JP7826066B2 - Medical information processing device and medical information learning device - Google Patents

Description

The embodiments disclosed in this specification and drawings relate to a medical information processing device and a medical information learning device.

By systematizing disease classifications and relationships as medical knowledge and building a database, it is expected that this knowledge will be useful for understanding complex diseases, discovering research hypotheses, and supporting medical treatment. For example, in the medical field, medical ontologies such as ICD-10, a disease classification system, and SNOMED-CT, which describes the semantic relationships between medical terms, have been developed. Furthermore, in recent years, data analysis methods using graph structures that can express the relationships between multiple medical events have been attracting attention.

While medical ontologies and graph structures can express relationships between medical events, the same structures cannot necessarily be used to explain different events between patients.

Japanese Patent Application Laid-Open No. 2018-170004 Japanese Patent Application Laid-Open No. 2018-067303

One of the problems that the embodiments disclosed in this specification and drawings attempt to solve is to visualize individual differences between patients. However, the problems that the embodiments disclosed in this specification and drawings attempt to solve are not limited to the above problem. Problems corresponding to the effects of each configuration shown in the embodiments described below can also be positioned as other problems.

A medical information processing device according to an embodiment has an acquisition unit, a storage unit, a calculation unit, and a display control unit. The acquisition unit acquires patient background information relating to background factors of one or more target patients. The storage unit stores a medical knowledge graph having nodes corresponding to medical events and edges representing relationships between the nodes, and graph features of the medical knowledge graph are expressed by a mathematical model characterized by the patient background information. The calculation unit calculates fluctuations in the graph features relating to the target patient based on the patient background information of the target patient and the mathematical model. The display control unit displays the fluctuations in the graph features.

FIG. 1 is a diagram showing an example of the configuration of a medical information system according to this embodiment. FIG. 2 is a conceptual diagram of a medical knowledge graph. FIG. 3 is a diagram showing a specific example of a medical knowledge graph. FIG. 4 is a diagram showing the concept of the relationship between patient background and patient graph. FIG. 5 is a diagram showing an example of the configuration of the medical information learning device shown in FIG. FIG. 6 is a diagram showing input and output of the machine learning model according to this embodiment. FIG. 7 is a diagram showing elements having fluctuations among elements involved in the machine learning model according to this embodiment. FIG. 8 is a diagram showing types of mathematical models that represent fluctuations. FIG. 9 is a diagram showing the flow of the learning process by the medical information learning device shown in FIG. FIG. 10 is a diagram showing an outline of a mathematical model relating to edge feature amounts in step SA2 of FIG. FIG. 11 is a diagram schematically showing the process of determining edge feature amounts using Granger causality estimation in step SA2 of FIG. FIG. 12 is a diagram illustrating an example of the edge feature amount DB. FIG. 13 is a diagram showing an outline of a mathematical model relating to node features and learning parameters in step SA3 of FIG. FIG. 14 is a diagram schematically showing the process of determining the parameters of the mathematical model related to the node features and learning parameters by machine learning in step SA3 of FIG. 9. FIG. 15 is a diagram illustrating an example of a node feature DB. FIG. 16 is a diagram illustrating an example of the weight parameter DB. FIG. 17 is a diagram illustrating an example of the configuration of the medical image processing apparatus shown in FIG. FIG. 18 is a diagram showing the flow of fluctuation display processing by the medical image processing apparatus shown in FIG. FIG. 19 is a diagram showing an example of a fluctuation display screen displayed in step SB4 of FIG. FIG. 20 is a diagram showing a fluctuation display screen when all patients who fall under the category of heart failure are designated as target patients. FIG. 21 is a diagram showing a display screen of fluctuations for each patient background factor. FIG. 22 shows a display screen of the treatment-response graph and fluctuations. FIG. 23 is a diagram illustrating an outline of graph features according to Application Example 1. FIG. 24 is a diagram illustrating an outline of graph features according to the second application example. FIG. 25 is a diagram illustrating an example of a network configuration of a machine learning model according to Application Example 2. FIG. 26 is a diagram schematically showing the relationship between the patient graph and spatial information according to Application Example 4. In FIG. FIG. 27 is a diagram showing the concept of a patient graph network according to Application Example 4. FIG. 28 is a diagram illustrating an outline of graph features according to Application Example 5.

Embodiments of a medical information processing device and a medical information learning device will be described in detail below with reference to the drawings.

Figure 1 is a diagram showing an example configuration of a medical information system 100 according to this embodiment. The medical information system 100 is a computer network system having a medical information storage device 1, a medical knowledge graph storage device 2, a medical information learning device 3, and a medical information processing device 4. The medical information storage device 1, the medical knowledge graph storage device 2, the medical information learning device 3, and the medical information processing device 4 are connected to a network so that they can communicate with each other.

The medical information storage device 1 is a computer with a storage device that stores medical information, etc., of multiple patients. The medical knowledge graph storage device 2 is a computer with a storage device that stores medical knowledge graphs, etc. The medical knowledge graph is expressed as a graph structure generated based on medical information, such as medical ontologies, of multiple patients. Hereinafter, the medical knowledge graph will be assumed to be a graph. The medical information learning device 3 is a computer that trains a machine learning model that uses the medical knowledge graph to estimate medical judgment information. The medical information learning device 3 also trains a mathematical model that defines the graph features of the medical knowledge graph. The medical information processing device 4 is a computer that calculates and presents fluctuations in graph features, etc., between patients.

Figure 2 is a conceptual diagram of a medical knowledge graph 20. The medical knowledge graph 20 is a graph that combines medical knowledge and the medical information of multiple patients. Medical information is information generated through medical treatment for a patient. Medical information is classified into information related to medical events (hereinafter referred to as medical event information), information related to diseases (hereinafter referred to as disease information), and information related to patient background factors (hereinafter referred to as patient background information). Medical events are divided into four categories: symptoms, findings, treatment, and response. Symptoms are physical and mental changes caused by diseases. Specifically, types of symptom-related medical events include swelling, nausea, and shortness of breath. Findings are the doctor's judgment of the symptoms. Findings may be further divided into physical findings and laboratory findings. Treatment is a medical procedure to cure or alleviate symptoms. Types of treatment-related medical events include, for example, the administration of cardiac stimulants and CRT implantation surgery. Response is the patient's physical and mental response to treatment. Responses may be further divided into a treatment response category and a side effect category. Examples of types of response-related medical events include diuresis of 40 mL/h or more or electrolyte abnormalities in response to administration of a diuretic. Disease information is information related to the name or symbol of the disease to which the patient is affected. Patient background information is information that broadly influences most of the patient's medical event information or acts as an interaction term in the relationship between medical event information. Medical event information and patient background information may be assigned based on the presence or absence of time information or the frequency of change. Typically, patient background information is information that does not have time information. Patient background factors that are factors in patient background information that do not have time information include, for example, gender, height, weight, BMI, lifestyle habits, medical history, surgical history, etc.

Medical information is collected by various hospital information systems, such as hospital information systems (HIS), radiology information systems (RIS), and picture archiving and communication systems (PACS). Each piece of medical event information and disease information contained in the medical information is associated with the date on which the information occurred.

As shown in FIG. 2, the medical knowledge graph 20 is a graph constructed from multiple nodes 21 and multiple edges 22. Nodes 21 correspond to medical events. Edges 22 indicate relationships between nodes 21. That is, edges 22 indicate relationships between medical events corresponding to connected nodes 21. Relationships between medical events refer to causal or correlational relationships between medical events. Causal relationships refer to a cause-and-effect relationship between medical events, while correlational relationships refer to relationships without causal relationships. That is, edges 22 representing causal relationships are directed, while edges 22 representing correlational relationships are undirected. Edges 22 can also connect medical events belonging to different categories. Medical events used in the medical knowledge graph 20 are selected from the medical information stored in the medical information storage device 1 based on medical knowledge, medical ontology, etc. The relationship between a certain medical event and another medical event is analyzed based on medical knowledge, medical ontology, etc., and if a relationship is confirmed, the two nodes 21 corresponding to the two medical events are connected by an edge 22.

The medical knowledge graph 20 does not have a node corresponding to the medical event of disease. A specific disease is extracted from a collection of a series of medical events, such as symptoms, findings, treatment, and response, that have occurred in a target patient. In other words, the medical knowledge graph 20 represents each disease as a chain of relationships between a series of medical events, such as symptoms, findings, treatment, and response. The medical knowledge graph 20 can also be said to be a graph representing the concept of disease, describing the chain of relationships between a series of medical events, such as symptoms, findings, treatment, and response.

Figure 3 shows a specific example of a medical knowledge graph. The medical knowledge graph shown in Figure 3 is a graph in which medical event information for a target patient is mapped onto a medical knowledge graph. A medical knowledge graph onto which medical event information for a target patient is mapped is also called a patient graph. As an example, the medical knowledge graph shown in Figure 3 shows a symptom graph limited to medical events belonging to a symptom category. As shown in Figure 3, the symptom graph is a graph including nodes representing symptoms and edges indicating the relationships between symptoms. The color intensity of the node represents the degree to which the target patient is affected by the medical event. In this way, a medical knowledge graph describes the chain of relationships between a series of medical events that occurred in the target patient, and from this chain of relationships between a series of medical events, it is possible to obtain medical judgments about diseases, etc. to which the target patient may be suffering.

Even for the same disease, the presence or strength of the relationship between medical events varies depending on the background of each individual patient. Inferring a patient's disease without taking into account the patient's background will not necessarily yield appropriate results.

Figure 4 is a diagram illustrating the concept of the relationship between patient background and patient graph. As shown in Figure 4, patient background, unlike nodes and edges, has an overall impact on the patient graph. Even when the same patient graph is constructed for patient A and patient B, the patient background, such as gender, age, BMI, and medical history, may differ. Thus, even if patient graphs are the same for patient A and patient B, if the patient backgrounds are different, there must be room for different medical judgments to be derived from the two patient graphs. This corresponds to the patient graph being modified and mutated by some of the nodes and/or edges of the patient graph being added to the context of the patient background. The medical information system 100 according to this embodiment presents the degree of modification between different patient backgrounds as fluctuation.

Figure 5 is a diagram showing an example configuration of a medical information learning device 3 according to this embodiment. As shown in Figure 5, the medical information learning device 3 has a processing circuitry 31, a memory 32, an input interface 33, a communication interface 34, and a display 35.

The processing circuitry 31 has processors such as a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit). By executing a learning program, the processing circuitry 31 realizes functions such as an acquisition function 311, a learning function 312, and a display control function 313. Note that each of the functions 311 to 313 does not necessarily have to be realized by a single processing circuit. A processing circuit may be configured by combining multiple independent processors, and each processor may realize each of the functions 311 to 313 by executing a program. Furthermore, each of the functions 311 to 313 may be a modular program that constitutes a learning program, or may be an individual program. These programs are stored in memory 32.

By implementing the acquisition function 311, the processing circuitry 31 acquires various information. For example, the processing circuitry 31 acquires patient background information related to the background factors of multiple patients from the medical information storage device 1. The processing circuitry 31 also acquires a medical knowledge graph from the medical knowledge graph storage device 2. As described above, the medical knowledge graph has nodes corresponding to medical events and edges representing the relationships between the nodes. The graph features of the medical knowledge graph are characterized by a mathematical model that includes patient background information.

By implementing the learning function 312, the processing circuitry 31 determines parameters of the mathematical model based on the patient background information of multiple patients and the mathematical model. The processing circuitry 31 also trains learning parameters of a machine learning model that estimates medical judgment information from the medical knowledge graph. The learning parameters are characterized by the mathematical model, which includes the patient background information. The processing circuitry 31 determines the parameters of the mathematical model based on the patient background information of multiple patients and the mathematical model, along with the learning parameters. The processing circuitry 31 determines the parameters of the mathematical model through statistical analysis or machine learning based on the patient background information of multiple patients.

By implementing the display control function 313, the processing circuit 31 displays various information on the display 35.

Memory 32 is a storage device such as a ROM (Read Only Memory), RAM (Random Access Memory), HDD (Hard Disk Drive), SSD (Solid State Drive), or integrated circuit storage device that stores various information. In addition to the above storage devices, memory 32 may also be a portable storage medium such as a CD (Compact Disc), DVD (Digital Versatile Disc), or flash memory, or a drive that reads and writes various information from and to semiconductor memory elements. Memory 32 may also be located in another computer connected to the medical information learning device 3 via a network. For example, memory 32 stores a medical knowledge graph obtained from the medical knowledge graph storage device 2. Memory 32 also stores a machine learning model.

The input interface 33 accepts various input operations from the operator, converts the accepted input operations into electrical signals, and outputs them to the processing circuit 31. Specifically, the input interface 33 is connected to input devices such as a mouse, keyboard, trackball, switch, button, joystick, touchpad, and touch panel display. The input interface 33 outputs electrical signals corresponding to the input operations to the input devices to the processing circuit 31. In addition, the input devices connected to the input interface 33 may be input devices provided on another computer connected via a network, etc.

The communication interface 34 is an interface for sending and receiving various information between other computers included in the medical information system 100, such as the medical information storage device 1, the medical knowledge graph storage device 2, and the medical information processing device 4.

The display 35 displays various information in accordance with the display control function 313 of the processing circuit 31. The display 35 may be, for example, a liquid crystal display (LCD), a cathode ray tube (CRT) display, an organic electroluminescence display (OELD), a plasma display, or any other display. A projector may also be provided in place of or in combination with the display 35.

Next, we will explain the learning process executed by the processing circuit 31 in accordance with the learning program. First, we will explain the machine learning model used in the learning process.

The machine learning model according to this embodiment is a neural network that estimates medical judgment information for a target patient from a patient graph onto which the medical event information for the target patient is mapped. The medical judgment information is information relating to at least one medical judgment among disease classification information, prognosis prediction information, and severity classification information. In the following examples, the medical judgment information is assumed to be disease classification information.

Figure 6 is a diagram showing the input and output of the machine learning model according to this embodiment. As shown in Figure 6, the machine learning model is a neural network whose learning parameters have been trained to input a patient graph onto which medical event information of target patients is mapped and output disease classification information. The machine learning model has a graph convolution layer 61, a readout layer 62, and a linear combination layer 63.

A patient graph is generated by mapping the medical event information of a target patient onto a medical knowledge graph. Medical event information is information that contains a combination of the type of medical event and the feature values of the medical event. The feature values of the medical event include, for example, at least one of the following: the degree to which the target patient corresponds to the medical event, the order in which the medical event occurred, the number of times the medical event occurred, and the degree to which the medical event occurred. Note that the degree of correspondence is typically represented by two values, corresponding or not corresponding, but may also be defined by three or more discrete or continuous values indicating the degree of correspondence.

As described above, a medical knowledge graph is a graph constructed from multiple nodes and multiple edges. A medical knowledge graph has graph features. Graph features include node types, node values, edge types, and edge values. Node types are the types or categories of medical events. Node values are features of the medical events mapped to the nodes. That is, node values indicate the degree to which the target patient corresponds to the medical event, the order in which the medical event occurs, the number of times the medical event occurs, and the degree to which the medical event occurs. Node types and values are collectively referred to as node features. Edge types are the combination of the causal direction (in the case of a directed graph) and the types of the two connected nodes. Edge values are the degree of causality/correlation. The degree of causality/correlation is typically represented by two values: present (directed) or absent (undirected), but may also be defined by three or more discrete or continuous values indicating the degree. Edge types and values are collectively referred to as edge features. The patient graph is a graph in which medical event information is mapped onto the medical knowledge graph, and therefore has graph features.

Note that in this embodiment, there is no need to clearly distinguish between "type" and "value." For example, node or edge values may be classified into types or categories by comparing them with a threshold value.

The graph convolutional layer 61 is a Graph Convolutional Network (GCN) that inputs a pre-convolution patient graph and outputs a post-convolution patient graph. The graph convolutional layer 61 performs convolution processing on the pre-convolution patient graph and calculates post-convolution graph features from the pre-convolution graph features. The post-convolution graph features of the processing target node are calculated based on the pre-convolution graph features of the processing target node, the pre-convolution graph features of nodes connected to the processing target node via edges (hereinafter referred to as adjacent nodes), and weight parameters of the graph convolutional layer 61. The weight parameters of the graph convolutional layer 61 are an example of learning parameters trained by the medical information learning device 3.

The readout layer 62 is a network layer that converts the graph features of the convoluted patient graph 20B into a feature vector 20C. The feature vector 20C is a vector with the same number of dimensions as the number of nodes included in the convoluted patient graph 20B.

The linear combination layer 63 is a network layer that converts the feature vector 20C into disease classification information 71. The linear combination layer 63 is also called a multilayer perceptron (MLP). The linear combination layer 63 performs a classification task or a regression task. The classification task may be a two-class classification that determines the presence or absence of a relevant disease, or a multi-class classification that identifies one disease from multiple disease candidates. It may also be a multi-label classification that allows multiple labels for the same dataset. The regression task outputs a numerical value indicating the probability of each of the relevant one or more diseases.

In this embodiment, the linear combination layer 63 has a classifier that performs multi-class classification, classifying the feature vector 20C into multiple classes corresponding to multiple diseases. In multi-class classification, a softmax function is calculated to output the probability that the feature vector 20C belongs to each disease (class). This probability of belonging to each disease is output as disease classification information 71. For example, as shown in FIG. 6, the probability of belonging to the disease "heart failure," the probability of belonging to the disease "renal failure," the probability of belonging to the disease "COPD," etc. are output as disease classification information 71. The probability of belonging to each disease in the disease classification information 71 is calculated based on the feature vector 20C after convolution and weight parameters of the linear combination layer 63. The weight parameters of the linear combination layer 63 are an example of learning parameters trained by the medical information learning device 3.

Fluctuations due to differences in patient backgrounds between patients also appear in the process of estimating disease classification information using the machine learning model shown in Figure 6.

Figure 7 is a diagram showing elements that have fluctuations among the elements involved in the machine learning model according to this embodiment. As shown in Figure 7, elements that have fluctuations are graph features including node type, node value, edge type, and edge value. For example, node features such as node type and value are values based on medical events, but the occurrence and severity of medical events are actually influenced by patient background. As a result, edge features such as edge type and value are also influenced by patient background. Other elements that have fluctuations include the weight parameters of the graph convolution layer 61 and the weight parameters of the linear combination layer 63.

Figure 8 shows the types of mathematical models that represent fluctuations. As shown in Figure 8, mathematical models that represent fluctuations include fixed-effect models, stratified models, mixed models, and functional models. Fixed-effect models do not distinguish between individual patients and express patient values as Patient Value = True Value + Error. Fixed-effect models cannot explain individual and intergroup differences. Stratified models are mathematical models that stratify patients by background information related to one or more background factors. Stratified models distinguish between individual patients and express Patient A's value as Patient A's Value = Patient A's True Value + Error, and express values for male patients (hereinafter simply referred to as "men") as Male Value = Male True Value + Error. Note that "value" refers to values that fluctuate, such as graph features and weight parameter values. Stratified models can somewhat explain individual and intergroup differences. Mixed models are mathematical models based on random effects due to patient background information related to one or more background factors. Mixed models are also called hierarchical models. In the mixed model, patient A's value is expressed as patient A's value = true value + patient A's random effect + error, and male's value is expressed as male's value = true value + male's random effect + error. The mixed model can well explain individual differences and inter-group differences. The functional model is a mathematical model based on a function whose value is uniquely determined by patient background information regarding one or more background factors. The functional model expresses patient A's value as patient A's value = true value + patient A's bias + error. The functional model can fully explain individual differences.

Next, we will explain the learning process executed by the processing circuit 31 in accordance with the learning program.

Figure 9 shows the flow of the learning process. As shown in Figure 9, the processing circuitry 31 acquires patient background information and medical event information for each of multiple patients by implementing the acquisition function 311 (step SA1).

When step SA1 is performed, the processing circuitry 31, by implementing the learning function 312, determines the parameters of a mathematical model related to edge features of the medical knowledge graph through statistical analysis based on patient background information, medical event information, and the mathematical model (step SA2). The mathematical model in this embodiment is expressed as either a stratified model, a mixed model (hierarchical model), or a functional model.

Fig. 10 is a diagram showing an outline of a mathematical model related to edge features. As shown in Fig. 10, the edge feature e _i is generalized by a mathematical model f using patient background information Z as an explanatory variable, as shown in the following formula (1). The patient background information Z has patient background factors Z ₁ , Z ₂ , .... Individual patient background factors include age, sex, BMI, etc.

e _i =f(Z ₁ ,Z ₂ ,…) (1)

Here, the mathematical model f is represented by a functional model. In this case, the mathematical model f consists of a combination of fixed effect terms that are independent of patient background information and random effect terms that are characterized by patient background factors.

Statistical analysis methods used to determine the parameters of a mathematical model related to edge features include analysis of variance, Granger causality estimation, or causal effect estimation. As an example, we will explain how to determine model parameters using Granger causality estimation.

Figure 11 is a diagram that schematically illustrates the process of determining edge features using Granger causality estimation. As shown in Figure 11, the causality from medical event x to medical event y can be represented by an edge from medical event x to medical event y in a directed graph. Here, we consider the determination of the edge feature of the edge from medical event x to medical event y, for example, the presence or absence of an edge. We assume that the causality from medical event x to medical event y is influenced by patient background factor z. Note that medical events x and y can be replaced with specific types of medical events, such as "swelling" and "shortness of breath."

The occurrence probability of the current medical event y is determined based on the occurrence probability of the past medical event x and the occurrence probability of the past medical event y. More specifically, as shown in FIG. 11 and the following equation (2), the occurrence probability λ _y (t) of the current medical event y is determined based on the sum of the error term μ, the second term representing the occurrence probability of the past medical event x, and the third term representing the occurrence probability of the past medical event y. t _i is a medical event, and i represents the index of the medical event. The second term, Ex, represents the set (sample space) of medical events x. _{t i} ∈ Ex means the occurrence time of each medical event x. The third term, Ey, represents the set (sample space) of medical events y. _{t i} ∈ Ey means the occurrence time of each medical event y. φ(t - t _i ) represents the occurrence probability of a medical event in the past t - t _i . That is, the second term represents the occurrence probability of the past medical event x, and the third term represents the occurrence probability of the past medical event y.

In the case of a functional model, the coefficient _αx of the second term is a parameter with patient background factors as variables, as expressed by the following equation (3). More specifically, the coefficient _αx is expressed by a function σ of the sum ( _WxZ + _bx ) of the product of each patient background factor and weight _Wx and the bias _bx . The fixed effect term that does not depend on patient background information is _bx , and the random effect term characterized by the patient background factors is _Wx . When the coefficient _αx is equal to zero, this indicates that there is no causality from medical event x to medical event y. The weight _Wx and bias _bx are examples of model parameters. The coefficient _αy of the third term, like the coefficient _αx , may be expressed taking into account patient background factors, or may be expressed without taking into account patient background factors. The coefficient _αx is an example of an edge feature of the edge from medical event x to medical event y. The coefficient _αy is an example of an edge feature of the edge from medical event y to medical event x.

α _x = σ (W _x Z + b _x ) (3)

The processing circuit 31 performs a Granger causality test on the edge connecting the medical event x and the medical event y based on the medical event information of multiple patients, and determines the weight _Wx and bias _bx , which are model parameters of a mathematical model representing the edge feature of the edge. As an example, Granger causality is tested using an F-test. For example, first, the occurrence probability _λy (t) of the current medical event y is estimated when the second term is included in equation (2) and when it is not. Next, the statistical difference in the estimation error of the occurrence probability _λy (t) between the two is determined based on the F-test. The weight _Wx and bias _bx included in the coefficient _αx are determined according to the statistical difference. If there is no statistical difference, the coefficient _αx is set to zero. Combinations of the medical event x and the medical event y are sequentially rearranged from the multiple medical events included in the medical event information of multiple patients, and Granger causality is determined for all combinations. As a result, model parameters of the mathematical model related to the edge feature are determined for all edges included in the medical knowledge graph. If it is determined that Granger causality does not exist between the medical events x and y, there is no edge between the medical events x and y, and therefore the edge feature is zero, and the model parameters, weight W _x and bias b _x, are also determined to be zero.

The processing circuitry 31 associates edge features with the model parameters of a mathematical model related to the edge features for each edge number. For example, the processing circuitry 31 creates a database such as a look-up table (LUT) that associates edge features with the model parameters of a mathematical model related to the edge features for each edge. Hereinafter, this database will be referred to as the edge feature DB.

Fig. 12 is a diagram showing an example of an edge feature DB. As shown in Fig. 12, the edge feature DB stores edge numbers, edge features, and model parameters in a searchable manner. For example, edge number "1" is associated with edge feature "e1" and model parameters "W _x 1" and "b _x 1". In the stratification model, the model parameters are determined for each patient background factor. The edge feature DB is stored in the medical knowledge graph storage device 2.

Note that equation (3) is designed assuming that the linearly coupled layer 63 is a single-layer linear neural network, but it may also be designed assuming a multi-layer linear neural network, or other neural networks such as a convolutional neural network (CNN) or a recurrent neural network (RNN).

In the above examples, the mathematical model is a functional model. However, the mathematical model according to this embodiment may be a mixed model. Like the functional model, the mixed model is composed of a combination of a fixed effect term that is independent of patient background information and a random effect term that is characterized by patient background factors. In the case of a mixed model, the coefficient α _x may be estimated using a generalized linear mixed model or a hierarchical Bayesian model that takes into account patient background information. The mathematical model may also be a stratified model. In the case of a stratified model, a Granger causality test may be performed for each patient group (e.g., a male group and a female group) stratified by patient background information, and a separate coefficient α _x may be calculated for each patient group. In this case, the coefficient α _x may be a general function rather than a function including a weight W _x and a bias b _x .

When step SA2 is performed, the processing circuitry 31, by implementing the learning function 312, determines the parameters of the mathematical model related to the node features and learning parameters through machine learning based on the patient background information, medical event information, and mathematical model (step SA3).

13 is a diagram showing an overview of a mathematical model related to node features and learning parameters. As shown in Fig. 13, the node feature v _i of a node included in the medical knowledge graph 20B is expressed by a mathematical model g with patient background information Z as a variable, as shown in the following equation (4). The patient background information Z has patient background factors Z ₁ , Z ₂ , ...

v _i =g(Z ₁ ,Z ₂ ,…) (4)

Hereinafter, mathematical model g will be represented as a functional model. In this case, mathematical model g consists of a combination of fixed effect terms that are independent of patient background information and random effect terms that are characterized by patient background factors.

13, the learning parameters _wi of the linear combination layer 63 are expressed by a mathematical model h with patient background information Z as a variable, as shown in the following equation (5). The patient background information Z has patient background factors Z ₁ , Z ₂ , . . .

w _i =h(Z ₁ ,Z ₂ ,…) (5)

Hereinafter, the mathematical model h will be represented as a functional model. In this case, the mathematical model h consists of a combination of a fixed effect term β that is independent of patient background information and a random effect term h' that is characterized by patient background factors, as shown in equation (6) below. As shown in equation (7) below, the expected value of the random effect term h' is zero.

h(Z ₁ ,Z ₂ ,…)=β+h′(Z ₁ ,Z ₂ ,…) (6)
E(h')=0 (7)

Machine learning is used to determine the parameters of the mathematical model related to node features and learning parameters. Specifically, when training the learning parameters of the machine learning model shown in Figure 6, the parameters of the mathematical model related to node features and learning parameters are also determined. This machine learning is described below.

Figure 14 is a diagram that schematically illustrates the process of determining the parameters of a mathematical model related to node features and learning parameters through machine learning. As shown in Figure 14, the learning parameters of a machine learning model 60 are trained so that the patient graph 20A is input and the disease label y is output. The machine learning model 60 has a graph convolution layer 61, a readout layer 62, and a linear combination layer 63. The graph convolution layer 61 inputs the patient graph 20A before convolution and outputs the patient graph 20B after convolution.

The node feature of each node included in the patient graph 20B is affected by the patient background information. For example, the node feature v _i of node number "i" is affected by the patient background information Z. The node feature v _i is expressed by the mathematical model shown in the following equation (8). As shown in equation (8), the node feature v _i is expressed as the sum of the product of the coefficient (α _v Z + β _v ) and the initial node feature x _i and the bias b _v . The fixed effect term that is independent of the patient background information Z is β _v , and the random effect term characterized by the patient background information Z is α _v . Note that the weight α _v , bias β _v , and bias b _v are determined for each node.

v _i = (α _v Z + β _v ) x _i +b _v (8)

The readout layer 62 converts the graph features of the convoluted patient graph 20B into a feature vector 20C. There is a one-to-one correspondence between the elements of the feature vector 20C and the nodes included in the patient graph 20B. For example, as shown in FIG. 14 , the node feature _v1 of the first node included in the patient graph 20B is assigned to the first element of the feature vector 20C.

The linear combination layer 63 converts the feature vector 20C into a disease label y. For example, the linear combination layer 63 performs a softmax operation based on the node features of multiple elements included in the feature vector 20C to calculate disease information representing the probability of each of multiple disease classes. FIG. 14 illustrates the softmax operation for a disease class y. In this operation, each node feature v is multiplied by a weight parameter w determined according to the combination of each element and the disease class y. The weight parameter w is affected by patient background information. For example, the weight parameter w _i for element number (node number) "i" is affected by the patient background information. The weight parameter w _i is expressed by the mathematical model shown in Equation (9) below. As shown in Equation (9), the weight parameter w _i is expressed as the sum of the product of the coefficient (α _w Z + β _w ) and the node feature v _i and the bias b _w . The fixed effect term independent of the patient background information Z is β _w , and the random effect term characterized by the patient background information Z is α _w . The weight α _w , bias β _w and bias b _w are determined for each node.

w _i = (α _w Z + β _w ) v _i +b _w (9)

The processing circuitry 31 maps medical event information onto the medical knowledge graph to generate a patient graph. The processing circuitry 31 trains the learning parameters of the machine learning model based on supervised learning, using the patient graph as an input sample and disease information as a teaching sample. The processing circuitry 31 calculates a loss function. The loss function is a function that evaluates the error between the disease classification information calculated by sequentially propagating the patient graph through the graph convolution layer, readout layer, and linear combination layer, and the disease information that is the teaching sample. The processing circuitry 31 updates the learning parameters of the machine learning model so as to minimize the loss function according to an arbitrary optimization method. Any optimization method, such as stochastic gradient descent or Adam (adaptive moment estimation), may be used.

The processing circuit 31 determines whether a stopping condition is satisfied. The stopping condition may be set, for example, when the number of updates of the learning parameters reaches a predetermined number, or when the update amount of the learning parameters is less than a threshold value. The processing circuit 31 repeatedly updates the learning parameters and calculates the loss function while changing the combination of input samples and teaching samples until the stopping condition is satisfied. Then, when the stopping condition is satisfied, the processing circuit 31 sets the learning parameters for the current number of updates as the final learning parameters. This ends the learning.

The learning parameters to be trained include the weight parameters of the graph convolution layer 61 as well as the node feature v _i and the weight parameter w _i . As shown in equation (9), the weight parameter w _i is expressed as the sum of the product of the coefficient (α _w Z + β _w ) and the node feature v _i and the bias b _w , and therefore, it is possible to determine the coefficient (α _w Z + β _w ) and the bias b _w from the weight parameter w _i and the node feature v _i . Also, it is possible to determine the parameters α _w and β _w from the coefficient (α _w Z + β _w ) and the patient background information Z. As shown in equation (8), the node feature v _i is expressed as the sum of the product of the coefficient (α _v Z + β _v ) and the initial node feature x _i and the bias b _v , and therefore, it is possible to determine the coefficient (α _v Z + β _v ) and the bias b _v from the initial node feature x _i and the node feature v _i . Furthermore, the parameters α _v and β _v can be determined from the coefficient (α _v Z+β _v ) and the patient background information Z.

The processing circuitry 31 associates, for each node, the node features with the model parameters of a mathematical model related to the node features. For example, the processing circuitry 31 creates a database such as an LUT that associates, for each node, the node features with the model parameters of a mathematical model related to the node features. Hereinafter, this database will be referred to as the node feature DB.

Fig. 15 is a diagram showing an example of a node feature DB. As shown in Fig. 15, the node feature DB records node numbers, node features, and model parameters in a searchable manner. For example, node number "1" is associated with node feature "v1" and model parameters "α _v 1,""β _v 1," and "b _v 1" related to the mathematical model of the node feature. The node feature DB is stored in the medical knowledge graph storage device 2.

Similarly, the processing circuitry 31 associates each weight parameter with the model parameters of the mathematical model related to that weight parameter. For example, the processing circuitry 31 creates a database such as an LUT that associates each weight parameter with the model parameters of the mathematical model related to that weight parameter for each weight parameter number. Hereinafter, this database will be referred to as the weight parameter DB.

Fig. 16 is a diagram showing an example of a weight parameter DB. As shown in Fig. 16, the weight parameter DB records weight parameter numbers, weight parameters, and model parameters in a searchable manner. For example, weight parameter number "1" is associated with weight parameter "w1" and model parameters related to the weight parameter, "α _w1 ,""β _w1 ," and "b _w1 ." The weight parameter DB is stored in the medical knowledge graph storage device 2.

Note that equations (8) and (9) are designed assuming that the linearly coupled layer 63 is a single-layer linear neural network, but they may also be designed assuming a multi-layer linear neural network, or other neural networks such as a convolutional neural network (CNN) or a recurrent neural network (RNN).

In the above examples, the mathematical model is a functional model. However, the mathematical model according to this embodiment may be a mixed model. Like the functional model, the mixed model is composed of a combination of a fixed effect term that is independent of patient background information and a random effect term that is characterized by the patient background information. In the case of a mixed model, the weights α _v and α _w may be estimated using a generalized linear mixed model or a hierarchical Bayesian model that takes into account the patient background information. The mathematical model may also be a stratified model. In the case of a stratified model, a Granger causality test may be performed for each patient group (e.g., a male group and a female group) stratified by patient background information, and separate weights α _v and α _w may be calculated for each patient group. In this case, the weights α _v and α _w may be general weights rather than functions including the weights W _x and bias b _x as in Equation (3).

When step SA3 is performed, the processing circuitry 31 outputs a medical knowledge graph and a learned model by implementing the learning function 312 (step SA4). In step SA4, the processing circuitry 31 outputs a medical knowledge graph to which model parameters of a mathematical model related to edge features and model parameters of a mathematical model related to graph features are assigned. Note that if the edge features are fixed regardless of the patient, a medical knowledge graph to which edge features are further assigned may be output. The processing circuitry 31 also outputs a machine learning model to which weight parameters and model parameters of a mathematical model related to the weight parameters are assigned. The output medical knowledge graph and machine learning model are stored in the medical knowledge graph storage device 2.

When step SA4 is performed, the learning process ends.

Note that the above learning process is an example, and various modifications are possible. For example, steps SA2 and SA3 may be performed in reverse order, or in parallel. Also, in step SA3, the parameters of the mathematical model related to node features and the parameters of the mathematical model related to learning parameters are determined based on supervised learning, but they may also be determined based on a hierarchical Bayesian model.

As described above, the medical information learning device 3 has a processing circuitry 31 and a memory 32. The processing circuitry 31 acquires patient background information related to the background factors of multiple patients. The memory 32 stores a medical knowledge graph having nodes corresponding to medical events and edges representing the relationships between the nodes. The graph features of the medical knowledge graph are characterized by a mathematical model that includes the patient background information. The processing circuitry 31 determines the model parameters of the mathematical model based on the patient background information of multiple patients and the mathematical model.

The above configuration makes it possible to define the graph features of a medical knowledge graph using a mathematical model that uses patient background information. This makes it possible to incorporate fluctuation components resulting from differences in patient background between patients into the graph features of a medical knowledge graph.

Next, we will explain the medical information processing device 4 related to this embodiment.

Figure 17 is a diagram showing an example configuration of the medical information processing device 4. As shown in Figure 17, the medical information processing device 4 has a processing circuitry 41, a memory 42, an input interface 43, a communication interface 44, and a display 45.

The processing circuitry 41 has processors such as a CPU and a GPU. By executing a medical information processing program to display fluctuations, the processing circuitry 41 realizes an acquisition function 411, an estimation function 412, a fluctuation calculation function 413, a visualized graph generation function 414, a selection function 415, and a display control function 416. Note that each of the functions 411 to 416 does not necessarily have to be realized by a single processing circuit. A processing circuit may be configured by combining multiple independent processors, and each processor may execute a program to realize each of the functions 411 to 416. Furthermore, each of the functions 411 to 416 may be a modular program that constitutes the medical information processing program, or may be an individual program. These programs are stored in memory 42.

By implementing the acquisition function 411, the processing circuitry 41 acquires various information. For example, the processing circuitry 41 acquires patient background information related to one or more patients to be displayed (hereinafter referred to as target patients) from the medical information storage device 1. The processing circuitry 41 also acquires a medical knowledge graph, an edge feature DB, a node feature DB, and a weight parameter DB.

By implementing the estimation function 412, the processing circuitry 41 uses the trained model generated by the medical information learning device 3 to estimate medical judgment information from the patient graph.

By implementing the fluctuation calculation function 413, the processing circuitry 41 calculates the fluctuation of graph features and/or learning parameters for a target patient based on the patient background information and a mathematical model. The graph features and learning parameters are represented by a mathematical model with model parameters as variables. The model parameters are determined by the medical information learning device 3 based on the patient background information.

By implementing the visualization graph generation function 414, the processing circuitry 41 generates a visualization graph that visualizes the medical knowledge graph related to the patient to be displayed. The graph and/or edges of the visualization graph have a display form that corresponds to the fluctuation of the graph features.

By implementing the selection function 415, the processing circuitry 41 selects various display objects, such as patients, medical event categories, and diseases, when displaying fluctuations. The processing circuitry 41 selects the display objects in accordance with user instructions via the input interface 43.

By implementing the display control function 416, the processing circuitry 41 displays various information on the display 45. For example, the processing circuitry 41 displays graph features and/or fluctuations in learning parameters.

The memory 42 is a storage device such as ROM, RAM, HDD, SSD, or integrated circuit storage device that stores various information. In addition to the above storage devices, the memory 42 may also be a portable storage medium such as a CD, DVD, or flash memory, or a drive that reads and writes various information to and from semiconductor memory elements. The memory 42 may also be located in another computer connected to the medical information processing device 4 via a network. For example, the memory 42 stores the medical knowledge graph, edge feature DB, node feature DB, and weight parameter DB acquired from the medical knowledge graph storage device 2.

The input interface 43 accepts various input operations from the operator, converts the accepted input operations into electrical signals, and outputs them to the processing circuit 41. Specifically, the input interface 43 is connected to input devices such as a mouse, keyboard, trackball, switch, button, joystick, touchpad, and touch panel display. The input interface 43 outputs electrical signals corresponding to the input operations to the input devices to the processing circuit 41. In addition, the input devices connected to the input interface 43 may be input devices provided on another computer connected via a network, etc.

The communication interface 44 is an interface for sending and receiving various information to and from other computers included in the medical information system 100, such as the clinical information storage device 1, the medical knowledge graph storage device 2, and the medical information learning device 3.

The display 45 displays various information in accordance with the display control function 414 of the processing circuit 41. The display 45 may be, for example, a liquid crystal display, a CRT display, an organic EL display, a plasma display, or any other display. A projector may also be provided in place of or in combination with the display 45.

Next, we will explain the fluctuation display processing executed by the processing circuitry 41 in accordance with the medical information display program.

Figure 18 is a diagram showing the flow of the fluctuation display process. As shown in Figure 18, the processing circuitry 41 selects one or more patients (hereinafter referred to as target patients) to be displayed by implementing the selection function 415 (step SB1). For example, in step SB1, a target patient selection screen is displayed on the display 45, and one or more target patients are selected in accordance with instructions from the user via the input interface 43. In this embodiment, it is assumed that one patient is selected as the display target. Note that the medical knowledge graph, edge feature DB, node feature DB, and weight parameter DB may be acquired from the medical knowledge graph storage device 2. The processing circuitry 41 appropriately reads edge features, node features, weight parameters, and model parameters from the edge feature DB, node feature DB, and weight parameter DB as needed.

As described below, patients with a specific disease may be selected as target patients. For example, a list of patients with heart failure may be displayed, and a specific patient may be selected from the list as the target patient.

When step SB1 is performed, the processing circuitry 41 acquires the patient background information of the target patient selected in step SB1 by implementing the acquisition function 411 (step SB2). In step SB2, the processing circuitry 41 acquires the patient background information of the target patient from the medical information storage device 1.

When step SB2 is performed, the processing circuitry 41 realizes the fluctuation calculation function 413 to calculate the fluctuation of the graph features and/or learning parameters based on the patient background information acquired in step SB2 (step SB3).

The fluctuation of a graph feature refers to the degree of correlation of the graph feature with patient background information. If the graph feature takes a constant value regardless of differences in patient background information, the fluctuation of the graph feature is small. Conversely, if the graph feature fluctuates greatly depending on differences in patient background information, the fluctuation of the graph feature is large. When there is one target patient, the fluctuation of the graph feature for that target patient can be defined based on graph features at multiple points in time with different patient background information. When there are multiple target patients, the fluctuation of the graph feature for those multiple people can be defined based on graph features at multiple points in time with different patient background information. The same applies to the fluctuation of learning parameters.

The method for calculating fluctuations will be described in detail below. Typically, the processing circuitry 41 calculates graph features for each of multiple target patients based on the patient background information and a mathematical model of graph features, and calculates an index of variation in graph features across these multiple target patients as the fluctuation of the graph features. The mathematical model of the graph features uses a mathematical model whose model parameters are determined by the medical information learning device 3. Standard deviation, variance, etc. can be calculated as an index of variation. Patients belonging to a statistical population are selected as target patients. The statistical population can be set arbitrarily depending on the survey subjects, such as all patients, patients suffering from a specific disease, male patients, female patients, or patients in a specific age group. Fluctuations can also be calculated for weight parameters in a similar manner. The mathematical model of the weight parameters uses a mathematical model whose model parameters are determined by the medical information learning device 3.

Since it is expected that the range of graph feature values may differ for each node or edge, the processing circuitry 41 may normalize the fluctuations. For example, when displaying fluctuations for only heart failure patients, normalization is possible by calculating the variance in graph features for heart failure patients among all target patients and dividing the "variation in graph features for heart failure patients" by the "variation in graph features for all patients." Fluctuations can also be normalized for weight parameters in a similar manner.

Regarding the node feature, the fluctuation of the node feature v _i defined by the mathematical model of equation (8) (v _i = (α _v Z + β _v ) x _i + b _v ) may be calculated, or the fluctuation of only the coefficient (α _v Z + β _v ) may be calculated. In the former case, the fluctuation of the node feature itself can be calculated, but if the node feature values x _i and v _i vary between patients, there is a risk that fluctuations not caused by patient background factors will also be taken into account, which may make interpretation more difficult than in the latter case. In the latter case, it is possible to obtain the fluctuation of the mapping to the node feature caused by patient background factors. The user can freely select whether to calculate the former fluctuation or the latter fluctuation. Similarly, for the weight parameter, the fluctuation of the weight parameter w defined by the mathematical model of equation (9) (w _i = (α _w Z + β _w ) v _i + b _w ) may be calculated, or the fluctuation of only the coefficient (α _w Z + β _w ) may be calculated. For the edge feature, the fluctuation of the edge feature α _x defined by the mathematical model of equation (3) (α _x = σ (W _x Z + b _x )) is calculated.

When there is one target patient, the processing circuitry 41 calculates graph features for the target patient for each of multiple time points based on the patient background information and a mathematical model of the graph features, and calculates an index of the variability of the graph features across these multiple time points as the fluctuation of the graph features. Standard deviation, variance, etc. may be calculated as an index of variability. Fluctuations can also be calculated for weight parameters in a similar manner. Furthermore, when there is one target patient, the processing circuitry 41 may calculate, as the fluctuation, the deviation value of the graph features or weight parameters for the target patient relative to the variability of the graph features or weight parameters calculated for the statistical population.

When step SB3 is performed, the processing circuit 41 displays the fluctuation calculated in step SB3 by implementing the display control function 416 (step SB4). The fluctuation is displayed on the display 45.

Figure 19 is a diagram showing an example of the fluctuation display screen I1. As shown in Figure 19, the display screen I1 has a display field I11 for the medical knowledge graph 81, a patient selection field I12, a disease selection field I13, and a medical event category selection field I14.

The patient's identifier (hereinafter referred to as the patient identifier) is displayed in selection field I12. The patient identifier may be a patient ID, a patient name, or the like. The patient identifier displayed in selection field I12 can be selected by the operator via the input interface 33, etc. The selected patient identifier should be highlighted. For example, in Figure 19, the patient identifier "00003" is highlighted. One or more patient identifiers can be selected.

The selection field I13 displays the identifier of the target disease (hereinafter referred to as the disease identifier). The disease identifier may be the name or symbol of the disease. The disease identifier displayed in the selection field I13 can be selected by the operator via the input interface 33 or the like. The selected disease identifier should be highlighted. For example, in Figure 19, the disease identifier "heart failure" is highlighted. When a disease identifier is selected, the patient identifier of the patient with the disease corresponding to that disease identifier is displayed in the selection field I12. The disease identifier and patient identifier are managed in association with each other in the medical information storage device 1 or the like. Note that one medical event category or multiple categories may be selected. Note that one or multiple disease identifiers can be selected.

Selection field I14 displays a list of medical event categories to be displayed. The displayed medical event category may be the name or symbol of the medical event category, or a simulated image or reduced image of a patient graph related to the medical event category. The selected medical event category should be highlighted. For example, in Figure 19, a check mark is placed next to the medical event category "Symptoms." Note that one or more medical event categories can be selected.

Display field I11 displays a visualized graph 81 that visualizes the medical knowledge graph of the target patient. The visualized graph 81 has nodes 82 corresponding to medical events and edges 83 connecting the nodes 82. The nodes 82 are displayed with the name or symbol of the corresponding medical event.

The visualized graph 81 is displayed, for example, in the following procedure. First, the processing circuitry 41 identifies the medical event category selected in selection field I14. If no medical event category is selected in selection field I14 or if the visualized graph 81 is initially displayed, all medical event categories are identified as the selected medical event category. Next, the processing circuitry 41 extracts a graph (hereinafter referred to as a subgraph) consisting of nodes belonging to the identified medical event category and edges connecting these nodes from the medical knowledge graph. For example, in Figure 19, the symptom category is selected in selection field I14, so a subgraph consisting of nodes belonging to the symptom category and edges connecting these nodes is extracted from the medical knowledge graph.

The processing circuitry 41 then displays a visualized graph 81 that visualizes the subgraph, displays the node 82 in a display format that corresponds to the fluctuation of the node feature, and displays the edge 83 in a display format that corresponds to the fluctuation of the edge feature. For example, in Figure 19, the greater the degree of fluctuation, the more blurred the outline of the node 82 is displayed. Also, the greater the degree of fluctuation, the more blurred the line representing the edge 83 is displayed. Note that fluctuation may be displayed for only either the node 82 or the edge 83.

To facilitate understanding of the degree of fluctuation, it is advisable to display object I15, which indicates the degree of blurring of the straight lines representing the edges in stages, and object I16, which indicates the degree of blurring of the node outline in stages, at any location on display screen I1.

In this way, by visualizing the fluctuations in node features on nodes 82 and the fluctuations in edge features on edges 83, it becomes possible to grasp the degree of fluctuation between the patient background information of the target patients. Furthermore, because the fluctuations are displayed for each node 82 and edge 83, it is possible to confirm the fluctuations for each node 82 and edge 83.

As shown in FIG. 19, the display screen I1 may be provided with a display field I17 for displaying the fluctuation of medical judgment information. Display field I17 displays the fluctuation of medical judgment information calculated for the target patient. For example, the fluctuation of the disease hit probability, which is an example of disease classification information, is displayed as the fluctuation of medical judgment information. For example, as shown in FIG. 19, the fluctuation of the disease hit probability is displayed as "70% ± 20%." 70% corresponds to the median of the disease hit probability, and ± 20% corresponds to the fluctuation. Note that the method of displaying the fluctuation of the disease hit probability is not limited to the above method, and a possible range of the disease hit probability, such as "50% - 90%," may also be displayed.

The fluctuation in the disease probability can be calculated based on the fluctuation in the weight parameters. Strictly speaking, the fluctuation in the disease probability depends not only on the fluctuation in the weight parameters, but also on the fluctuation in the node features and edge features.

Specifically, the fluctuation in the disease association probability is calculated as follows. First, the processing circuitry 41 calculates the disease association probability for each of a plurality of target patients based on the patient background information of the target patient, the patient graph of the target patient, and the machine learning model, and calculates an index of the variation in the disease association probability across the plurality of target patients as the fluctuation in the disease association probability. The patient graph is generated by mapping the medical event information of the target patient to a medical knowledge model to which graph features defined by a mathematical model whose model parameters are determined by the medical information learning device 3 are assigned. A weight parameter w _i defined by a mathematical model whose model parameters are determined by the medical information learning device 3 is assigned to the linear combination layer of the machine learning model. As an index of variation, standard deviation, variance, etc. may be calculated.

Note that in addition to the display format according to the fluctuation, the nodes 82 and edges 83 may also be displayed in a display format according to the graph features. In this case, the processing circuitry 41 assigns a node feature to the node 82 and an edge feature to the edge 83. When the node feature is defined by a single-dimensional scalar value such as applicable or not applicable, the processing circuitry 41 assigns the scalar value to the node. When the node feature is defined by a multi-dimensional vector such as the number of occurrences, the order of occurrence, and the degree of occurrence, the processing circuitry 41 aggregates the vector, converts it into a scalar value, and assigns the scalar value to the node 82. For example, the larger the node feature, the darker the display color of the node 82.

Similarly, when the edge feature amount is defined by a single-dimensional scalar value, such as presence or absence, the processing circuitry 31 assigns the scalar value to the edge. When the edge feature amount is defined by a multi-dimensional vector, the processing circuitry 31 aggregates the vector, converts it into a scalar value, and assigns the scalar value to the edge 83. The processing circuitry 41 then displays the node 82 in a display format corresponding to the assigned node, and displays the edge 83 in a display format corresponding to the assigned edge. For example, the larger the edge feature amount, the thicker the edge 83 will be displayed.

Note that the method of displaying fluctuations in graph features is not limited to the above. For example, fluctuations in node features may be displayed by attaching a numerical value or category representing the degree of fluctuation to the node, and fluctuations in edge features may be displayed by attaching a numerical value or category representing the degree of fluctuation to the edge. Furthermore, numerical values or categories representing the fluctuations of node features and edge features may be displayed in a list along with the node number and edge number.

Fluctuations in the learning parameters may be displayed on the display screen I1. Fluctuations in the learning parameters may affect, for example, the disease classification information estimated by the trained model, and therefore may be displayed together with the disease classification information of the target patient. For example, a numerical value or category indicating the degree of fluctuation may be displayed.

Once step SB4 is performed, the flicker display process ends.

Note that the flickering display process described above is just one example, and various modifications are possible.

As an example, although the above example describes a single target patient, there may be multiple target patients. Below, as variant example 1, we will explain an example of fluctuation display when all patients with heart failure are designated as target patients.

Figure 20 shows the fluctuation display screen I2 when all patients with heart failure are designated as target patients. As shown in Figure 20, a medical knowledge graph 81 for all patients with heart failure is displayed in display field I11. Each node 82 is displayed with a degree of blurring that corresponds to the fluctuation of the node feature for all patients with heart failure, and each edge 83 is displayed with a degree of blurring that corresponds to the fluctuation of the edge feature for all patients with heart failure.

According to variant example 1, by displaying the fluctuations in graph features for all patients with a specific disease, it becomes possible to check the fluctuations between patients.

As another example, fluctuations may be displayed for each patient background factor. Below, as variant example 2, an example of displaying fluctuations for each patient background factor is described. Note that displaying fluctuations for each patient background factor can be implemented when model parameters are determined for each patient background factor, such as when the mathematical model is a stratified model or a mixed model.

Figure 21 is a diagram showing display screen I3 of fluctuations for each patient background factor. As shown in Figure 21, display screen I3 displays GUI buttons for selecting patient background factors. For example, in Figure 21, an age button for selecting the patient background factor "age," a gender button for selecting the patient background factor "gender," and a BMI button for selecting the patient background factor "BMI" are displayed in display field I11. These GUI buttons are displayed so that the user can select them via input interface 43.

When a GUI button is selected, the processing circuitry 41 highlights and displays nodes and edges that are strongly fluctuating due to the patient background factor corresponding to the selected GUI button. The nodes and edges are highlighted, for example, by the following method. First, the processing circuitry 41 identifies the patient background factor corresponding to the selected GUI button as a factor of interest. The processing circuitry 41 identifies a model parameter associated with the factor of interest among the model parameters of each node or edge. For example, the model parameters identified are the weight α _v in equation (8) and the weight α _w in equation (9), which are coefficients of the identified patient background factor. If the identified model parameter has a value equal to or greater than a threshold, the processing circuitry 41 highlights the node or edge associated with the model parameter as a node or edge that is strongly fluctuating due to the factor of interest. As an alternative method, the processing circuitry 41 may fix the values of patient background factors other than the factor of interest and significantly change the value of the factor of interest to highlight the fluctuation of the calculated node or edge.

For example, as shown in Figure 21, when BMI is selected as the patient background factor of interest, nodes 82 and edges 83 in the visualized graph 81 that fluctuate strongly due to BMI are displayed in an emphasized manner. Other nodes 82 and edges 83 are displayed lighter than the strongly fluctuating nodes 82 and edges 83. This makes it possible to visualize only the fluctuations of nodes and edges related to the patient background factor of interest.

Variation 2 allows the display of fluctuations limited to the patient background factors that the user is interested in. This allows the user to check the fluctuations of individual patient background factors.

As another example, two or more medical event categories may be selected. Below, as variant example 3, an example of the display of fluctuations when a treatment category and a response category are selected will be described. The visualized graph relating to the treatment category and response category is called a treatment-response graph. Fluctuations are displayed together with the treatment-response graph.

Figure 22 is a diagram showing the treatment-response graph and fluctuation display screen I3. As shown in Figure 22, treatment and response are selected as the medical event category to be displayed. Furthermore, patient "00001," patient "00002," and patient "00003," who are all suffering from heart failure, are selected as the target patients. In this case, a treatment-response graph, which is a visualized graph that visualizes a medical knowledge graph showing the relationship between treatment events and response events, is displayed in display field I11 for patient "00001," patient "00002," and patient "00003." For example, the node for the treatment event "Diuretic," which represents the administration of a diuretic, the node for the response (more specifically, treatment response) event "Diuresis 40 mL/h or greater," which represents diuresis of 40 mL/h or greater, and the node for the response (more specifically, side effect) event "Electrolyte Abnormality," which represents the occurrence of an electrolyte abnormality in the urine, are connected via edges.

Each edge is assigned a numerical value that represents the degree of fluctuation in the edge feature. For example, in the treatment-response graph for patient "00001," the edge from the treatment event "diuretic" to "diuresis 40 mL/h or more" displays a numerical value of "90%" as the degree of correlation or causality of the edge, and the edge from the treatment event "diuretic" to "electrolyte abnormality" displays a numerical value of "5%" as the degree of correlation or causality of the edge. Note that, as shown in Figure 22, differences in the degree of correlation or causality between patients indicate the occurrence of fluctuation. The degree of correlation or causality of an edge is not limited to being represented numerically, but may also be represented in the form of edge thickness, as shown in Figure 22.

According to variant 3, it is possible to display fluctuations together with the treatment-response graph. By displaying the treatment-response graph, it is possible to predict the treatment effect, but at this time, by displaying fluctuations in edge features, etc., it is also possible to evaluate the fluctuations in treatment effect between patients.

As described above, the medical information processing device 4 according to this embodiment has a processing circuitry 41 and a memory 42. The processing circuitry 41 acquires patient background information relating to the background factors of one or more target patients. The memory 42 stores a medical knowledge graph having nodes corresponding to medical events and edges representing the relationships between the nodes. The graph features of the medical knowledge graph are expressed by a mathematical model characterized by the patient background information. The processing circuitry 41 calculates the fluctuation of the graph features relating to the target patient based on the patient background information and the mathematical model. The processing circuitry 41 displays the fluctuation of the graph features.

The above configuration makes it possible to display the fluctuations in graph features between patient background information. Displaying the fluctuations is useful when explaining a disease to patients, etc. For example, it allows them to understand that some people with the same disease may experience certain symptoms while others do not, or that some people with the same disease may respond well to treatment while others do not. Displaying the fluctuations is also useful when discovering research hypotheses. For example, it can help identify new subtype classification candidates and understand the eligibility criteria for research design. Displaying the fluctuations is also useful when estimating disease candidates, predicting treatment responses, and predicting prognosis. For example, it allows for estimations that take individual differences into account.

Furthermore, by visualizing fluctuations in graph features on the nodes and edges of a medical knowledge graph, it is possible to visualize individual patient differences while maintaining common concepts and interpretability regarding the relationships between medical events.

Various application examples of the medical information system 100 according to this embodiment are described below. The following application examples can be applied as appropriate to the various embodiments described above.

(Application example 1)
FIG. 23 is a diagram illustrating an overview of graph features according to Application Example 1. As shown in FIG. 23 , medical event information such as a relevant/non-relevant feature, a temporal feature, and/or a location feature is assigned to each node of the patient graph 20A as graph features. The relevant/non-relevant feature is information indicating the degree to which the node corresponds to the medical event. The temporal feature is information regarding the time when the medical event occurred. Specifically, the temporal feature is the date and time of the occurrence of the medical event, the order of occurrence, the number of occurrences, etc. The location feature is information regarding the location where the medical event occurred. Specifically, the location information is the location or location of the occurrence of the medical event. The location of occurrence is sensor information such as GPS (Global Positioning System) that indicates the point where the medical event occurred. The location of occurrence is the address or name of the medical institution, department, home, or hospital where the medical event was diagnosed.

As described above, the processing circuitry 41, by implementing the fluctuation calculation function 413, calculates the fluctuation of graph features related to the target patient based on the patient background information and mathematical model of the target patient. The relevant/non-relevant features, temporal features, and location features are each expressed mathematically using the mathematical model of equation (8). The fluctuations of the relevant/non-relevant features, temporal features, and location features may be calculated, for example, as an index of variation in the feature between multiple patients. The processing circuitry 41, by implementing the visualization graph generation function 414, generates a visualization graph having a display format corresponding to the fluctuation of each feature. The feature to be displayed for the fluctuation can be selected arbitrarily from the relevant/non-relevant features, temporal features, and location features. The processing circuitry 41, by implementing the display control function 416, displays the generated visualization graph in a display format corresponding to the fluctuation.

The processing circuit 41 may calculate the overall fluctuation of the relevant/non-relevant feature, the temporal feature, and the location feature. As an example, the processing circuit 41 may calculate the fluctuation of each of the relevant/non-relevant feature, the temporal feature, and the location feature, and calculate the statistical value (average value, minimum value, maximum value, etc.) of the fluctuation of each feature as the overall fluctuation.

(Application Example 2)
FIG. 24 is a diagram showing an overview of graph features related to Application Example 2. As shown in FIG. 24, time-series medical event information consisting of multiple pieces of medical event information with different occurrence times is assigned to each node of the patient graph 20A as a graph feature. The value of the medical event information changes over time. For example, whether a medical event is relevant or not changes over time. Occurrence time information is assigned to the medical event information as meta information. The occurrence time information is a timestamp of the medical event information, and is, for example, the diagnosis date and time or recording date and time of the medical event. For example, as shown in FIG. 24, medical event information at time t, medical event information at time t+1, and medical event information at time t+2 are assigned to each node as graph features.

The processing circuit 41 in application example 2 calculates the fluctuation of graph features based on the time-series graph features.

Figure 25 is a diagram showing an example network configuration of a machine learning model 60A related to Application Example 2. As shown in Figure 25, the machine learning model 60A has a graph convolution layer 61, a readout layer 62, and an RNN (Recurrent Neural Network) layer 64. The RNN layer 64 outputs disease classification information 71 from the time-series graph features 20C.

As shown in Figure 25, assume that the processing targets are patient graph 20A0, which is assigned a graph feature at time t, patient graph 20A1, which is assigned a graph feature at time t+1, and patient graph 20A2, which is assigned a graph feature at time t+2. The processing circuit 31 inputs patient graphs 20A0, 20A1, and 20A2 to the graph convolution layer 61, outputs a convolved patient graph corresponding to patient graph 20A0, a convolved patient graph corresponding to patient graph 20A1, and a convolved patient graph corresponding to patient graph 20A2, and inputs each convolved patient graph to the readout layer 62, which outputs feature vector 20C0 at time t, feature vector 20C1 at time t+1, and feature vector 20C2 at time t+2. Feature vector 20C0 is a vector representation of the graph feature at time t, feature vector 20C1 is a vector representation of the graph feature at time t+1, and feature vector 20C2 is a vector representation of the graph feature at time t+2.

As shown in Figure 25, the processing circuitry 31 inputs the feature vectors 20C0, 20C1, and 20C2 to the RNN layer 64 and outputs a single disease classification information 71. By using the RNN layer 64, it is possible to obtain the probability of various diseases by using medical event information at multiple sequential times. The weight parameters of the RNN layer 64 are affected by patient background information. The weight parameters w _i of the RNN layer 64 can be expressed by a mathematical model with coefficients, node features, and biases as elements, similar to equation (9).

In Application Example 2, temporal changes in patient background are incorporated into the fluctuation components. The patient background information in Application Example 2 is assigned time-series patient background information consisting of multiple pieces of patient background information with different occurrence times. Each piece of patient background information has patient background factors such as age, gender, BMI, and medical history. As described above, the elements that have fluctuations among the elements involved in the machine learning model include node features and edge features.

An example of calculating a node feature according to Application Example 2 will be described. As shown in FIG. 14 and equation (8) above, the node feature v _i is expressed as the sum of the product of the coefficient (α _v Z+β _v ) and the initial node feature x _i and the bias b _v . In Application Example 2, the node feature v _i can be calculated by replacing the patient background information Z in equation (8) with patient background information Z' shown in equation (10) below.

Z′=RNN(Z[t]) (10)

Specific algorithms for the RNN (Z[t]) shown in equation (10) can include various algorithms such as Elman networks, Jordan networks, LSTM (long short-term memory), and GRU (gated recurrent units).

The weighting parameter w _i of the RNN layer 64 expressed by equation (9) can also be calculated by replacing the patient background information Z in equation (9) with the patient background information Z′ shown in equation (10).

An example of calculating an edge feature amount according to Application Example 2 will be described below. As illustrated in the above-described Fig. 11 and formulas (2) and (3), the processing circuitry 31 performs a Granger causality test on an edge connecting a medical event x and a medical event y based on medical event information of a plurality of patients, determines a weight _Wx and a bias _bx , which are model parameters of a mathematical model representing the edge feature amount of the edge, and determines a coefficient _αx, which is an edge feature amount, based on the weight _Wx and the bias _bx . In Application Example 2, the weight _Wx and the bias _bx can be determined by replacing the patient background information Z in formula (3) with the patient background information Z' shown in formula (10).

(Application Example 3)
In Application Example 2, a time-series graph feature is assigned to each node. In Application Example 3, the patient graph itself changes in time series.

The processing circuitry 31 of the medical information learning device 3 in Application Example 3 continuously acquires learning samples (clinical information) with different occurrence times, even after generating a trained model. As described above, a learning sample is a combination of medical event information and disease information. The processing circuitry 31 performs continuous learning of the learning parameters of the machine learning model and the model parameters of the mathematical model on a regular or irregular basis based on the acquired multiple pieces of medical information. This continuous learning generates a time-series trained model with different occurrence times. The learning parameters include an adjacency matrix and/or learning parameters of the machine learning model. The patient graph before convolution at each time point is applied to the trained model at each time point, generating a patient graph after convolution at each time point. This makes it possible to generate a time-series patient graph in which the graph features of each node and the connection relationships between nodes change over time.

The trained model for each occurrence time constituting the time series trained model is trained according to the processing procedure shown in FIG. 9 based on the training samples acquired up to that time. At this time, in order to suppress a large change in the characteristics of the patient graph before and after updating the adjacency matrix A and/or the weight parameter W, a regularization term may be provided in the loss function as shown in the following equations (11) and (12). Equation (11) is the normalization term L _reg related to the update of the adjacency matrix A, and is expressed by the Kullback-Leibler divergence between the adjacency matrix A at the processing reference time and the adjacency matrix A′ at another time. Equation (12) is the normalization term L _reg related to the update of the weight parameter W ^(l) , and is expressed by the Kullback-Leibler divergence between the weight parameter W ^(l) at the processing reference time and the weight parameter W′ ^(l) at another time.

(Application Example 4)
In some of the above-described embodiments, spatial information such as location features is assigned to each node of the patient graph as a graph feature. In Application Example 4, the concept of space is added to the patient graph itself.

Figure 26 is a diagram schematically illustrating the relationship between patient graphs and spatial information in Application Example 4. As shown in Figure 26, spatial information related to the target patient of each patient graph is assigned to that patient graph. The spatial information includes locational information and biological information as its elements. The locational information may be sensor information such as GPS indicating the current location of the target patient, or the location of the medical institution where the target patient visits, the target patient's home address, or the hospital room where the target patient is hospitalized. The biological information includes the target patient's blood relationship, medical history, gene sequence information, etc. By embedding spatial information in the patient graph, it is possible to construct a network (hereinafter referred to as a patient graph network) in which multiple patient graphs are arranged according to the presence or degree of spatial information relationship.

Figure 27 is a diagram illustrating the concept of the patient graph network 200. As shown in Figure 27, the patient graph network 200 is composed of multiple patient graphs that are connected according to the presence or absence or degree of spatial information relationship. Figure 27 illustrates four patient graphs 201, 202, 203, and 204, but the number of patient graphs is not particularly limited as long as it is two or more.

Patient graphs connected to each other by an edge (e.g., patient graph 201 and patient graph 202) have a spatial information relationship. Conversely, patient graphs not connected to each other by an edge (e.g., patient graph 201 and patient graph 204) do not have a spatial information relationship. The distance between patient graphs connected by an edge represents the degree of spatial information relationship. The processing circuitry 31 can evaluate the degree of spatial information relationship based on locational information, such as the home addresses of the two patients or the distance between their hospital rooms. The processing circuitry 31 can evaluate biological information based on the degree of kinship between the two patients, the medical relationship in their medical history, and the degree of match in their gene sequences. The processing circuitry 31 can determine whether or not there is a spatial information relationship based on a comparison between the degree evaluated by the above method and a threshold. Edges may be formed based on a comprehensive evaluation of the above elements of spatial information, or may be formed for each of the above elements.

By assigning spatial information to patient graphs, it is possible to easily search for patient graphs whose spatial information is close to that of the target patient. As an example, the processing circuitry 31 can extract patient graphs connected by edges to the patient graph of the target patient as patient graphs of patients whose spatial information is close to that of the target patient.

Note that spatial information is not limited to including both location information and biological information, and may include only either location information or biological information.

(Application Example 5)
Application example 5 utilizes the spatial information according to application example 4. The processing circuitry 31 of the medical information processing device 3 according to application example 5 convolves graph features of a patient graph of a patient different from the target patient onto the patient graph of the target patient.

Figure 28 is a diagram showing an overview of graph features related to Application Example 5. As shown in Figure 28, medical event information such as relevant/non-relevant features, temporal features, location features, and/or spatially proximate patient features is assigned as graph features to each node of the patient graph 20B. The relevant/non-relevant features, temporal features, and location features are as described in Application Example 1. Spatially proximate patient features are medical event information of other patients whose spatial information is proximate to that of the target patient. For example, the node for the medical event "headache" is assigned the relevant/non-relevant features, temporal features, and location features of the target patient, as well as the relevant/non-relevant features, temporal features, and location features of the target patient's father and/or mother.

The processing circuitry 31 of Application Example 5 maps the medical event information of Application Example 5 onto a medical knowledge graph to generate a patient graph 20A for the target patient. The medical event information of Application Example 5 includes, for each medical event, relevant/non-relevant features, temporal features, location features, and spatially proximate patient features. The relevant/non-relevant features, temporal features, and location features are medical event information for the target patient. The spatially proximate patient features are medical event information for other patients whose spatial information is proximate to the target patient. In other words, the processing circuitry 31 of Application Example 5 maps the medical event information of the target patient and the medical event information of other patients onto each node as the node features of that node. Through the mapping, a patient graph 20B is generated in which relevant/non-relevant features, temporal features, location features, and spatially proximate patient features are assigned to each node.

Then, the processing circuitry 31 applies the patient graph having nodes assigned with relevant/non-relevant features, temporal features, location features, and spatially proximate patient features to the machine learning model 60 to estimate medical judgment information such as disease classification information. According to Application Example 5, it is possible to estimate medical judgment information by taking into account the relevant/non-relevant features, temporal features, location features, and spatially proximate patient features of each medical event. This makes it possible to estimate medical judgment information that takes into account the medical information (graph features) of not only the target patient, but also patients whose spatial information is proximate to the target patient. Note that each node does not need to be assigned all of the relevant/non-relevant features, temporal features, location features, and spatially proximate patient features; it may be assigned any two or one of the spatially proximate patient features and the relevant/non-relevant features, temporal features, and location features.

(Application Example 6)
In some of the above embodiments, the processing circuitry 31 displays a visualized graph based on a partial patient graph that includes only nodes and edges belonging to some of all medical event categories, including symptoms, physical findings, test findings, treatments, treatment responses, and side effects. However, the processing circuitry 31 may also display a visualized graph based on the entire patient graph across all medical event categories, as shown in FIG. 2 . In this case, the processing circuitry 31 may display the visualized graph so that each category can be distinguished by a visual effect such as color. This makes it possible to get an overview of the entire patient graph.

At least one of the embodiments described above makes it possible to visualize individual differences between patients.

The term "processor" used in the above description refers to circuits such as a CPU, GPU, or application-specific integrated circuit (ASIC), programmable logic device (e.g., simple programmable logic device (SPLD), complex programmable logic device (CPLD), and field programmable gate array (FPGA)). A processor realizes its functions by reading and executing a program stored in a memory circuit. Note that instead of storing a program in a memory circuit, the processor may be configured to directly incorporate the program into its circuit. In this case, the processor realizes its functions by reading and executing the program embedded in the circuit. Furthermore, rather than executing a program, the function corresponding to the program may be realized by combining logic circuits. Note that each processor in this embodiment is not limited to being configured as a single circuit per processor; it may also be configured as a single processor by combining multiple independent circuits to realize its function. Furthermore, multiple components in Figures 1, 5, and 17 may be integrated into a single processor to realize its function.

While several embodiments have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments may be embodied in a variety of other forms, and various omissions, substitutions, modifications, and combinations of embodiments may be made without departing from the spirit of the invention. These embodiments and their variations are included within the scope and spirit of the invention, as well as within the scope of the invention and its equivalents as set forth in the claims.

REFERENCE SIGNS LIST 1 Medical information storage device 2 Medical knowledge graph storage device 3 Medical information learning device 4 Medical information processing device 31 Processing circuit 32 Memory 33 Input interface 34 Communication interface 35 Display 41 Processing circuit 42 Memory 43 Input interface 44 Communication interface 45 Display 100 Medical information system

Claims (22)

an acquisition unit that acquires patient background information regarding one or more background factors of a target patient;
a storage unit for storing a medical knowledge graph having nodes corresponding to medical events and edges representing relationships between the nodes, wherein graph features of the medical knowledge graph are expressed by a mathematical model characterized by patient background information;
a calculation unit that calculates fluctuations of the graph feature amount related to the target patient based on patient background information of the target patient and the mathematical model;
a display control unit that displays the fluctuation of the graph feature;
A medical information processing device comprising: The medical image processing device of claim 1, wherein the graph features include node types, node values, edge types, and/or edge values. The medical information processing device of claim 2, wherein the display control unit displays a visualization graph that visualizes the medical knowledge graph, and displays nodes and/or edges included in the visualization graph in a display format that corresponds to the fluctuation. a selection unit for selecting background factors to be displayed from the patient background information;
the display control unit displays the fluctuation of the graph feature related to the background factor of the selected display object.
The medical information processing device according to claim 1. The medical information processing device of claim 1, wherein the medical events include events belonging to at least one of the following categories: symptoms, examinations, treatments, treatment responses, and side effects. a selection unit for selecting a category of the medical event to be displayed;
the display control unit extracts, from the medical knowledge graph, a subgraph including nodes belonging to a category to be displayed and edges connecting the nodes, displays a visualized graph that visualizes the subgraph, and displays the nodes and/or edges included in the visualized graph in a display form according to the fluctuation.
The medical information processing device according to claim 1. the graph features of the medical knowledge graph are represented by the mathematical model with model parameters as variables;
The model parameters are determined based on the patient background information.
The medical information processing device according to claim 1. The medical information processing device of claim 1, wherein the medical knowledge graph is a graph generated based on medical information of multiple patients or a medical ontology. the storage unit further stores a machine learning model that estimates medical judgment information from the medical knowledge graph, and learning parameters of the machine learning model are expressed by the mathematical model characterized by the patient background information;
the calculation unit calculates fluctuations of the learning parameters related to the target patient based on the patient background information of the target patient and the mathematical model;
the display control unit displays the fluctuation of the learning parameter.
The medical information processing device according to claim 1. the patient background information is time-series patient background information occurring at different times,
the calculation unit calculates the fluctuation of the graph feature related to the target patient based on the time-series patient background information of the target patient and the mathematical model.
The medical information processing device according to claim 1. an acquisition unit that acquires patient background information regarding background factors of a plurality of patients;
a storage unit that stores a medical knowledge graph having nodes corresponding to medical events and edges representing relationships between the nodes, wherein graph features of the medical knowledge graph are characterized by a mathematical model including patient background information;
a learning unit that determines model parameters of the mathematical model based on patient background information of a plurality of patients and the mathematical model;
A medical information learning device comprising: The medical information learning device of claim 11, wherein the mathematical model is a stratified model stratified by patient background information, a mixed model having random effects based on patient background information, or a function whose value is uniquely determined by patient background information. The medical information learning device of claim 11, wherein the mathematical model has terms independent of patient background information and terms characterized by patient background information. The medical information learning device of claim 11, wherein the model parameters are determined by statistical analysis or machine learning based on patient background information of the multiple patients. the graph feature is an edge type and/or an edge value,
The model parameters are determined based on analysis of variance, Granger causality estimation, or causal effect estimation.
The medical information learning device according to claim 11. The medical information learning device of claim 11, wherein the model parameters are determined based on a hierarchical Bayesian model or supervised learning. the storage unit further stores a machine learning model that estimates medical judgment information from the medical knowledge graph, and learning parameters of the machine learning model are expressed by the mathematical model characterized by the patient background information;
the learning unit determines the learning parameters based on the patient background information of the plurality of patients and the mathematical model.
The medical information learning device according to claim 11. the machine learning model has a linear combination layer that estimates medical judgment information from the graph features of the medical knowledge graph;
The learning parameters are weight parameters of the linear combination layer.
18. The medical information learning device according to claim 17. The medical information learning device of claim 11, wherein the patient background information does not include time information. The medical information learning device of claim 11, wherein the medical knowledge graph is a graph generated based on medical information of multiple patients or a medical ontology. The medical information learning device of claim 11, wherein the medical events include events belonging to at least one of the following categories: symptoms, examinations, treatments, treatment responses, and side effects. the patient background information is time-series patient background information,
the learning unit determines the model parameters based on time-series patient background information of a plurality of patients and the mathematical model.
The medical information learning device according to claim 11.