Artificial image data generation device, prediction device, artificial image data generation method, prediction method, and program
Patent Information
- Application Number
- JP2023128031
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-08-04
- Publication Date
- 2025-02-17
Smart Images
Figure 2025023685000001_ABST
Abstract
Description
[Technical Field]
[0001] The present invention relates to an artificial image data generating device, a prediction device, an artificial image data generating method, a prediction method, and a program. [Background technology]
[0002] In recent years, various predictions have been made based on machine learning. For example, in epidemiological studies, disease predictions have been made using machine learning. In epidemiological studies, linear models have often been used as machine learning models (e.g., Non-Patent Document 1). When a linear model is used in epidemiological studies, a disease outcome is predicted based on the relationship between factors and the outcome (probability of developing a disease, or the presence or absence of a disease). Predictions based on these linear models are called generalized linear regressions. Generalized linear regressions include linear regression, logistic regression, and Cox regression. [Prior art documents] [Non-patent literature]
[0003] [Non-Patent Document 1] Methods in Molecular Biology, 2009, Vol. 471, pp. 179-195 Summary of the Invention [Problem to be solved by the invention]
[0004] However, prediction accuracy using generalized linear regression tends to be low. Possible reasons for this low accuracy include the difficulty in describing correlations between multiple variables, the fact that performance is easily impaired when variables are multicollinear, and the loss of information when simply selecting variables. There is a need to improve the prediction accuracy of predictive models using machine learning.
[0005] The present invention has been made in consideration of the above points, and provides an artificial image data generation device, a prediction device, an artificial image data generation method, a prediction method, and a program that can improve the prediction accuracy of a prediction model using machine learning. [Means for solving the problem]
[0006] The present invention has been made to solve the above-mentioned problems, and one aspect of the present invention is an artificial image data generation device that includes a target dataset acquisition unit that acquires a target dataset consisting of multiple types of data about a target, a pixel value data generation unit that generates pixel value data by converting each of the multiple types of data that make up the target dataset into pixel values, and an artificial image data generation unit that generates artificial image data by arranging the multiple pixel values that make up the pixel value data in a predetermined order.
[0007] In addition, one aspect of the present invention is that, in the above-mentioned artificial image data generation device, the order is an order, among the orders in which the multiple pixel values constituting the pixel value data are arranged, based on the accuracy of the prediction result by a learning model that outputs a prediction result for the target when the artificial image data is input.
[0008] Another aspect of the present invention is a prediction device comprising: an artificial image data acquisition unit that acquires artificial image data generated by converting each of multiple types of data about a prediction target into pixel values and arranging them in a predetermined order; and a prediction unit that makes a prediction about the prediction target from the artificial image data based on a trained model that is a training model that has been trained to output a prediction result about the training target when training artificial image data generated by converting each of the multiple types of data about the training target into pixel values and arranging them in the predetermined order is input, wherein the order is an order based on the accuracy of the prediction result by the training model, among the orders in which the pixel values converted from each of the multiple types of data about the training target are arranged.
[0009] In addition, one aspect of the present invention is that in the above-mentioned prediction device, the order is an order among the orders in which pixel values converted from the multiple types of data are arranged that results in the highest accuracy of the prediction results by the learning model.
[0010] Furthermore, one aspect of the present invention is that in the above-mentioned prediction device, the order is an order randomly selected from the orders in which the pixel values converted from the multiple types of data are arranged, which order results in the highest accuracy of the prediction results by the learning model.
[0011] Another aspect of the present invention is an artificial image data generation method including a target dataset acquisition step of acquiring a target dataset consisting of multiple types of data about a target, a pixel value data generation step of generating pixel value data by converting each of the multiple types of data that make up the target dataset into pixel values, and an artificial image data generation step of generating artificial image data by arranging the multiple pixel values that make up the pixel value data in a predetermined order.
[0012] Furthermore, one aspect of the present invention is a prediction method comprising: an artificial image data acquisition step of acquiring artificial image data generated by converting each of multiple types of data about a prediction target into pixel values and arranging them in a predetermined order; and a prediction step of making a prediction about the prediction target from the artificial image data based on a trained model, which is a training model trained to output a prediction result about the training target when training artificial image data generated by converting each of the multiple types of data about the training target into pixel values and arranging them in the predetermined order is input, wherein the order is an order based on the accuracy of the prediction result by the training model, among the orders in which the pixel values converted from each of the multiple types of data about the training target are arranged.
[0013] Another aspect of the present invention is a program for causing a computer to execute a target dataset acquisition step of acquiring a target dataset consisting of multiple types of data about a target, a pixel value data generation step of generating pixel value data by converting each of the multiple types of data that make up the target dataset into pixel values, and an artificial image data generation step of generating artificial image data by arranging the multiple pixel values that make up the pixel value data in a predetermined order.
[0014] Another aspect of the present invention is a program for causing a computer to execute an artificial image data acquisition step of acquiring artificial image data generated by converting each of multiple types of data about a prediction target into pixel values and arranging them in a predetermined order, and a prediction step of making a prediction about the prediction target from the artificial image data based on a trained model, which is a training model trained to output a prediction result about the training target when training artificial image data generated by converting each of the multiple types of data about the training target into pixel values and arranging them in the predetermined order is input, wherein the order is an order based on the accuracy of the prediction result by the training model, among the orders in which the pixel values converted from each of the multiple types of data about the training target are arranged. [Effects of the Invention]
[0015] According to the present invention, it is possible to improve the prediction accuracy of a prediction model using machine learning. [Brief explanation of the drawings]
[0016] [Figure 1] FIG. 1 is a diagram illustrating an overview of a method for generating an artificial image according to an embodiment of the present invention. [Figure 2] FIG. 10 is a diagram illustrating an example of image recognition of handwritten characters. [Figure 3] 1 is a diagram illustrating an example of a functional configuration of a prediction device 1 according to an embodiment of the present invention. [Figure 4] FIG. 2 is a diagram illustrating an example of the configuration of a learning model B2 according to an embodiment of the present invention. [Figure 5] FIG. 10 is a diagram showing an example of the flow of an order determination process according to an embodiment of the present invention. [Figure 6] FIG. 10 is a diagram illustrating an example of a flow of a prediction process according to an embodiment of the present invention. [Figure 7] FIG. 10 is a diagram illustrating an example of a frequency distribution of AUCs for pixel value order candidates according to the embodiment of the present invention. [Figure 8] FIG. 1 is a diagram illustrating an example of a comparison result of prediction performance between a conventional deep learning model and a linear model. DETAILED DESCRIPTION OF THE INVENTION
[0017] (Embodiment) Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings. Fig. 1 is a diagram showing an overview of a method for generating an artificial image according to this embodiment. The artificial image is used in the prediction process of a prediction device 1 according to this embodiment.
[0018] The target dataset 2 is composed of multiple types of data (also referred to as variables or features) about the prediction target. In this embodiment, the prediction target is, as an example, a medical checkup recipient. The prediction device 1 predicts the presence or absence of a disease for the medical checkup recipient. The target dataset 2 is one or more pieces of data collected through medical checkups, questionnaires, tests, etc. The target dataset 2 is, for example, data indicating features such as the age, sex, height, weight, obesity level, and lifestyle habits of the medical checkup recipient. The data constituting the target dataset 2 may be either numerical values or categorical variables. Note that categorical variables are converted into numerical values.
[0019] The variables (features) that make up the target dataset 2 are converted into numbers between 0 and 255 (integers equal to or greater than 0). The variables converted into these numbers are treated as pixel values of the pixels that make up the digital image. The pixel value converted feature 3 shown in Figure 1 is an example of a feature converted into a pixel value.
[0020] The variables constituting the target data set 2 are converted into pixel values ranging from 0 to 255 based on, for example, equation (1).
[0021]
number
[0022] In equation (1), X is the value of a certain feature. min is the minimum value of the feature contained in the target dataset 2. X max is the maximum value of the feature included in the target dataset 2. For example, if the feature is age and the feature values are "20 years old," "30 years old," "40 years old," "50 years old," and "60 years old," "30 years old" is converted into a pixel value of "64" based on equation (1). Note that a conversion other than that of equation (1) may be used to convert features into pixel values. Also, the range of numbers from 0 to 255 representing pixel values is an example, and other ranges may be used.
[0023] Next, we will explain the order in which pixels are arranged to generate an image from the features converted into pixel values. There are as many ways to arrange pixels as there are types of features. If the number of types of features is f, then there are f! (factorial of f) ways to arrange pixels.
[0024] In this embodiment, only square images are considered as images to be generated. Here, rotation (90, 180, or 270 degrees) and inversion (up and down, left and right, or diagonal) of the image are not considered. In other words, the image generated by these rotation and inversion operations is considered to be the same as the image before the operations. The number of orders for arranging pixels is the factorial of f divided by 8 (f! / 8).
[0025] The pixel array 4 shown in Figure 1 is an example of a case where feature quantities converted into pixel values are arranged in a square. Here, the number of types of feature quantities is set to 100. Therefore, the pixel array 4 is a square matrix with 10 rows and 10 columns. The artificial image 5 is an image generated based on the pixel array 4. The pixel array 4 corresponds to image data for generating the artificial image 5. Therefore, the pixel array 4 is an example of artificial image data.
[0026] By using the above-described method, as many artificial images as there are pixel arrangement orders (arrangement of the number of types of feature amounts) can be generated from one target data set 2. If an artificial image configured in a certain order most easily recognizes the presence or absence of a disease, that artificial image is the optimal artificial image for predicting the presence or absence of a disease.
[0027] Here, the reason why there is an optimal artificial image for predicting the presence or absence of a disease will be explained with reference to Fig. 2, taking image recognition of handwritten characters as an example. Fig. 2 is a diagram showing an example of image recognition of handwritten characters. In Fig. 2, pixel values are represented by shades of gray.
[0028] In image recognition, each pixel is treated as a variable. If the pixels in the image to be recognized are not arranged in a specific order, the visibility of the image will be impaired. If the visibility of the image is impaired, it will be difficult for the image recognition to correctly recognize the image.
[0029] For example, image 61 shown in FIG. 2 has the number "6" handwritten. Image 62 is an image generated by changing the order of pixels in image 61 for some of the pixels. Compared to image 61, image 62 has less legibility of the number. Image 63 is an image generated by further changing the order of pixels in image 62 for all of the pixels. Compared to image 62, image 63 has even less legibility of the number.
[0030] In epidemiological studies, it is also possible that an optimal order exists for the pixels in an image generated by converting variables into pixel values and arranging them. In this case, it is possible to generate an optimal artificial image by converting those variables into pixels. When machine learning is used to predict the presence or absence of disease using the optimal artificial image, the prediction accuracy may exceed that of conventional linear models.
[0031] 3 is a diagram showing an example of the functional configuration of the prediction device 1 according to this embodiment. The prediction device 1 includes a control unit 10 and a storage unit 11. The prediction device 1 is, for example, a personal computer (PC).
[0032] The control unit 10 performs various controls on the prediction device 1. The control unit 10 includes an artificial image data generation function unit 100, an artificial image data acquisition unit 105, a prediction unit 106, and an output unit 107. These functional units are realized, for example, by a CPU loading a program from a ROM (Read Only Memory) into a RAM (Random Access Memory) and executing processing in accordance with the program. The ROM and RAM are included in the storage unit 11.
[0033] The artificial image data generation function unit 100 includes a target data set acquisition unit 101 , a pixel value data generation unit 102 , an artificial image data generation unit 103 , and an order determination unit 104 .
[0034] The target dataset acquisition unit 101 acquires a target dataset. The target dataset is composed of multiple types of data about a target. The target dataset acquisition unit 101 acquires the target dataset from, for example, an external server. The target dataset may be pre-stored in the storage unit 11, and the target dataset acquisition unit 101 may acquire the target dataset from the storage unit 11.
[0035] The pixel value data generating unit 102 generates pixel value data by converting each of the multiple types of data constituting the target data set into pixel values.
[0036] The artificial image data generator 103 generates artificial image data by arranging a plurality of pixel values that make up the pixel value data in a predetermined order.
[0037] The order determination unit 104 determines a predetermined order. The predetermined order is an order based on the accuracy of the prediction result by the learning model B2, among the orders in which pixel values obtained by converting multiple types of data are arranged. The predetermined order is, for example, an order in which the accuracy of the prediction result by the learning model B2 is highest. When artificial image data is input, the learning model B2 outputs a prediction result for the prediction target. A specific example of the learning model B2 will be described later. The order determination unit 104 stores the determined order as array information A1 in the storage unit 11. Note that the order determination unit 104 stores the array information A1 in the storage unit 11 before the artificial image data generation unit 103 executes the process of generating artificial image data.
[0038] The artificial image data acquisition unit 105 acquires artificial image data. The artificial image data is image data generated by converting multiple types of data about the prediction target into pixel values and arranging the pixel values in a predetermined order.
[0039] The prediction unit 106 makes a prediction about a prediction target from artificial image data based on the trained model B1. The trained model B1 is a training model B2 that has been trained to output a prediction result about the training target when training artificial image data generated by converting multiple types of data about the training target into pixel values and arranging them in a predetermined order is input. The trained model B1 is stored in the storage unit 11 in advance.
[0040] The output unit 107 outputs the prediction result by the prediction unit 106. For example, the output unit 107 displays the prediction result on a display unit (not shown). Alternatively, the output unit 107 may output the prediction result to an external device (such as a server).
[0041] The storage unit 11 stores various types of information. For example, the storage unit 11 stores array information A1 and a trained model B1. The storage unit 11 is configured using a storage device such as a magnetic hard disk drive or a semiconductor storage device.
[0042] The prediction device 1 may be realized as a virtual server. The functional units of the prediction device 1 may be distributed across multiple servers. The prediction device 1 may be realized as a cloud server.
[0043] For example, the artificial image data generation function unit 100 may be provided in a device separate from the prediction device 1. In this case, the device is also referred to as an artificial image data generation device. The device may be a PC, a virtual server, multiple servers, or a cloud server.
[0044] Here, the learning model B2 will be described with reference to FIG. 4. FIG. 4 is a diagram showing an example of the configuration of the learning model B2 according to this embodiment. In this embodiment, the learning model B2 is, as an example, a deep learning model. A deep learning model is a multi-layered neural network. The trained model B1 includes an input layer, an intermediate layer (also referred to as a hidden layer), and an output layer.
[0045] In Figure 4, "X j (j=1, 2, . . ., i: i is the number of features) each indicate a feature input to the input layer. In this embodiment, the feature is each pixel value of the artificial image. i j (i=1, 2, . . ., n: n is the number of nodes in each hidden layer, j=1, 2, . . ., h: h is the number of layers that make up the hidden layer) indicates the nodes that make up the hidden layer. In learning model B2, for example, the number of layers that make up the hidden layer is five. j " (j=0, 1) respectively indicate the variables output from the output layer. The variables output from the output layer indicate the presence or absence of a disease. In the learning process, the variables output from the output layer indicate the response to the input. In the inference process, the variables output from the output layer indicate the prediction result.
[0046] Next, an order determination process for determining the order in which pixel values are arranged will be described with reference to Fig. 5. Fig. 5 is a diagram showing an example of the flow of the order determination process according to this embodiment. The order determination process is executed by the control unit 10. In the following description, the order in which pixel values are arranged will also be referred to as a pixel value order. A pixel value order that is a candidate before being determined will also be referred to as a pixel value order candidate.
[0047] Step S10: The order determination unit 104 starts a learning process and a test process for each pixel value order candidate. The learning process and the test process are the processes from step S20 to step S60 described below. Therefore, the order determination unit 104 repeatedly executes the processes from step S20 to step S60 for all pixel value order candidates. As described above, there are as many pixel value order candidates as there are ways to arrange the variables included in the target dataset.
[0048] Step S20: The order determination unit 104 causes the target dataset acquisition unit 101 to acquire the target dataset.
[0049] Step S30: The order determination unit 104 randomly divides the acquired target data set into training data, validation data, and test data. The order determination unit 104 randomly divides the target data set so that 70% of the target data set becomes training data, 10% becomes validation data, and 20% becomes test data.
[0050] Step S40: The order determination unit 104 corrects imbalances between categories in the target data set. For example, the order determination unit 104 increases (oversamples) data for categories with a small number of samples based on the Synthetic Minority Over-sampling Technique (SMOTE). That is, the order determination unit 104 increases the number of samples in categories with diseases based on SMOTE.
[0051] Step S50: The order determination unit 104 performs training of the training model B2 using the training data and the verification data. Here, the order determination unit 104 causes the pixel value data generation unit 102 to generate pixel value data by converting data used as feature quantities, among the multiple types of data constituting the training data, into pixel values. The order determination unit 104 causes the artificial image data generation unit 103 to generate training artificial image data by arranging the multiple pixel values constituting the pixel value data generated by the pixel value data generation unit 102 in the order of the pixel value order candidates. The training artificial image data is artificial image data used for training.
[0052] The order determination unit 104 pairs the artificial image data for learning with data used as labels and uses the paired data as teacher data (training data). The order determination unit 104 generates artificial image data for the verification data in the same way as for the training data. The order determination unit 104 pairs the artificial image data generated from the verification data with data used as labels, and uses the paired data to verify the learning efficiency of the learning model B2.
[0053] The order determination unit 104 performs learning of the learning model B2 using the training data. The learning efficiency is verified using the validation data. The order determination unit 104 terminates the learning when the value of the loss function when the validation data is used no longer changes.
[0054] Step S60: The order determination unit 104 uses test data to test the learned learning model B2 (trained model). Here, the order determination unit 104 causes the pixel value data generation unit 102 to generate pixel value data by converting data used as feature quantities, among the multiple types of data constituting the test data, into pixel values. The order determination unit 104 causes the artificial image data generation unit 103 to generate artificial image data by arranging the multiple pixel values constituting the pixel value data generated by the pixel value data generation unit 102 in the order of the pixel value order candidates.
[0055] The order determination unit 104 inputs the artificial image data generated by the artificial image data generation unit 103 to the learning model B2 that has undergone training. The order determination unit 104 acquires the output from the learning model B2 that has undergone training. The order determination unit 104 evaluates the prediction accuracy of the learning model B2 for each piece of training data based on the output from the learning model B2 and the label.
[0056] Step S70: The order determination unit 104 ends the learning process and the test process.
[0057] Step S80: The order determination unit 104 determines the order of pixel value order candidates that maximizes the prediction accuracy of the learning model B2. Therefore, in this embodiment, the predetermined order for arranging the multiple pixel values that constitute the pixel value data is the order for arranging the pixel values converted from multiple types of data that maximizes the prediction accuracy of the learning model B2.
[0058] The order determination unit 104 determines the pixel value order candidate that is determined to have the highest AUC as the pixel value order, and stores the determined pixel value order in the storage unit 11 as array information A1. With this, the control unit 10 ends the order determination process.
[0059] In this embodiment, the predetermined order in which the multiple pixel values constituting the pixel value data are arranged is described as an example of an order in which the accuracy of the prediction results by learning model B2 is highest among the orders in which pixel values converted from multiple types of data are arranged, but this is not limited to this. The predetermined order for arranging the pixel values constituting the pixel value data may be any order based on the accuracy of the prediction result by the learning model B2 that outputs a prediction result for the prediction target when artificial image data is input, among the orders for arranging pixel values obtained by converting each of the multiple types of data. For example, the predetermined order may be any order in which the accuracy of the prediction result by the learning model B2 is equal to or higher than a predetermined level, among the orders for arranging pixel values obtained by converting each of the multiple types of data.
[0060] Next, the prediction process performed by the prediction device 1 will be described with reference to FIG. 6 is a diagram showing an example of the flow of the prediction process according to this embodiment. The prediction process is executed by the control unit 10. Note that, before the prediction process is executed, the array information A1 and the trained model B1 are generated in advance and stored in the storage unit 11.
[0061] Step S110: The target dataset acquisition unit 101 acquires a target dataset.
[0062] Step S120: The pixel value data generating unit 102 generates pixel value data by converting each of the multiple types of data constituting the target data set into pixel values.
[0063] Step S130: The artificial image data generation unit 103 generates artificial image data by arranging the pixel values constituting the pixel value data in a predetermined order. Here, the artificial image data generation unit 103 reads out the array information A1 from the storage unit 11. The artificial image data generation unit 103 generates the artificial image data by arranging the pixel values in the order indicated by the read out array information A1.
[0064] Step S140: The artificial image data acquisition unit 105 acquires the artificial image data generated by the artificial image data generation unit 103.
[0065] Step S150: The prediction unit 106 makes a prediction about the prediction target from the artificial image data based on the trained model B1.
[0066] Step S160: The output unit 107 outputs the prediction result by the prediction unit . With this, the control unit 10 ends the prediction process.
[0067] (Example) In this example, the target dataset includes 76 types of data (variables). One of the 76 variables indicates an answer to the question of whether or not a person has a disease (for example, schizophrenia). The answer to the question of whether or not a person has a disease is data used as a label (correct label). The remaining 75 of the 76 variables indicate answers to questions about demographic attributes (age, sex, place of residence, educational background, income, etc.), health status (height, weight, obesity level, lifestyle habits, etc.), and physical comorbidities, mental comorbidities, and social comorbidities. These 75 variables are data used as features.
[0068] The number of samples is 2000. Of the 2000 target datasets, 1777 samples are samples of screened individuals without the disease, and the remaining 223 samples are samples of screened individuals with the disease.
[0069] In this case, the target dataset has a small number of samples (223) of diseased screening recipients compared to the number of samples (1777) of disease-free screening recipients. In other words, the target dataset has an imbalance between the number of samples in the disease-free category and the number of samples in the diseased category. For example, based on SMOTE, the imbalance between the categories can be corrected by increasing the number of samples in the diseased category.
[0070] The number of ways to arrange variables is the factorial of the number of variables. In one example of this embodiment, there are 75! / 8 possible pixel value orders. Because 75! / 8 is an enormous number, in this embodiment, 10,000 possible pixel value orders are randomly extracted from the 75! / 8 possible pixel value orders.
[0071] Furthermore, even if a predetermined number of pixel value order candidates are randomly extracted from all pixel value order candidates, if the predetermined number is sufficiently large, it is considered that the distribution of accuracy of prediction results for the extracted pixel value order candidates will sufficiently reproduce the distribution of accuracy of prediction results for all pixel value order candidates.
[0072] In this example, the learning model B2 has five hidden layers. The numbers of nodes in the five hidden layers are 128, 64, 32, 16, and 8, respectively. HeNormal was used to initialize the weights of the learning model B2. Rectified Linear Unit (ReLU) was used as the activation function for the hidden layers. A sigmoid function was used as the activation function for the output layer.
[0073] The output from learning model B2 indicates the probability that a certain medical examinee has a disease. Since the prediction result indicates the presence or absence of a disease as a binary value, in this embodiment, the AUC (Area Under the Curve) is used to determine a threshold value for converting the probability into the binary value.
[0074] The order determination unit 104 calculates a receiver operating characteristic curve (ROC curve) for each piece of training data based on the output from the learning model B2 and the label. The order determination unit 104 calculates the AUC from the ROC curve. The closer the AUC is to 1, the higher the prediction accuracy of the learning model B2. The AUC is used as the loss function.
[0075] FIG. 7 shows an example of the frequency distribution of AUC for pixel value order candidates. FIG. 7 shows the frequency distribution of AUC for 10,000 pixel value order candidates. As shown in FIG. 7, the AUC differs for each pixel value order candidate. Therefore, it can be seen that the prediction accuracy of learning model B2 changes depending on the pixel value order candidate. This suggests that there is an optimal pixel value order that maximizes the prediction accuracy of learning model B2.
[0076] As shown in Figure 7, the AUC for most of the pixel value order candidates was distributed around 0.88. A few pixel value order candidates with AUCs between 0.5 and 0.7 were also identified. In addition, pixel value order candidates with AUCs higher than 0.95 were also identified very rarely. The pixel value order candidate that provides the highest AUC is determined as the pixel value order.
[0077] As described above, in this embodiment, the predetermined order in which the multiple pixel values constituting the pixel value data are arranged is the order in which the accuracy of the prediction results by learning model B2 is highest among the orders randomly selected from the orders in which pixel values converted from multiple types of data are arranged (candidate pixel value orders). Alternatively, a pixel value order candidate with an AUC of 0.95 or more may be determined as the pixel value order.
[0078] Here, we will explain the relationship between learning model B2, which is a deep learning model according to this embodiment, and conventional deep learning models. In conventional deep learning models, the order of feature quantities is not taken into consideration, and feature quantities are input to a neural network as numerical data without being converted into pixel values. If the order of feature quantities is not taken into consideration, even if pixel value data obtained by converting multiple types of data (features) constituting a target dataset into pixel values is input to a deep learning model, the prediction accuracy of the model is considered to be no different from that of conventional deep learning models. This is because simply converting feature quantities into pixel values is considered to merely change the normalization of the numerical values input to the neural network.
[0079] For example, in conventional deep learning models, feature values are normalized to fall within the range of 0 to 1 or -1 to +1. In contrast, converting feature values to pixel values is equivalent to converting the feature values to a range of 0 to 255. In other words, simply converting feature values to pixel values without considering the order of feature values is considered to merely change the normalization of the numbers input to the neural network compared to conventional deep learning models.
[0080] As described above, in learning model B2, which is a deep learning model according to this embodiment, the multiple pixel values constituting the pixel value data are arranged in a predetermined order. The predetermined order may be an order that maximizes the accuracy of the prediction result by learning model B2, or a randomly selected order that maximizes the accuracy of the prediction result by learning model B2. Therefore, learning model B2 according to this embodiment, which takes the order of feature quantities into consideration, is considered to have higher prediction accuracy than conventional deep learning models in which the order of feature quantities is not considered and feature quantities are input to a neural network as numerical data.
[0081] As described above, the prediction accuracy of learning model B2, which is a deep learning model according to this embodiment, is considered to be higher than that of conventional deep learning models. Therefore, if it is confirmed that the prediction accuracy of conventional deep learning models is higher than that of conventional linear models, the prediction accuracy of learning model B2, which is a deep learning model according to this embodiment, is considered to be higher than that of conventional linear models. Therefore, the prediction performance of the conventional deep learning model and the conventional linear model will be compared.
[0082] In this comparison, predictions were made when the order of the multiple types of data (features) that make up the target dataset was used as is, and the data was input into a deep learning model without being converted into pixel value data, and the predictions were compared with those made by a conventional linear model.
[0083] FIG. 8 shows an example of the results of a comparison of the prediction performance between a conventional deep learning model and a linear model. Logistic regression is used as an example of a conventional linear model. In FIG. 8, learning model B2 is referred to as an "artificial neural network."
[0084] To evaluate generalizability, we compared the AUC of a conventional deep learning model and logistic regression using the above validation data set with the bootstrap method 10,000 times. To evaluate predictive performance, we also compared the AUC using k-fold cross-validation, dividing the target dataset into 10 parts.
[0085] As shown in Figure 8, the AUC values for the conventional deep learning model were higher than those for logistic regression in both the bootstrap method and k-fold cross-validation. The AUC comparison results confirmed that the conventional deep learning model exceeded the predictive accuracy of logistic regression, a conventional machine learning algorithm. This confirmed result strongly suggests that the predictive accuracy of learning model B2, which is the deep learning model according to this embodiment, is higher than that of conventional linear models.
[0086] In the present embodiment, an example in which the machine learning algorithm is deep learning has been described, but the present invention is not limited to this. The machine learning algorithm may be any supervised machine learning algorithm that can be applied to image recognition.
[0087] In this embodiment, an example in which all of the multiple types of data constituting the target dataset are arranged in a predetermined order has been described, but the present invention is not limited to this. Some of the multiple types of data constituting the target dataset may be arranged in a predetermined order, and the order of the remaining data may remain unchanged, as it is in the order included in the target dataset.
[0088] As described above, the artificial image data generation device of this embodiment (in this embodiment, the artificial image data generation function unit 100) includes a target dataset acquisition unit 101, a pixel value data generation unit 102, and an artificial image data generation unit 103. The target dataset acquisition unit 101 acquires a target dataset that is made up of multiple types of data about a target. The pixel value data generating unit 102 generates pixel value data by converting each of the multiple types of data constituting the target data set into pixel values. The artificial image data generating unit 103 generates artificial image data by arranging the plurality of pixel values that make up the pixel value data in a predetermined order.
[0089] With this configuration, the artificial image data generation device of this embodiment (in this embodiment, the artificial image data generation function unit 100) can generate artificial image data from multiple types of data, and can therefore generate data for making predictions based on image recognition from multiple types of data.
[0090] The prediction device 1 according to this embodiment also includes an artificial image data acquisition unit 105 and a prediction unit . The artificial image data acquisition unit 105 acquires artificial image data generated by converting multiple types of data about the prediction target (in this embodiment, a medical examination recipient) into pixel values and arranging them in a predetermined order (in this embodiment, the order indicated by the array information A1). The prediction unit 106 makes a prediction (the presence or absence of a disease (in this embodiment, schizophrenia as an example)) about the prediction target (in this embodiment, the screening recipient) from the artificial image data based on the trained model B1, which is a learning model trained to output a prediction result about the learning target (in this embodiment, the screening recipient) when training artificial image data generated by converting multiple types of data (in this embodiment, the training data) about the learning target (in this embodiment, the screening recipient) into pixel values and arranging them in that order (in this embodiment, the order indicated by the array information A1) is input. The order (in this embodiment, the order indicated by the array information A1) is an order based on the accuracy of the prediction result by the learning model B2, among the orders in which pixel values obtained by converting multiple types of data about the learning object are arranged.
[0091] With this configuration, the prediction device 1 according to this embodiment can convert multiple types of data about the prediction target into artificial image data and make predictions about the prediction target from the artificial image data, thereby improving the prediction accuracy of a prediction model using machine learning. Improving the prediction accuracy of a prediction model using machine learning means improving the prediction accuracy compared to, for example, generalized linear regression (linear model).
[0092] The accuracy of predictions using conventional generalized linear regression tends to be low. The following reasons may explain the low accuracy of predictions using generalized linear regression. One reason is that generalized linear regression has difficulty describing the correlation between multiple variables. Another reason is that generalized linear regression's performance is easily impaired when variables are multicollinear. In other words, in generalized linear regression, the coefficients of each variable can explain the relationship between the feature and the response only if the relationship is monotonic. Another reason is that in generalized linear regression, information is lost when simply selecting variables.
[0093] The artificial image data generation device (in this embodiment, the artificial image data generation function unit 100) and the prediction device 1 according to this embodiment can perform prediction using artificial images (artificial image data) even when the relationship between the feature and the response is not monotonic but has complex nonlinearity.
[0094] Artificial images also offer new concepts for interpreting predictive models. For example, feature importance and correlation between features can be explained by their location within the artificial image. Features located as pixels near the center of the image are considered to be more important than features located as pixels near the periphery of the image. Features located as adjacent pixels in the image are considered to be highly correlated.
[0095] The prediction device 1 according to this embodiment is suitable for various applications, for example, for the purpose of predicting the presence or absence of a disease. For example, the prediction device 1 may be used to predict diseases that prohibit driving when a driver's license is renewed. In another example, the prediction device 1 may be used to predict employee diseases in employee health management by a company's health insurance association. In another example, the prediction device 1 may be used by an insurance company or the like as an onset risk prediction tool for predicting the onset of various diseases.
[0096] In addition, in the present embodiment, an example in which the prediction device 1 is used to predict the presence or absence of a disease has been described, but the present invention is not limited to this. The prediction device 1 may be used to make various predictions in fields other than epidemiology. For example, the prediction device 1 may be applied to analyze data that appears chaotic.
[0097] Note that parts of the prediction device 1 in the above-described embodiment, such as the artificial image data generation function unit 100, the artificial image data acquisition unit 105, and the prediction unit 106, may be implemented by a computer. In this case, a program for implementing this control function may be recorded on a computer-readable recording medium, and the program may be loaded into a computer system and executed. Note that the term "computer system" as used herein refers to a computer system built into the prediction device 1, including hardware such as an operating system (OS) and peripheral devices. Furthermore, the term "computer-readable recording medium" refers to portable media such as flexible disks, optical magnetic disks, ROMs (Read Only Memory), and CD-ROMs (Compact Disc-Read Only Memory), as well as storage devices such as hard disks built into a computer system. Furthermore, the term "computer-readable recording medium" may also include devices that dynamically store programs for a short period of time, such as communication lines used when transmitting programs via a network such as the Internet or a communication line such as a telephone line, or devices that store programs for a certain period of time, such as volatile memory within a computer system that serves as a server or client in such cases. Furthermore, the above program may be one that realizes part of the functions described above, or may be one that can realize the functions described above in combination with a program already recorded in the computer system. Furthermore, part or all of the prediction device 1 in the above-described embodiment may be realized as an integrated circuit such as an LSI (Large Scale Integration). Each functional block of the prediction device 1 may be individually implemented as a processor, or part or all of them may be integrated into a processor. Furthermore, the integrated circuit implementation method is not limited to LSI, and may be implemented using a dedicated circuit or a general-purpose processor. Furthermore, if an integrated circuit implementation technology that can replace LSI emerges due to advances in semiconductor technology, an integrated circuit based on that technology may be used.
[0098] One embodiment of the present invention has been described in detail above with reference to the drawings, but the specific configuration is not limited to that described above, and various design changes and the like are possible within the scope that does not deviate from the gist of the present invention. [Explanation of symbols]
[0099] 1... prediction device, 100... artificial image data generation function unit, 101... target data set acquisition unit, 102... pixel value data generation unit, 103... artificial image data generation unit, 105... artificial image data acquisition unit, 106... prediction unit, A1... array information, B1... trained model, B2... training model
Claims
1. A target dataset acquisition unit that acquires a target dataset consisting of multiple types of data about a target; a pixel value data generating unit that generates pixel value data by converting each of the plurality of types of data constituting the target data set into a pixel value; an artificial image data generating unit that generates artificial image data by arranging a plurality of pixel values that constitute the pixel value data in a predetermined order; An artificial image data generating device comprising:
2. The order is an order based on the accuracy of a prediction result by a learning model that outputs a prediction result for the target when the artificial image data is input, among the orders of arranging the multiple pixel values that constitute the pixel value data.
2. The artificial image data generating device according to claim 1.
3. an artificial image data acquisition unit that acquires artificial image data generated by converting a plurality of types of data for a prediction target into pixel values and arranging the pixel values in a predetermined order; a prediction unit that performs prediction on the prediction target from the artificial image data based on a trained model, which is a learning model trained to output a prediction result on the learning target when training artificial image data generated by converting each of the multiple types of data on the learning target into pixel values and arranging them in the order is input; and Equipped with The order is an order based on the accuracy of the prediction result by the learning model, among the orders in which pixel values obtained by converting the plurality of types of data for the learning object are arranged. Prediction device.
4. The order is an order in which the accuracy of the prediction result by the learning model is highest among the orders in which the pixel values obtained by converting each of the plurality of types of data are arranged. The prediction device according to claim 3 .
5. The order is an order that provides the highest accuracy of the prediction result by the learning model among the orders randomly selected from the orders in which pixel values obtained by converting each of the multiple types of data are arranged. The prediction device according to claim 4 .
6. A target dataset acquisition step of acquiring a target dataset consisting of multiple types of data about a target; a pixel value data generating step of generating pixel value data by converting each of the plurality of types of data constituting the target data set into a pixel value; an artificial image data generating step of generating artificial image data by arranging a plurality of pixel values constituting the pixel value data in a predetermined order; The artificial image data generating method includes the steps of:
7. an artificial image data acquisition step of acquiring artificial image data generated by converting each of a plurality of types of data for a prediction target into pixel values and arranging the pixel values in a predetermined order; a prediction step of making a prediction about the prediction target from the artificial image data based on a trained model, which is a training model trained to output a prediction result about the training target when training artificial image data generated by converting each of the multiple types of data about the training target into pixel values and arranging them in the order is input; having The order is an order based on the accuracy of the prediction result by the learning model, among the orders in which pixel values obtained by converting the plurality of types of data for the learning object are arranged. Forecasting methods.
8. On the computer, A target dataset acquisition step of acquiring a target dataset consisting of multiple types of data about a target; a pixel value data generating step of generating pixel value data by converting each of the plurality of types of data constituting the target data set into a pixel value; an artificial image data generating step of generating artificial image data by arranging a plurality of pixel values constituting the pixel value data in a predetermined order; A program for executing the above.
9. On the computer, an artificial image data acquisition step of acquiring artificial image data generated by converting each of a plurality of types of data for a prediction target into pixel values and arranging the pixel values in a predetermined order; a prediction step of making a prediction about the prediction target from the artificial image data based on a trained model, which is a training model trained to output a prediction result about the training target when training artificial image data generated by converting each of the multiple types of data about the training target into pixel values and arranging them in the order is input; A program for executing The order is an order based on the accuracy of the prediction result by the learning model, among the orders in which pixel values obtained by converting the plurality of types of data for the learning object are arranged. program.