Prediction device, prediction method, prediction program, prediction system

Abstract

To provide a technology capable of predicting the prognosis of a patient with acute myocardial infarction with a higher degree of accuracy.SOLUTION: A prediction device 1 comprises: an acquisition unit that acquires a patient's at least one feature amount; a prediction unit that predicts a prognosis on the basis of the at least one feature amount by using a learned prediction model; and an output unit that outputs the result of the prediction of the prognosis.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present disclosure relates to a prediction device, a prediction method, a prediction program, and a prediction system.

Background Art

[0002] Acute myocardial infarction (AMI: Acute Myocardial Infarction, hereinafter also referred to as "AMI") is a major cause of death worldwide. To date, several methods have been proposed to predict the short-term and long-term prognosis of AMI patients.

[0003] For example, Patent Document 1 discloses risk scoring methods such as the TIMI (Thrombolysis In Myocardial Ischemia) risk index and the GRACE (Global Registries of Acute Coronary Events) score as methods for predicting the prognosis of AMI.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] The conventional prediction methods disclosed in Patent Document 1 have existed since before percutaneous coronary intervention by catheter became the main treatment method for AMI. For this reason, it is difficult to say that the conventional prediction methods are suitable for the current standard medical treatment in which percutaneous coronary intervention has become the main treatment method, and there is a need for a prediction method that can predict the prognosis of AMI patients with higher accuracy.

[0006] The present disclosure has been made to solve such problems, and an object thereof is to provide a technique capable of predicting the prognosis of patients with acute myocardial infarction with higher accuracy.

Means for Solving the Problems

[0007] A prediction device according to an aspect of the present disclosure predicts the prognosis of a patient with acute myocardial infarction. The prediction device includes an acquisition unit, a prediction unit, and an output unit. The acquisition unit acquires at least one feature amount of the patient. The prediction unit predicts the prognosis based on at least one feature amount using a learned prediction model. The output unit outputs the prediction result of the prognosis. At least one feature quantity is 10 types of feature quantities including the maximum value of creatine phosphokinase in blood, the amount of hemoglobin in blood, heart rate, the concentration of creatinine in blood, systolic blood pressure, blood glucose level, age, the result of classification regarding the severity of cardiac dysfunction in acute myocardial infarction, the number of white blood cells in blood, and the amount of C-reactive protein in blood.

[0008] A prediction method according to another aspect of the present disclosure is a method for predicting the prognosis of a patient with acute myocardial infarction by a computer. The prediction method includes: (a) a step of acquiring at least one feature amount of the patient; (b) a step of predicting the prognosis based on at least one feature amount using a learned prediction model; and (c) a step of outputting the prediction result of the prognosis. At least one feature quantity is 10 types of feature quantities including the maximum value of creatine phosphokinase in blood, the amount of hemoglobin in blood, heart rate, the concentration of creatinine in blood, systolic blood pressure, blood glucose level, age, the result of classification regarding the severity of cardiac dysfunction in acute myocardial infarction, the number of white blood cells in blood, and the amount of C-reactive protein in blood.

[0009] A prediction program according to another aspect of the present disclosure is a program for predicting the prognosis of a patient with acute myocardial infarction. The prediction program causes a computer to execute: (a) a step of acquiring at least one feature amount of the patient; (b) a step of predicting the prognosis based on at least one feature amount using a learned prediction model; and (c) a step of outputting the prediction result of the prognosis. At least one feature quantity is 10 types of feature quantities including the maximum value of creatine phosphokinase in blood, the amount of hemoglobin in blood, heart rate, the concentration of creatinine in blood, systolic blood pressure, blood glucose level, age, the result of classification regarding the severity of cardiac dysfunction in acute myocardial infarction, the number of white blood cells in blood, and the amount of C-reactive protein in blood.

[0010] A prediction system according to another aspect of the present disclosure predicts the prognosis of a patient with acute myocardial infarction. The prediction system includes a user terminal and a server device configured to communicate with the user terminal. The server device includes an acquisition unit, a prediction unit, and an output unit. The acquisition unit acquires at least one feature amount of the patient from the user terminal. The prediction unit predicts the prognosis based on at least one feature amount using a learned prediction model. The output unit outputs the prediction result of the prognosis to the user terminal. At least one feature quantity is 10 types of feature quantities including the maximum value of creatine phosphokinase in blood, the amount of hemoglobin in blood, heart rate, the concentration of creatinine in blood, systolic blood pressure, blood glucose level, age, the result of classification regarding the severity of cardiac dysfunction in acute myocardial infarction, the number of white blood cells in blood, and the amount of C-reactive protein in blood. [Effect of the Invention]

[0011] According to the present disclosure, since the prognosis can be predicted based on the characteristics of a patient using a learned prediction model, the prognosis of a patient with acute myocardial infarction can be predicted with higher accuracy. [Brief Description of the Drawings]

[0012]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Figure 13

Figure 14

Figure 15

Figure 16

Embodiments for Carrying Out the Invention

[0013] This embodiment will be described in detail with reference to the drawings. For the same or corresponding parts in the drawings, the same reference numerals are given, and the description thereof will not be repeated in principle.

[0014] <Prediction Device According to Embodiment 1> The prediction device 1 according to Embodiment 1 will be described with reference to FIGS. 1 to 15.

[0015] [Configuration of Prediction Device] FIG. 1 is a diagram for explaining an application example of the prediction device 1 according to Embodiment 1. The prediction device 1 is an information terminal that executes predetermined information processing, such as a desktop PC (personal computer), a laptop PC, a smartphone, a smartwatch, a wearable device, and a tablet PC. In Embodiment 1, the prediction device 1 is used by a user such as a doctor and executes a process for predicting the prognosis of an AMI patient (hereinafter, also referred to as "prediction process").

[0016] Specifically, the prediction device 1 includes a prediction model 121 for predicting the prognosis of an AMI patient. The prediction model 121 is generated by machine learning based on at least one feature amount belonging to the patient and the presence or absence of in-hospital death of the patient. The prediction device 1 predicts the prognosis of the patient based on the feature amount of the AMI patient by using the prediction model 121. The user can provide the prediction result by the prediction device 1 to the patient, and perform medical treatment with higher accuracy than before for the patient.

[0017] As shown in FIG. 1, the prediction device 1 includes a control device 10, a display device 20, a keyboard 31, and a mouse 32. The display device 20 displays an image on the display 25 according to the control of the control device 10. Each of the keyboard 31 and the mouse 32 receives a user operation and inputs information based on the received user operation to the control device 10. Note that the display device 20 may include a touch panel (not shown). The user may input information to the control device 10 by operating the touch panel.

[0018] FIG. 2 is a diagram showing the configuration of the control device 10 included in the prediction device 1 according to Embodiment 1. As shown in FIG. 2, the control device 10 includes an arithmetic device 11, a storage device 12, an input interface 13, and an output interface 14.

[0019] The arithmetic device 11 is an example of a "prediction unit" and is an arithmetic entity (computer) that executes various processes according to various programs. The arithmetic device 11 includes, for example, at least one of a CPU (Central Processing Unit), an FPGA (Field Programmable Gate Array), a GPU (Graphics Processing Unit), and an MPU (Multi Processing Unit). Further, the arithmetic device 11 may include a volatile memory such as a DRAM (Dynamic Random Access Memory) and an SRAM (Static Random Access Memory), and a non-volatile memory such as a ROM (Read Only Memory) and a flash memory. Note that the arithmetic device 11 may be composed of an arithmetic circuit (Processing Circuitry).

[0020] The storage device 12 includes non-volatile memories such as HDD (Hard Disk Drive) and SSD (Solid State Drive). The storage device 12 stores various programs and data such as the prediction model 121 and the prediction program 122. The prediction program 122 is a program for the arithmetic unit 11 to execute prediction processing using the prediction model 121.

[0021] The input interface 13 is an example of an "acquisition unit" and acquires information input from the outside. For example, the input interface 13 acquires information input from each of the keyboard 31 and the mouse 32.

[0022] The output interface 14 is an example of an "output unit" and outputs information to the outside according to the control of the arithmetic unit 11. For example, the output interface 14 outputs information for displaying an image to the display device 20.

[0023] In the control device 10 configured as described above, the input interface 13 acquires at least one feature amount of the AMI patient from the outside. The arithmetic unit 11 executes prediction processing according to the prediction program 122, and predicts the prognosis of the patient based on the feature amount of the AMI patient and the prediction model 121. The output interface 14 outputs image information including the prediction result obtained by the arithmetic unit 11 to the display device 20. The display device 20 displays the prediction result on the display 25 based on the image information acquired from the control device 10.

[0024] Thereby, the user can predict the prognosis of the patient using the learned prediction model 121 by inputting at least one feature amount of the AMI patient into the prediction device 1.

[0025] [Generation of Prediction Model] The generation of the prediction model 121 will be described with reference to FIGS. 3 to 8.

[0026] FIG. 3 is a diagram for explaining the generation of the prediction model 121 according to Embodiment 1. The prediction model 121 according to Embodiment 1 is generated by supervised learning. As shown in FIG. 3, the designer of the prediction model 121 first prepares a plurality of sample data to be used when generating the prediction model 121. For example, the designer prepares data on patients with AMI after percutaneous coronary intervention has become the main treatment method for AMI. In this example, the designer prepares data on 2,553 patients with ST-segment elevation acute myocardial infarction registered at medical institutions during the period from 2009 to 2015 as sample data.

[0027] Each of the plurality of sample data includes at least one feature amount and correct answer data. The at least one feature amount includes the patient's clinical data and biological data, etc. The correct answer data includes information on whether the patient died in the hospital.

[0028] Here, with reference to FIGS. 4 and 5, the sample data will be specifically described. FIGS. 4 and 5 are diagrams for explaining an example of the sample data. In FIGS. 4 and 5, for each feature amount, data on patients who survived and data on patients who died in the hospital are shown.

[0029] As shown in FIGS. 4 and 5, out of the 2,553 patients, the number of patients who survived is 2,358, and the number of patients who died in the hospital is 195. That is, out of the 2,553 sample data, the number of sample data (Survival) assigned survival as the correct answer data is 2,358, and the number of sample data (Dead) assigned in-hospital death as the correct answer data is 195. The sample data includes data corresponding to various feature amounts for each of the patients who survived and the patients who died in the hospital.

[0030] As shown in FIG. 4, the characteristic quantities include, for example, age, gender, body mass index (BMI), presence or absence of hypertension, presence or absence of diabetes, presence or absence of dyslipidemia, presence or absence of smoking, presence or absence of a family with a history of AMI, presence or absence of a past episode of AMI, presence or absence of a past episode of cardiovascular disease, presence or absence of treatment by hemodialysis, the result of the Killip classification regarding the severity of myocardial dysfunction in AMI, systolic blood pressure (BPs), heart rate (HR), the number of white blood cells (WBC) in the blood, the amount of hemoglobin (Hb) in the blood, blood sugar level (BS), the maximum value of creatine phosphokinase in the blood (max CPK: max Creatine Phosphorus Kinase, hereinafter also referred to as "maxCPK"), the concentration of creatinine (Cr) in the blood, the amount of C-reactive protein (CRP) in the blood, each TIMI grade before and after treatment, presence or absence of treatment using a stent, and presence or absence of treatment by thrombus aspiration, etc.

[0031] Here, creatine phosphokinase (CPK: Creatine Phosphorus Kinase, hereinafter also referred to as "CPK") is an enzyme present in muscles such as skeletal muscle, myocardial muscle, and smooth muscle, and plays an important role in the energy metabolism of muscle cells. When abnormalities occur in these muscle cells, CPK flows out into the blood, so the amount of CPK in the blood increases. When AMI occurs, the value of CPK shown by a blood test becomes higher than the normal value and peaks out within 24 hours after the onset of AMI. maxCPK is the value of CPK at the peak. Note that the unit for representing the value of CPK is "IU / L" or "U / L". Regarding these units, for example, the amount of enzyme that can change 1 μmol of substrate per minute at a predetermined temperature (for example, 30 degrees, 37 degrees in the measurement by an automatic biochemical analyzer for clinical tests) in 1 L of a sample is defined as 1 unit.

[0032] Furthermore, as shown in FIG. 5, the feature quantities include, for example, the site where AMI occurred, the responsible blood vessel where AMI occurred, the number of lesions, the season when AMI occurred, the time of onset of AMI, the time required from when the patient arrived at the hospital until blood flow reperfusion, the time required from when AMI occurred until the patient arrived at the hospital, and the method of coming to the hospital. As shown in parentheses in FIGS. 4 and 5 respectively, some of the plurality of feature quantities also include information on the ratio (%) or the range of data variation (IQR: Interquartile Range).

[0033] Returning to FIG. 3, the designer generates and evaluates the prediction model 121 using the hold-out method. Specifically, the designer divides a plurality of sample data into training data for generating the prediction model 121 and test data for evaluating the prediction model 121. For example, the designer uses 80% of the plurality of sample data as training data and 20% of the plurality of sample data as test data. Note that the ratio of each of the training data and the test data may be set appropriately by the designer.

[0034] Furthermore, the designer generates the prediction model 121 by K-fold cross-validation using a plurality of training data. Specifically, the designer divides the plurality of training data into K pieces (K = 5 in this example), and out of the K pieces of training data, uses 1 piece of training data as verification data and the remaining (K - 1) pieces of training data as learning data. Then, the designer learns the prediction model using the learning data, and then evaluates the prediction model using the verification data. Next, the designer uses 1 piece of training data different from the training data already used as verification data out of the K pieces of training data as new verification data, and the remaining (K - 1) pieces of training data as new learning data. Then, the designer learns the prediction model using the learning data, and then evaluates the prediction model using the verification data.

[0035] The designer repeats the set of machine learning and evaluation for the prediction model as described above K times until all the training data is used as validation data. In this way, the designer generates a prediction model using a plurality of training data.

[0036] As the machine learning of the prediction model, ensemble learning is applied. For example, in Embodiment 1, random forest is used as the machine learning algorithm. Random forest is a method that generates a plurality of decision trees based on at least one feature randomly selected from a plurality of features included in the training data, and improves the prediction accuracy by taking the average or majority vote of the prediction results of each of the plurality of decision trees.

[0037] By using a random forest classifier as the learning algorithm, the designer can select the prediction model with the highest prediction accuracy among the plurality of generated prediction models, and can also select at least one feature with the highest prediction accuracy. Then, the designer evaluates the selected prediction model using test data. At this time, the designer uses the feature with the highest prediction accuracy as the feature of the test data. In this way, the designer can generate the final prediction model 121.

[0038] Note that the machine learning algorithm of the prediction model 121 is not limited to random forest, and known learning algorithms such as support vector machine, neural network, logistic regression, and xgboost (Extreme Gradient Boosting) may be applied. Further, the designer may generate the prediction model 121 using the same algorithm as the algorithm used to extract at least one feature amount, or may generate the prediction model 121 using an algorithm different from the algorithm used to extract at least one feature amount. For example, when the designer selects at least one feature amount with the highest prediction accuracy from the learning data, the random forest algorithm is used, while when generating the final prediction model 121 using the selected at least one feature amount, the prediction model 121 may be generated using another algorithm different from the random forest described above.

[0039] FIG. 6 is a diagram showing the evaluation results of AUC (Area Under the Curve) for each number of types of feature amounts. FIG. 6(A) shows the evaluation results of AUC in the prediction model generated by random forest. FIG. 6(B) shows the evaluation results of AUC in the prediction model generated by xgboost. FIG. 6(C) shows the evaluation results of AUC in the prediction model generated by logistic regression. FIG. 6 shows a graph with the number of types of feature amounts on the horizontal axis and AUC on the vertical axis for each learning algorithm.

[0040] As shown in FIG. 9 described later, the AUC is an evaluation index represented by the area under the ROC (Receiver Operating Characteristic) curve plotted in a graph with the false positive rate (1 - specificity) on the horizontal axis and the sensitivity on the vertical axis. The closer the AUC is to 1, the higher the prediction accuracy of the prediction model can be said to be. The false positive rate indicates the ratio of the labels of sample data associated with a negative label (for example, survival) as the correct data being mispredicted by the prediction model as a positive label (for example, in-hospital death). The sensitivity indicates the ratio of the labels of sample data associated with a positive label (for example, in-hospital death) as the correct data being correctly predicted as a positive label (for example, in-hospital death) by the prediction model.

[0041] As shown in FIG. 6, for any prediction model generated by any of the learning algorithms of random forest, xgboost, and logistic regression, when using six or more types of feature quantities as input, an AUC with a high value of 0.85 or more can be obtained. Furthermore, for any prediction model generated by any learning algorithm, when using ten types of feature quantities as input, the highest value of AUC can be obtained.

[0042] FIGS. 7 and 8 are diagrams for explaining the importance of each feature quantity. FIGS. 7 and 8 show graphs with the importance of the feature quantity on the horizontal axis and the type of feature quantity on the vertical axis. The importance of the feature quantity is calculated by a learner using random forest as the learning algorithm. The higher the importance of the feature quantity used as the input data of the prediction model, the higher the prediction accuracy of the prediction model.

[0043] As shown in FIG. 7, the ten types of feature quantities that can obtain the highest value of AUC are maxCPK, the amount of hemoglobin, heart rate, creatinine concentration, systolic blood pressure, blood glucose level, age, the result of Killip classification, the number of white blood cells, and the amount of C-reactive protein.

[0044] As shown in FIG. 8, when all the features are selected as the input data of the prediction model, in order from the most important features, maxCPK, systolic blood pressure, age, creatinine concentration, Killip classification result, hemoglobin amount, white blood cell count, heart rate, blood glucose level, body mass index, and C-reactive protein amount, etc. can be mentioned.

[0045] In the graphs shown in both FIGS. 7 and 8, the importance of maxCPK is the highest, and among all the features, maxCPK can enhance the prediction accuracy of the prediction model the most.

[0046] Thus, the inventor of the present invention, who is the designer of the prediction model 121, has found that when predicting the prognosis of AMI patients, the prediction accuracy can be maintained high by using six or more types of features as the patient features used, the prediction accuracy can be maximized by using ten types of features as shown in FIG. 7, and the prediction accuracy can be maximized by using maxCPK among all the features. Note that Python code was used to generate the prediction model 121, but the prediction model 121 may be generated using other programming languages.

[0047] [Evaluation of Prediction Model] With reference to FIGS. 9 to 13, the evaluation of the prediction model 121 will be described. FIGS. 9 to 13 are diagrams showing the comparison results of the prediction models in this embodiment and the comparative examples. In the evaluation results shown in FIGS. 9 to 13, the prediction results when ten types of features as shown in FIG. 7 are input to the prediction model 121 generated by random forest as this embodiment are used. The prediction results by the conventional TIMI risk index are used as Comparative Example 1. The prediction results by the conventional GRACE score are used as Comparative Example 2.

[0048] Fig. 9 shows a graph of the ROC (Receiver Operating Characteristic) curve for each of this Example, Comparative Example 1, and Comparative Example 2, with the false positive rate on the horizontal axis and the sensitivity on the vertical axis. In the graph, the area of the portion below the ROC curve is referred to as the AUC (Area Under the Curve) and takes a value from 0 to 1. The closer the value of the AUC is to 1, the higher the prediction accuracy.

[0049] As shown in Fig. 9, for the prediction result using the prediction model 121 of this Example, the AUC is closer to 1 than the prediction results using each of Comparative Example 1 and Comparative Example 2.

[0050] Thus, according to the comparison using the ROC curve, it can be seen that the prediction model 121 of this Example has higher prediction accuracy than each of Comparative Example 1 and Comparative Example 2.

[0051] Fig. 10 shows the confusion matrix when predicting the prognosis using 511 sample data for each of this Example, Comparative Example 1, and Comparative Example 2. The confusion matrix is used when evaluating the prediction accuracy and includes true negative (TN:True Negative), false positive (FP:False Positive), false negative (FN:False Negative), and true positive (TP:True Positive).

[0052] True negative indicates the number of sample data labels correctly predicted as the negative label (e.g., survival) when the correct data is associated with the negative label. The higher the prediction accuracy, the larger the value. False positive indicates the number of sample data labels mispredicted as the positive label (e.g., in-hospital death) when the correct data is associated with the negative label. The higher the prediction accuracy, the smaller the value. False negative indicates the number of sample data labels mispredicted as the negative label (e.g., survival) when the correct data is associated with the positive label. The higher the prediction accuracy, the smaller the value. True positive indicates the number of sample data labels correctly predicted as the positive label (e.g., in-hospital death) when the correct data is associated with the positive label. The higher the prediction accuracy, the larger the value.

[0053] As shown in Figure 10, for the prediction results of this example, the value of true negative (TN) is larger than that of each of the prediction results of Comparative Example 1 and Comparative Example 2. For the prediction results of this example, the value of false positive (FP) is smaller than that of each of the prediction results of Comparative Example 1 and Comparative Example 2. For the prediction results of this example, the value of false negative (FN) is smaller than that of the prediction result of Comparative Example 2, and its value is the same as that of the prediction result of Comparative Example 1. For the prediction results of this example, the value of true positive (TP) is larger than that of the prediction result of Comparative Example 2, and its value is the same as that of the prediction result of Comparative Example 1.

[0054] Thus, according to the comparison using the confusion matrix, it can be seen that the prediction model 121 of this example has higher prediction accuracy than each of Comparative Example 1 and Comparative Example 2.

[0055] Figure 11(A) shows the results of comparing the sensitivities for each of the present example, Comparative Example 1, and Comparative Example 2. Figure 11(B) shows the results of comparing the specificities for each of the present example, Comparative Example 1, and Comparative Example 2. Figure 11(C) shows the results of comparing the precisions for each of the present example, Comparative Example 1, and Comparative Example 2. Figure 11(D) shows the results of comparing the F-values for each of the present example, Comparative Example 1, and Comparative Example 2. Figure 11(E) shows the results of comparing the accuracies for each of the present example, Comparative Example 1, and Comparative Example 2.

[0056] The sensitivity (recall rate) is calculated as TP / (TP + FN), and the larger the value, the higher the prediction accuracy. The specificity is calculated as TN / (TN + FP), and the larger the value, the higher the prediction accuracy. The precision is calculated as TP / (TP + FP), and the larger the value, the higher the prediction accuracy. The F-value is calculated as (2 * recall rate * precision) / (recall rate + precision), and the larger the value, the higher the prediction accuracy. The accuracy is calculated as (TP + TN) / (TN + FP + FN + FP), and the larger the value, the higher the prediction accuracy.

[0057] As shown in Figure 11, for each of the sensitivity, specificity, precision, F-value, and accuracy, the prediction results of the present example are larger than those of each of Comparative Example 1 and Comparative Example 2.

[0058] Thus, according to the comparison using the sensitivity, specificity, precision, F-value, and accuracy, it can be seen that the prediction model 121 of the present example has higher prediction accuracy than each of Comparative Example 1 and Comparative Example 2.

[0059] Figure 12 shows a graph of the PR (Precision Recall) curve with the recall rate on the horizontal axis and the precision on the vertical axis for each of the present example, Comparative Example 1, and Comparative Example 2. In the graph, the area of the part below the PR curve is referred to as AUPRC (Area Under the Precision Recall Curve) and takes a value from 0 to 1. The higher the prediction accuracy, the closer the value of AUPRC is to 1.

[0060] As shown in FIG. 12, the prediction result of this embodiment has an AUPRC closer to 1 than the prediction results of Comparative Example 1 and Comparative Example 2 respectively.

[0061] Thus, according to the comparison using the PR curve, it can be seen that the prediction model 121 of this embodiment has higher prediction accuracy than each of Comparative Example 1 and Comparative Example 2.

[0062] FIG. 13 summarizes the comparison results of this embodiment, Comparative Example 1, and Comparative Example 2. As shown in FIG. 13, the prediction model 121 of this embodiment has higher prediction accuracy than each of Comparative Example 1 and Comparative Example 2.

[0063] The TIMI risk index of Comparative Example 1 and the GRACE score of Comparative Example 2 have existed since before percutaneous coronary intervention became the main treatment method for AMI, so it is difficult to say that they conform to the current standard medical treatment. In contrast, the prediction model 121 of this embodiment is generated using sample data of AMI patients after percutaneous coronary intervention became the main treatment method for AMI, so it conforms to the current standard medical treatment.

[0064] The prediction model 121 of this embodiment uses machine learning by AI (Artificial Intelligence), while the TIMI risk index of Comparative Example 1 and the GRACE score of Comparative Example 2 do not use AI.

[0065] The prediction model 121 of this embodiment can predict the prognosis of AMI patients with higher accuracy than the TIMI risk index of Comparative Example 1 and the GRACE score of Comparative Example 2 respectively by the user inputting the feature quantities of AMI patients into the prediction device 1, so it is excellent in convenience and practicality.

[0066] [Display of prediction results] With reference to FIG. 14, an example of the display of the prediction result by the prediction device 1 will be described. FIG. 14 is a diagram for explaining an example of the display of the prediction result by the prediction device 1 according to the first embodiment.

[0067] As shown in FIG. 14, the display device 20 of the prediction device 1 displays, on the display 25, an image 21 and an image 22 for the user to input the value of at least one feature amount, and an image 23 including the prediction result calculated based on the value of at least one feature amount.

[0068] The image 21 includes an icon for the user to input the result of the Killip classification. The image 22 includes icons for the user to input each of max CPK, age, creatinine concentration, amount of C-reactive protein, amount of hemoglobin, number of white blood cells, blood glucose level, systolic blood pressure, and heart rate. The image 23 includes information indicating, for example, the probability of in-hospital death as the prediction result.

[0069] The user can obtain the prediction result displayed by the image 23 by inputting at least one feature amount into the prediction device 1 via the images 21 and 22 using a keyboard 31, a mouse 32, etc.

[0070] Note that the prediction device 1 may display, on the display 25 of the display device 20, information indicating the probability of survival as the prediction result. Further, the prediction device 1 may present the prediction result by points calculated according to a predetermined criterion instead of probabilities.

[0071] [Processing of Prediction Device] With reference to FIG. 15, the processing of the prediction device 1 will be described. FIG. 15 is a flowchart of the prediction process executed by the prediction device 1 according to the first embodiment. The processing steps shown in FIG. 15 (hereinafter abbreviated as "S") are realized by the arithmetic unit 11 executing the prediction program 122.

[0072] As shown in FIG. 15, the prediction device 1 acquires at least one feature quantity through the input interface 13 (S1). The prediction device 1 predicts the prognosis of the patient based on the at least one acquired feature quantity and the learned prediction model 121 (S2). The prediction device 1 outputs the predicted prognosis result to the display device 20 through the output interface 14.

[0073] As described above, according to the prediction device 1 according to Embodiment 1, since the prognosis can be predicted based on the feature quantity of the patient and the learned prediction model 121, the prognosis of AMI patients can be predicted with higher accuracy. Therefore, the prediction device 1 can also be used for informed consent in the clinical field, and the user can improve the quality of medical treatment for AMI patients by using the prediction device 1.

[0074] <Prediction System According to Embodiment 2> Hereinafter, only the parts different from the prediction device 1 according to Embodiment 1 will be described for the prediction system 1000 according to Embodiment 2. FIG. 16 is a diagram showing the prediction system 1000 according to Embodiment 2.

[0075] As shown in FIG. 16, the prediction system 1000 includes a user terminal 200 and a server device 300. The server device 300 is configured to communicate with the user terminal 200 via the network 500.

[0076] The user terminal 200 is an information terminal that executes predetermined information processing, such as a desktop PC (personal computer), a laptop PC, a smartphone, a smartwatch, a wearable device, and a tablet PC.

[0077] Similar to the control device 10 according to the first embodiment, the server device 300 includes an acquisition unit 330, a prediction unit 310, and an output unit 340. The acquisition unit 330 and the output unit 340 can be configured by communication devices for communicating with the user terminal 200 via, for example, the network 500. The prediction unit 310 can be configured by an arithmetic device such as a CPU, similar to the control device 10 for example.

[0078] In the prediction system 1000 having such a configuration, the user inputs at least one feature amount of an AMI patient to the user terminal 200. The user terminal 200 outputs at least one feature amount to the server device 300 via the network 500. The server device 300 acquires at least one feature amount acquired from the user terminal 200, and predicts the prognosis of the patient based on the acquired at least one feature amount and the learned prediction model. The server device 300 outputs the prediction result to the user terminal 200 via the network 500. The user terminal 200 displays the prediction result acquired from the server device 300 on the display 250.

[0079] Thereby, the user can predict the prognosis of the patient using the server device 300 by inputting the feature amount of the AMI patient to the user terminal 200.

[0080] Note that the server device 300 may be configured to communicate not only with one user terminal 200 but also with a plurality of user terminals 200. For example, in the prediction system 1000, the server device 300 may exist in the form of cloud computing. In this case, the server device 300 may predict the prognosis of the patient based on at least one feature amount acquired from each of the plurality of user terminals 200, and output the prediction result to each of the plurality of user terminals 200. In this way, for example, it becomes possible to transmit the prediction result from the server device 300 of the core hospital to each user terminal 200 in response to the input from the user terminals 200 installed in each of a plurality of hospitals in the region.

[0081] The embodiments disclosed herein should be considered as illustrative in all respects and not restrictive. The scope of the present disclosure is indicated by the scope of the claims rather than the description of the above-described embodiments, and is intended to include all modifications within the meaning and scope equivalent to the scope of the claims.

Description of Reference Numerals

[0082] 1 Prediction device, 10 Control device, 11 Arithmetic device, 12 Storage device, 13 Input interface, 14 Output interface, 20 Display device, 21, 22, 23 Images, 25, 250 Displays, 31 Keyboard, 32 Mouse, 121 Prediction model, 122 Prediction program, 200 User terminal, 300 Server device, 310 Prediction unit, 330 Acquisition unit, 340 Output unit, 500 Network, 1000 Prediction system.

Claims

1. A prediction device for predicting the prognosis of a patient with acute myocardial infarction, comprising: an acquisition unit that acquires at least one feature amount of the patient; a prediction unit that predicts the prognosis based on the at least one feature amount using a learned prediction model; and an output unit that outputs a prediction result of the prognosis. The at least one feature amount is ten types of feature amounts including the maximum value of creatine phosphokinase in blood, the amount of hemoglobin in blood, heart rate, the concentration of creatinine in blood, systolic blood pressure, blood glucose level, age, the result of classification regarding the severity of cardiac dysfunction in the acute myocardial infarction, the number of white blood cells in blood, and the amount of C-reactive protein in blood. The prediction device.

2. The prediction device according to claim 1, wherein the prediction model is generated by machine learning based on the ten types of feature amounts and the presence or absence of in-hospital death of the patient.

3. The prediction device according to claim 2, wherein the machine learning algorithm includes random forest.

4. further comprising a display device, the output unit outputs the prediction result to the display device, and the display device displays an image for inputting the values of the ten types of feature amounts and an image including the prediction result. The prediction device according to any one of claims 1 to 3.

5. A prediction method for predicting the prognosis of a patient with acute myocardial infarction by a computer, comprising: a step of acquiring at least one feature amount of the patient; a step of predicting the prognosis based on the at least one feature amount using a learned prediction model; and a step of outputting a prediction result of the prognosis. The at least one feature amount is ten types of feature amounts including the maximum value of creatine phosphokinase in blood, the amount of hemoglobin in blood, heart rate, the concentration of creatinine in blood, systolic blood pressure, blood glucose level, age, the result of classification regarding the severity of cardiac dysfunction in the acute myocardial infarction, the number of white blood cells in blood, and the amount of C-reactive protein in blood. The prediction method.

6. A prediction program for predicting the prognosis of a patient with acute myocardial infarction, causing a computer to execute a step of acquiring at least one feature amount of the patient; execute a step of predicting the prognosis based on the at least one feature amount using a learned prediction model; and execute a step of outputting a prediction result of the prognosis. The prediction program, wherein the at least one feature quantity is ten types of feature quantities including the maximum value of creatine phosphokinase in blood, the amount of hemoglobin in blood, heart rate, the concentration of creatinine in blood, systolic blood pressure, blood glucose level, age, the result of classification regarding the severity of cardiac dysfunction in the acute myocardial infarction, the number of white blood cells in blood, and the amount of C-reactive protein in blood. **Claim 7** A prediction system for predicting the prognosis of a patient with acute myocardial infarction, comprising: a user terminal; and a server device configured to communicate with the user terminal, wherein the server device includes: an acquisition unit configured to acquire at least one feature quantity of the patient from the user terminal; a prediction unit configured to predict the prognosis based on the at least one feature quantity by using a learned prediction model; and an output unit configured to output the prediction result of the prognosis to the user terminal, wherein the at least one feature quantity is ten types of feature quantities including the maximum value of creatine phosphokinase in blood, the amount of hemoglobin in blood, heart rate, the concentration of creatinine in blood, systolic blood pressure, blood glucose level, age, the result of classification regarding the severity of cardiac dysfunction in the acute myocardial infarction, the number of white blood cells in blood, and the amount of C-reactive protein in blood.

Images