately determine the classification class to which a test subject who is likely to have an injury or disease belongs.(Item 6) The diagnostic support device according to any one of items 1 to 5, wherein a first boundary value for the three or more classification classes is determined based on a negative likelihood ratio of the AI ​​model, a second boundary value for the three or more classification classes is determined based on a positive likelihood ratio of the AI ​​model, and a third boundary value for the three or more classification classes is determined based on a value that is based on the sensitivity and specificity of the AI ​​model and is different from the likelihood ratio, and the third boundary value is greater than the first boundary value and less than the second boundary value. This item allows for accurate determination of the classification class to which the test subject belongs based on the likelihood of having an injury or illness. (Item 7) The diagnostic support device according to any one of items 1 to 6, further comprising a display unit that displays the classification class to which the risk score is determined to belong. This item allows medical professionals to understand the classification class. (Item 8) The diagnostic support device according to item 7, wherein the display unit further displays a predetermined number of positive likelihood ratios and negative likelihood ratios distributed over the entire range of the classification class to which the risk score is determined to belong. According to this item, a medical professional can understand the relationship between classification classes and likelihood ratios. (Item 9) The diagnostic support device according to item 7 or 8, wherein the display unit does not display classification classes, of the three or more classification classes, to which the risk score has not been determined to belong. According to this item, it is easier for a medical professional to understand information about classification classes to which the risk score has been determined to belong. (Item 10) The diagnostic support device according to item 9, wherein the display unit further displays at least one of: a first indicator representing the relative position of the risk score with respect to the upper and lower limits of the classification class to which the risk score has been determined to belong; a second indicator representing the relative position of a predetermined positive likelihood ratio with respect to the upper and lower limits of the classification class to which the risk score has been determined to belong; and a third indicator representing the relative position of a predetermined negative likelihood ratio with respect to the upper and lower limits of the classification class to which the risk score has been determined to belong. According to this item, it is easier for a medical professional to understand the relationship between classification classes and likelihood ratios.(Item 11) The diagnosis support device according to any one of items 1 to 10, further comprising an action presenting unit that presents recommended actions based on the classification class to which the risk score is determined to belong. According to this item, it is possible to efficiently classify classes using an index with high diagnostic ability according to the condition of the test subject, and present appropriate actions to the medical professional. (Item 12) The diagnosis support device according to item 11, in which the action presenting unit determines a posterior probability value based on a prior probability value and a positive likelihood ratio or a negative likelihood ratio corresponding to the risk score, and determines the recommended action based on the posterior probability value. According to this item, it is possible for a medical professional to determine the post-test probability of the test subject having an injury or disease using a pre-test probability value and a likelihood ratio corresponding to the calculated risk score, thereby more easily selecting a corresponding recommended action. (Item 13) The diagnosis support device according to any one of items 1 to 12, in which the medical data is electrocardiogram data, and the teacher medical data includes electrocardiogram data of a patient with paroxysmal atrial fibrillation. This item makes it easier for medical professionals to quantitatively determine the possibility that a test subject has paroxysmal atrial fibrillation. (Item 14) A program for causing a computer to: acquire medical data about a test subject; calculate a risk score representing the possibility of the test subject having an injury or disease for the medical data using an AI model created by learning using teacher medical data; and determine the classification class to which the risk score belongs from three or more classification classes using a diagnostic performance index of the AI ​​model. This item makes it easier for medical professionals to determine the probability that a test subject will have an injury or disease after testing. (Item 15) A diagnostic support system including a user terminal and a server, wherein the server includes: a data acquisition unit that acquires medical data about the test subject; and a score calculation unit that calculates a risk score representing the possibility of the test subject having an injury or disease for the medical data using an AI model created by learning using teacher medical data; and the user terminal includes: a determination unit that determines the classification class to which the risk score belongs from three or more classification classes using the diagnostic performance index of the AI ​​model.This item allows medical personnel to more easily determine the probability that a test subject will have a post-test illness or injury.

[0124] The invention is not limited to the above-described embodiment, and various modifications and variations are possible within the scope of the gist of the invention.

[0125] This application claims priority based on Japanese Patent Application No. 2024-102812, filed on June 26, 2024, the entire contents of which are incorporated herein by reference.

Claims

1. A diagnostic support device comprising: a data acquisition unit that acquires medical data on a subject; a score calculation unit that calculates a risk score representing the possibility of the subject having an illness or injury for the medical data using an AI model created by learning using training medical data; and a determination unit that determines the classification class to which the risk score belongs from three or more classification classes using a diagnostic performance index of the AI ​​model.

2. The diagnostic support device according to claim 1, wherein two or more boundary values ​​of the three or more classification classes are determined based on different diagnostic performance indexes of the AI ​​model.

3. The diagnostic support device according to claim 1 or 2, wherein the first boundary values ​​of the three or more classification classes are determined based on a negative likelihood ratio of the AI ​​model.

4. A diagnostic support device according to any one of claims 1 to 3, wherein the second boundary values ​​of the three or more classification classes are determined based on a positive likelihood ratio of the AI ​​model.

5. A diagnostic support device according to any one of claims 1 to 4, wherein the third boundary value of the three or more classification classes is determined based on a value that is based on the sensitivity and specificity of the AI ​​model and is different from a likelihood ratio.

6. A diagnostic support device as described in any one of claims 1 to 5, wherein first boundary values ​​of the three or more classification classes are determined based on a negative likelihood ratio of the AI ​​model, second boundary values ​​of the three or more classification classes are determined based on a positive likelihood ratio of the AI ​​model, and third boundary values ​​of the three or more classification classes are determined based on a value that is based on the sensitivity and specificity of the AI ​​model and is different from the likelihood ratio, and the third boundary value is greater than the first boundary value and less than the second boundary value.

7. A diagnostic support device according to any one of claims 1 to 6, further comprising a display unit that displays the classification class to which the risk score is determined to belong.

8. The diagnostic support device according to claim 7, wherein the display unit further displays a predetermined number of positive likelihood ratios and negative likelihood ratios distributed over the entire range of the classification class to which the risk score is determined to belong.

9. A diagnostic support device according to claim 7 or 8, wherein the display unit does not display any of the three or more classification classes to which the risk score is not determined to belong.

10. The diagnostic support device of claim 9, wherein the display unit further displays at least one of: a first indicator representing the relative position of the risk score with respect to the upper and lower limits of the classification class to which the risk score is determined to belong; a second indicator representing the relative position of a predetermined positive likelihood ratio with respect to the upper and lower limits of the classification class to which the risk score is determined to belong; and a third indicator representing the relative position of a predetermined negative likelihood ratio with respect to the upper and lower limits of the classification class to which the risk score is determined to belong.

11. A diagnostic support device according to any one of claims 1 to 10, further comprising an action presentation unit that presents recommended actions determined based on the classification class to which the risk score is determined to belong.

12. The diagnostic support device of claim 11, wherein the action presentation unit determines a posterior probability value based on a prior probability value and a positive likelihood ratio or a negative likelihood ratio corresponding to the risk score, and determines the recommended action based on the posterior probability value.

13. A diagnostic support device according to any one of claims 1 to 12, wherein the medical data is electrocardiogram data, and the teacher medical data includes electrocardiogram data of a patient with paroxysmal atrial fibrillation.

14. A program for causing a computer to perform the following steps: acquire medical data on a subject; calculate a risk score representing the possibility of the subject having an illness or injury for the medical data using an AI model created by learning using training medical data; and determine the classification class to which the risk score belongs from three or more classification classes using a diagnostic performance index of the AI ​​model.

15. A diagnostic support system including a user terminal and a server, wherein the server includes: a data acquisition unit that acquires medical data on a subject; and a score calculation unit that calculates a risk score representing the possibility of the subject having an injury or illness for the medical data using an AI model created by learning using teacher medical data; and the user terminal includes: a determination unit that determines the classification class to which the risk score belongs from three or more classification classes using a diagnostic performance index of the AI ​​model.

Images