Causal relationship inference device, causal relationship inference method, and causal relationship inference program

Abstract

To infer a value of an allocation variable with high accuracy.SOLUTION: In a causal relation inference device comprising a processor which executes a program, and a storage device which stores the program, the processor executes: first calculation processing for calculating an internal vector on the basis of feature quantity vectors of a plurality of samples, and a first learning parameter; second calculation processing for calculating a reallocation vector on the basis of the internal vector calculated by the first calculation processing, and a second learning parameter; and third calculation processing for calculating point-wise weighting vectors for each of the plurality of samples on the basis of the reallocation vector calculated by the second calculation processing, and a third learning parameter.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to a causal relationship inference device, a causal relationship inference method, and a causal relationship inference program for inferring causal relationships present in data. [Background technology]

[0002] To accurately estimate the effects of a treatment or drug, it is necessary to compare cases where the intervention is administered to the same patient with and without it. However, there are cases where it is not possible to conduct experiments in which the same patient is administered the drug repeatedly, such as in the case of anticancer drug treatments that have strong side effects. As such, there is a method called propensity score analysis, which is a convenient way to estimate the effects of an intervention by comparing a group of patients who have undergone an intervention with a group of patients who have not undergone an intervention. Propensity score analysis estimates the effects of an intervention by comparing similar patients in the intervention group and the non-intervention group.

[0003] A propensity score is the probability that a patient belongs to the intervention group. By estimating the effects of treatment or medication for patients with a similar tendency to belong to the intervention group, it is possible to estimate similar effects when intervening on the same patient. A statistical method called logistic regression is generally used to calculate propensity scores, but because its prediction accuracy is low, it is difficult to identify similar patients.

[0004] Deep learning (neural network) is one of the technologies that realizes AI (Artificial Intelligence). Deep learning can achieve high prediction accuracy. Each element of the feature vector, which is the input item used for prediction, undergoes a weighted product-sum operation with other elements each time it passes through multiple perceptrons. Therefore, in principle, it is difficult to know the importance of each element of the feature vector on its own. This is a fatal flaw when using deep learning in the medical field.

[0005] For example, if two patients have similar propensity scores, but the AI ​​cannot explain how it determined that they were similar, it will be extremely difficult for a doctor to determine whether the values ​​were the same by chance or whether the patients truly have similar characteristics.

[0006] Non-Patent Document 1 discloses a new method of learning linear regression or logistic regression so as to explain the classification results of machine learning methods such as deep learning that do not have a function for calculating the importance of features. Logistic regression is a machine learning model equivalent to perceptron, and is most widely used in all fields. For example, the logistic regression shown on page 119 of Non-Patent Document 1 below has a function for calculating the importance of features for the entire data sample.

[0007] Non-Patent Document 2 discloses a new explanation method that explains classifier predictions interpretably and faithfully by learning an interpretable model that locally changes the predictions. This explanation method explains a variety of models for text (e.g., random forests) and image classification (e.g., neural networks).

[0008] Non-Patent Document 3 discloses a 30-day risk prediction model for re-hospitalization in heart failure patients discharged from hospital. This risk prediction model uses a deep unified network, a new mesh-like network structure of deep learning designed to avoid overfitting.

[0009] Non-Patent Document 4 discloses a method for evaluating the effectiveness of vocational training programs. [Prior art documents] [Non-patent literature]

[0010] [Non-Patent Document 1] Friedman J, Trevor H, Robert T. The elements of statistical learning. second edition. New York: Springer series in statistics, 2001. [Non-Patent Document 2] Ribeiro, Marco Tulio, Sameer Singh, and Carlos Guestrin. "Why should I trust you?: Explaining the predictions of any classifier." Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, 1016. [Non-Patent Document 3] Golas, Sara Bersche, et al. "A machine learning model to predict the risk of 30-day readmissions in patients with heart failure: a retrospective analysis of electronic medical records data." BMC medical informatics and decision making 18.1 (1018): 44. [Non-Patent Document 4] Dehejia, Rajeev H., and Sadek Wahba. "Causal effects in nonexperimental studies: Reevaluating the evaluation of training programs." Journal of the American statistical Association 94.448 (1999): 1053-1062. Summary of the Invention [Problem to be solved by the invention]

[0011] The method of Non-Patent Document 2 merely attempts to provide a retroactive explanation using linear regression, and even when attempting to explain normal fully connected deep learning, there is no mathematical guarantee that the importance of features used by deep learning when making predictions can be completely calculated. If completely linear regression could achieve the same prediction accuracy as deep learning, the initial deep learning itself would no longer be necessary. Therefore, the method of Non-Patent Document 1 has a contradiction in its constructive concept.

[0012] The present invention has been made in view of the above, and has an object to infer values ​​of assigned variables with high accuracy. [Means for solving the problem]

[0013] A causal relationship inference device according to one aspect of the invention disclosed in the present application is a causal relationship inference device having a processor that executes a program and a storage device that stores the program, wherein the processor performs a first calculation process to calculate an internal vector based on feature vectors of a plurality of samples and a first learning parameter, a second calculation process to calculate a reallocation vector based on the internal vector calculated by the first calculation process and a second learning parameter, and a third calculation process to calculate a point-wise weight vector for each of the plurality of samples based on the reallocation vector calculated by the second calculation process and a third learning parameter; and repeatedly executing, for each sample, a fourth calculation process of calculating a predicted value for an allocation variable to be assigned to the sample based on the point-wise weight vector calculated by the third calculation process and the feature vector, and an update process of updating the first learning parameters, the second learning parameters, and the third learning parameters using the predicted value for the allocation variable calculated by the fourth calculation process and the allocation variables. In the first calculation process, the processor calculates the internal vector based on the feature vector and the first learning parameters after update by the update process. In the second calculation process, the processor calculates the reallocation vector based on the internal vector and the second learning parameters after update by the update process. In the third calculation process, the processor calculates the point-wise weight vector based on the reallocation vector and the third learning parameters after update by the update process. It is characterized by: Effect of the Invention

[0014] According to the representative embodiment of the present invention, the value of the assigned variable can be inferred with high accuracy. Problems, configurations and effects other than those described above will become apparent from the following description of the embodiment. [Brief description of the drawings]

[0015] [Figure 1] FIG. 1 is a block diagram showing an example of the system configuration of a causal relationship inference system. [Diagram 2] FIG. 2 is an explanatory diagram showing an example of the structure of a neural network. [Diagram 3] FIG. 3 is a flowchart showing an example of a causal relationship inference processing procedure performed by the causal relationship inference device. [Figure 4] FIG. 4 is a flowchart showing a detailed example of a processing procedure for generating learning parameters (step S302). [Diagram 5] FIG. 5 is an explanatory diagram showing the experimental results. [Figure 6] FIG. 6 is an explanatory diagram showing the importance distribution for each feature amount obtained by using kernel density estimation in the experimental results. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS EXAMPLES

[0016] In the first embodiment, a causal relationship inference device is described as an example, which estimates the effect or efficacy of a drug and outputs the patient background and factors on which multiple patients are identified. The first embodiment enables doctors to accurately estimate the effect of a drug. This contributes to a quick recovery of patients and an improvement in the quality of medical care, and also leads to a reduction in the country's rapidly increasing medical expenses by suppressing the use of ineffective treatments.

[0017] <System configuration example> Fig. 1 is a block diagram showing an example of the system configuration of a causal relationship inference system. In Fig. 1, a server-client type causal relationship inference system 1 is taken as an example for explanation, but a standalone type may also be used. (A) is a block diagram showing an example of the hardware configuration of the causal relationship inference system 1, and (B) is a block diagram showing an example of the functional configuration of the causal relationship inference system 1. The same components in (A) and (B) are given the same reference numerals.

[0018] The causal relationship inference system 1 has a configuration in which a client terminal 100 and a causal relationship inference device 120, which is a server, are connected via a network 110 so as to be able to communicate with each other.

[0019] In (A), a client terminal 100 has a hard disk drive (HDD) 101 which is an auxiliary storage device, a memory 102 which is a main storage device, a processor 103, an input device 104 which is a keyboard and a mouse, and a monitor 105. A causal relationship inference device 120 has a hard disk drive (HDD) 121 which is an auxiliary storage device, a memory 122 which is a main storage device, a processor 123, an input device 124 which is a keyboard and a mouse, and a monitor 125. The main storage device, the auxiliary storage device, and a portable storage medium (not shown) are collectively referred to as a storage device. The storage device stores a neural network 200 shown in FIG. 2 and learning parameters 165 thereof.

[0020] In (B), the client terminal 100 has a client database (DB) 151. The client DB 151 is stored in a storage device such as the HDD 101 or the memory 102. The client DB 151 stores a propensity score 152 and an intervention effect 153.

[0021] The propensity score 152 is an assigned variable, for example, a predicted value indicating the presence or absence of an intervention, and is an example of a predicted value indicating a causal relationship. The intervention effect 153 is information indicating the effect of an intervention. An intervention refers to, for example, an act of controlling the presence or absence or the degree of factors that affect various events related to human health (including actions that lead to the maintenance and promotion of health, and medication, examinations, etc. for the prevention, diagnosis, or treatment of injuries and illnesses in medical care). The intervention effect 153 is data obtained from the intervention effect calculation unit 162 via the network 110. In addition, in the case of a server-client type, there are one or more client terminals 100.

[0022] The causal relationship inference device 120 includes a propensity score calculation unit 161, an intervention effect calculation unit 162, and a server database (DB) 163. The propensity score calculation unit 161 calculates the propensity score 152 and generates a learning parameter 165.

[0023] The intervention effect calculation unit 162 uses the learning parameters 165 to construct a neural network 200, and by providing test data to the neural network 200, executes a prediction process and outputs the intervention effect 153 to the client terminal 100. The propensity score calculation unit 161 and the intervention effect calculation unit 162 realize their functions by causing the processor 123 to execute a program stored in a storage device such as the HDD 121 or the memory 122. The server DB 163 stores the analysis data 164 and the learning parameters 165.

[0024] The causal relationship inference device 120 may be configured with a plurality of devices. For example, in order to distribute the load, a plurality of causal relationship inference devices 120 may exist. The causal relationship inference device 120 may be configured with a plurality of devices for each function. For example, the causal relationship inference device 120 may be configured with a first server including the propensity score calculation unit 161 and the server DB 163, and a second server including the intervention effect calculation unit 162 and the server DB 163. The causal relationship inference device may be configured with a first causal relationship inference device including the propensity score calculation unit 161 and the intervention effect calculation unit 162, and a second causal relationship inference device including the server DB 163. The causal relationship inference device may be configured with a first causal relationship inference device including the propensity score calculation unit 161, a second causal relationship inference device including the intervention effect calculation unit 162, and a third causal relationship inference device including the server DB 163.

[0025] <Example of neural network structure> 2 is an explanatory diagram showing an example of the structure of a neural network. In the neural network 200, the feature vector 201 is a D-dimensional (D is an integer of 1 or more) real-valued vector, and includes features that are factor values ​​such as medication information and test values ​​for age and gender. In the feature vector 201, features that are factor values ​​of D=3,512 dimensions are used, as in Non-Patent Document 3.

[0026] Moreover, the neural network 200 has internal vectors 202 to 204. The neural network 200 performs a product-sum operation 221 between the feature vector 201 and the learning parameter 211, and outputs the operation result to the internal vector 202.

[0027] Furthermore, the neural network 200 performs a product-sum operation 222 between the operation result of the internal vector 202 and the learning parameter 212 , and outputs the operation result to the internal vector 203 .

[0028] Furthermore, the neural network 200 performs a product-sum operation 223 between the operation result of the internal vector 203 and the learning parameter 213 , and outputs the operation result to the internal vector 204 .

[0029] Furthermore, the neural network 200 performs a product-sum operation 224 between the result of the operation of the internal vector 204 and the learning parameter 214 to calculate a reallocation vector 205 which is the result of the operation.

[0030] Furthermore, the neural network 200 calculates a point-wise weight vector 206, which is the result of a Hadamard operation 225 between the reallocation vector 205 and the learning parameters 215. The point-wise weight vector 206 is a weight vector in which the weight value differs for each patient, which is a sample.

[0031] In addition, the neural network 200 calculates a predetermined action, for example, an assignment variable Z indicating the presence or absence of an intervention, for each patient n by a product-sum operation 226 between the feature vector 201 and the point-wise weight vector 206. (n) A predicted value 207, i.e., a propensity score 152, for is calculated.

[0032] <Example of causal inference processing procedure> 3 is a flowchart showing an example of a causal relationship inference process procedure performed by the causal relationship inference device. In FIG. 3, steps S301 to S302 are processes performed by the propensity score calculation unit 161, and steps S303 to S307 are processes performed by the intervention effect calculation unit 162.

[0033] The causal reasoning device 120 reads known analysis data 164 (step S301). The analysis data 164 includes a feature vector 201 for each patient n and an assigned variable Z (n) and the intervention effect y for patient n (n) Here, n={1, . . . , N} is an index for specifying data of a certain patient, and in the first embodiment, N=30,000.

[0034] In order to facilitate understanding of the first embodiment, the feature vector 201 will be described using an example of three dimensions (D=3) of {age, sex, white blood cell count}. (n) is the target variable, i.e., the correct label, and takes two values, "1" or "0". Specifically, for example, Z (n) =1 means "intervention" and Z (n) =0 means "no intervention."

[0035] In addition, the intervention effect (n) is a treatment result when an intervention is performed, such as blood pressure after administration of a hypertensive drug. Note that the causal relationship inference device 120 in the first embodiment calculates, for example, an average treatment effect in each of a patient population with and without an intervention.

[0036] Next, the causal relationship reasoning device 120 generates the learning parameters 165 (step S302). Here, the details of the learning parameter generation (step S302) will be described with reference to FIG.

[0037] 4 is a flowchart showing a detailed example of a processing procedure for generating learning parameters (step S302). First, the causal relationship reasoning device 120 calculates an internal vector by the following formula (1) (step S401).

[0038]

number

[0039] Vector h in the above formula (1) l ,h l+1 are internal vectors 202, 203, and 204. l=1,...,L is an index indicating the number of layers of the internal neuron group. In the first embodiment, L=2.

[0040] W on the right side of the above formula (1) l is the learning parameter 165 (211, 212, 213). However, when l=1, the internal vector h l becomes the feature vector 201, and W1 becomes the learning parameter 211.

[0041] The σ on the right side of the above formula (1) is a sigmoid function. Inside the sigmoid function σ are product-sum operations 221, 222, 223, 224, and 226 between the learning parameter 165, which is a matrix, and an internal vector.

[0042] In addition, the bias parameter and the patient index n are removed from the above formula (1) in order to make it easier to understand the configuration of Example 1. The above formula (1) shows a fully connected neural network, but it is also possible to use the calculations of neural networks such as LSTM (Long Short-Term Memory) and CNN (Convolutional Neural Network).

[0043] Next, the causal relationship reasoning device 120 calculates the reallocation vector 205 by the following formula (2) (step S402).

[0044]

number

[0045] The left side of the above formula (2) is η(∈RD ) is the reallocation vector 205. The internal vector h L+1 is the internal vector 203(h L ) and learning parameters 213(W L ) is the result of a multiplication and addition operation.

[0046] Next, the causal relationship reasoning device 120 calculates the point-wise weight vector 206 by the following formula (3) (step S403).

[0047]

number

[0048] The vector ξ(∈R D ) is the point-wise weight vector 206. The vector w(∈R D ) is a learning parameter 215. The operator on the right side of the above equation (3) is a Hadamard product, which performs a Hadamard operation 225.

[0049] Next, the causal reasoning device 120 calculates the propensity score 152, which is a predicted value 207 indicating the presence or absence of an intervention, by the following formula (4) (step S404). (Step S403).

[0050]

number

[0051] The z (0≦z≦1) on the left side of the above formula (4) is the allocation variable Z (n) , that is, the propensity score 152. In addition, in the above formula (4), when solving the intervention effect 153 of multiple classes, a softmax function is used instead of the sigmoid function σ.

[0052] Next, the causal reasoning device 120 uses the statistical gradient method to calculate the assignment variable Z (n)(with or without intervention) and propensity score 152, z (n) (Assignment variable Z indicating whether or not an intervention was performed (n) Using the predicted value 207 for (a), the learning parameters 165 (211 to 215) are updated and optimized so as to minimize the cross entropy shown in the following formula (5) (step S405). The initial values ​​of the learning parameters 165 (211 to 215) are set by random numbers.

[0053]

number

[0054] The causal relationship reasoning device 120 stores the generated learning parameters 165 (211 to 215) in the server DB 163. This ends the generation of the learning parameters (step S302), and the process proceeds to step S303.

[0055] The learning parameters 165 updated by the learning parameter generation (step S302) are applied to the learning parameter generation (step S302) when new analysis data 164 is subsequently read (step S301).

[0056] Returning to FIG. 3, the causal reasoning device 120 performs patient stratification (step S303). Specifically, for example, the causal reasoning device 120 applies k-means clustering (where k>0) to the reallocation vector 205 (η). In the first embodiment, k=5 classes, and a patient group C (k) ={C (1) ,C (2) ,…,C (5)}, which is called patient stratification. (k) = {1, ..., 33, ..., 200, ...}. In the first embodiment, the L2 distance is used as the distance measure when performing k-means clustering, but the distance measure is not limited to the L2 distance.

[0057] In this way, by clustering using the reallocation vector 205, a patient group C with similar feature vectors 201 to which the learning parameters 165 are applied is classified into a group of patients. (k) can be generated.

[0058] In addition, the causal relationship inference device 120 uses, instead of the reallocation vector 205, a point-wise weight vector 206, the feature vector 201, the propensity score 152, and a patient-specific importance vector p (n) Alternatively, patient stratification may be performed by performing k-means clustering using either of the above.

[0059]

number

[0060] In this way, by clustering using the point-wise weight vector 206, a patient group C is clustered into a group of patients having similar point-wise weight vectors 206. (k) This allows the generation of patient population C (k) In the case of the relocation vector 205, the patients with similar pointwise weight vectors 206 have a higher propensity score for each patient 152 compared to the case of the relocation vector 205. (n) A predictive model that generates similar populations (Equation 4) will generate similar populations.

[0061] In addition, by clustering using the feature vector 201, a patient group C with similar feature vectors 201 is identified. (k) This allows the generation of patient population C (k) In this case, the original characteristics of patients having similar feature vectors 201 are more consistent than when the reallocation vector 205 is used.

[0062] In addition, by clustering using the propensity score 152, patient population C was divided into groups of patients with similar propensity scores 152. (k)This allows the generation of patient population C (k) Allocation variable Z indicating the presence or absence of intervention (n) Conceptually, patients with similar predicted values ​​207 for will be grouped together as patients with similar intervention probabilities compared to when the reallocation vector 205 is used.

[0063] Also, the importance vector p (n) By clustering using the importance vector p (n) Patient group C is a group of patients with similar (k) This allows the generation of patient population C (k) In the importance vector p (n) In comparison with the case where the reallocation vector 205 is used, patients having similar characteristics will be grouped together as patients having similar characteristics when calculating the propensity score 152.

[0064] Patient stratification (step S303) is an optional process, and after learning parameter generation (step S302), patient stratification (step S303) may be skipped and the process may proceed to importance calculation (step S304).

[0065] Next, the causal relationship inference device 120 calculates the importance (step S304). Specifically, for example, the causal relationship inference device 120 calculates the importance of the patient group C (k) Patient-specific importance vector p´ for each (m) Calculate the importance vector p´ (m) (k) is patient population C (k) The importance of each factor in the feature vector 201 of patient m in the above is included.

[0066]

number

[0067] N(k) on the right side of the above equation (7) is the patient population C (k) The number of patients with index m is the number of patients in the patient group C (k)In the patient-specific importance vector p´ (m) (k) is calculated. Note that for patient population C (k) Let p(k) be the average importance vector of k.

[0068] Next, the causal reasoning device 120 calculates the intervention effect 153 (step S305). Specifically, for example, the causal reasoning device 120 calculates the intervention effect 153 of the patient population C (k) The average intervention effect ATE(k) for each intervention is calculated as the intervention effect 153. (n) However, if the treatment outcome is an intervention, such as blood pressure after administration of a hypertensive drug, the average intervention effect ATE(k) is (k) This is the average blood pressure after taking hypertension medication for each period.

[0069]

number

[0070] Then, the causal relationship inference device 120 calculates the patient population C (k) Average intervention effect ATE(k) for all patients in C (1) ,C (2) ,…,C (k) The average value AATE is calculated.

[0071]

number

[0072] N on the right side of the above formula (9) k is the number of patients in class k.

[0073] Next, the causal reasoning device 120 calculates the propensity score 152{ξ (n) ,p (n) ,p´ (m) (k),z (n) ,C (k)} and the intervention effect 153 {ATE(k), AATE} are stored in the client DB 151 as the analysis results (step S306).

[0074] The causal relationship inference device 120 is (k) For each factor of the feature vector 201, z (n) (Assignment variable Z indicating whether or not an intervention was performed (n) Based on the predicted value for (k) In particular, the causal reasoning device 120 may classify a patient population C into an intervention group and a non-intervention group. (k) Within z (n) (Assignment variable Z indicating whether or not an intervention was performed (n) The patient n whose predicted value 207) for the factor is equal to or greater than a threshold may be classified as an intervention group for the factor, and the patient n whose predicted value 207 is less than the threshold may be classified as a non-intervention group for the factor. The causal reasoning device 120 may classify this into a patient group C in which a specific intervention effect 153 was obtained. (k) Then, the causal relationship reasoning device 120 may compare the intervention group and the non-intervention group with the same importance for each factor, for example, by using FIG.

[0075] Next, the causal relationship inference device 120 outputs the intervention effect 153 to the client terminal 100. The client terminal 100 displays the intervention effect 153 on the monitor 105.

[0076] Thus, according to the first embodiment, the causal reasoning device 120 calculates a point-wise weight vector 206(ξ) for each patient n. This allows the causal reasoning device 120 to calculate z (n) (Assignment variable Z indicating whether or not an intervention was performed (n) Therefore, the predicted value of z for patients in the intervention group can be estimated with high accuracy. (n) (prediction value of 207 indicating the presence or absence of intervention) and z (n) (Assignment variable Z indicating whether or not an intervention was performed (n) If the predicted value for is equivalent to (the values ​​are the same or the difference in values ​​is within an acceptable range), then it can be determined that the patients in the non-intervention group will be the same as the patients in the intervention group, for example, that the patients in the non-intervention group will be given the same intervention as the patients in the intervention group.

[0077] In particular, by stratifying patients, the patient population becomes multiple patient groups C (k) In order to converge to the above, the implementation of the same intervention to the patients in the non-intervention group as to the patients in the intervention group was performed for patient group C. (k) This is possible every time.

[0078] In addition, the assignment variable Z (n) and z (n) (Assignment variable Z indicating whether or not an intervention was performed (n) By learning from the predicted values ​​for causal relationships207), it is possible to improve the accuracy of inferring causal relationships.

[0079] In addition, the causal reasoning device 120 uses the point-wise weight vector 206 for each patient n to calculate the feature vector 201 for each patient n, which is the cause, and the intervention effect y (n) Patient-specific importance vector p´ indicating the causal relationship with (m) (k) can be inferred. This allows us to estimate the propensity score z (n) (Assignment variable Z indicating whether or not an intervention was performed (n) This makes it easier to explain the analytical assumptions behind the predicted values207) for

[0080] In addition, the causal reasoning device 120 calculates the feature vector 201, which is the cause, and the intervention effect y (n) and the assignment variable Z indicating whether or not there was an intervention. (n) and z (n) (Assignment variable Z indicating whether or not an intervention was performed (n) Using the predicted values ​​for (k) and (A) it is possible to predict outcomes in patient populations (intervention effects (ATE(k), AATE)) with high accuracy.

[0081] In other words, the intervention effect of the patients in the intervention group is highly accurate, so the z (n) (Assignment variable Z indicating whether or not an intervention was performed (n) If the same intervention is given to patients in the intervention group who have a value equivalent to the predicted value of β207), it can be estimated that the same intervention effect will be obtained as the intervention effect of patients in the intervention group. EXAMPLES

[0082] Next, a second embodiment will be described. The second embodiment shows another processing example of the intervention effect calculation (step S305). The intervention effect calculation (step S305) according to the second embodiment is performed for one patient group that is not clustered in the patient stratification (step S303) or a clustered patient group C (k) In each of the above, the intervention effect 153 is calculated by directly searching for patients to compare in the intervention group and the non-intervention group using the following formula (10). Here, the patient population C (k) This will be explained using an example.

[0083]

number

[0084] According to the above formula (10), the causal reasoning device 120 calculates the reallocation vector 205 (η (i) ) and the reallocation vector 205(η (j) ) is calculated for patient j with the smallest L2 distance. If the L2 distance is equal to or greater than a threshold, patient j is determined to be inapplicable. In the second embodiment, the threshold is set to 3 sigma or more of the average distance between reallocation vectors in different pairs. Note that a set of pairs consisting of tuples (i, j) of patient i and patient j is represented as pairs.

[0085] By calculating the intervention effect using the above formula (10) (step S305), it is possible to make it easier for doctors and analysts to intuitively understand the analysis procedure of the causal relationship inference device 120 compared to the first embodiment.

[0086] In addition, the causal relationship inference device 120 uses, instead of the reallocation vector 205 in the above formula (10), a point-wise weight vector 206, the feature vector 201, the propensity score 152, and the patient-specific importance vector p calculated by the above formula (6). (n)The intervention effect calculation (step S305) may be performed using any one of the above.

[0087] In this way, by executing the intervention effect calculation (step S305) using the point-wise weight vector 206, the intervention effect 153 can be obtained by placing emphasis on the point-wise weight vector 206, which is the weight for each patient. (k) In this way, it is easy to associate the point-wise weight vector 206 with the intervention effect 153.

[0088] In addition, by executing the intervention effect calculation (step S305) using the feature vector 201, the intervention effect 153 can be obtained by placing emphasis on the feature vector 201. (k) In this manner, it becomes easy to associate the feature vector 201 with the intervention effect 153.

[0089] In addition, by performing the intervention effect calculation (step S305) using the propensity score 152, the intervention effect 153 can be obtained by placing emphasis on the predicted value 207 regarding the allocation variable Z(n) indicating the presence or absence of an intervention. (k) In this way, it becomes easier to link propensity scores152 with intervention effects153.

[0090] Also, the importance vector p (n) By performing the intervention effect calculation (step S305) using (n) By focusing on the intervention effect 153, the user can obtain the intervention effect 153. (k) In the importance vector p (n) and intervention effects153.

[0091] In addition, the causal relationship inference device 120 uses pairs to calculate the patient population C (k) For each pair, the matched average intervention effect MATE(k) is calculated (step S305). Note that |pairs| on the right side of the following formula (11) is the number of elements in pairs.

[0092]

number

[0093] <Experiment> Here, an experimental example will be described. Non-Patent Document 4 is a document on a method for evaluating the effect of a vocational training program. In order to evaluate the performance of the causal reasoning device 120 according to the first embodiment, the data in Non-Patent Document 4 was used. To explain the relationship between the symbols in the first embodiment and the data in Non-Patent Document 4, the feature vector 201 is {age, education, Black (1 if black, 0 otherwise), Hispanic (1 if Hispanic, 0 otherwise), married (1 if married, 0 otherwise), nodegree (1 if no degree, 0 otherwise), RE75 (earnings in 1975)}, and the intervention effect y(n) is RE78 (earnings in 1978).

[0094] Also, whether or not there is intervention (n) is a treatment indicator (1 if treated (treatment), 0 if not treated (control)). The exact treatment effect depending on the presence or absence of intervention is calculated using the NSW data in Non-Patent Document 4. The data input to the causal relationship inference device 120 according to Example 1 is CPS-3 in Non-Patent Document 4.

[0095] A comparison was made between the case where the ATE was calculated after calculating the propensity score using a standard method called logistic regression and the case where the ATE output by the causal reasoning device 120 in Examples 1 and 2 (AATE in Example 1, MATE in Example 2) was calculated.

[0096] 5 is an explanatory diagram showing the experimental results. In the experimental result 500, the ATE value of $1,794 for NSW was set as the true value. The ATE value of logistic regression was "$110", the ATE (AATE) value of Proposed 1 (the causal relationship inference device 120 according to the first embodiment) was "$1,828", and the ATE (MATE) value of Proposed 2 (the causal relationship inference device 120 according to the second embodiment) was "$1,843".

[0097] Therefore, the causal reasoning device 120 according to the first and second embodiments allows doctors and analysts to estimate the intervention effect 153 better than when using the standard method (logistic regression).

[0098] 6 is an explanatory diagram showing the importance distribution for each factor of the feature amount using kernel density estimation in the experimental result 500. The importance distribution 601 is the importance vector p (n) The importance distribution 602 is a distribution of the importance vector p (n) is the distribution of values.

[0099] The importance distribution 603 is the importance vector p (n) The importance distribution 604 is a distribution of the importance vector p (n) The importance distribution 605 is the distribution of the importance vector p (n) is the distribution of values.

[0100] The importance distribution 606 is the importance vector p (n) The importance distribution 607 is the distribution of the importance vector p (n)The importance distribution 608 is the distribution of the importance vector p (n) is the distribution of values.

[0101] In the importance distributions 601 to 608, the horizontal axis represents each importance vector p (n) The vertical axis shows the value of , and the vertical axis shows the kernel density estimator (generally, frequency). "Treatment" shows the samples that received the intervention (intervention group), and "Control" shows the samples that did not receive the intervention (non-intervention group).

[0102] For example, according to the importance distribution 606 of not graduating from high school (nodegree), a value of "1" for not graduating from high school (nodegree) is understood to increase the probability of receiving treatment (intervention). The participants belonging to the control group did not actually receive the intervention, but if they had the opportunity to receive the intervention, sample n with a value of "1" for not graduating from high school (nodegree) is likely to have been a group that was likely to have received the intervention.

[0103] In this way, the causal reasoning device 120 according to the first and second embodiments can realize the propensity score analysis with high accuracy and with an easily explainable analysis process.

[0104] The causal relationship inference device 120 according to the first and second embodiments can also be configured as follows (1) to (13).

[0105] (1) In a causal relationship inference device 120 having a processor 123 for executing a program and a storage device (memory 122) for storing the program, the processor 123 performs a first calculation process (step S401) for calculating internal vectors 202 to 204 based on a feature vector 201 of a plurality of samples (e.g., patients) and first learning parameters 211 to 213, and calculates a reallocation vector 205 (η(n) ), and the reallocation vector 205 (η (n) ) and a third learned parameter 215 for each of the plurality of samples. (n) ) is calculated (step S403).

[0106] (2) In the causal relationship inference device 120 described above in (1), the processor 123 calculates the point-wise weight vector 206(ξ (n) ) and the feature vector 201, an assignment variable Z (n) Prediction value for (propensity score 152, z (n) ) and a fourth calculation process is executed.

[0107] (3) In the causal relationship inference device 120 described above in (2), the processor 123 calculates the assigned variable Z (n) Prediction value for (propensity score 152, z (n) ) and the assigned variable Z (n) and executes an update process (step S405) for updating the first learning parameters 211 to 213, the second learning parameters 214, and the third learning parameters 215 using the above. In the first calculation process, the processor 123 calculates the internal vectors 202 to 204 based on the feature vector 201 and the first learning parameters 211 to 213 after the update by the update process. In the second calculation process, the processor 123 calculates the reallocation vector 205 (η (n) ), and in the third calculation process, the processor 123 calculates the reallocation vector 205 (η (n)), and the third learning parameters 215 updated by the update process, the point-wise weight vector 206(ξ (n) ) is calculated.

[0108] (4) In the causal relationship inference device 120 of (2) above, the processor 123 calculates the assigned variable Z (n) Prediction value for (propensity score 152, z (n) ) and the assigned variable Z (n) and the assigned variable Z (n) The result obtained for the sample is y (n) and a fifth calculation process (step S305) for calculating the effect (intervention effect 153 {ATE(k), AATE}) obtained by the plurality of samples based on the results.

[0109] (5) In the causal relationship inference device 120 described above in (4), the processor 123 executes a clustering process (step S303: patient stratification) for clustering the plurality of samples, and in the fifth calculation process, the processor 123 clusters the plurality of sample groups C obtained by the clustering process. (1) ,C (2) ,…,C (k) For each of the assignment variables Z (n) Prediction value for (propensity score 152, z (n) ) and the assigned variable Z (n) and the assigned variable Z (n) The result obtained for the sample is y (n) Based on the sample population C (k) The effect gained by the intervention (intervention effect 153 {ATE(k), AATE}) is calculated.

[0110] (6) In the causal reasoning device 120 of (5) above, in the clustering process, the processor 123 calculates the reallocation vector 205 (η (n) ) the plurality of samples are clustered based on the

[0111] (7) In the causal relationship inference device 120 of (2) above, the processor 123 selects the assigned variable Z (n) a first sample having a value of the first assignment value (intervention), and (n) A pair of the first sample and the second sample is generated based on the distance between the first sample and the second sample, and a first result y obtained by the first sample at the first assigned value is calculated based on the distance between the first sample and the second sample. (i) and the second result y obtained by the second sample at the second assigned value. (j) and a sixth calculation process (step S305) for calculating the effect (intervention effect 153 {ATE(k), AATE}) obtained by the plurality of samples based on the results.

[0112] (8) In the causal reasoning device 120 described above in (7), the processor 123 executes a clustering process (step S303: patient stratification) for clustering the plurality of samples, and in the sixth calculation process, the processor 123 clusters the plurality of sample groups C obtained by the clustering process. (1) ,C (2) ,…,C (k) For each of the samples, a pair of the first sample and the second sample is generated, and based on the first result and the second result in each pair, the sample population C is (k) The effect gained by the intervention (intervention effect 153 {ATE(k), AATE}) is calculated.

[0113] (9) In the causal reasoning device 120 of (7) above, in the sixth calculation process, the processor 123 calculates the reallocation vector 205 (η (n) ) and the reallocation vector 205 (η (n) ) and generate the pair based on the distance between them.

[0114] (10) In the causal relationship inference device 120 of (1) above, the processor 123 calculates an importance vector p′ indicating the importance of the feature vector for each of the plurality of samples.(m) Then, a seventh calculation process (step S304, equation (7)) is executed to calculate (k).

[0115] (11) In the causal reasoning device 120 of (10) above, the processor 123 executes a clustering process (step S303: patient stratification) for clustering the plurality of samples, and in the seventh calculation process, the processor 123 clusters the plurality of sample groups C obtained by the clustering process. (1) ,C (2) ,…,C (k) For each of the importance vectors p' (m) Calculate (k).

[0116] (12) In the causal reasoning device 120 of (2) above, the assignment variable indicates an intervention in the sample.

[0117] (13) In the causal reasoning device 120 of (2) above, the effect obtained by the sample is a therapeutic effect.

[0118] The present invention is not limited to the above-described embodiments, and includes various modified examples and equivalent configurations within the spirit of the appended claims. For example, the above-described embodiments have been described in detail to clearly explain the present invention, and the present invention is not necessarily limited to those having all of the configurations described. Also, a part of the configuration of one embodiment may be replaced with a configuration of another embodiment. Also, a configuration of another embodiment may be added to a configuration of one embodiment. Also, a part of the configuration of each embodiment may be added, deleted, or replaced with another configuration.

[0119] In addition, each of the above-mentioned configurations, functions, processing units, processing means, etc. may be realized in hardware, for example by designing some or all of them as an integrated circuit, or may be realized in software by a processor interpreting and executing a program that realizes each function.

[0120] Information such as programs, tables, files, etc. that realize each function can be stored in a storage device such as a memory, a hard disk, or an SSD (Solid State Drive), or in a recording medium such as an IC (Integrated Circuit) card, an SD card, or a DVD (Digital Versatile Disc).

[0121] In addition, the control lines and information lines shown are those considered necessary for the explanation, and do not necessarily show all the control lines and information lines necessary for implementation. In reality, it can be considered that almost all components are connected to each other. [Explanation of symbols]

[0122] 202~204 Internal Vector 211-213 First learning parameter 120 Causal Inference Device 122 Memory 123 processors 152 Propensity Score 153 Intervention Effects 161 Propensity score calculation part 162 Intervention Effect Calculation Section 164 Analysis Data 165 Learning parameters 200 Neural Networks 201 Feature Vector 205 Reallocation Vector 206 pointwise weight vectors 207 Prediction 214 Second Learning Parameter 215 Third Learning Parameter Z (n) Allocation variables

Claims

1. A causal relationship inference device having a processor that executes a program and a storage device that stores the program, The processor, a first calculation process of calculating an internal vector based on feature vectors of a plurality of samples and a first learning parameter; a second calculation process for calculating a reallocation vector based on the internal vector calculated by the first calculation process and a second learning parameter; a third calculation process for calculating a point-wise weight vector for each of the plurality of samples based on the reallocation vector calculated by the second calculation process and a third learning parameter; a fourth calculation process for calculating, for each sample, a predicted value for an assignment variable to be assigned to the sample based on the point-wise weight vector calculated by the third calculation process and the feature vector; repeatedly performing an update process for updating the first learning parameter, the second learning parameter, and the third learning parameter by using the predicted value for the allocation variable calculated by the fourth calculation process and the allocation variable; In the first calculation process, the processor calculates the internal vector based on the feature vector and a first learning parameter updated by the update process; In the second calculation process, the processor calculates the reallocation vector based on the internal vector and a second learning parameter updated by the update process; In the third calculation process, the processor calculates the point-wise weight vector based on the reallocation vector and a third learning parameter updated by the update process. A causal relationship inference device comprising:

2. A causal relationship inference device having a processor that executes a program and a storage device that stores the program, The processor, a first calculation process of calculating an internal vector based on feature vectors of a plurality of samples and a first learning parameter; a second calculation process for calculating a reallocation vector based on the internal vector calculated by the first calculation process and a second learning parameter; a third calculation process for calculating a point-wise weight vector for each of the plurality of samples based on the reallocation vector calculated by the second calculation process and a third learning parameter; a fourth calculation process for calculating, for each sample, a predicted value for an assignment variable to be assigned to the sample based on the point-wise weight vector calculated by the third calculation process and the feature vector; a fifth calculation process for calculating, for each of the plurality of samples, effects obtained by the plurality of samples based on the predicted values ​​for the allocation variables calculated by the fourth calculation process, the allocation variables, and results obtained by the sample in the allocation variables; A causal relationship inference device that executes the above-mentioned procedure.

3. A causal relationship inference device as claimed in claim 2, A clustering process is performed to cluster the plurality of samples; In the fifth calculation process, the processor calculates an effect obtained by each of a plurality of sample populations obtained by the clustering process based on a predicted value for the allocation variable, the allocation variable, and a result obtained by the sample for the allocation variable. A causal relationship inference device comprising:

4. A causal relationship inference device as claimed in claim 3, In the clustering process, the processor clusters the plurality of samples based on the reallocation vector. A causal relationship inference device comprising:

5. A causal relationship inference device having a processor that executes a program and a storage device that stores the program, The processor, a first calculation process of calculating an internal vector based on feature vectors of a plurality of samples and a first learning parameter; a second calculation process for calculating a reallocation vector based on the internal vector calculated by the first calculation process and a second learning parameter; a third calculation process for calculating a point-wise weight vector for each of the plurality of samples based on the reallocation vector calculated by the second calculation process and a third learning parameter; a fourth calculation process for calculating, for each sample, a predicted value for an assignment variable to be assigned to the sample based on the point-wise weight vector calculated by the third calculation process and the feature vector; a sixth calculation process for generating pairs of the first and second samples based on a distance between a first sample of the plurality of samples, the value of which of the allocation variable is a first allocation value, and a second sample of the plurality of samples, the value of which of the allocation variable is a second allocation value, and for the pairs, calculating effects obtained by the plurality of samples based on a first result obtained by the first sample at the first allocation value and a second result obtained by the second sample at the second allocation value; A causal relationship inference device that executes the above-mentioned procedure.

6. A causal relationship inference device as claimed in claim 5, The processor, A clustering process is performed to cluster the plurality of samples; In the sixth calculation process, the processor generates a pair of the first sample and the second sample for each of a plurality of sample populations obtained by the clustering process, and calculates an effect obtained by the sample population based on the first result and the second result for each pair. A causal relationship inference device comprising:

7. A causal relationship inference device as claimed in claim 5, In the sixth calculation process, the processor generates the pair based on a distance between the reallocation vector of a first sample and the reallocation vector of a second sample. A causal relationship inference device comprising:

8. A causal relationship inference device having a processor that executes a program and a storage device that stores the program, The processor, a first calculation process of calculating an internal vector based on feature vectors of a plurality of samples and a first learning parameter; a second calculation process for calculating a reallocation vector based on the internal vector calculated by the first calculation process and a second learning parameter; a third calculation process for calculating a point-wise weight vector for each of the plurality of samples based on the reallocation vector calculated by the second calculation process and a third learning parameter; a seventh calculation process of calculating an importance vector indicating an importance of the feature quantity vector for each of the plurality of samples; A causal relationship inference device that executes the above-mentioned procedure.

9. A causal relationship inference device as claimed in claim 8, The processor, A clustering process is performed to cluster the plurality of samples; In the seventh calculation process, the processor calculates the importance vector for each of a plurality of sample groups obtained by the clustering process. A causal relationship inference device comprising:

10. A causal relationship inference device having a processor that executes a program and a storage device that stores the program, The processor, a first calculation process of calculating an internal vector based on feature vectors of a plurality of samples and a first learning parameter; a second calculation process for calculating a reallocation vector based on the internal vector calculated by the first calculation process and a second learning parameter; a third calculation process for calculating a point-wise weight vector for each of the plurality of samples based on the reallocation vector calculated by the second calculation process and a third learning parameter; a fourth calculation process for calculating, for each sample, a predicted value for an assignment variable to be assigned to the sample based on the point-wise weight vector calculated by the third calculation process and the feature vector; The assignment variable indicates an intervention in the sample. A causal relationship inference device comprising:

11. A causal relationship inference device as claimed in claim 2, The effect obtained by the sample is a therapeutic effect. A causal relationship inference device comprising:

12. A causal relationship inference method using a causal relationship inference device having a processor that executes a program and a storage device that stores the program, comprising: The processor, a first calculation process of calculating an internal vector based on feature vectors of a plurality of samples and a first learning parameter; a second calculation process for calculating a reallocation vector based on the internal vector calculated by the first calculation process and a second learning parameter; a third calculation process for calculating a point-wise weight vector for each of the plurality of samples based on the reallocation vector calculated by the second calculation process and a third learning parameter; a fourth calculation process for calculating, for each sample, a predicted value for an assignment variable to be assigned to the sample based on the point-wise weight vector calculated by the third calculation process and the feature vector; repeatedly performing an update process for updating the first learning parameter, the second learning parameter, and the third learning parameter by using the predicted value for the allocation variable calculated by the fourth calculation process and the allocation variable; In the first calculation process, the processor calculates the internal vector based on the feature vector and a first learning parameter updated by the update process; In the second calculation process, the processor calculates the reallocation vector based on the internal vector and a second learning parameter updated by the update process; In the third calculation process, the processor calculates the point-wise weight vector based on the reallocation vector and a third learning parameter updated by the update process. A causal reasoning method comprising the steps of:

13. A causal relationship inference method using a causal relationship inference apparatus having a processor that executes a program and a storage device that stores the program, comprising: The processor, a first calculation process of calculating an internal vector based on feature vectors of a plurality of samples and a first learning parameter; a second calculation process for calculating a reallocation vector based on the internal vector calculated by the first calculation process and a second learning parameter; a third calculation process for calculating a point-wise weight vector for each of the plurality of samples based on the reallocation vector calculated by the second calculation process and a third learning parameter; a fourth calculation process for calculating, for each sample, a predicted value for an assignment variable to be assigned to the sample based on the point-wise weight vector calculated by the third calculation process and the feature vector; a fifth calculation process for calculating, for each of the plurality of samples, effects obtained by the plurality of samples based on the predicted values ​​for the allocation variables calculated by the fourth calculation process, the allocation variables, and results obtained by the sample in the allocation variables; A causal reasoning method comprising the steps of:

14. A processor comprising: a first calculation process of calculating an internal vector based on feature vectors of a plurality of samples and a first learning parameter; a second calculation process for calculating a reallocation vector based on the internal vector calculated by the first calculation process and a second learning parameter; a third calculation process for calculating a point-wise weight vector for each of the plurality of samples based on the reallocation vector calculated by the second calculation process and a third learning parameter; a fourth calculation process for calculating, for each sample, a predicted value for an assignment variable to be assigned to the sample based on the point-wise weight vector calculated by the third calculation process and the feature vector; repeatedly executing an update process for updating the first learning parameter, the second learning parameter, and the third learning parameter by using the assignment variable and the predicted value for the assignment variable calculated by the fourth calculation process; In the first calculation process, the processor calculates the internal vector based on the feature vector and a first learning parameter updated by the update process; In the second calculation process, the processor is caused to calculate the reallocation vector based on the internal vector and a second learning parameter updated by the update process; In the third calculation process, the processor is caused to calculate the point-wise weight vector based on the reallocation vector and a third learning parameter updated by the update process. A causal relationship inference program comprising:

15. The processor: a first calculation process of calculating an internal vector based on feature vectors of a plurality of samples and a first learning parameter; a second calculation process for calculating a reallocation vector based on the internal vector calculated by the first calculation process and a second learning parameter; a third calculation process for calculating a point-wise weight vector for each of the plurality of samples based on the reallocation vector calculated by the second calculation process and a third learning parameter; a fourth calculation process for calculating, for each sample, a predicted value for an assignment variable to be assigned to the sample based on the point-wise weight vector calculated by the third calculation process and the feature vector; a fifth calculation process for calculating, for each of the plurality of samples, effects obtained by the plurality of samples based on the predicted values ​​for the allocation variables calculated by the fourth calculation process, the allocation variables, and results obtained by the sample in the allocation variables; A causal relationship inference program characterized by executing the above.

Images