Predictive model generation method, information processing device, and predictive model generation program

Abstract

To enable generation of a highly accurate prediction model with a small amount of computation.SOLUTION: A prediction model generation method is provided, comprising accumulating multiple combinations 103 of parameters representing branching conditions and branching probabilities for one or more conditional branch parts in a workflow, and generating a model for predicting a branching probability corresponding to a parameter used when using a particular conditional branch part of the one or more conditional branch parts using the accumulated multiple combinations 103.SELECTED DRAWING: Figure 7

Description

Technical Field

[0001] The present invention relates to a method for generating a prediction model, an information processing apparatus, and a prediction model generation program.

Background Art

[0002] In the administration of countries, local governments, etc., measures for health services for residents may be formulated.

[0003] For example, there is a technology that vectorizes healthcare data and estimates the effect when a measure is implemented based on the learning result with the ratio of target achievement and the actual measured effect of measure information as the target variable.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] However, at the site of measure formulation, measure plans are made based on intuition, experience, and assumptions, and there is a risk that it may be difficult to achieve the target of measure indicators (KPIs).

[0006] In one aspect, it aims to generate a high-precision prediction model with a small amount of calculation.

Means for Solving the Problems

[0007] In one aspect, the prediction model generation method accumulates a plurality of combinations of parameters indicating branch conditions and branch probabilities for one or more conditional branch components in a workflow, and uses the accumulated plurality of combinations to generate a model that predicts the branch probability corresponding to the parameter used when a specific conditional branch component among the one or more conditional branch components is used. In the process of accumulating the multiple combinations, in addition to the branch probability and the parameter, the multiple combinations are also accumulated in association with path information indicating the path that passes through the multiple conditional branch components. In the process of calculating a first branch probability indicating an output value that satisfies the accepted first parameter when a first parameter is input at a specific conditional branch component, the first branch probability is calculated when the path to the specific conditional branch component satisfies the path information. [Effects of the Invention]

[0008] In one respect, it allows for the generation of highly accurate predictive models with less computational effort. [Brief explanation of the drawing]

[0009] [Figure 1] This diagram illustrates policy planning in related examples. [Figure 2] This figure illustrates a predictive model as an embodiment. [Figure 3] This diagram illustrates the propagation of influences between parts in a predictive model as an embodiment. [Figure 4] This is a diagram illustrating the service implementation components included in the predictive model as an embodiment. [Figure 5] This diagram illustrates the calculation of the overall effectiveness and cost of the measures in a predictive model as an embodiment of the system. [Figure 6] This figure illustrates the accumulated information used for learning each component in a predictive model as an embodiment. [Figure 7] This figure illustrates an example of the information storage process as an embodiment. [Figure 8] This figure illustrates the process of generating a regression model from conditional branching components included in a prediction model as an embodiment. [Figure 9] This figure illustrates the process of generating a regression model from service implementation components included in a predictive model as an embodiment. [Figure 10] This diagram illustrates the accumulated information used for learning each component in the predictive model, which is the first modified example. [Figure 11] This figure illustrates the storage process for stored information as a first modified example. [Figure 12] This is a flowchart illustrating the process of building a predictive model in the embodiment. [Figure 13] This is a flowchart illustrating the probability calculation process using the prediction model in the embodiment. [Figure 14]This flowchart illustrates the first example of the process for building a predictive model in the first modified example. [Figure 15] This flowchart illustrates the first example of the probability calculation process using the prediction model in the first modified example. [Figure 16] This flowchart illustrates a second example of the process for building a predictive model in the first modified example. [Figure 17] This flowchart illustrates a second example of the probability calculation process using the prediction model in the first modified example. [Figure 18] This is a block diagram schematically showing an example of the hardware configuration of the information processing device in the embodiment and each modified example. [Modes for carrying out the invention]

[0010] [A] Related examples In order to extend the healthy life expectancy of citizens, the national government is urging local governments to formulate data-driven health plans and implement health programs efficiently.

[0011] In terms of insurance programs, there are measures such as CKD (chronic kidney disease) follow-up programs that provide specialist consultations and health guidance to citizens who are identified as being at high risk of disease during health checkups.

[0012] Based on the numerical values ​​of the test results from the health checkup, a referral to a specialist may be made. For example, if urine protein is 2+ or higher, or if the eGFR is 60 ml / min / 1.73 m³, the patient may be referred for further examination. 2 If the result is less than [amount missing], examination by a specialist is recommended.

[0013] In formulating this policy, changing the numerical values ​​used as criteria for judging these test results may alter the impact of the policy.

[0014] We would like to relax the criteria for referring patients to specialists to increase the number of people who need to see a specialist, but if we relax the criteria too much and the number of people who need to see a specialist becomes too large, there is a risk that specialists will be unable to cope.

[0015] The targeted measures are represented, for example, in the form of a workflow consisting of multiple conditional branches and multiple service implementations. Based on the status data of the individual targeted by the measures, the multiple conditional branches are followed in order to determine "which service to assign".

[0016] Evaluation indicators for policies may include effectiveness, cost, and the number of people who receive the service. For example, in CKD follow-up policies, effectiveness could include how much the number of new chronic kidney disease cases (= number of citizens × incidence rate) has decreased; cost could include the amount borne by local governments for conducting health checkups and providing treatment at individual doctors; and the number of people who receive the service could include the number of people receiving health checkups and the services provided by individual doctors.

[0017] Figure 1 is a diagram illustrating policy planning in related examples.

[0018] There is a technology that predicts the KPIs of policy proposals from integrated individual data of various types, thereby supporting decision-making in policy formulation.

[0019] In Figure 1, the policy candidate setting indicated by symbol A1 is set as a policy target to reduce the "adult obesity rate" in the target area to 25%, and the policy candidate is set as having internists provide exercise and nutrition guidance to young residents of the area, with the adult obesity rate in the area set as the KPI.

[0020] In code A2, policy evaluations are performed on the selected policy candidates.

[0021] In code A3, policy KPI predictions are performed. For example, if a candidate policy is implemented, the probability (categorized into high, medium, and low) of each person in the target area becoming obese within three years is predicted.

[0022] In code A4, integrated data for predicting policy KPIs is referenced. This integrated data may include, for example, a history of daily health information (such as steps taken), a history of diagnostic results, a history of lifestyle habits, and a history of test results for each individual resident of the region.

[0023] In code A5, policy evaluation is made visible.

[0024] This allows for the calculation of the probability of each individual becoming obese, and the obesity rate within the region to be calculated.

[0025] However, there is a risk that the calculation time required to predict the effects of changing policies will be long. Since the probability of each individual becoming obese is calculated, predictions using data from hundreds of thousands of citizens may take several minutes to tens of minutes to be immediately available.

[0026] Policymakers sometimes want to experiment with different changes to a particular part of a policy during the policy revision phase to determine the most useful revisions. They also want to shorten calculation time and quickly iterate through the revision cycle.

[0027] [B] Embodiment An embodiment will be described below with reference to the drawings. However, the embodiment shown below is merely illustrative, and there is no intention to exclude various modifications or applications of techniques not explicitly shown in the embodiment. In other words, this embodiment can be implemented in various ways without departing from its spirit. Furthermore, each figure is not intended to represent only the components shown in the figure, but may include other functions, etc. In the following, the same reference numerals in each figure have similar functions, and their explanation may be omitted.

[0028] Figure 2 illustrates a predictive model as an embodiment.

[0029] The initiative is modeled as a workflow composed of a combination of components such as conditional branching and service implementation. Then, using a model that has learned from accumulating information and parameters on how people flow based on past usage data for each conditional branching component, the number of people who will receive each service is predicted.

[0030] In the example shown in Figure 2, the input for code B1 is "number of people N=1000".

[0031] In code B2, part #1, which is service component A, is set to "health check".

[0032] In code B3, component #2, which is conditional branch component B, is set to "eGFR<α".

[0033] As shown in code B31, the conditional branching component learns the probability of going to each branch. α is set as the parameter for the conditional decision, the probability of going to the upper branch is p, and the probability of going to the lower branch is 1-p.

[0034] If the condition "eGFR < α" is not met (see NO route in code B3), then, as shown in code B4, it is determined that there is "no intervention" by a specialist for the citizen in question.

[0035] On the other hand, if "eGFR < α" is satisfied (see the YES route in code B3), then, as shown in code B5, component #3 as conditional branch component C is set to "HbA1c < β".

[0036] If the condition "HbA1c < β" is met (see the YES route in code B5), then, as shown in code B6, component #4 as conditional branch component D is set to "nephrologist," and it is determined that intervention by a "nephrologist" is necessary for the citizen in question.

[0037] On the other hand, if the condition "HbA1c < β" is not met (see NO route in code B5), it is determined that intervention by a "diabetes specialist" is necessary for the citizen, as shown in code B7.

[0038] In the example shown in Figure 2, the number of people flowing through parts #1, #2, #3, and #4 is predicted, as indicated by the dotted arrows.

[0039] Figure 3 illustrates the propagation of influences between parts in a predictive model as an embodiment.

[0040] Downstream components can be affected by upstream components. The downstream components are influenced by the parameter values ​​of each component used upstream.

[0041] In code C1, "health check" is set as component #1.

[0042] In code C2, "eGFR<60" is set as component #2.

[0043] If "eGFR<60" is satisfied (see YES route in code C2), then "HbA1c<6.5" is set as component #3 in code C3.

[0044] If the condition "HbA1c < 6.5" is met (see YES route in code C3), then in code C4, "nephrologist" is set as component #4, and it is determined that intervention by a "nephrologist" is necessary for the citizen in question.

[0045] On the other hand, if the condition "HbA1c < 6.5" is not met (see NO route in code C3), then in code C5, it is determined that intervention by a "diabetes specialist" is necessary for the citizen in question.

[0046] If the condition "eGFR<60" is not met in code C2 (see NO root of code C2), then "eGFR<80" is set as component #5 in code C6.

[0047] If the eGFR < 80 condition is met (see YES route in code C6), then in code C7, it is determined that "health guidance" is necessary for the citizen in question.

[0048] On the other hand, if the "eGFR < 80" condition is not met (see NO route in code C6), then code C8 determines that there is "no intervention" for the citizen in question.

[0049] As shown in symbol C61, the proportion of people flowing through the YES route and the NO route of part #5 is affected by the allocation of people in part #2.

[0050] As shown in symbol C62, changing the judgment threshold (60) of the parameter of part #2 changes the range of people who are eligible for the YES route of part #5.

[0051] Therefore, the number of people receiving each service may be predicted using a model that has been trained by accumulating information and parameters on the flow of people from past usage data for each component, as well as the combinations of components used at the time of use.

[0052] In other words, information on how people flow can be learned from past usage data. Since information on how people flow changes depending on the combination of parts and parameters used, it can be stored together and used to predict how people flow according to the combination of parts and parameters used at the time of use.

[0053] Figure 4 illustrates the service implementation components included in the predictive model as an embodiment.

[0054] The components indicated by reference numerals D1 to D7 in Figure 4 are the same as the components indicated by reference numerals B1 to B7 in Figure 2.

[0055] For each component, the impact (e.g., cost and effectiveness) can be predicted using a model that has been trained by accumulating information on the impact (e.g., cost and effectiveness) from past usage data, the combination of components used, and parameters.

[0056] As shown in symbol D21, the effect E when service is performed on one person for each serviced part. A and cost C A This should be learned.

[0057] Effect E A and cost C A Since it changes depending on the combination of parts and parameters used, it is stored together and adjusted according to the combination of parts and parameters used, Effect E A and cost C A It is acceptable to make it possible to predict this.

[0058] FIG. 5 is a diagram for explaining the calculation of the effect and cost of all measures in the prediction model as an embodiment.

[0059] The effect and cost per person are determined using the regression models of service implementation components #4, #5, and #6.

[0060] In the example shown in FIG. 5, at reference numeral E1, the number of people N is input.

[0061] At reference numeral E2, component #1 is set to "health check", and the total number of people flowing into component #1 and the effect are calculated. The effect per person is E A = f A_E Let the cost per person be C A = f A_C Then, when the number of people = N, the total effect = N × E A and the total cost = N × C A results.

[0062] At reference numeral E3, component #2 is set to "eGFR <α (= 50)", and the number of people N is input. The branch probability p flowing through the YES route of E3 when using the parameter α = 50 in the order of component #1 → component #2 B is calculated using the prediction model f B (α) learned in component #2, so p B = f B (α = 50), and the number of people = N × p B results.

[0063] When "eGFR <α" is not satisfied (see the NO route of E3), as shown at reference numeral E4, component #6 is set to "no intervention", and it is determined that there is "no intervention" by a specialist for the citizen in question.

[0064] On the other hand, when "eGFR <α" is satisfied (see the YES route of E3), as shown at reference numeral E5, component #3 is set to "HbA1c <β (= 6.5)". The branch probability p flowing through the YES route of E5 when using the parameters α = 50 and β = 6.5 in the order of component #1 → component #2 → component #3 CThis is the predictive model f trained in part #2. C Using (α,β), p C =f C (α=50, β=6.5) and the number of people = N × p B ×p C This is the result.

[0065] If the condition "HbA1c < β" is met (see the YES route in code E5), then, as shown in code E6, component #4 is set to "nephrologist," and it is determined that intervention by a "nephrologist" is necessary for the citizen in question. The total number of people and the effect of component #4 are also calculated. The effect per person is E A =f E_E The cost per person is C A =f E_C Therefore, the number of people is N p =N×p A ×p B Therefore, the sum of the effects = N p ×E p Therefore, the total cost = N p ×C D The same calculation is performed for parts #5 and #6, summing the number of people and their effects.

[0066] On the other hand, if the condition "HbA1c < β" is not met (see NO root of code E5), as shown in code E7, component #5 is set to "diabetes specialist," and it is determined that intervention by a "diabetes specialist" is necessary for the citizen in question.

[0067] The overall effect of the measure = the sum of the effects on people who go to part #1 + the sum of the effects on people who go to part #4 + the sum of the effects on people who go to part #5 + the sum of the effects on people who go to part #6. Also, the overall cost of the measure = the sum of the costs on people who go to part #1 + the sum of the costs on people who go to part #4 + the sum of the costs on people who go to part #5 + the sum of the costs on people who go to part #6.

[0068] Figure 6 is a diagram illustrating the stored information 103 accumulated for learning each component in a predictive model as an embodiment.

[0069] Codes F1 to F5 indicate components used in past CKD follow-up measures implemented in other cities in City B.

[0070] In code F1, component #1 is set to "health check".

[0071] In code F2, component #2 is set to "eGFR<α (=60)".

[0072] If "eGFR < α (=60)" is satisfied (see the YES route for code F2), then in code F3, component #3 is set to "HbA1c < β (=6.5)".

[0073] If the condition "HbA1c < β (=6.5)" is met (see the YES route in code F3), then in code F4, component #4 is set to "nephrologist," and it is determined that intervention by a "nephrologist" is necessary for the citizen in question.

[0074] On the other hand, if the condition "HbA1c < β (=6.5)" is not met (see NO root in code F4), then, as shown in code F5, it is determined that intervention by a "diabetes specialist" is necessary for the citizen in question.

[0075] As information for the execution of the policy flow, everyone's health check data 101 and treatment data 102 are stored. Then, accumulated information 103 is generated and stored from the information for the execution of the policy flow.

[0076] In the stored information 103, for each component #1 to #4, information on the flow of people and its impact (cost and effect), along with the parameters of each component, are registered in association.

[0077] Component #1 contains information on the flow of people and its impact, including the cost per person when implemented (C). A and disease risk reduction effect E A It will be registered.

[0078] Component #2 registers the probability pB of people flowing along the YES route as information about the flow of people and its impact, and registers α=60 as a parameter.

[0079] Component #3 registers the probability pC of people flowing along the YES route as information about the flow of people and its impact, and registers α=60 and β=6.5 as parameters.

[0080] Component #4 contains information on the flow of people and its impact, specifically the cost per person when implemented (C). A and disease risk reduction effect E A The data is registered, and parameters α=60 and β=6.5 are registered.

[0081] Figure 7 is a diagram illustrating the storage process of stored information 103 as an embodiment.

[0082] Code G1 shows past CKD follow-up measures in City B, and code G2 shows past CKD follow-up measures in City C.

[0083] Then, as shown in code G3, past performance data for part #2 is accumulated, for example. For example, the probability p of flowing through the YES route. B And the parameter α for conditional judgment is accumulated.

[0084] In the example shown in Figure 7, stored information 103 contains p B =0.6, α=60 was registered, and in City C, p B The values ​​=0.5 and α=50 are registered.

[0085] Figure 8 illustrates the process of generating a regression model from conditional branching components included in a prediction model as an embodiment.

[0086] Based on past usage data, the branching probability p B Regression model f B(α) is learned. For example, a one-dimensional Gaussian process regression may be used. The input variable is set to the parameter α of part #2 shown in Figure 7, and the target variable y is the branching probability p of part #2 shown in Figure 7. B It may be set to this.

[0087] Figure 9 illustrates the process of generating a regression model from service implementation components included in a prediction model as an embodiment.

[0088] As the information aggregated in part #4 shown in Figure 7, aggregated information 104 shown in Figure 9 is generated. In the example shown in Figure 9, in City B, cost C D =100, Effect E D =0.2, the path from the start to part #4 is "Start → #1 → #2 → #3 → #4", and the parameters for each part α=60, β=6.5 are registered. Also, in City C, cost C D =100, Effect E D =0.3, the path from the start to part #4 is "Start → #1 → #2 → #3 → #4", and the parameters for each part α=50 and β=8.0 are registered.

[0089] Based on past usage data, Effect E D Regression model f D_E (α,β) and cost C D Regression model f D_C (α,β) is learned. For example, a two-dimensional Gaussian process regression may be used. The input variables are set to the parameter α of part #2 and the parameter β of part #3 shown in Figure 7, and the target variable is the effect E D and cost C D It may be set to this.

[0090] Figure 10 illustrates the stored information 103a accumulated for learning each component in the prediction model as the first modified example.

[0091] Codes J1 to J8 indicate past CKD follow-up measures implemented in other cities in City B.

[0092] In code J1, component #1 is set to "health check".

[0093] In code J2, component #2 is set to "eGFR<α (=60)".

[0094] If "eGFR < α (=60)" is satisfied (see the YES route in code J2), then in code J3, component #3 is set to "HbA1c < β (=6.5)".

[0095] If the condition "HbA1c < β (=6.5)" is met (see the YES route in code J3), then in code J4, component #4 is set to "nephrologist," and it is determined that intervention by a "nephrologist" is necessary for the citizen in question.

[0096] On the other hand, if the condition "HbA1c < β (=6.5)" is not met (see NO route in code J4), it is determined that intervention by a "diabetes specialist" is necessary for the citizen, as shown in code J5.

[0097] If the condition "eGFR < α (=60)" is not met in code J2 (see NO root of code J2), then in code J6, component #5 is set to "eGFR < α (=80)".

[0098] If the condition "eGFR < α (=80)" is met (see the YES route in code J6), then code J6 determines that "health guidance" is necessary for the citizen in question.

[0099] On the other hand, if the condition "eGFR < α (=80)" is not met (see NO route in code J6), then, as shown in code J8, it is determined that there is "no intervention" for the citizen in question.

[0100] As information for the execution of the policy flow, everyone's health check data 101 and treatment data 102 are stored. Then, from the information for the execution of the policy flow, accumulated information 103a is generated and stored.

[0101] The stored information 103a registers, for each component #1 to #4, information on the flow of people and its impact (cost and effect), the combination of components (path from the start to the component), and the parameters of each component, all associated with each other.

[0102] Component #1 contains information on the flow of people and its impact, specifically the cost per person when implemented (C). A and disease risk reduction effect E A The combination of parts "Start → #1" is registered.

[0103] Component #2 contains information about the flow of people and its impact, including the probability pE of people flowing along the YES route, the parameter α=60, and the component combination "Start → #1 → #2".

[0104] Component #3 contains information about the flow of people and its impact, including the probability pC of flowing along the YES route, the parameters α=60 and β=6.5, and the component combination "Start → #1 → #2 → #3".

[0105] Component #4 contains information on the flow of people and its impact, specifically the cost per person when implemented (C). A and disease risk reduction effect E A The following is registered, with α=60 and β=6.5 registered as parameters, and the component combination "Start → #1 → #2 → #3 → #4" is registered.

[0106] Figure 11 is a diagram illustrating the storage process of stored information 103a as a first modified example.

[0107] Code K1 shows past CKD follow-up measures in City B, and code K2 shows past CKD follow-up measures in City C.

[0108] Then, as shown by code K3, for example, past performance data for part #2 is accumulated. For example, the probability p of flowing through the YES route. B The path from the start to part #2 (in other words, path information) and the parameter α for conditional judgment are accumulated.

[0109] In the example shown in Figure 11, the stored information 103 contains p B =0.6, route "start → #1 → #2", α=60 is registered, and in City C p B =0.5, route "Start → #1 → #2", α=50 is registered.

[0110] Based on past experience when the route "Start → #1 → #2" was used in that order, the branching probability p B Regression model f B (α) (see Figure 8) is learned. For example, a one-dimensional Gaussian process regression may be used. The input variable is set to the parameter α of part #2 shown in Figure 11, and the target variable is the branching probability p of part #2 shown in Figure 11. B It may be set to this.

[0111] Based on past usage of the route "Start → #1 → #2 → #3 → #4" (see aggregated information 104 in Figure 9), Effect E P Regression model f D_E (α,β) and cost C D Regression model f D_C (α,β) is learned. For example, a two-dimensional Gaussian process regression may be used. The input variables are set to the parameter α of part #2 and the parameter β of part #3 shown in Figure 11, and the target variable is the effect E D and cost C D It may be set to this.

[0112] Using the regression model shown in Figure 8, the branch probability and effect values ​​can be obtained with confidence intervals. The confidence interval is μ * -σ * <y<μ * +σ * This corresponds to the interval. Using the confidence interval information for the branch probability and effect values ​​of each component, the overall cost and effect of the measure may also be predicted with confidence intervals.

[0113] The process of building the prediction model in the embodiment will be explained according to the flowchart (steps S1, S2) shown in Figure 12.

[0114] The component is incorporated into the policy flow, and parameters and the probability of people flowing to the branch destination when used in other municipalities are collected and stored (Step S1).

[0115] Using the accumulated data, the input variables are defined as parameters, and the target variable as the branching probability. The branching probability is then calculated from the parameters, and the regression model f is trained (step S2). The process of building the prediction model is then completed.

[0116] Next, the process of calculating probabilities using the prediction model in the embodiment will be explained according to the flowchart (steps S11 to S17) shown in Figure 13.

[0117] For each service component of each flow, the following steps S12 to S17 are repeatedly executed (step S11).

[0118] The route to the service part is determined (Step S12).

[0119] The number of people z is set to the total number of people N (step S13).

[0120] For each component along the path, the following steps S15 to S17 are repeatedly executed (step S14).

[0121] It is determined whether the component is a conditional branching component (step S15).

[0122] If the component is not a conditional branch component (see NO route in step S15), the repetition of steps S15 to S17 (see step S14) is skipped, and the next repetition process is performed.

[0123] On the other hand, if the component is a conditional branching component (see the YES route in step S15), the learned regression model f and the parameters used in the flow are used to calculate the branching probability p to that path (step S16).

[0124] The number of people z is updated and set to number of people z = number of people z × p (step S17). Then, once the repetition of steps S15 to S17 (see step S14) and steps S12 to S17 (see step S11) is complete, the probability calculation process using the prediction model is finished.

[0125] Next, a first example of the process for building the prediction model in the first modified example will be explained according to the flowchart (steps S21-S24) shown in Figure 14.

[0126] The components are incorporated into the policy flow, and parameters from when they were used in other municipalities, the probability of people going to the branch destination, and the combination of components used are collected and stored (Step S21).

[0127] For each combination of parts used, the following steps S23 and S24 are repeatedly executed (step S22).

[0128] The parameters and the probability of people moving are extracted only for the corresponding combination pattern (step S23).

[0129] Using the accumulated data, the input variables are set as parameters and the target variable as the branching probability. The branching probability is then calculated from the parameters, and the regression model f is trained (step S24). Once steps S23 to S24 are repeated (see step S22), the first example of the prediction model construction process in the first modified example is complete.

[0130] Next, a first example of the probability calculation process using the prediction model in the first modified example will be explained according to the flowchart (steps S31-S38) shown in Figure 15.

[0131] For each service component of each flow, the following steps S32 to S38 are repeatedly executed (step S31).

[0132] The route to the service part is determined (step S32).

[0133] The number of people z is set to represent the total number of people N (step S33).

[0134] For each component along the path, the following steps S35 to S37 are repeatedly executed (step S34).

[0135] It is determined whether the component is a conditional branching component (step S35).

[0136] If the component is not a conditional branch component (see NO route in step S35), the process proceeds to step S38 once the repetition of steps S35 to S37 (see step S34) is complete.

[0137] On the other hand, if a component is a conditional branching component (see the YES route in step S35), the branching probability p to that path is calculated using the learned regression model f that matches the combination of components in the flow and the parameters used in the flow (step S36).

[0138] The number of people z is updated, and the formula number of people z = number of people z × p is set (step S37). Once steps S35 to S37 are repeated (see step S34), the process proceeds to step S38.

[0139] The number of people passing through the parts, known as number z, is set (step S38). Once steps S32 to S38 are repeated (see step S31), the first example of the probability calculation process using the prediction model in the first modified example is complete.

[0140] Next, a second example of the process for building the prediction model in the first modified example will be explained according to the flowchart (steps S41-S44) shown in Figure 16.

[0141] The components are incorporated into the policy flow, and the parameters, costs, effects, and combinations of components used when they were used by other municipalities are collected and stored (Step S41).

[0142] For each combination of parts used, the following steps S43 and S44 are repeatedly executed (step S42).

[0143] The parameters and cost-benefit analysis for only the matching combination patterns are extracted (step S43).

[0144] Using the accumulated data, with input variables = parameters and target variables = cost and effect, the cost and effect are determined from the parameters, and the regression model f is learned (step S44). Then, once steps S43 and S44 are repeated (see step S42), the second example of the prediction model construction process in the first modified example is completed.

[0145] Next, a second example of the probability calculation process using the prediction model in the first modified example will be explained according to the flowchart (steps S51-S60) shown in Figure 17.

[0146] For each service component of each flow, the following steps S52 to S59 are repeatedly executed (step S51).

[0147] The route to the service part is determined (step S52).

[0148] The number of people z is set to the total number of people N (step S53).

[0149] For each component along the path, the following steps S55 to S57 are repeatedly executed (step S54).

[0150] It is determined whether the component is a conditional branching component (step S55).

[0151] If the component is not a conditional branch component (see NO route in step S55), the process proceeds to step S58 once the repetition of steps S55 to S57 (see step S54) is complete.

[0152] On the other hand, if a component is a conditional branching component (see the YES route in step S55), the branching probability p to that path is calculated using the learned regression model f that matches the combination of components in the flow and the parameters used in the flow (step S56).

[0153] The number of people z is updated, and the formula number of people z = number of people z × p is set (step S57). Once steps S55 to S57 are repeated (see step S54), the process proceeds to step S58.

[0154] Using the trained regression model f that matches the combination of parts in the flow, and the parameters used in the flow, the per capita effect and cost of that part are calculated (Step S58).

[0155] The sum of the effects and costs of the service-provided parts is calculated (step S59). The sum of the effects of the parts = number of people z × effect per person, and the sum of the costs of the parts = number of people z × cost per person. Once steps S52 to S57 are repeated (see step S51), the process proceeds to step S60.

[0156] The overall effect of the measure is set to equal the sum of the effects of all service components, and the overall cost of the measure is set to equal the sum of the costs of all service components (step S60). Then, the second example of the probability calculation process using the predictive model in the first modified example is completed.

[0157] Figure 18 is a schematic block diagram showing an example of the hardware configuration of the information processing device 1 in the embodiment and the first modified example.

[0158] As shown in Figure 18, the information processing device 1 comprises a CPU 11, a memory unit 12, a display control unit 13, a storage device 14, an input interface (IF) 15, an external recording medium processing unit 16, and a communication IF 17.

[0159] The memory unit 12 is an example of a storage unit, and exemplifies it as Read Only Memory (ROM) and Random Access Memory (RAM). The ROM of the memory unit 12 may contain programs such as a Basic Input / Output System (BIOS). The software programs in the memory unit 12 may be read and executed by the CPU 11 as appropriate. The RAM of the memory unit 12 may be used as temporary storage memory or working memory.

[0160] The display control unit 13 is connected to the display device 131 and controls the display device 131. The display device 131 is a liquid crystal display, an organic light-emitting diode (OLED) display, a cathode ray tube (CRT), an electronic paper display, etc., and displays various information to the operator, etc. The display device 131 may be combined with an input device, for example, a touch panel.

[0161] The storage device 14 is a storage device with high I / O performance, and may include, for example, Dynamic Random Access Memory (DRAM), SSD, Storage Class Memory (SCM), or HDD.

[0162] Input IF15 is connected to an input device such as a mouse 151 or a keyboard 152, and may control such an input device. The mouse 151 and keyboard 152 are examples of input devices, and the operator performs various input operations through these input devices.

[0163] The external recording medium processing unit 16 is configured to accommodate a recording medium 160. When the recording medium 160 is mounted, the external recording medium processing unit 16 is configured to read the information recorded on the recording medium 160. In this example, the recording medium 160 is portable. For example, the recording medium 160 may be a flexible disk, optical disk, magnetic disk, magneto-optical disk, or semiconductor memory.

[0164] Communication IF17 is an interface that enables communication with external devices.

[0165] The CPU 11 is an example of a processor (in other words, a computer), and is a processing unit that performs various controls and calculations. The CPU 11 realizes various functions by executing the Operating System (OS) and programs loaded into the memory unit 12. The CPU 11 may be a multiprocessor containing multiple CPUs, a multicore processor having multiple CPU cores, or a configuration having multiple multicore processors.

[0166] The device for controlling the overall operation of the information processing device 1 is not limited to the CPU 11, but may be, for example, one of the following: MPU, DSP, ASIC, PLD, or FPGA. Furthermore, the device for controlling the overall operation of the information processing device 1 may be a combination of two or more types of CPU, MPU, DSP, ASIC, PLD, and FPGA. Note that MPU is an abbreviation for Micro Processing Unit, DSP is an abbreviation for Digital Signal Processor, and ASIC is an abbreviation for Application Specific Integrated Circuit. Also, PLD is an abbreviation for Programmable Logic Device, and FPGA is an abbreviation for Field Programmable Gate Array.

[0167] [C] Effect The prediction model generation method, information processing device, and prediction model generation program described above can, for example, achieve the following effects.

[0168] CPU11 stores multiple combinations of parameters indicating branching conditions and branching probabilities for one or more conditional branching components in the workflow. Using the stored multiple combinations, CPU11 generates a model that predicts the branching probability corresponding to the parameters used when a particular conditional branching component is used among the one or more conditional branching components.

[0169] This allows for the generation of highly accurate predictive models with minimal computational effort. Specifically, because it's not necessary to predict each individual, it avoids the need for hundreds of thousands of calculations, and the number of people flowing into each service can be predicted with computational effort several times less than the number of components.

[0170] When a specific conditional branching component receives input for a first parameter, the CPU 11 calculates a first branch probability that indicates an output value that satisfies the received first parameter.

[0171] This makes it possible to efficiently and accurately predict the number of people using each service.

[0172] In the process of accumulating multiple combinations, the CPU 11 stores multiple combinations, associating them with path information indicating the path that passes through multiple conditional branching components, in addition to the branch probability and the parameters. In the process of calculating the first branch probability, the CPU 11 calculates the first branch probability when the path to a specific conditional branching component satisfies the path information.

[0173] This allows for predictions that take into account the influence of other conditional branching components upstream in the flow when the allocation of people based on the flow's conditional branching components is affected by those components.

[0174] In the process of accumulating multiple combinations, CPU 11 stores multiple combinations by associating branch probabilities and parameters with path information indicating a path that passes through one or more conditional branch components, and the impact values ​​when using one or more conditional branch components and one or more service execution components included in the path information. In the process of generating a model, CPU 11 generates a model that predicts the impact values ​​when using a specific conditional branch component, in addition to the branch probability, when the path to a specific conditional branch component satisfies the path information.

[0175] This eliminates the need to make predictions for each individual, thus avoiding the need for hundreds of thousands of calculations. The impact of a measure (e.g., cost and effectiveness) can be predicted with a computation amount that is only a few times the number of components.

[0176] [D] Other The disclosed technology is not limited to the embodiments described above and can be implemented in various modifications without departing from the spirit of this embodiment. Each configuration and process of this embodiment can be selected or combined as needed.

[0177] In the embodiments described above, a predictive model for the workflow in administrative health services was generated, but the embodiments are not limited to this. The embodiments described above may be used to generate predictive models for various workflows, such as tasks with conditional branching, tests, and questionnaires. In this case as well, the same effects and advantages as in the embodiments described above can be obtained.

[0178] [E] Note The following additional information is disclosed regarding the embodiments described above.

[0179] (Note 1) For one or more conditional branching components in the workflow, multiple combinations of parameters indicating the branching condition and branching probability are accumulated. Using the accumulated multiple combinations, a model is generated that predicts the branching probability corresponding to the parameters used when a specific conditional branching component is used among the one or more conditional branching components. A method for generating predictive models in which a computer performs the processing.

[0180] (Note 2) In the aforementioned specific conditional branching component, when a first parameter is input, a first branching probability is calculated that indicates an output value satisfying the received first parameter. The predictive model generation method described in Appendix 1, wherein the computer performs the processing.

[0181] (Note 3) In the process of accumulating the above-mentioned multiple combinations, in addition to the branch probability and the parameters, the multiple combinations are also accumulated in association with path information indicating the path that passes through the multiple conditional branch components. In the process of calculating the first branch probability, the first branch probability is calculated when the path to the specific conditional branch component satisfies the path information. The predictive model generation method described in Appendix 2, wherein the computer performs the processing.

[0182] (Note 4) In the process of accumulating the above-mentioned multiple combinations, in addition to the branch probability and the parameters, the above-mentioned multiple combinations are accumulated by associating them with path information indicating a path that passes through one or more conditional branch components and one or more service implementation components included in the path information, In the process of generating the aforementioned model, if the path to the specific conditional branch component satisfies the path information, a model is generated that predicts the influence value when using the specific conditional branch component, in addition to the branching probability. A method for generating a predictive model as described in any one of the appendices 1 to 3.

[0183] (Note 5) For one or more conditional branching components in the workflow, multiple combinations of parameters indicating the branching condition and branching probability are accumulated. Using the accumulated multiple combinations, a model is generated that predicts the branching probability corresponding to the parameters used when a specific conditional branching component is used among the one or more conditional branching components. An information processing device equipped with a processor.

[0184] (Note 6) The aforementioned processor, In the aforementioned specific conditional branching component, when a first parameter is input, a first branching probability is calculated that indicates an output value satisfying the received first parameter. The method for generating the predictive model described in Appendix 5.

[0185] (Note 7) The aforementioned processor, In the process of accumulating the above-mentioned multiple combinations, in addition to the branch probability and the parameters, the multiple combinations are also accumulated in association with path information indicating the path that passes through the multiple conditional branch components. In the process of calculating the first branch probability, the first branch probability is calculated when the path to the specific conditional branch component satisfies the path information. The information processing device described in Appendix 6.

[0186] (Note 8) The aforementioned processor, In the process of accumulating the above-mentioned multiple combinations, in addition to the branch probability and the parameters, the above-mentioned multiple combinations are accumulated by associating them with path information indicating a path that passes through one or more conditional branch components and one or more service implementation components included in the path information, In the process of generating the aforementioned model, if the path to the specific conditional branch component satisfies the path information, a model is generated that predicts the influence value when using the specific conditional branch component, in addition to the branching probability. An information processing device as described in any one of the appendices 5 to 7.

[0187] (Note 9) For one or more conditional branching components in the workflow, multiple combinations of parameters indicating the branching condition and branching probability are accumulated. Using the accumulated multiple combinations, a model is generated that predicts the branching probability corresponding to the parameters used when a specific conditional branching component is used among the one or more conditional branching components. A predictive model generation program that uses a computer to perform processing.

[0188] (Note 10) In the aforementioned specific conditional branching component, when a first parameter is input, a first branching probability is calculated that indicates an output value satisfying the received first parameter. A predictive model generation program, as described in Appendix 9, that causes a computer to perform the processing.

[0189] (Note 11) In the process of accumulating the above-mentioned multiple combinations, in addition to the branch probability and the parameters, the multiple combinations are also accumulated in association with path information indicating the path that passes through the multiple conditional branch components. In the process of calculating the first branch probability, the first branch probability is calculated when the path to the specific conditional branch component satisfies the path information. A predictive model generation program, as described in Appendix 10, that causes a computer to perform the processing.

[0190] (Note 12) In the process of accumulating the above-mentioned multiple combinations, in addition to the branch probability and the parameters, the above-mentioned multiple combinations are accumulated by associating them with path information indicating a path that passes through one or more conditional branch components and one or more service implementation components included in the path information, In the process of generating the aforementioned model, if the path to the specific conditional branch component satisfies the path information, a model is generated that predicts the influence value when using the specific conditional branch component, in addition to the branching probability. A predictive model generation program described in any one of the appendices 9 to 11, which causes a computer to perform the processing. [Explanation of Symbols]

[0191] 1: Information Processing Device 11: CPU 12: Memory section 13: Display Control Unit 14:Storage device 16: External recording medium processing unit 101: Health checkup data 102: Treatment data 103,103a: Stored information 103a: Stored information 104: Aggregated Information 131:Display device 151: Mouse 152: Keyboard 160: Recording media 15: Input IF 17: Communication Interface

Claims

1. For one or more conditional branching components in the workflow, multiple combinations of parameters indicating the branching condition and branching probability are accumulated. Using the accumulated multiple combinations, a model is generated that predicts the branching probability corresponding to the parameters used when a specific conditional branching component is used among the one or more conditional branching components. The computer performs the process, In the process of accumulating the above-mentioned multiple combinations, in addition to the branch probability and the parameters, the multiple combinations are also accumulated in association with path information indicating the path that passes through the multiple conditional branch components. In the process of calculating a first branch probability that indicates an output value satisfying the received first parameter when a first parameter is input to the aforementioned specific conditional branch component, the first branch probability is calculated when the path to the aforementioned specific conditional branch component satisfies the path information. Methods for generating predictive models.

2. In the process of accumulating the plurality of combinations, in addition to the branch probability and the parameters, the plurality of combinations are accumulated by associating route information indicating a path that passes through one or more conditional branch components with the influence values ​​when using one or more conditional branch components and one or more service implementation components included in the route information indicating a path that passes through one or more conditional branch components, In the process of generating the aforementioned model, if the path to the specific conditional branch component satisfies path information indicating a path that passes through one or more conditional branch components, a model is generated that predicts the influence value when using the specific conditional branch component, in addition to the branch probability. The method for generating a predictive model according to claim 1.

3. For one or more conditional branching components in the workflow, multiple combinations of parameters indicating the branching condition and branching probability are accumulated. Using the accumulated multiple combinations, a model is generated that predicts the branching probability corresponding to the parameters used when a specific conditional branching component is used among the one or more conditional branching components. Equipped with a processor, The aforementioned processor, In the process of accumulating the above-mentioned multiple combinations, in addition to the branch probability and the parameters, the multiple combinations are also accumulated in association with path information indicating the path that passes through the multiple conditional branch components. In the process of calculating a first branch probability that indicates an output value satisfying the received first parameter when a first parameter is input to the aforementioned specific conditional branch component, the first branch probability is calculated when the path to the aforementioned specific conditional branch component satisfies the path information. Information processing device.

4. For one or more conditional branching components in the workflow, multiple combinations of parameters indicating the branching condition and branching probability are accumulated. Using the accumulated multiple combinations, a model is generated that predicts the branching probability corresponding to the parameters used when a specific conditional branching component is used among the one or more conditional branching components. Let the computer perform the process, In the process of accumulating the above-mentioned multiple combinations, in addition to the branch probability and the parameters, the multiple combinations are also accumulated in association with path information indicating the path that passes through the multiple conditional branch components. In the process of calculating a first branch probability that indicates an output value satisfying the received first parameter when a first parameter is input to the aforementioned specific conditional branch component, the first branch probability is calculated when the path to the aforementioned specific conditional branch component satisfies the path information. The computer is made to perform the process. A predictive model generation program.

Images