Abnormality prediction method and abnormality prediction program

Abstract

To improve the accuracy of predicting abnormality.SOLUTION: An abnormality prediction device acquires a text feature amount 1601 indicating a first feature amount extracted from medical record data pertaining to a patient and representing the level of possibility of diseases (brain tumor, anemia, renal failure, infection, cerebral hemorrhage). The abnormality prediction device acquires a numeric feature amount 1602 indicating a second feature amount extracted from test value data pertaining to a patient and representing the level of possibility of diseases. The abnormality prediction device multiplies the text feature amount 1601 and an importance degree 1603 together to the numeric feature amount 1602 regarding each disease and thereby corrects the numeric feature amount 1602. The importance degree 1603 indicates the degree of importance of each disease corresponding to test items included in the test data pertaining to a patient. The abnormality prediction device determines a disease for the patient on the basis of the text feature amount 1601 and the corrected numeric feature amount 1602.SELECTED DRAWING: Figure 16

Description

[Technical Field]

[0001] The present invention relates to an abnormality prediction method and an abnormality prediction program. [Background technology]

[0002] In the past, when differentiating diseases in emergency situations, there has been a need to make judgments from among possible disease candidates without overlooking serious diseases that may require hospitalization. For example, it is not desirable to send home a patient with a disease that may suddenly worsen before the hospital is closed, so judgments are required that do not overlook such diseases. Differential diagnosis is performed using various types of data (multimodal data), such as text and numerical values.

[0003] Prior art includes, for example, a multimodal deep learning model for the early detection of Alzheimer's disease (AD), which comprehensively analyzes image, genetic (single nucleotide polymorphism), and clinical test data to classify patients into AD, MCI (mild cognitive impairment), and control (see, for example, Non-Patent Document 1 below). [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] Janani Venugopalan,Li Tong,Hamid Reza Hassanzadeh & May D.Wang “Multimodal deep learning models for early detection of Alzheimer's disease stage” Scientific Reports 2021 Summary of the Invention [Problem to be solved by the invention]

[0005] However, when conventional techniques use multiple types of data, such as text and numerical values, to predict suspected abnormalities (e.g., diseases), they may not be able to predict serious abnormalities that should not be overlooked.

[0006] In one aspect, the present invention aims to improve the accuracy of abnormality prediction. [Means for solving the problem]

[0007] In one embodiment, there is provided an anomaly prediction method including: acquiring first feature quantities representing the likelihood of each anomaly included in a plurality of candidate anomalies, extracted from first data of a first type related to an anomaly prediction target; acquiring second feature quantities representing the likelihood of each anomaly, extracted from second data of a second type different from the first type related to the anomaly prediction target; correcting the acquired second feature quantities based on the importance of each anomaly defined in accordance with the information content included in the second data, by referring to a storage unit that stores the importance of each anomaly defined in accordance with the information content included in the second data; and determining an anomaly for the anomaly prediction target from among the plurality of anomalies, based on the acquired first feature quantities and the corrected second feature quantities. [Effects of the Invention]

[0008] Advantageous Effects of Invention According to one aspect of the present invention, an effect is achieved in that it is possible to improve the accuracy of predicting an abnormality. [Brief explanation of the drawings]

[0009] [Figure 1] FIG. 1 is an explanatory diagram illustrating an example of an abnormality prediction method according to an embodiment. [Figure 2] FIG. 2 is an explanatory diagram showing an example of the system configuration of the information processing system 200. As shown in FIG. [Figure 3] FIG. 3 is a block diagram showing an example of the hardware configuration of the abnormality prediction device 201. As shown in FIG. [Figure 4] FIG. 4 is an explanatory diagram showing an example of the contents stored in the importance table 220. As shown in FIG. [Figure 5] FIG. 5 is an explanatory diagram showing a specific example of text data. [Figure 6] FIG. 6 is an explanatory diagram showing a specific example of test value data. [Figure 7] FIG. 7 is a block diagram showing an example of the functional configuration of the abnormality prediction device 201. As shown in FIG. [Figure 8] FIG. 8 is an explanatory diagram (part 1) showing an example of learning importance. [Figure 9] FIG. 9 is an explanatory diagram (part 2) showing an example of learning importance. [Figure 10] FIG. 10 is an explanatory diagram (part 1) showing an example of the operation of the abnormality prediction device 201. [Figure 11] FIG. 11 is an explanatory diagram (part 2) showing an example of the operation of the abnormality prediction device 201. [Figure 12] FIG. 12 is an explanatory diagram (part 3) showing an example of the operation of the abnormality prediction device 201. [Figure 13] FIG. 13 is an explanatory diagram (part 4) showing an example of the operation of the abnormality prediction device 201. [Figure 14] FIG. 14 is an explanatory diagram (part 5) showing an example of the operation of the abnormality prediction device 201. [Figure 15] FIG. 15 is an explanatory diagram (part 6) showing an example of the operation of the abnormality prediction device 201. [Figure 16] FIG. 16 is an explanatory diagram (part 1) showing an example of disease prediction. [Figure 17] FIG. 17 is an explanatory diagram (part 2) showing an example of disease prediction. [Figure 18] FIG. 18 is a flowchart showing an example of an abnormality prediction procedure of the abnormality prediction device 201. [Figure 19] FIG. 19 is a flowchart illustrating an example of a specific processing procedure of the importance learning processing. DETAILED DESCRIPTION OF THE INVENTION

[0010] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of an abnormality prediction method and an abnormality prediction program according to the present invention will be described in detail below with reference to the accompanying drawings.

[0011] (Embodiment) 1 is an explanatory diagram illustrating an example of an abnormality prediction method according to an embodiment. In FIG. 1, an information processing device 101 is a computer that predicts an abnormality in an abnormality prediction target based on multiple types of data related to the abnormality prediction target. The abnormality prediction target is, for example, a patient, a device, or a facility that is the target of abnormality prediction.

[0012] Anomaly prediction is the process of determining an anomaly corresponding to an anomaly prediction target, for example, an anomaly that has occurred in the anomaly prediction target or that is likely to occur in the anomaly prediction target. The anomaly prediction target may be, for example, a patient in a medical institution. In this case, the anomaly may be, for example, a plurality of candidate diseases for the patient.

[0013] The multiple types of data are information created according to different standards and data formats, such as text data, numerical data, image data, etc. For example, if the abnormality prediction target is a "patient," the multiple types of data include medical record data (text data) and test value data (numerical data) contained in the patient's electronic medical record.

[0014] For example, in emergency medical care, the diagnosis of diseases (differential diagnosis of diseases) is performed for emergency patients who are brought in. Disease diagnosis is sometimes performed using various types of data, such as text data and numerical data. By using multiple types of data (so-called multimodal data), it becomes possible to determine diseases from various perspectives.

[0015] However, when differential diagnosis is performed using multiple types of data, it is necessary to comprehensively and multifacetedly assess the patient's condition from various types of data. For this reason, it is difficult for young doctors with little experience to make differential diagnosis using multiple types of data, and they may overlook serious diseases.

[0016] For example, if a serious illness that requires hospitalization is overlooked in the differential diagnosis of emergency medical care, the patient's condition may suddenly worsen after returning home, leading to a serious situation. While it is possible to hospitalize all patients with illnesses to avoid such a situation, this is not realistic due to the shortage of hospital beds and the high costs involved.

[0017] If a computer could predict suspected diseases based on multiple types of patient data, it would be possible to assist in the differential diagnosis of diseases by inexperienced junior doctors.

[0018] For example, there is a conventional technique for predicting suspected abnormalities by extracting feature amounts (e.g., the degree of abnormality) from each of a plurality of types of data, integrating the feature amounts extracted from each data, and inputting the result into a prediction model. Non-Patent Document 1, for example, can be referred to as an example of the conventional technique.

[0019] However, with conventional technology, if a serious anomaly that should not be overlooked has only a slight feature value, it may not be included in the candidates for suspected anomalies and may be overlooked. For example, if there is another noticeable anomaly, the feature value of that anomaly will become dominant, and the serious anomaly with a slight feature value will not be raised as a candidate.

[0020] As an example, let us consider a case where a disease of a patient is predicted using the patient's medical record data and test value data. Here, the multiple candidate abnormalities are "Disease A, Disease B, Disease C, and Disease D." The feature values ​​for each of Diseases A to D represent the likelihood of each of Diseases A to D. Specifically, the feature values ​​for each of Diseases A to D take values ​​between 0 and 1, with larger values ​​indicating a higher likelihood of the disease.

[0021] Here, the correct disease (the disease that the patient is likely to develop) is called "Disease C." Disease C is a serious disease that requires hospitalization. Furthermore, the feature quantities for each of diseases A to D extracted from the medical record data are "A=0.2, B=0.25, C=0.2, D=0.25." Furthermore, the feature quantities for each of diseases A to D extracted from the test value data are "A=0.1, B=0.2, C=0.05, D=0.5."

[0022] In conventional technology, suspected abnormalities (diseases) are predicted by integrating the feature quantities for each disease A to D extracted from each data (medical record data, test value data) and inputting them into a prediction model. For example, when the feature quantities for each disease A to D extracted from each data are integrated by adding them together, the integrated feature quantities for each disease A to D are "A=0.3, B=0.45, C=0.25, D=0.75."

[0023] Here, it is assumed that the disease with the highest feature value after integration is presented as the prediction result among diseases A to D. In this case, with conventional technology, "Disease D" would be presented as the prediction result, and "Disease C," which should have been listed as a suspected disease, would be overlooked. For example, if a junior doctor were to diagnose the patient as having disease D based on the prediction result and send the patient home without hospitalization, the patient's condition could suddenly worsen after the patient returns home, leading to a serious incident.

[0024] It is possible to apply a process to emphasize features for specific diseases that require hospitalization. However, if a process to emphasize features for specific diseases is applied, only specific diseases will be presented as prediction results, which may result in predictions that are almost meaningless.

[0025] For this reason, it is desirable to have a system in which, even when there is a noticeable abnormality, the feature value of that abnormality becomes dominant so that abnormalities that should be suspected are not overlooked.

[0026] Here, for each disease, similar characteristics may appear in test value (numerical) data, etc. For example, in the case of renal failure (disease), unique characteristics may appear in the value of the test item "body temperature" included in the test value data. Such characteristics can serve as an indicator (degree of importance) of which information included in the test value data should be focused on when determining a certain disease. Such characteristics can be identified from past test value data using machine learning, and may also be accumulated as knowledge by experienced doctors, etc.

[0027] Therefore, in this embodiment, an anomaly prediction method is described in which the importance of each anomaly is defined in advance according to the information content included in a certain type of data, and the feature extracted from the type of data is optimized using the importance, thereby improving the accuracy of anomaly prediction. Here, an example of processing by the information processing device 101 is described.

[0028] Here, the abnormality prediction target is referred to as "abnormality prediction target 102," and the multiple candidate abnormalities are referred to as "abnormalities A, B, C, and D." The abnormality prediction target 102 is, for example, a patient. The abnormalities A to D are, for example, candidate diseases.

[0029] (1) The information processing device 101 acquires first feature quantities f1 to f4 that represent the likelihood of each of abnormalities A to D, extracted from first data d1 related to the abnormality prediction target 102. Here, the first data d1 is a first type of data, such as text data. The first feature quantities f1 to f4 are expressed by values ​​between 0 and 1, and the larger the value, the higher the likelihood of each of abnormalities A to D.

[0030] For example, the larger the first feature value f1, the higher the possibility of abnormality A. For example, when the first data d1 is "text data," the first features f1 to f4 are extracted from the first data d1 using existing technology such as NLP (Natural Language Processing).

[0031] (2) The information processing device 101 acquires second feature amounts f5 to f8 that represent the likelihood of each of the abnormalities A to D, extracted from second data d2 related to the abnormality prediction target 102. The second data d2 is a second type of data different from the first type, such as numerical data. The second feature amounts f5 to f8 are expressed by values ​​between 0 and 1, with larger values ​​representing higher likelihoods of each of the abnormalities A to D.

[0032] For example, the larger the second feature value f5, the higher the possibility of anomaly A. For example, when the first data d2 is "numerical data," the second feature values ​​f5 to f8 are extracted from the second data d2 using existing techniques such as deep learning and decision trees.

[0033] (3) The information processing device 101 corrects the acquired second feature amounts f5 to f8 based on the importance of each of the abnormalities A to D defined according to the information content included in the second data d2, with reference to the storage unit 110. Here, the storage unit 110 stores the importance of each of the abnormalities A to D defined according to the information content included in the second type of data.

[0034] For example, suppose the second type of data is "text data." In this case, the information content included in the second type of data may be specified, for example, by characters, symbols, words, sentences, etc. included in the second type of data. Also, suppose the second type of data is "numeric data." In this case, the information content included in the second type of data may be specified, for example, by data items included in the second type of data.

[0035] Here, the importance of each of the abnormalities A to D, which is determined from the storage unit 110 and corresponds to the information content included in the second data d2, is expressed as "importance k1 to k4." The importance k1 to k4 is information for weighting the second feature amounts f5 to f8. The importance k1 to k4 is expressed, for example, by a value between 0 and 1, with a larger value indicating a higher degree of importance.

[0036] For example, in the case of abnormality A, the stronger the tendency for a specific feature to appear in the information included in the second data d2, the higher the value of the importance k1 for abnormality A. Specifically, for example, the information processing device 101 refers to the storage unit 110 and identifies the importance k1 to k4 for each of abnormalities A to D according to the content of the information included in the second data d2.

[0037] Then, the information processing device 101 corrects the second feature amounts f5 to f8 by multiplying the second feature amounts f5 to f8 by the identified importance levels k1 to k4 for each of the anomalies A to D. Here, the corrected second feature amounts f5 to f8 are referred to as "optimized second feature amounts f5' to f8'."

[0038] (4) The information processing device 101 determines an abnormality for the abnormality prediction target 102 from among the abnormalities A to D based on the acquired first feature amounts f1 to f4 and the optimized second feature amounts f5' to f8'. An abnormality for the abnormality prediction target 102 is a suspected abnormality for the abnormality prediction target 102, such as an abnormality that has occurred in the abnormality prediction target 102 or that is highly likely to occur in the abnormality prediction target 102.

[0039] Specifically, for example, the information processing device 101 calculates an integrated feature by adding together the first feature amounts f1 to f4 and the optimized second feature amounts f5' to f8' for each of the abnormalities A to D. Then, the information processing device 101 determines whether the abnormality prediction target 102 is abnormal by referring to the integrated feature amounts for each of the abnormalities A to D.

[0040] To explain in more detail, for example, the information processing device 101 may input the integrated feature values ​​of each of the anomalies A to D into an existing prediction model, and thereby determine the anomaly predicted by the prediction model as an anomaly for the anomaly prediction target 102. The prediction model is a model that takes, for example, the integrated feature values ​​of each of the anomalies A to D as input and outputs (predicts) a suspected anomaly.

[0041] Furthermore, the information processing device 101 may determine that an abnormality among the abnormalities A to D, whose integrated feature amount is equal to or greater than a threshold, is an abnormality in the abnormality prediction target 102. Furthermore, the information processing device 101 may determine that an abnormality among the abnormalities A to D, whose integrated feature amount is the highest, is an abnormality in the abnormality prediction target 102.

[0042] In this way, according to the information processing device 101, the second features f5 to f8 extracted from the second type of data (e.g., second data d2) can be optimized using the importance of each abnormality (e.g., abnormalities A to D) defined according to the information content included in the second type of data.

[0043] This enables the information processing device 101 to improve the accuracy of predicting an anomaly for the anomaly prediction target 102. For example, even if a feature amount of a serious anomaly that should be suspected is slight, the information processing device 101 can prevent a serious anomaly from being overlooked due to the feature amount of another anomaly becoming dominant by weighting the feature amount using the importance.

[0044] (System configuration example of information processing system 200) Next, a system configuration example of an information processing system 200 including the information processing device 101 shown in Fig. 1 will be described. Here, the case where the information processing device 101 shown in Fig. 1 is applied to an abnormality prediction device 201 in the information processing system 200 will be described as an example.

[0045] Fig. 2 is an explanatory diagram showing an example of the system configuration of an information processing system 200. In Fig. 2, the information processing system 200 includes an abnormality prediction device 201 and a client device 202. In the information processing system 200, the abnormality prediction device 201 and the client device 202 are connected via a wired or wireless network 210. The network 210 is, for example, the Internet, a LAN (Local Area Network), or a WAN (Wide Area Network).

[0046] Here, the abnormality prediction device 201 is a computer that has an importance table 220 and predicts an abnormality for an abnormality prediction target. The abnormality prediction target may be, for example, a patient. The abnormality prediction target may also be, for example, a device such as a multifunction peripheral or medical equipment. The abnormality prediction device 201 is, for example, a server. The contents stored in the importance table 220 will be described later with reference to FIG. 4.

[0047] The client device 202 is a computer used by a user of the information processing system 200. The user may be, for example, a doctor at a medical institution or a customer engineer (CE) who performs maintenance and inspection of equipment. The client device 202 may be, for example, a personal computer (PC), a tablet PC, or a smartphone.

[0048] Although the anomaly prediction device 201 and the client device 202 are provided separately here, this is not limiting. For example, the anomaly prediction device 201 may be realized by the client device 202. Furthermore, the information processing system 200 may include multiple client devices 202.

[0049] (Example of hardware configuration of abnormality prediction device 201) Next, an example of the hardware configuration of the abnormality prediction device 201 will be described.

[0050] Fig. 3 is a block diagram showing an example of the hardware configuration of the abnormality prediction device 201. In Fig. 3, the abnormality prediction device 201 has a CPU (Central Processing Unit) 301, a memory 302, a disk drive 303, a disk 304, a communication I / F (Interface) 305, a portable recording medium I / F 306, and a portable recording medium 307. Furthermore, each component is connected to each other by a bus 300.

[0051] Here, the CPU 301 is responsible for overall control of the anomaly prediction device 201. The CPU 301 may have multiple cores. The memory 302 includes, for example, a read-only memory (ROM) and a random access memory (RAM). A program stored in the memory 302 is loaded into the CPU 301, causing the CPU 301 to execute coded processing.

[0052] The disk drive 303 controls reading / writing of data from / to the disk 304 under the control of the CPU 301. The disk 304 stores data written under the control of the disk drive 303. The disk 304 is, for example, a magnetic disk, an optical disk, or the like.

[0053] The communication I / F 305 is connected to the network 210 via a communication line, and is connected to an external computer (for example, the client device 202 shown in FIG. 2) via the network 210. The communication I / F 305 serves as an interface between the network 210 and the inside of the device, and controls input and output of data from and to the external computer. The communication I / F 305 is, for example, a modem or a LAN adapter.

[0054] The portable recording medium I / F 306 controls reading and writing of data from and to the portable recording medium 307 under the control of the CPU 301. The portable recording medium 307 stores data written under the control of the portable recording medium I / F 306. The portable recording medium 307 is, for example, a CD (Compact Disc)-ROM, a DVD (Digital Versatile Disk), or a USB (Universal Serial Bus) memory.

[0055] In addition to the above-described components, the abnormality prediction device 201 may also have, for example, an input device, a display, etc. Furthermore, the abnormality prediction device 201 may not have, for example, the portable recording medium I / F 306 and the portable recording medium 307 among the above-described components. Furthermore, the client device 202 shown in FIG. 2 can also be realized by the same hardware configuration as the abnormality prediction device 201. However, in addition to the above-described components, the client device 202 also has, for example, an input device, a display, etc.

[0056] (Storage contents of importance table 220) Next, the contents stored in the importance table 220 of the abnormality prediction device 201 will be described with reference to Fig. 4. The importance table 220 is realized by a storage device such as the memory 302 and the disk 304 shown in Fig. 3, for example.

[0057] In the following description, a specific example may be described in which the target of abnormality prediction is a "patient" and the predicted abnormality is a "disease."

[0058] Fig. 4 is an explanatory diagram showing an example of the contents stored in the importance table 220. In Fig. 4, the importance table 220 has fields for disease name, body temperature, and blood pressure, and stores importance information (for example, importance information 400-1 to 400-4) as records by setting information in each field.

[0059] Here, the disease name is a name that identifies the disease that is causing the abnormality. Body temperature is one piece of information (examination item) included in the test value data. Blood pressure is one piece of information (examination item) included in the test value data. Test value data is an example of numeric data (type: numeric).

[0060] For example, importance information 400-1 indicates an importance of "0.1" for the disease "anemia" defined for the test item "body temperature" included in the test value data. Also, importance information 400-1 indicates an importance of "0.7" for the disease "anemia" defined for the test item "blood pressure" included in the test value data. Importance is expressed as a value between 0 and 1, with larger values ​​indicating higher importance.

[0061] The contents stored in the importance table 220 may be manually set by someone with domain knowledge (such as the knowledge of an experienced doctor or medical textbooks). The importance table 220 may also store learning results obtained by the abnormality prediction device 201 (for example, the learning unit 705, which will be described later). The importance table 220 may also store information obtained by manually adjusting the learning results obtained by the abnormality prediction device 201. For example, the importance of each disease may be adjusted taking into account the mortality rate of each disease. The target of abnormality prediction may be defined as "equipment," and the predicted abnormality may be defined as "failure." In this case, the importance of each failure may be adjusted taking into account the number of tools and parts required for each failure, the magnitude of the recovery cost, and the like.

[0062] (Examples of multiple types of data) Next, a specific example of multiple types of data related to an abnormality prediction target will be described with reference to Figures 5 and 6. Here, the multiple types of data are "text data" and "numeric data."

[0063] Fig. 5 is an explanatory diagram showing a specific example of text data. In Fig. 5, medical record data 500 is an example of text data related to a patient (an abnormality prediction target), and is information describing the patient's condition, progress, etc. The medical record data 500 describes the patient's symptom of "dizziness."

[0064] Fig. 6 is an explanatory diagram showing a specific example of laboratory value data. In Fig. 6, laboratory value data 600 is an example of numerical data related to a patient (a subject for which abnormality is predicted), and includes laboratory value data 601 and 602. Each of the laboratory value data 601 and 602 represents the patient's test results (test values ​​for test items).

[0065] Test value data 601 represents a test value of "39.5° C." for the test item "body temperature." Test value data 602 represents a test value of "74-118" for the test item "blood pressure."

[0066] (Example of functional configuration of abnormality prediction device 201) Next, an example of the functional configuration of the abnormality prediction device 201 will be described.

[0067] FIG. 7 is a block diagram showing an example of the functional configuration of the abnormality prediction device 201. In FIG. 7, the abnormality prediction device 201 includes an acquisition unit 701, an extraction unit 702, a correction unit 703, a prediction unit 704, a learning unit 705, an output unit 706, and a storage unit 710. The acquisition unit 701 to the output unit 706 are functions that constitute the control unit 700, and specifically, the functions are realized by, for example, having the CPU 301 execute a program stored in a storage device such as the memory 302, the disk 304, or the portable recording medium 307 shown in FIG. 3, or by the communication I / F 305. The processing results of each functional unit are stored in a storage device such as the memory 302 or the disk 304. The storage unit 710 is realized by, for example, a storage device such as the memory 302 or the disk 304. Specifically, the storage unit 710 stores the importance table 220 shown in FIG. 4, for example.

[0068] The acquiring unit 701 acquires multiple types of data related to an abnormality prediction target. The abnormality prediction target is, for example, a patient in a medical institution. In this case, the abnormality is, for example, multiple disease candidates for the patient (infection, anemia, renal failure, etc.). The abnormality prediction target may also be a device such as a multifunction peripheral. In this case, the abnormality is, for example, multiple malfunction candidates for the device (breakage, deterioration, defects, etc. of each part).

[0069] The multiple types of data include a first type of first data and a second type of second data. The first type is, for example, any of text, image, and numerical value. The second type is, for example, any of text, image, and numerical value that is different from the first type. The first data is, for example, the medical record data 500 shown in FIG. 5. The second data is, for example, the test value data 600 shown in FIG. 6.

[0070] Specifically, for example, the acquiring unit 701 may acquire multiple types of data (e.g., medical record data 500, test value data 600) included in the abnormality prediction request by receiving the abnormality prediction request from the client device 202 shown in Fig. 2. Alternatively, the acquiring unit 701 may acquire multiple types of data included in the abnormality prediction request by accepting the abnormality prediction request through a user's operation input using an input device (not shown).

[0071] The abnormality prediction request is a request to predict an abnormality for an abnormality prediction target. The abnormality prediction request includes multiple types of data (e.g., medical record data 500, test value data 600). The abnormality prediction request may also include information that identifies the abnormality prediction target (e.g., patient name, device name, etc.).

[0072] The extraction unit 702 extracts, from the acquired first data of the first type, a first feature amount representing the likelihood of each anomaly included in a plurality of candidate anomalies for the target of anomaly prediction. The extraction unit 702 also extracts, from the acquired second data of the second type, a second feature amount representing the likelihood of each anomaly included in a plurality of candidate anomalies for the target of anomaly prediction.

[0073] Each feature amount (first feature amount, second feature amount) represents the likelihood of each abnormality. Each feature amount is represented by a value between 0 and 1, for example, and the larger the value, the higher the possibility of each abnormality. Any existing technology, such as NLP or machine learning, may be used to extract each feature amount (first feature amount, second feature amount).

[0074] In the following explanation, multiple candidate anomalies for an anomaly prediction target will be referred to as "anomalies #1 to #n" (n is a natural number equal to or greater than 2), and any one of anomalies #1 to #n will be referred to as "anomaly #i" (i = 1, 2, ..., n). Also, the first feature quantity indicating the likelihood of anomaly #i will be referred to as the "first feature quantity (#i) of anomaly #i." Also, the second feature quantity indicating the likelihood of anomaly #i will be referred to as the "second feature quantity (#i) of anomaly #i."

[0075] Here, the first data is "medical record data 500 (text data)." In this case, the extraction unit 702 extracts a first feature (#i) of each abnormality #i from the medical record data 500 using, for example, an analysis technique such as NLP. Also, the second data is "lab value data 600 (numeric data)." In this case, the extraction unit 702 extracts a second feature (#i) of each abnormality #i from each laboratory value data 601, 602 in the laboratory value data 600 using, for example, machine learning such as deep learning or decision tree. Also, if the first data or the second data is "image data," the extraction unit 702 may extract a feature of each abnormality #i from the image data using, for example, machine learning such as deep learning or decision tree.

[0076] Here, the anomaly prediction device 201 extracts the feature quantities (first feature quantity (#i), second feature quantity (#i)) of each anomaly #i from each data (first data, second data), but this is not limiting. For example, the extraction unit 702 may obtain, from another computer (such as the client device 202), information representing the feature quantities of each anomaly #i extracted from each data in the other computer.

[0077] The correction unit 703 corrects the second feature amount (#i) of each abnormality #i based on the importance of each abnormality #i according to the information content included in the second data, with reference to the storage unit 710. Here, the storage unit 710 stores the importance of each abnormality #i defined according to the information content included in the second type of data.

[0078] The importance of an abnormality #i corresponds to an index of how much importance (attention) should be given to the information included in the second type of data when determining an abnormality #i. The importance of an abnormality #i is expressed, for example, by a value between 0 and 1, with a larger value indicating a higher degree of importance. The importance of an abnormality #i is used to optimize the second feature (anomaly #i).

[0079] For example, suppose the second type of data is "text data." In this case, the importance of each anomaly #i is defined according to the information content specified by, for example, characters, symbols, words, sentences, etc. contained in the text data. Also, suppose the second type of data is "numeric data." In this case, the importance of each anomaly #i is defined according to the information content specified by, for example, data items contained in the numeric data. Also, suppose the second type of data is "image data." In this case, the importance of each anomaly #i is defined according to the information content specified by, for example, characters, symbols, words, sentences, objects, etc. contained in the image data.

[0080] Here, the second type of data is defined as "numerical data," and the second data is defined as "test value data 600 (601, 602)." In this case, the correction unit 703, for example, refers to the importance table 220 shown in FIG. 4 to identify the importance of each abnormality #i according to the test item (body temperature, blood pressure) included in each test value data 601, 602.

[0081] For example, if abnormality #i is "anemia," the importance of the abnormality "anemia" corresponding to the test item "body temperature" included in test value data 601 is "0.1." Also, the importance of the abnormality "anemia" corresponding to the test item "blood pressure" included in test value data 602 is "0.7."

[0082] Then, the correction unit 703 corrects the second feature amount (#i) of each abnormality #i based on the importance of each identified abnormality #i. More specifically, for each abnormality #i, the correction unit 703 corrects the second feature amount (#i) of each abnormality #i by multiplying the second feature amount (#i) of each abnormality #i by the identified importance of each abnormality #i.

[0083] Here, the case where the second feature amount (#i) of each anomaly #i is corrected by multiplying it by the importance of each anomaly #i has been described as an example, but this is not limiting. For example, the correction unit 703 may correct the second feature amount (#i) of each anomaly #i by adding a value obtained by multiplying the identified importance of each anomaly #i by a constant α. The constant α is a value smaller than 1 and can be set arbitrarily.

[0084] In addition, although the importance of an abnormality #i is assumed to be higher as the numerical value increases, this is not limiting. For example, the importance of an abnormality #i is assumed to be higher as the numerical value decreases. In this case, the correction unit 703 may correct the second feature amount (#i) of each abnormality #i by dividing the identified importance of each abnormality #i by the second feature amount (#i) of each abnormality #i.

[0085] Furthermore, the correction unit 703 may correct the second feature amount (#i) of each abnormality #i based on the importance of each abnormality #i according to the information content included in the second data and the first feature amount (#i) of each abnormality #i. Here, among the abnormalities #i, for example, there may be some that are easy to predict from the first type of data and some that are easy to predict from the second type of data.

[0086] For example, some diseases can be predicted with higher accuracy using text data than using numerical data, while others can be predicted with higher accuracy using numerical data than using text data. Therefore, when optimizing the second feature (#i) of each abnormality #i, correcting the second feature (#i) without considering the first feature (#i) of each abnormality #i may not result in a sufficient improvement in prediction accuracy.

[0087] Therefore, the correction unit 703 may correct the second feature amount (#i) of each anomaly #i by taking into account the weight of suspicion of each anomaly #i identified from the first type of data. More specifically, for example, the correction unit 703 may correct the second feature amount (#i) of each anomaly #i by multiplying the importance of each identified anomaly #i by the first feature amount (#i) of each anomaly #i.

[0088] Here, the case where the second feature quantity (#i) of each anomaly #i is corrected by multiplying it by the first feature quantity (#i) of each anomaly #i has been described as an example, but this is not limiting. For example, the correction unit 703 may correct the second feature quantity (#i) of each anomaly #i by multiplying it by the importance of each anomaly #i, or by adding the value obtained by multiplying the first feature quantity (#i) of each anomaly #i by a constant β. The constant β is a value smaller than 1 and can be set arbitrarily.

[0089] The prediction unit 704 determines whether an abnormality is present among the abnormalities #1 to #n for the abnormality prediction target based on the first feature amount (#i) of each abnormality #i and the corrected second feature amount (#i) of each abnormality #i. Specifically, for example, the prediction unit 704 calculates an integrated feature amount for each abnormality #i by adding together the first feature amount (#i) of each abnormality #i and the corrected second feature amount (#i) of each abnormality #i. Then, the prediction unit 704 determines whether an abnormality is present for the abnormality prediction target based on the calculated integrated feature amount for each abnormality #i.

[0090] To explain in more detail, for example, the prediction unit 704 may input the integrated feature of each anomaly #i into a prediction model, and determine that the anomaly predicted by the prediction model is an anomaly for the anomaly prediction target. The prediction model is, for example, a model that inputs the integrated feature of each anomaly #i and outputs (predicts) a suspected anomaly. As the prediction model, any existing prediction model such as kNN (k-Nearest Neighbor method) or SVM (Support Vector Machine) may be used.

[0091] Furthermore, the prediction unit 704 may determine that an anomaly having an integrated feature amount equal to or greater than a threshold value among the anomalies #1 to #n is an anomaly for the anomaly prediction target. The threshold value may be set arbitrarily. Furthermore, the prediction unit 704 may determine that the top N anomalies having the highest integrated feature amounts among the anomalies #1 to #n are anomalies for the anomaly prediction target. N may be set arbitrarily, for example, to about 1 to 3.

[0092] For example, by setting N to "N=3", the prediction unit 704 can determine the top three most probable anomalies among anomalies #1 to #n as anomalies for the anomaly prediction target. In this case, the prediction unit 704 may rank the three anomalies determined for the anomaly prediction target in descending order of the integrated feature amount. The higher the rank, the more suspicious (probable) the anomaly is.

[0093] The extraction unit 702 may extract features representing the likelihood of each anomaly #i from other types of data than the first data of the first type and the second data of the second type among the multiple types of data. In this case, the prediction unit 704 may further use the features extracted from the other types of data when determining an anomaly for the anomaly prediction target.

[0094] Specifically, for example, the prediction unit 704 may calculate an integrated feature for each anomaly #i by adding together the first feature (#i) for each anomaly #i, the corrected second feature (#i) for each anomaly #i, and features extracted from other types of data. Then, the prediction unit 704 may determine an anomaly for the anomaly prediction target based on the calculated integrated feature for each anomaly #i.

[0095] The output unit 706 outputs the determination result. The output format of the output unit 706 may be, for example, storage in a storage device such as the memory 302 or the disk 304, transmission to another computer via the communication I / F 305, display on a display (not shown), or printout on a printer (not shown).

[0096] Specifically, for example, the output unit 706 may output an abnormality prediction result including the determined determination result as a response to the abnormality prediction request. The abnormality prediction result includes information identifying the abnormality prediction target (e.g., patient name, device name, etc.) and information identifying the abnormality determined for the abnormality prediction target (e.g., disease name, fault name). Furthermore, when N abnormalities are determined for the abnormality prediction target, the abnormality prediction result may include information identifying the ranking of each abnormality. A higher ranking indicates a more suspicious (highly likely) abnormality.

[0097] The learning unit 705 learns the importance of each abnormality #i according to the information content included in the second type of data based on the second type of past data for each abnormality #i. Any existing method such as a decision tree may be used to learn the importance of each abnormality #i.

[0098] For example, the second type of data may be "numerical data." In this case, the acquiring unit 701 acquires the numerical data (past data) for each abnormality #i as learning data. Then, the learning unit 705 may calculate the importance of each abnormality #i according to the data item by learning (supervised learning), for example, how much a data item included in the acquired learning data contributes to predicting each abnormality #i. More specifically, for example, the learning unit 705 may learn which data item has a characteristic value when an abnormality #i occurs, and calculate the importance of the abnormality #i such that the value increases for a data item with a characteristic value.

[0099] The second type of data is defined as "text data." In this case, the acquiring unit 701 acquires text data (past data) for each abnormality #i as learning data. Then, the learning unit 705 may calculate the importance of each abnormality #i according to the words contained in the acquired learning data, for example, by learning (supervised learning) how much the words contained in the acquired learning data contribute to predicting each abnormality #i. More specifically, for example, the learning unit 705 may learn which words are frequently included in the abnormality #i, and calculate the importance of the abnormality #i such that the more frequently a word appears, the larger the value.

[0100] The learned importance of each abnormality #i (e.g., disease) is associated with the information content (e.g., test item) included in the second type of data (e.g., test value data) and stored in the storage unit 710. In this case, the correction unit 703 may refer to the importance of each abnormality #i according to the information content included in the learned second type of data, and correct the second feature amount (#i) of each abnormality #i based on the importance of each abnormality #i according to the information content included in the second data.

[0101] An example of learning the importance of each abnormality #i will be described later with reference to FIGS.

[0102] The functional units (acquisition unit 701 to output unit 706) of the anomaly prediction device 201 may be realized by a plurality of computers (for example, the anomaly prediction device 201 and the client device 202) in the information processing system 200. In this case, communication between the functional units of different computers is performed by, for example, transmission and reception between the functional units via the network 210.

[0103] (Example of learning the importance of each anomaly #i) Next, an example of learning the importance of each abnormality #i will be described using Figures 8 and 9. First, an example of learning the importance of each abnormality #i will be described using Figure 8, taking as an example a case where the second type of data is "numerical data." Here, the abnormality prediction target is a "patient," the abnormality #i is a "disease," and the numerical data is "lab value data." Furthermore, the test items included in the laboratory value data are "body temperature" and "blood pressure."

[0104] FIG. 8 is an explanatory diagram (part 1) showing an example of learning importance. In FIG. 8, an acquisition unit 701 acquires past data 800. The past data 800 includes laboratory test data (numerical data) of a patient with a certain disease (abnormality #i). The laboratory test data is, for example, the laboratory test data 600 shown in FIG. 6.

[0105] The past data 800 includes a group 801 of test value data of a patient with the disease "anemia," a group 802 of test value data of a patient with the disease "cerebrovascular disease," a group 803 of test value data of a patient with the disease "renal failure," and a group 804 of test value data of a patient with the disease "infectious disease."

[0106] The learning unit 705 learns the importance of each disease corresponding to the test item included in the test value data based on the acquired past data 800. Specifically, for example, the learning unit 705 learns the importance of the test item "body temperature" for the disease "anemia" based on the test value data group 801. The learning unit 705 also learns the importance of the test item "blood pressure" for the disease "anemia" based on the test value data group 801.

[0107] The importance of the learned disease "anemia" is associated with, for example, the test item "body temperature" and stored in the importance table 220. The importance of the learned disease "anemia" is also associated with the test item "blood pressure" and stored in the importance table 220. As a result, importance information 400-1 is stored in the importance table 220 as a record.

[0108] According to the importance information 400-1, when determining the disease "anemia," the test item "blood pressure" is more important than the test item "body temperature." Similarly, for the diseases "cerebrovascular disease," "renal failure," and "infectious disease," the importance corresponding to each test item "body temperature" and "blood pressure" is learned, and importance information 400-2 to 400-4 is stored in the importance table 220.

[0109] In this way, the abnormality prediction device 201 can learn the importance of each disease according to the test items “body temperature” and “blood pressure” included in the test value data, based on past test value data for each disease (abnormality #i).

[0110] Next, an example of learning the importance of each abnormality #i will be explained using Fig. 9, taking the case where the second type of data is "text data" as an example. Here, the abnormality prediction target is a "patient," the abnormality #i is a "disease," and the text data is "medical record data."

[0111] FIG. 9 is an explanatory diagram (part 2) showing an example of learning importance. In FIG. 9, the acquisition unit 701 acquires past data 900. The past data 900 includes medical record data (text data) of a patient at the time of a certain disease (abnormality #i). The medical record data is, for example, the medical record data 500 shown in FIG. 5.

[0112] The past data 900 includes a group 901 of medical record data of patients with the disease "anemia," a group 902 of medical record data of patients with the disease "cerebrovascular disease," a group 903 of medical record data of patients with the disease "renal failure," and a group 904 of medical record data of patients with the disease "infectious disease."

[0113] The learning unit 705 learns the importance of each disease according to the information content contained in the medical record data based on the acquired past data 900. Here, the information content contained in the medical record data is assumed to be words. The words are identified, for example, by morphological analysis of the medical record data. Here, "body temperature" and "headache" are used as examples of words contained in the medical record data.

[0114] Specifically, for example, the learning unit 705 learns the importance of the disease "anemia" according to the word "body temperature" based on the medical record data group 901. The learning unit 705 also learns the importance of the disease "anemia" according to the word "headache" based on the medical record data group 901.

[0115] The learned importance of the disease "anemia" is associated with, for example, the word "body temperature" and stored in the importance table 220. The learned importance of the disease "anemia" is also associated with the word "headache" and stored in the importance table 220. As a result, importance information 900-1 is stored as a record in the importance table 220.

[0116] According to the importance information 900-1, when determining the disease "anemia," the word "headache" is considered to be more important than the word "body temperature." Similarly, for the diseases "cerebrovascular disease," "renal failure," and "infectious disease," the importance corresponding to the words "body temperature" and "headache" is learned, and importance information 900-2 to 900-4 is stored in the importance table 220.

[0117] In this way, the abnormality prediction device 201 can learn the importance of each disease according to words (for example, body temperature, headache) contained in the medical record data based on past medical record data at the time of each disease (abnormality #i).

[0118] When the second type of data is "medical record data," for example, multiple importance levels may be identified from the importance information 900-1 when correcting (optimizing) the second feature value of the disease "anemia." For example, if the medical record data (second data) includes the words "body temperature" and "headache," the importance information 900-1 may identify an importance level of "0.1" corresponding to the word "body temperature" and an importance level of "0.7" corresponding to the word "headache." In this case, the correction unit 703 may correct the second feature value of the disease "anemia" using, for example, the highest importance level of "0.7" among the identified multiple importance levels. This allows the correction unit 703 to correct the second feature value of the disease "anemia" by taking into account the word with the highest importance level included in the medical record data (second data).

[0119] (Example of operation of the abnormality prediction device 201) Next, an example of the operation of the abnormality prediction device 201 will be described with reference to Figs. 10 to 15. Here, the target of abnormality prediction is "Patient X" and the abnormality #i is "Disease". Assume that Patient X suffers from the disease "renal failure" and the correct answer of the abnormality prediction result is "renal failure". Also, the first data of the first type is "medical record data 500", and the second data of the second type is "test value data 600".

[0120] 10 to 15 are explanatory diagrams showing an example of the operation of the abnormality prediction device 201. In Fig. 10, the abnormality prediction device 201 extracts, from the medical record data 500, a first feature amount that indicates the likelihood of each disease included in a plurality of candidate diseases for patient X. The plurality of diseases include, for example, anemia, cerebrovascular disease, renal failure, and infectious disease.

[0121] The text feature 1000 indicates a first feature that represents the likelihood of each disease (anemia, cerebrovascular disease, renal failure, infectious disease, etc.). For example, the first feature that represents the likelihood of the disease "anemia" is "0.2."

[0122] 11, the abnormality prediction device 201 extracts, from the test value data 600 (601, 602), a second feature that indicates the likelihood of each disease included in a plurality of candidate diseases for patient X. The plurality of diseases include, for example, anemia, cerebrovascular disease, renal failure, and infectious diseases.

[0123] The body temperature (numerical value) feature 1110 indicates a second feature that indicates the likelihood of each disease (anemia, cerebrovascular disease, renal failure, infectious disease, etc.) extracted from the test value data 601. For example, the second feature that indicates the likelihood of the disease "anemia" is "0.1."

[0124] Furthermore, the blood pressure (numerical value) feature 1120 indicates a second feature that indicates the likelihood of each disease (anemia, cerebrovascular disease, renal failure, infectious disease, etc.) extracted from the test value data 602. For example, the second feature that indicates the likelihood of the disease "anemia" is "0.15."

[0125] 12, the abnormality prediction device 201 optimizes the body temperature (numeric value) feature 1110. Specifically, for example, the abnormality prediction device 201 first refers to the importance table 220 (see, for example, FIG. 4) to identify the importance of each disease corresponding to the test item "body temperature" included in the test value data 601. Here, it is assumed that "importance 1201" is identified as the importance of each disease corresponding to the test item "body temperature."

[0126] Then, for each disease (anemia, cerebrovascular disease, renal failure, infectious disease, etc.), the abnormality prediction device 201 corrects the body temperature (numeric value) feature 1110 by multiplying the identified importance 1201 and the text feature 1000. Hereinafter, the corrected body temperature (numeric value) feature 1110 is referred to as the "optimized feature (body temperature) 1202."

[0127] For example, in the optimized feature (body temperature) 1202, the optimized second feature for the disease "anemia" is "0.002 (= 0.2 × 0.1 × 0.1)". Furthermore, the optimized second feature for the disease "cerebrovascular disease" is "0.015 (= 0.25 × 0.3 × 0.2)". Furthermore, the optimized second feature for the disease "renal failure" is "0.005 (= 0.2 × 0.5 × 0.05)". Furthermore, the optimized second feature for the disease "infectious disease" is "0.075 (= 0.25 × 0.6 × 0.5)".

[0128] As a result, the abnormality prediction device 201 can optimize the body temperature (numerical) feature 1110 by taking into account the weight of each suspected disease identified from the text feature 1000 and weighting it in consideration of the degree of importance of each disease according to the test item ``body temperature.''

[0129] 13, the abnormality prediction device 201 optimizes the blood pressure (numerical value) feature 1120. Specifically, for example, the abnormality prediction device 201 first refers to the importance table 220 (see, for example, FIG. 4) to identify the importance of each disease corresponding to the test item "blood pressure" included in the test value data 601. Here, it is assumed that "importance 1301" is identified as the importance of each disease corresponding to the test item "blood pressure."

[0130] Then, for each disease (anemia, cerebrovascular disease, renal failure, infectious disease, etc.), the abnormality prediction device 201 corrects the blood pressure (numeric) feature 1120 by multiplying the identified importance 1301 and the text feature 1000. Hereinafter, the corrected blood pressure (numeric) feature 1120 is referred to as the "optimized feature (blood pressure) 1302."

[0131] For example, in the optimized feature (blood pressure) 1302, the optimized second feature for the disease "anemia" is "0.0003 (= 0.2 × 0.01 × 0.15)". The optimized second feature for the disease "cerebrovascular disease" is "0.0005 (= 0.25 × 0.01 × 0.2)". The optimized second feature for the disease "renal failure" is "0.024 (= 0.2 × 0.8 × 0.15)". The optimized second feature for the disease "infectious disease" is "0.000625 (= 0.25 × 0.01 × 0.25)".

[0132] As a result, the abnormality prediction device 201 can optimize the blood pressure (numerical) feature 1120 by taking into account the weight of each suspected disease identified from the text feature 1000 and weighting it in consideration of the degree of importance of each disease according to the test item "blood pressure."

[0133] 14, the abnormality prediction device 201 calculates an integrated feature for each disease (anemia, cerebrovascular disease, renal failure, infectious disease, etc.). Specifically, for each disease, the abnormality prediction device 201 calculates an integrated feature 1400 by adding together a text feature 1000, an optimized feature (body temperature) 1202, and an optimized feature (blood pressure) 1302.

[0134] Although not shown in the figure, there may be optimized features for other test items other than body temperature and blood pressure. In this case, the abnormality prediction device 201 further adds up the optimized features to calculate the integrated feature 1400. Furthermore, in addition to the text feature 1000, the body temperature (numeric) feature 1110, and the blood pressure (numeric) feature 1120, there may be image features extracted from the image data. In this case, the abnormality prediction device 201 further adds up the image features to calculate the integrated feature 1400.

[0135] The integrated feature 1400 indicates the integrated feature of each disease (anemia, cerebrovascular disease, renal failure, infectious disease, etc.). In the integrated feature 1400, among multiple candidate diseases (#1 to #n), the integrated feature of the disease "renal failure" is the highest.

[0136] 15, the abnormality prediction device 201 determines a disease for patient X based on the integrated feature 1400. Specifically, for example, the abnormality prediction device 201 may input the integrated feature 1400 (integrated feature for each disease) into a prediction model, and thereby determine that the disease predicted by the prediction model is the disease for patient X.

[0137] Furthermore, the abnormality prediction device 201 may determine that the disease with the highest integrated feature value among multiple candidate diseases (anemia, cerebrovascular disease, renal failure, infectious disease, etc.) is the abnormality for patient X. Here, as shown in disease certainty 1500, the higher the integrated feature value of the disease, the higher the certainty that the patient has that disease.

[0138] In this case, the abnormality prediction device 201 determines that the disease with the highest integrated feature amount, "renal failure," is the disease of patient X. Then, the abnormality prediction device 201 outputs an abnormality prediction result 1501. The abnormality prediction result 1501 includes the determined disease, "renal failure," of patient X. Since the abnormality prediction result 1501 matches the disease, "renal failure," that patient X actually suffers from, it can be said to be a highly accurate prediction result.

[0139] For example, an experienced doctor will perform differential diagnosis in the following order (i) to (iii).

[0140] (i) Based on the results of interviews with patients (e.g., equivalent to 500 medical records), list several likely diseases. (ii) Blood tests, urine tests, etc. are performed (e.g., equivalent to the test value data 600), and attention is paid to test values ​​that are closely related to the diseases listed in (i) above. (iii) Based on a comprehensive assessment of the medical record data and test value data, diseases with low certainty are excluded from the differential diagnosis.

[0141] In optimizing the numerical feature quantities (1110, 1120) obtained from the test value data 600 (601, 602), the abnormality prediction device 201 can weight the numerical feature quantities (1110, 1120) as described in (ii) above. Then, by integrating the text feature quantity 1000 and the optimized feature quantities (1202, 1302), the abnormality prediction device 201 can comprehensively determine the feature quantities of each data and diagnose the disease of patient X as described in (iii) above. Therefore, the abnormality prediction device 201 can predict abnormalities with the same high accuracy as an experienced doctor.

[0142] (Example of disease prediction) Next, an example of disease prediction will be described. First, an example of disease prediction for patient Y will be described with reference to FIG. 16. Here, it is assumed that patient Y has symptoms of "headache" and "fever." It is also assumed that patient Y has the disease "brain tumor," and the correct answer of the abnormality prediction result is "brain tumor."

[0143] Fig. 16 is an explanatory diagram (part 1) showing an example of disease prediction. In Fig. 16, text feature 1601 indicates a first feature (e.g., equivalent to "likelihood") that indicates the likelihood of each disease (brain tumor, anemia, renal failure, infection, cerebral hemorrhage) extracted from medical record data (not shown) related to patient Y. In text feature 1601, since patient Y has symptoms of "headache" and "fever," the first feature for the diseases "brain tumor, anemia, renal failure, infection" is higher than that for the disease "cerebral hemorrhage."

[0144] Furthermore, the numerical feature 1602 indicates a second feature (e.g., equivalent to "likelihood") that indicates the likelihood of each disease (brain tumor, anemia, renal failure, infection, intracerebral hemorrhage) extracted from the test value data (not shown) related to patient Y. In the numerical feature 1602, since patient Y has the symptom of "fever," the second feature for the disease "infectious disease" is extremely high compared to the other diseases.

[0145] Furthermore, the importance level 1603 indicates the importance level for each disease (brain tumor, anemia, renal failure, infection, intracerebral hemorrhage) according to the test item included in the test value data (not shown) for patient Y. In the importance level 1603, the diseases "brain tumor" and "intracerebral hemorrhage" have relatively high importance levels.

[0146] The abnormality prediction device 201 corrects the numerical feature 1602 by multiplying the numerical feature 1602 by the text feature 1601 and the importance 1603 for each disease (brain tumor, anemia, renal failure, infection, intracerebral hemorrhage).

[0147] This allows weighting to be performed taking into account the weight of each suspected disease identified from the text feature 1601, while also considering the degree of importance according to the test item included in the test value data (not shown), thereby optimizing the numerical feature 1602.

[0148] The abnormality prediction device 201 calculates an integrated feature 1604 by adding together the text feature 1601 and the corrected numerical feature 1602 for each disease (brain tumor, anemia, renal failure, infection, and intracerebral hemorrhage). The integrated feature 1604 indicates the integrated feature for each disease. In the integrated feature 1604, the integrated feature for the disease "brain tumor" is the highest.

[0149] The abnormality prediction device 201 determines the disease of patient Y based on the calculated integrated feature 1604. In the integrated feature 1604, the integrated feature for the disease "brain tumor" is the highest, and "brain tumor" is predicted as the suspected disease of patient Y.

[0150] In this way, the abnormality prediction device 201 can predict the disease "brain tumor" as the disease suspected of patient Y. The disease "brain tumor" matches the disease "brain tumor" that patient Y actually has. Here, although the second feature value of the disease "brain tumor" in the numerical feature value 1602 is low, the second feature value is weighted using the importance 1603 while taking into account the weight of the suspicion of the disease "brain tumor" identified from the text feature value 1601.

[0151] As a result, the likelihood of the disease "infectious disease" in the numerical feature 1602 decreases, and the likelihood of the disease "infectious disease" also decreases in the integrated feature 1604. Therefore, the abnormality prediction device 201 can relatively increase the likelihood of the disease "brain tumor" in the integrated feature 1604, and can prevent the serious disease "brain tumor" from being overlooked due to the feature of another disease (for example, infection) becoming dominant.

[0152] Next, an example of disease prediction for patient Z will be described with reference to FIG. 17. Here, it is assumed that patient Z has a symptom of "fever." It is also assumed that patient Z has the disease "infectious disease," and the correct answer of the abnormality prediction result is "infectious disease."

[0153] Fig. 17 is an explanatory diagram (part 2) showing an example of disease prediction. In Fig. 17, text feature 1701 indicates a first feature that indicates the likelihood of each disease (brain tumor, anemia, renal failure, infection, cerebral hemorrhage) extracted from medical record data (not shown) related to patient Z. In text feature 1701, since patient Z has the symptom of "fever," the first feature for the disease "infectious disease" is higher than for the other diseases.

[0154] Furthermore, the numerical feature 1702 indicates a second feature that indicates the likelihood of each disease (brain tumor, anemia, renal failure, infection, intracerebral hemorrhage) extracted from the test value data (not shown) for patient Z. In the numerical feature 1702, since patient Z has the symptom of "fever," the second feature for the disease "infectious disease" is higher than for the other diseases.

[0155] 16, importance 1703 indicates the importance of each disease (brain tumor, anemia, renal failure, infection, intracerebral hemorrhage) according to the test item included in the test value data (not shown) for patient Z. In importance 1703, similar to importance 1603 shown in FIG. 16, the diseases "brain tumor" and "intracerebral hemorrhage" have relatively high importance.

[0156] The abnormality prediction device 201 corrects the numerical feature 1702 by multiplying the numerical feature 1702 by the text feature 1701 and the importance 1703 for each disease (brain tumor, anemia, renal failure, infection, intracerebral hemorrhage).

[0157] This allows weighting to be performed taking into account the weight of each suspected disease identified from the text feature 1701, while also considering the degree of importance according to the test item included in the test value data (not shown), thereby optimizing the numerical feature 1702.

[0158] The abnormality prediction device 201 calculates an integrated feature 1704 by adding together the text feature 1701 and the corrected numerical feature 1702 for each disease (brain tumor, anemia, renal failure, infection, and intracerebral hemorrhage). The integrated feature 1704 indicates the integrated feature for each disease. In the integrated feature 1704, the integrated feature for the disease "infectious disease" is the highest.

[0159] The abnormality prediction device 201 determines the disease of patient Z based on the calculated integrated feature 1704. In the integrated feature 1704, the integrated feature for the disease "infectious disease" is the highest, and "infectious disease" is predicted as the suspected disease of patient Z.

[0160] In this way, the abnormality prediction device 201 can predict the disease "infectious disease" as the disease suspected for patient Z. The disease "infectious disease" matches the disease "infectious disease" that patient Z actually has. Here, it can be seen that even when the numerical feature 1702 is weighted using the importance 1703 while taking into account the weight of the suspicion of each disease identified from the text feature 1701, the disease for patient Z can be predicted with high accuracy.

[0161] (Procedure for Abnormality Prediction Processing by Abnormality Prediction Device 201) Next, the abnormality prediction process procedure of the abnormality prediction device 201 will be described with reference to FIG. 18. Here, the target of abnormality prediction is a "patient," and abnormality #i is a "disease." Furthermore, the first type of first data is "medical record data," and the second type of second data is "test value data."

[0162] Fig. 18 is a flowchart showing an example of an abnormality prediction procedure of the abnormality prediction device 201. In the flowchart of Fig. 18, first, the abnormality prediction device 201 refers to the importance table 220 to determine whether or not there is importance information (step S1801).

[0163] If there is no importance information (step S1801: No), the abnormality prediction device 201 executes importance learning processing (step S1802) and proceeds to step S1803. The importance learning processing is processing for learning the importance of each candidate disease. The specific processing procedure of the importance learning processing will be described later with reference to FIG. 19.

[0164] On the other hand, if there is importance information (step S1801: Yes), the abnormality prediction device 201 determines whether or not an abnormality prediction request has been received from the client device 202 (step S1803). Here, the abnormality prediction device 201 waits to receive the abnormality prediction request (step S1803: No).

[0165] When the abnormality prediction device 201 receives the abnormality prediction request (step S1803: Yes), the abnormality prediction device 201 extracts, from the medical record data included in the abnormality prediction request, a first feature amount that indicates the likelihood of each of a plurality of candidate diseases for the patient (the target of abnormality prediction) (step S1804). The medical record data is an example of a first type of first data related to the patient. The first feature amount is extracted using, for example, an analysis method such as NLP.

[0166] Next, the abnormality prediction device 201 extracts second feature values ​​representing the likelihood of each disease from the test value data included in the abnormality prediction request (step S1805). The test value data is an example of a second type of second data related to the patient. The second feature values ​​are extracted using a machine learning method such as Random Forest or Metric Learning. Note that the processing of step S1804 and the processing of step S1805 may be executed in reverse order or may be executed in parallel.

[0167] Next, the abnormality prediction device 201 refers to the importance table 220 to identify the importance for each disease according to the test item included in the test value data (step S1806). Note that if there are multiple test value data with different test items, the abnormality prediction device 201 performs the processes of steps S1805 and S1806 for each of the multiple test value data, for example.

[0168] Then, the abnormality prediction device 201 corrects the extracted second feature amount representing the likelihood of each disease based on the extracted first feature amount representing the likelihood of each disease and the identified importance of each disease using a weighting function (step S1807). The weighting function can be defined, for example, by the following formula (1).

[0169] F2 opt =F1 W F2 (1)

[0170] However, F2 opt indicates the second feature value that indicates the likelihood of each disease after correction (optimization). F1 indicates the first feature value that indicates the likelihood of each disease. F2 indicates the second feature value that indicates the likelihood of each disease. W indicates the importance of each disease.

[0171] Next, the abnormality prediction device 201 calculates an integrated feature for each disease by adding together the first feature representing the likelihood of each disease and the second feature representing the likelihood of each disease after correction (optimization) (step S1808).

[0172] Next, the abnormality prediction device 201 determines the disease of the patient (the target of abnormality prediction) from among the multiple candidate diseases based on the calculated integrated feature of each disease (step S1809).The abnormality prediction device 201 then transmits the abnormality prediction result including the determination result to the client device 202 (step S1810), and ends the series of processes according to this flowchart.

[0173] This allows the abnormality prediction device 201 to accurately predict suspected diseases of a patient (subject of abnormality prediction) from multiple types of data related to the patient.

[0174] Next, a specific processing procedure for the importance learning processing in step S1802 will be described with reference to FIG.

[0175] 19 is a flowchart showing an example of a specific processing procedure for the importance learning process. In the flowchart of FIG. 19, first, the abnormality prediction device 201 acquires past data (step S1901). The past data is learning data for learning the importance of each disease included in multiple candidate diseases, and is test value data (numeric data) of patients collected when they contracted each disease.

[0176] Next, the abnormality prediction device 201 learns the importance of each disease corresponding to the test item included in the test value data based on the acquired past data (step S1902).The abnormality prediction device 201 then stores the learned importance of each disease in the importance table 220 in association with the test item included in the test value data (step S1903), and returns to the step that called the importance learning process.

[0177] This allows the abnormality prediction device 201 to store in the importance table 220 importance information used for optimizing the feature amount (second feature amount) extracted from the test value data.

[0178] As described above, the anomaly prediction device 201 according to the embodiment can acquire first feature quantities representing the likelihood of each anomaly #i included in candidate anomalies #1 to #n, extracted from a first type of first data related to an anomaly prediction target. The anomaly prediction device 201 can also acquire second feature quantities representing the likelihood of each anomaly #i, extracted from a second type of second data related to an anomaly prediction target that is different from the first type. The first type is, for example, text, image, or numerical value. The second type is, for example, any of text, image, and numerical value that is different from the first type. The anomaly prediction device 201 can also correct the acquired second feature quantities based on the importance of each anomaly #i according to the information content included in the second data, with reference to the storage unit 710. The storage unit 710 stores the importance of each anomaly #i defined according to the information content included in the second type of data. The anomaly prediction device 201 can then determine an anomaly from among the anomalies #1 to #n for the anomaly prediction target based on the acquired first feature quantities and the corrected second feature quantities.

[0179] As a result, the anomaly prediction device 201 can optimize the second feature extracted from the second type of second data using the importance for each anomaly #i defined according to the information content included in the second type of data, thereby improving the accuracy of predicting an anomaly for the anomaly prediction target 102. For example, even if a feature of a serious anomaly that should normally be suspected is slight, the anomaly prediction device 201 can prevent a serious anomaly from being overlooked because the feature of another anomaly becomes dominant by weighting it using the importance.

[0180] Furthermore, according to the abnormality prediction device 201, the acquired second feature amount can be corrected based on the importance of each abnormality #i according to the information content included in the second data and the acquired first feature amount.

[0181] As a result, the anomaly prediction device 201 can further optimize the second feature extracted from the second type of second data by taking into account the weight of suspicion identified from the first feature extracted from the first type of first data. As a result, the anomaly prediction device 201 can further improve the accuracy of predicting an anomaly for the anomaly prediction target 102.

[0182] Furthermore, the anomaly prediction device 201 can refer to the storage unit 710 to identify the importance of each anomaly #i according to the information content included in the second data, and for each anomaly #i, multiply the identified importance of each anomaly #i by the first feature amount to correct the second feature amount.The anomaly prediction device 201 can then calculate the integrated feature amount for each anomaly #i by adding the acquired first feature amount and the corrected second feature amount for each anomaly #i, and can determine an anomaly for the anomaly prediction target based on the calculated integrated feature amount for each anomaly #i.

[0183] As a result, the abnormality prediction device 201 can accurately predict a suspected abnormality in a target for abnormality prediction by comprehensively considering the characteristics of each piece of data from a plurality of types of data related to the target for abnormality prediction.

[0184] Furthermore, the abnormality prediction device 201 can learn the importance of each abnormality #i according to the information content included in the second type of data based on the second type of past data at the time of each abnormality #i. Then, the abnormality prediction device 201 can correct the acquired second feature amount based on the importance of each abnormality #i according to the information content included in the second data, by referring to the importance of each abnormality #i according to the learned information content included in the second type of data.

[0185] This allows the anomaly prediction device 201 to learn the importance of each anomaly #i by analyzing information content in which similar features appear for each anomaly #i from data collected when the anomaly #i occurred in the past. Therefore, the anomaly prediction device 201 can obtain the importance used to optimize the feature amount of each anomaly #i without the need for a person with domain knowledge to manually set it.

[0186] Furthermore, the abnormality prediction device 201 can output the determination result.

[0187] As a result, the anomaly prediction device 201 can output the determined result as an anomaly prediction result in response to a request to predict an anomaly for an anomaly prediction target (anomaly prediction request). For example, the anomaly prediction device 201 can provide a service such as predicting a suspected anomaly for an anomaly prediction target.

[0188] From these facts, the abnormality prediction device 201 can improve the prediction accuracy when predicting an abnormality from multiple types of data on the target for abnormality prediction.

[0189] This abnormality prediction method is applicable, for example, to a disease prediction service in a medical institution that predicts suspected abnormalities in a patient in response to a request from a doctor. Specifically, for example, the abnormality prediction device 201 can treat the target of abnormality prediction as a "patient" and abnormalities #1 to #n as "multiple diseases." The first type of data is, for example, medical record data (text data). The second type of data is test value data (numeric data). In this case, even if a serious disease that should be suspected has only a small feature value, the abnormality prediction device 201 can prevent the feature value of another disease from dominating and causing the serious disease to be overlooked by weighting it using importance, etc.

[0190] Furthermore, this anomaly prediction method can be applied to, for example, a failure prediction service that predicts suspected equipment failures in response to a request from a CE during equipment maintenance and inspection work. Specifically, for example, the anomaly prediction device 201 can treat the target of anomaly prediction as "equipment" and abnormalities #1 to #n as "multiple failures." The first type of data is, for example, daily report data (text data). The second type of data is test value data (numeric data). The daily report data is, for example, a subjective record of the equipment status by a CE. The test value data is, for example, information measured by various sensors installed in the equipment (e.g., drum rotation speed of a multifunction peripheral, remaining toner amount, etc.). In this case, even if a serious failure that should be suspected has only a small feature value, the anomaly prediction device 201 can prevent the serious failure from being overlooked because the feature value of other failures dominates by weighting the data using importance, etc.

[0191] Furthermore, this anomaly prediction method can be applied to, for example, a training method prediction service that predicts an efficient training method in response to a request from a trainer in supporting the development of an athlete. Specifically, for example, the anomaly prediction device 201 can set the target of anomaly prediction as an "athlete" and anomalies #1 to #n as "multiple training methods." The first type of data is, for example, training record data (text data). The second type of data is physical data (numerical data). The training record data is, for example, a trainer's subjective record of the athlete's condition (e.g., "looks in good shape," "has good physical tone," etc.). The physical data may include, for example, the athlete's weight, body fat percentage, weights of training equipment, lap times, etc. In this case, even if an efficient training method that should be recommended has only a small feature value, the anomaly prediction device 201 can prevent the efficient training method from being overlooked due to the dominance of the feature values ​​of other training methods by weighting the recommended efficient training method using importance, etc.

[0192] Furthermore, this anomaly prediction method can be applied to, for example, a sales technique prediction service that predicts effective sales techniques for customers in response to requests from sales representatives in business support for sales activities. Specifically, for example, the anomaly prediction device 201 can set the target of anomaly prediction as a "customer" and anomalies #1 to #n as "plural sales techniques." The first type of data is, for example, a daily sales report (text data). The second type of data is budget and sales data (numerical data). The daily sales report is, for example, a sales representative's subjective record of the content and impressions of sales operations for customers. The budget and sales data is, for example, budgets and sales revenues related to customers. In this case, even if an effective sales technique that should be recommended has only a small feature value, the anomaly prediction device 201 can prevent the effective sales technique from being overlooked due to the dominance of the feature values ​​of other sales techniques by weighting it using importance, etc.

[0193] The anomaly prediction method described in this embodiment can be realized by executing a prepared program on a computer such as a personal computer or a workstation. The anomaly prediction program is recorded on a computer-readable recording medium such as a hard disk, flexible disk, CD-ROM, DVD, or USB memory, and is executed by being read from the recording medium by the computer. The anomaly prediction program may also be distributed via a network such as the Internet.

[0194] In addition, the information processing device 101 (abnormality prediction device 201) described in this embodiment can also be realized by an application-specific IC such as a standard cell or a structured ASIC (Application Specific Integrated Circuit) or a PLD (Programmable Logic Device) such as an FPGA.

[0195] The following additional notes are provided regarding the above-described embodiment.

[0196] (Supplementary Note 1) A first feature value is obtained that represents a likelihood of each anomaly included in a plurality of candidate anomalies, the first feature value being extracted from a first type of first data related to an anomaly prediction target; acquiring second feature amounts that represent a degree of possibility of each of the abnormalities, the second feature amounts being extracted from second data of a second type that is different from the first type and that is related to the abnormality prediction target; referring to a storage unit that stores a level of importance for each of the abnormalities defined in accordance with information content included in the second type of data, correcting the acquired second feature amount based on the level of importance for each of the abnormalities in accordance with information content included in the second data; determining an abnormality for the abnormality prediction target from among the plurality of abnormalities based on the acquired first feature amount and the corrected second feature amount; An abnormality prediction method characterized in that processing is executed by a computer.

[0197] (Appendix 2) The correction process is correcting the acquired second feature amount based on the importance of each anomaly according to information content included in the second data and the acquired first feature amount; 2. The abnormality prediction method according to claim 1,

[0198] (Appendix 3) The correction process is referring to the storage unit, identifying the importance of each of the abnormalities according to the information content included in the second data; correcting the second feature amount by multiplying the identified importance of each anomaly by the first feature amount; The determining process includes: calculating an integrated feature amount for each of the anomalies by adding together the acquired first feature amount and the corrected second feature amount for each of the anomalies; 3. The anomaly prediction method according to claim 2, further comprising determining an anomaly for the anomaly prediction target based on the calculated integrated feature values ​​of each anomaly.

[0199] (Appendix 4) The computer executes a process of learning the importance of each of the abnormalities according to information content included in the second type of data based on the second type of past data at the time of each of the abnormalities; The correction process includes: The abnormality prediction method according to any one of appendices 1 to 3, characterized in that the acquired second feature amount is corrected based on the importance of each of the abnormalities corresponding to the information content included in the second data, with reference to the importance of each of the abnormalities corresponding to the information content included in the learned second type of data.

[0200] (Appendix 5) The first type is one of text, image, and numerical value; 5. The abnormality prediction method according to any one of appendices 1 to 4, wherein the second type is any one of text, image, and numerical value that is different from the first type.

[0201] (Appendix 6) The abnormality prediction target is a patient, 6. The abnormality prediction method according to any one of appendices 1 to 5, wherein the plurality of abnormalities are a plurality of candidate diseases for the patient.

[0202] (Supplementary Note 7) The abnormality prediction target is a device, 6. The abnormality prediction method according to any one of appendices 1 to 5, wherein the plurality of abnormalities are a plurality of candidate failures for the device.

[0203] (Appendix 8) Output the judged result. 8. The abnormality prediction method according to any one of claims 1 to 7, wherein the processing is executed by the computer.

[0204] (Appendix 9) A first feature value is obtained that represents a likelihood of each anomaly included in a plurality of candidate anomalies, the first feature value being extracted from the first type of first data related to the anomaly prediction target; acquiring second feature amounts that represent a degree of possibility of each of the abnormalities, the second feature amounts being extracted from second data of a second type that is different from the first type and that is related to the abnormality prediction target; referring to a storage unit that stores a level of importance for each of the abnormalities defined in accordance with information content included in the second type of data, correcting the acquired second feature amount based on the level of importance for each of the abnormalities in accordance with information content included in the second data; determining an abnormality for the abnormality prediction target from among the plurality of abnormalities based on the acquired first feature amount and the corrected second feature amount; An abnormality prediction program characterized by causing a computer to execute processing. [Explanation of symbols]

[0205] 101 Information processing equipment 102 Abnormality prediction target 110,710 storage section 200 Information Processing Systems 201 Abnormality Prediction Device 202 Client device 210 Network 220 Importance Table 300 Bus 301 CPU 302 memory 303 Disk Drive 304 Disk 305 Communication I / F 306 Portable Recording Media I / F 307 Portable Recording Media 500 medical record data 600,601,602 Test value data 700 control section 701 Acquisition Department 702 Extraction part 703 Correction Unit 704 Prediction Department 705 Learning Department 706 Output section 800,900 Historical Data 801, 802, 803, 804 Test value data set 901,902,903,904 Medical record data group 1000,1601,1701 Text features 1110 Body temperature (numerical value) feature 1120 Blood Pressure (Numerical Value) Feature 1201,1301,1603,1703 Importance 1400,1604,1704 Integrated features 1500 Disease Certainty 1501 Anomaly Prediction Results 1602,1702 Numerical features d1 First data d2 Second data #1~#n,#i Abnormal

Claims

1. acquiring a first feature value that indicates a likelihood of each anomaly included in a plurality of candidate anomalies, the first feature value being extracted from a first type of first data related to an anomaly prediction target; acquiring second feature amounts that represent a degree of possibility of each of the abnormalities, the second feature amounts being extracted from second data of a second type that is different from the first type and that is related to the abnormality prediction target; referring to a storage unit that stores a level of importance for each of the abnormalities defined in accordance with information content included in the second type of data, correcting the acquired second feature amount based on the level of importance for each of the abnormalities in accordance with information content included in the second data; determining an abnormality for the abnormality prediction target from among the plurality of abnormalities based on the acquired first feature amount and the corrected second feature amount; An abnormality prediction method characterized in that processing is executed by a computer.

2. The correction process includes: correcting the acquired second feature amount based on the importance of each anomaly according to information content included in the second data and the acquired first feature amount; The method for predicting an abnormality according to claim 1 .

3. The correction process includes: referring to the storage unit, identifying the importance of each of the abnormalities according to the information content included in the second data; correcting the second feature amount by multiplying the identified importance of each anomaly by the first feature amount; The determining process includes: calculating an integrated feature amount for each of the anomalies by adding together the acquired first feature amount and the corrected second feature amount for each of the anomalies; The abnormality prediction method according to claim 2 , further comprising: determining whether an abnormality exists in the abnormality prediction target based on the calculated integrated feature amount of each abnormality.

4. the computer executes a process of learning a level of importance of each of the abnormalities according to information content included in the second type of data based on the second type of past data at the time of each of the abnormalities; The correction process includes:

2. The abnormality prediction method according to claim 1, further comprising: correcting the acquired second feature amount based on the importance of each of the abnormalities corresponding to the information content included in the second data, with reference to the importance of each of the abnormalities corresponding to the information content included in the learned second type of data.

5. the first type is one of text, image, and numeric; 2. The method of claim 1, wherein the second type is any one of text, image, and numerical value that is different from the first type.

6. the abnormality prediction target is a patient, The abnormality prediction method according to claim 1 , wherein the plurality of abnormalities are a plurality of candidate diseases for the patient.

7. Output the judged result.

7. The abnormality prediction method according to claim 1, wherein the processing is executed by the computer.

8. acquiring a first feature value that indicates a likelihood of each anomaly included in a plurality of candidate anomalies, the first feature value being extracted from a first type of first data related to an anomaly prediction target; acquiring second feature amounts that represent a degree of possibility of each of the abnormalities, the second feature amounts being extracted from second data of a second type that is different from the first type and that is related to the abnormality prediction target; referring to a storage unit that stores a level of importance for each of the abnormalities defined in accordance with information content included in the second type of data, correcting the acquired second feature amount based on the level of importance for each of the abnormalities in accordance with information content included in the second data; determining an abnormality for the abnormality prediction target from among the plurality of abnormalities based on the acquired first feature amount and the corrected second feature amount; An abnormality prediction program characterized by causing a computer to execute processing.

Images