Information processing method and information processing system

Abstract

The present technology relates to an information processing method and an information processing system that make it possible to improve the prediction accuracy of a prediction model constructed by federated learning. An information processing method according to the present technology includes: applying soft clustering to a plurality of pieces of data used in training a prediction model and extracting a feature amount for each class pertaining to prediction items of the prediction model; and determining the optimal number of classes to perform training of the prediction model on the basis of the feature amount. The present technology can be applied, for example, to a training system that constructs, by federated learning, a prediction model capable of predicting a future cell state.

Description

Information processing method and information processing system

[0001] The present technology relates to an information processing method and an information processing system, and more particularly to an information processing method and an information processing system that are capable of improving the prediction accuracy of a prediction model constructed by federated learning.

[0002] In immunotherapy, which uses cells as therapeutic agents, achieving a stable response rate requires a certain degree of uniformity in the cell state during culture. Because the initial cell state depends on the patient, collection date, treatment history, and other factors, even if cells are cultured under the same conditions, the final cell composition, cell state, cell number, etc. will not be consistent. Therefore, achieving uniformity in cell state requires effort and cost.

[0003] There is a need to build a predictive model that can predict future cell states based on initial cell states, culture conditions, etc. When building such a predictive model, it is necessary to train it while ensuring the confidentiality of patient data.

[0004] Therefore, we consider using federated learning, in which learning using patient data is carried out at each medical institution and the prediction model is updated on a server, to obtain a prediction model that can predict future cell states with high accuracy while ensuring the confidentiality of patient data.

[0005] In federated learning, optimizing the number of classes into which data is classified and the class labels assigned to each data leads to improved prediction accuracy of the constructed prediction model. For example, Patent Literature 1 describes a technology for generating a model with high learning accuracy by performing machine learning while integrating similar classes into one class.

[0006] JP 2018-41300 A

[0007] The elbow method, which is commonly used to determine the number of classes, can be difficult to determine the optimal number of classes because the bending points in the graph are sometimes hard to see. Also, the elbow method requires separate class labels to be assigned to each data.

[0008] When applying the technology described in Patent Document 1 to federated learning, the optimal number of classes is determined within the machine learning system, so there is a possibility that the number of classes during learning will not match at each medical institution that performs the learning.

[0009] The present technology has been developed in light of these circumstances, and makes it possible to improve the prediction accuracy of prediction models constructed through federated learning.

[0010] An information processing method according to one aspect of the present technology includes applying soft clustering to a plurality of data sets used to train a predictive model to extract features for each class of prediction items of the predictive model, and determining an optimal number of classes for training the predictive model based on the features.

[0011] An information processing system according to one aspect of the present technology includes a soft clustering unit that applies soft clustering to a plurality of data sets used to train a predictive model and extracts features for each class for prediction items of the predictive model, and an optimal class number determination unit that determines the optimal number of classes for training the predictive model based on the features.

[0012] In one aspect of the present technology, soft clustering is applied to multiple data used to train a predictive model, features for each class of the prediction items of the predictive model are extracted, and the optimal number of classes for training the predictive model is determined based on the features.

[0013] 1 is a diagram illustrating an example of the configuration of a learning system according to an embodiment of the present technology. FIG. 1 is a first diagram illustrating the flow of associative learning performed by the learning system of FIG. 1. FIG. 2 is a second diagram illustrating the flow of associative learning performed by the learning system of FIG. 1. FIG. 3 is a diagram illustrating an example of processing including soft clustering performed by a server. FIG. 4 is a diagram illustrating an example of a matrix to which soft clustering is applied. FIG. 5 is a diagram illustrating examples of record items corresponding to each column of an initial cell state matrix. FIG. 6 is a diagram illustrating examples of record items corresponding to each column of a stimulus condition matrix. FIG. 7 is a diagram illustrating examples of record items corresponding to each column of a cell state matrix after stimulus administration. FIG. 8 is a diagram illustrating soft clustering in jNMF. FIG. 9 is a diagram illustrating examples of an initial cell state matrix and a stimulus condition matrix. FIG. 10 is a diagram illustrating an example of a cell state matrix after stimulus administration. FIG. 11 is a diagram illustrating examples of basis matrices and coefficient matrices. FIG. 12 is a diagram illustrating examples of features for prediction items. FIG. 13 is a diagram illustrating an example of temporal changes in feature values ​​for each class for CD8+ cell count. FIG. 14 is a diagram illustrating an example of correlation coefficients of temporal changes in feature values ​​between seven classes. FIG. 15 is a diagram illustrating an example of temporal changes in feature values ​​for each class after integration for CD8+ cell count. FIG. 16 is a diagram illustrating an example of correlation coefficients of temporal changes in feature values ​​between classes after integration. FIG. 17 is a diagram illustrating details of the initial class labeling process. 25. is a diagram showing an example of the prediction accuracy of a prediction model when the number of classes is 4 and when the number of classes is 7. is a block diagram showing an example of the functional configuration of a server. is a block diagram showing an example of the functional configuration of an information processing device. is a flowchart explaining processing performed by a server. is a flowchart explaining processing for determining the optimal number of classes performed in step S2 of FIG. 22. is a flowchart explaining processing for initial class labeling performed in step S3 of FIG. 22. is a flowchart explaining processing performed by an information processing device. is a flowchart explaining class integration processing performed in step S62 of FIG. 25. is a flowchart explaining processing for initial class labeling performed in step S63 of FIG. 25. is a diagram showing another example of processing including soft clustering performed by a server. is a diagram explaining details of processing for initial data labeling and group labeling.1 is a diagram illustrating an example of the prediction accuracy of a prediction model when group labels based on clustering results are assigned and when random group labels are assigned. FIG. 2 is a block diagram illustrating an example of the hardware configuration of a computer.

[0014] Hereinafter, embodiments of the present technology will be described in the following order: 1. Overview of the learning system 2. Configuration and operation of each device 3. Modified examples

[0015] 1. Overview of Learning System Configuration of Learning System FIG. 1 is a diagram illustrating an example configuration of a learning system according to an embodiment of the present technology.

[0016] The learning system in Figure 1 is an information processing system that builds a predictive model capable of predicting future cellular states while ensuring the confidentiality of patient data through federated learning, in which learning using patient data is performed at each medical institution and the predictive model is updated on a server. Hereinafter, cellular states are defined to include the state, composition, number, etc. of cells.

[0017] The learning system in FIG. 1 is configured by connecting a server 101 and information processing devices 102 installed in medical institutions 1 to n via a network such as the Internet.

[0018] The server 101 (first information processing device) is configured, for example, by a single computer or by multiple computers working together. The server 101 performs training using a dataset including multiple pieces of data collected in advance, and generates a prediction model. The server 101 also integrates the prediction models retrained by the information processing devices 102 of each medical institution.

[0019] The information processing device 102 (second information processing device) is configured, for example, by one computer or by multiple computers working together. The information processing device 102 retrains the prediction model generated by the server 101 using a dataset including multiple patient data collected at each medical institution.

[0020] Next, the flow of the associative learning performed by the learning system of FIG. 1 will be described with reference to FIGS.

[0021] In performing federated learning, the server 101 first applies soft clustering to data included in a dataset collected in advance to determine the optimal number of classes for performing federated learning (optimal number of classes) and assigns an initial class label to each piece of data. The class label indicates which class the data to which the class label is assigned is classified.

[0022] The server 101 receives data with initial class labels and the number of optimal classes as input, and performs learning while updating the class labels so as to reduce the difference between the predicted value and the true value, thereby ultimately generating prediction models equal to the number of optimal classes. As shown in the upper part of Figure 2, the server 101 generates, for example, four prediction models A ０ , B ０ , C ０ , D ０ Generate.

[0023] Next, as shown by arrow #1 in FIG. 2, the server 101 ０ , B ０ , C ０ , D ０ and information indicating the optimum number of classes are transmitted to each of medical institutions 1 to n.

[0024] Next, as shown in the lower left of Figure 3, the information processing device 102 of the medical institution 1 performs soft clustering on the local data (patient data) included in dataset 1 collected at the medical institution 1, and merges classes until the optimal number of classes determined by the server 101 is reached. After that, the information processing device 102 of the medical institution 1 assigns an initial class label to each patient data item based on the results of the soft clustering.

[0025] The information processing device 102 of the medical institution 1 uses the dataset 1 including the patient data to which the initial class labels are assigned to calculate the four prediction models A transmitted from the server 101. ０ , B ０ , C ０ , D ０Specifically, the information processing device 102 of the medical institution 1 performs re-learning while updating the class labels so as to reduce the difference between the predicted value and the true value.

[0026] Next, as shown by arrow #21-1 in FIG. 3, the information processing device 102 of the medical institution 1 uses the re-trained prediction model A １ , B １ , C １ , D １ (model parameters) to the server 101.

[0027] 3, the information processing device 102 of the medical institution 2 performs soft clustering on the local data (patient data) included in the dataset 2 collected at the medical institution 2, and merges the classes until the optimal number of classes determined by the server 101 is reached. After that, the information processing device 102 of the medical institution 2 assigns an initial class label to each patient data item based on the results of the soft clustering.

[0028] The information processing device 102 of the medical institution 2 uses the dataset 2 including the patient data to which the initial class labels are assigned to calculate the four prediction models A transmitted from the server 101. ０ , B ０ , C ０ , D ０ Specifically, the information processing device 102 of the medical institution 1 performs re-learning while updating the class labels so as to reduce the difference between the predicted value and the true value.

[0029] As shown by arrow #21-2 in FIG. 3, the information processing device 102 of the medical institution 2 uses the re-trained prediction model A ２ , B ２ , C ２ , D ２ (model parameters) to the server 101.

[0030] 3, the information processing device 102 of the medical institution n performs soft clustering on the local data (patient data) included in the dataset n collected at the medical institution n, and merges the classes until the optimal number of classes is reached as determined by the server 101. Thereafter, the information processing device 102 of the medical institution n assigns an initial class label to each patient data item based on the results of the soft clustering.

[0031] The information processing device 102 of the medical institution n uses the data set n including the patient data to which the initial class labels are assigned to calculate the four prediction models A transmitted from the server 101. ０ , B ０ , C ０ , D ０ Specifically, the information processing device 102 of the medical institution 1 performs re-learning while updating the class labels so as to reduce the difference between the predicted value and the true value.

[0032] As shown by the arrow #21-n in FIG. 3, the information processing device 102 of the medical institution n uses the re-trained prediction model A ｎ , B ｎ , C ｎ , D ｎ (model parameters) to the server 101.

[0033] The server 101 receives the prediction model transmitted from each medical institution. As shown in the upper part of FIG. 3, the server 101 receives the prediction model A. １ From A ｎ By integrating these, a new prediction model A is generated, and a prediction model B is generated. １ From B ｎ The server 101 generates a new prediction model B by integrating the prediction models C. １ From C ｎ By integrating these, a new prediction model C is generated, and a prediction model D １ From D ｎ A new prediction model D is generated by integrating the above.

[0034] Next, as indicated by arrow #22 in Figure 3, the server 101 transmits the four prediction models A, B, C, and D to the information processing devices 102 of medical institutions 1 to n. The information processing devices 102 of each medical institution re-learn the prediction models A, B, C, and D transmitted from the server 101, and transmit the re-learned prediction models to the server 101. The server 101 integrates the prediction models transmitted from the information processing devices 102 of each medical institution.

[0035] By repeating this process, the predictive model is trained using a large amount of patient data collected at each medical institution, and is updated to a predictive model with high prediction accuracy, without the need to consolidate patient data outside the medical institution (e.g., server 101).

[0036] FIG. 4 is a diagram showing an example of processing including soft clustering performed by the server 101.

[0037] Before generating a prediction model, the server 101 first performs soft clustering on the pre-collected data, as shown in #11 in Fig. 4. Through soft clustering, feature values ​​for each class of the prediction item (e.g., CD8+ cell count) of the prediction model and the contribution of each data item to each class are extracted.

[0038] Next, the server 101 determines the optimal number of classes in the federated learning, as shown in #12 of Fig. 4. Specifically, the server 101 determines the optimal number of classes by merging multiple classes into one class based on the similarity of the features extracted by soft clustering.

[0039] Next, the server 101 assigns an initial class label to each piece of data, as shown in #13 of Fig. 4. Specifically, the server 101 combines the classes integrated in determining the optimal number of classes into one class, and then determines a class label to be assigned to each piece of data based on the contribution of each piece of data to each class extracted by soft clustering.

[0040] 5 to 19, details of each process performed by the server 101 before generating a prediction model will be described. Below, a process flow will be described for the case where the purpose is to construct a prediction model capable of predicting the number of CD8+ cells on the final day of culture by performing associative learning using cell culture data.

[0041] Cell culture data is data in which time-series data of various measurement values, etc., are compiled as a single data set for each sample. Cell culture data is created by culturing cells for several days while varying conditions for each sample, such as the combination of stimuli (including antibody stimulation and cytokine stimulation) and the amount of stimulation administered, and recording the amount of stimulation administered and FCM (Flow Cytometer) measurements, etc., periodically during the culture period (on days 0, 4, 7, 11, and 14 of culture).

[0042] Soft Clustering Here, an example will be described in which soft clustering called jNMF (joint non-negative tensor factorization) is applied to cell culture data.

[0043] When applying jNMF to cell culture data, first, multiple cell culture data are divided into three matrices, for example, an initial cell state matrix, a stimulus condition matrix, and a cell state matrix after stimulus administration, as shown in FIG.

[0044] The initial cell state matrix records FCM measurement values ​​on day 0 of culture (when no stimuli have been administered). The stimulation condition matrix records the administered stimulation value and cumulative administered stimulation value. The post-stimulus cell state matrix records FCM measurement values ​​on days 4, 7, 11, and 14 of culture (when stimuli have been administered).

[0045] In each of the three matrices, each row represents data for each sample, and each column represents data for each record item. In the stimulation condition matrix and post-stimulus cell state matrix, which contain data for four time points, the data for the same record item at the four time points are arranged row-wise, as shown in the speech bubble in Figure 5.

[0046] In each of the three matrices, the data for the predicted item, CD8+ cell count, is preprocessed to boost its value by 10 times so that its contribution to clustering in jNMF is higher than that of the data for the other recorded items.

[0047] FIG. 6 is a diagram showing an example of record items corresponding to each column of the initial cell state matrix.

[0048] For example, the first column of the initial cell state matrix records whether or not each sample is derived from a patient who underwent chemotherapy.

[0049] FIG. 7 is a diagram showing an example of recording items corresponding to each column of the stimulation condition matrix.

[0050] For example, the concentrations of antibody stimulating OKT3 in the culture medium on days 4, 7, 11, and 14 of culture are recorded in columns 1 to 4 of the stimulation condition matrix.

[0051] FIG. 8 is a diagram showing an example of record items corresponding to each column of the post-stimulus cell state matrix.

[0052] For example, the first to fourth columns of the post-stimulus cell state matrix record the proportion of lymphocytes in total leukocytes on days 4, 7, 11, and 14 of culture.

[0053] FIG. 9 is a diagram illustrating soft clustering in jNMF.

[0054] In jNMF, as shown in FIG. 9, a data matrix X1 to XM with n rows is decomposed into a basis matrix W and coefficient matrices H1 to HM.

[0055] The data matrix X1 has n rows and p １ The data matrix X2 has n rows and p ２ The data matrix XM has n rows and p Ｍ The basis matrix W is a matrix showing the clustering results of each sample, and is a matrix with n rows and k columns. The coefficient matrix H1 is a matrix showing the extraction results of the feature quantities of the data matrix X1, and is a matrix with k rows and p １ The coefficient matrix H2 is a matrix that indicates the extraction results of the feature quantities of the data matrix X2, and has k rows and p ２The coefficient matrix HM is a matrix that indicates the extraction results of the feature quantities of the data matrix XM, and has k rows and p Ｍ It is a column matrix.

[0056] In jNMF, HM is calculated from the basis matrix W and the coefficient matrix H1 so that the sum of squares of the norms of the data matrix Xm and the reconstruction matrix WHm shown in the following equation (1) is minimized.

[0057] For example, consider applying jNMF to data matrices X1 to X3, where the initial cell state matrix shown on the left side of Fig. 10 is data matrix X1, the stimulus condition matrix shown on the right side of Fig. 10 is data matrix X2, the cell state matrix after stimulus administration shown in Fig. 11 is data matrix X3, and the number of classes (number of bases) is 7. By applying jNMF, the data matrices X1 to X3 are decomposed into basis matrix W and coefficient matrices H1 to H3 shown in Fig. 12.

[0058] The basis matrix W expresses the degree to which each sample contributes to each class, and the coefficient matrices H1 to H3 express the degree to which each recorded item contributes to each class.

[0059] Determining the Optimal Number of Classes To determine the optimal number of classes, first, features for the CD8+ cell count, which is the prediction item, are extracted from the coefficient matrix H3, which represents the feature extraction results for the post-stimulus cell state matrix (data matrix X3). In the example of Figure 13, the matrix from columns 53 to 56 is extracted, as shown in the gray box. In the extracted matrix, each row represents the feature for each class, and each column represents the feature for each date. The extracted matrix represents the temporal change (trend) in the feature for each class for CD8+ cell count.

[0060] 14 is a diagram showing an example of temporal changes in feature values ​​for each class of CD8+ cell count, in which the horizontal axis represents the date and the vertical axis represents the feature value.

[0061] FIG. 14 shows the temporal changes in feature amounts from the 4th to 14th days of culture for each of classes 1 to 7.

[0062] 15, the correlation coefficient of the temporal changes in the feature values ​​is calculated between each of the seven classes 1 to 7. Next, among classes 1 to 7, combinations of classes whose correlation coefficients are equal to or greater than a threshold are combined into one class.

[0063] For example, if the threshold is set to 0.7, in the example of Figure 15, the correlation coefficient between class 1 and class 5 is 0.9835, so class 1 and class 5 are merged into one class. Also, the correlation coefficient between class 2 and class 4 is 0.9933, the correlation coefficient between class 4 and class 7 is 0.8899, and the correlation coefficient between class 7 and class 2 is 0.9362, so class 2, class 4, and class 7 are merged into one class. Hereinafter, the class resulting from the merge of class 1 and class 5 will also be referred to as class 1·5, and the class resulting from the merge of class 2, class 4, and class 7 will also be referred to as class 2·4·7.

[0064] When classes are combined, the features of each combined class are averaged for each day.

[0065] 16 is a diagram showing an example of temporal changes in feature values ​​for each class after integration for CD8+ cell counts, in which the horizontal axis represents the date and the vertical axis represents the feature values.

[0066] In FIG. 16, the temporal changes in the feature amounts from the 4th day to the 14th day of culture are shown for each of classes 1 and 5, classes 2, 4 and 7, class 3, and class 7.

[0067] As described above, after a series of processes, including calculating the correlation coefficient between classes, comparing the correlation coefficient with a threshold, integrating classes whose correlation coefficient is equal to or greater than the threshold, and averaging the features of the integrated classes, is performed once, the series of processes is repeated until there are no more combinations of classes whose correlation coefficient is equal to or greater than the threshold.

[0068] For example, as shown in Fig. 17, the correlation coefficient of the time-series changes in the feature quantities is calculated between each of classes 1 and 5, classes 2, 4 and 7, class 3, and class 7. In the example of Fig. 17, there is no combination with a correlation coefficient of 0.7 or more, so the series of processes ends.

[0069] The number of classes when there are no more class combinations with correlation coefficients equal to or greater than the threshold is determined to be the optimal number of classes in associative learning. In the example of FIG. 17, the optimal number of classes is 4. After the optimal number of classes is determined, class numbers (class names) are reassigned to each class. For example, class numbers 1 and 5 are assigned to class 1, 2, 4, and 7 to class 2, 3 to class 3, and 4 to class 6.

[0070] Initial Class Labeling When assigning an initial class label to each cell culture data, first, as shown by the arrow Ar1 in FIG. 18 , the z-score of each data (contribution) of the basis matrix W, which indicates the clustering result of each sample, is calculated and normalized to generate a matrix W1.

[0071] Next, as indicated by arrow Ar2 in Fig. 18 , the class with the highest z-score for each sample is extracted, and a matrix W2 is generated in which data of 1 is recorded for the class with the highest z-score for each sample and data of 0 is recorded for the other classes. The colored classes in matrix W2 in Fig. 18 indicate the classes with the highest z-score for each sample. For example, for sample 1, class 5 of classes 1 to 7 has the highest z-score.

[0072] Next, as indicated by arrow Ar3 in FIG. 18, the combination of classes that was integrated into one class in determining the optimal number of classes is also integrated into one class in matrix W2, thereby generating matrix W3.

[0073] In determining the optimal number of classes, classes 1 and 5 before the integration were integrated into class 1, and therefore, in matrix W3, for samples with the largest z-score of class 1 or class 5 before the integration, the class with the largest z-score is set to class 1. In determining the optimal number of classes, classes 2, 4, and 7 before the integration were integrated into class 2, and therefore, in matrix W3, for samples with the largest z-score of class 2, class 4, or class 7 before the integration, the class with the largest z-score is set to class 2.

[0074] In determining the optimal number of classes, the class number of class 3 was reassigned to class 3 before the integration, so in matrix W3, for the sample with the highest z-score of class 3 before the integration, the class with the highest z-score is set to class 3. In determining the optimal number of classes, the class number of class 4 was assigned to class 6 before the integration, so in matrix W3, for the sample with the highest z-score of class 6 before the integration, the class with the highest z-score is set to class 4.

[0075] Next, for each sample in matrix W3, the class label indicating the class with the largest z-score is assigned as the initial class label to the cell culture data of that sample. For example, the class label of class 1 is assigned as the initial class label to the cell culture data of sample 1.

[0076] In this way, the classes integrated in determining the optimal number of classes are combined into one class, and a class label indicating the class with the greatest contribution in each sample is assigned as an initial class label to the cell culture data of that sample. For example, the same class label is assigned as the initial class label to the cell culture data of sample 4, which has the highest z-score in class 1 before integration, and the cell culture data of sample 1, which has the highest z-score in class 5 before integration.

[0077] - Effect of optimizing the initial class labels and number of classes As described above, by applying soft clustering to each data, the optimal number of classes in associative learning was determined to be 4, and initial class labels from class 1 to class 4 were assigned to the cell culture data of each sample.

[0078] The present applicant extracted nine data (data on the final day of culture) from multiple cell culture data sets as test data, and used the remaining data (data on the samples other than the final day of culture used as test data and data on all days of the other samples) as training data, causing the server 101 to perform learning and calculate prediction accuracy. This learning was performed by inputting multiple cell culture data sets to which initial class labels had been assigned by the initial class labeling process and the optimal number of classes (=4) determined by the optimal number of classes determination process.

[0079] The present applicant created a linear regression model using four cumulative dose stimuli (the cumulative doses of OKT3 and IC3 / 28 as antibody stimuli, and the cumulative doses of IL2 and IL7 as cytokine stimuli) as explanatory variables and the CD8+ cell count as the target variable, ultimately obtaining four prediction models. The present applicant used these prediction models to predict the CD8+ cell count in test data and evaluated the prediction models by comparing them with the true values ​​of the test data. For comparison purposes, the present applicant also prepared a prediction model with seven classes (with the same conditions except for the number of classes and initial class labels) to confirm the effectiveness of the model, and evaluated the prediction model in the same way.

[0080] FIG. 19 is a diagram showing an example of the prediction accuracy of the prediction model when the number of classes is set to 4 and when the number of classes is set to 7.

[0081] 19, the error rate between the predicted value and the true value calculated for each of the nine test data is shown in a box plot. In the box plot of FIG. 19, the horizontal axis represents the number of classes, and the vertical axis represents the error rate.

[0082] When comparing the prediction accuracy of the prediction model when the number of classes is set to 4 (optimal number of classes) with the prediction accuracy of the prediction model when the number of classes is set to 7 (non-optimal number of classes), the error rate is smaller overall and the prediction accuracy of the prediction model is higher when the number of classes is set to 4.

[0083] In this way, by determining the optimal number of classes in the federated learning before the federated learning and assigning optimal initial class labels to the data used to learn (generate) the predictive model, it is possible to improve the predictive accuracy of the predictive model constructed by the federated learning.

[0084] 2. Configuration and Operation of Each Device Server Configuration FIG. 20 is a block diagram showing an example of the functional configuration of the server 101. As shown in FIG.

[0085] As shown in FIG. 20, the server 101 includes a data acquisition unit 151 , a soft clustering unit 152 , an optimum class number determination unit 153 , a labeling unit 154 , a learning unit 155 , a communication unit 156 , and an integration unit 157 .

[0086] The data acquisition unit 151 acquires a data set used for training (generation) a prediction model, and supplies it to the soft clustering unit 152 and the labeling unit 154 .

[0087] The soft clustering unit 152 applies, for example, jNMF to each piece of data included in the dataset supplied from the data acquisition unit 151, and extracts the contribution of each piece of data to each class and the feature amount for each record item. Specifically, the soft clustering unit 152 calculates a basis matrix indicating the clustering result and a coefficient matrix indicating the feature amount extraction result.

[0088] The soft clustering unit 152 supplies the feature extraction results to the optimum class number determination unit 153 and supplies the clustering results to the labeling unit 154 .

[0089] Optimal class number determination unit 153 determines the optimal number of classes in the associative learning based on the clustering result by soft clustering unit 152, and notifies learning unit 155 and communication unit 156 of the determined optimal number of classes. In addition, when determining the optimal number of classes, optimal class number determination unit 153 supplies flag information indicating into which class each basis vector (each class before integration) has been integrated to labeling unit 154.

[0090] The labeling unit 154 assigns an initial class label to each piece of data included in the data set supplied from the data acquisition unit 151, based on the clustering result by the soft clustering unit 152 and the flag information supplied from the optimal class number determination unit 153. The labeling unit 154 supplies the data set including the data to which the initial class labels have been assigned to the learning unit 155.

[0091] The learning unit 155 generates a prediction model for the optimal number of classes by performing learning using the optimal number of classes determined by the optimal class number determination unit 153 and the data set supplied from the labeling unit 154 as input. In other words, the learning unit 155 performs learning of the prediction model using multiple pieces of data classified into the optimal number of classes. Specifically, the learning unit 155 generates the prediction model by performing learning while updating the class labels assigned to each piece of data so as to reduce the difference between the predicted value and the true value. The learning unit 155 supplies the generated prediction model to the communication unit 156.

[0092] The communication unit 156 transmits the prediction model supplied from the learning unit 155 and information indicating the optimal number of classes determined by the optimal number of classes determination unit to the information processing device 102 of each medical institution. The communication unit 156 also receives the prediction model transmitted from the information processing device 102 of each medical institution and supplies it to the integrating unit 157. Furthermore, the communication unit 156 transmits the prediction model supplied from the integrating unit 157 to the information processing device 102 of each medical institution.

[0093] The integration unit 157 integrates the prediction models re-trained by the information processing devices 102 of the medical institutions to generate a new prediction model, and supplies the new prediction model to the communication unit 156 .

[0094] Note that flag information, initial class labels assigned to each data, and the update status of the class labels during learning by the learning unit 155 may be presented to the administrator (user) of the server 101. For example, this information is displayed on the display unit of the server 101. The administrator can check how the prediction model is being constructed by looking at this information.

[0095] Furthermore, flag information may be transmitted to the information processing device 102 of each medical institution, or flag information on the information processing device 102 side may be transmitted from the information processing device 102 of each medical institution. The flag information on the information processing device 102 side indicates, when classes are integrated in the information processing device 102, into which class each basis vector (each class before integration) has been integrated.

[0096] When the flag information on the server 101 side is transmitted to the information processing device 102, the information processing device 102 can perform clustering in the same way as during learning by the server 101, for example, by integrating classes based on the flag information on the server 101 side. When the flag information on the information processing device 102 side is transmitted from the information processing device 102 of each medical institution to the server 101, the flag information on the information processing device 102 side may be presented to the administrator together with the flag information on the server 101 side, etc.

[0097] Configuration of Information Processing Apparatus FIG. 21 is a block diagram showing an example of the functional configuration of the information processing apparatus 102.

[0098] As shown in FIG. 21, the information processing device 102 includes a data acquisition unit 171 , a soft clustering unit 172 , a class integration unit 173 , a labeling unit 174 , a learning unit 175 , and a communication unit 176 .

[0099] The data acquisition unit 171 acquires a dataset including patient data used for retraining the prediction model, and supplies it to the soft clustering unit 172 and the labeling unit 174 .

[0100] The soft clustering unit 172 applies, for example, jNMF to each patient data included in the dataset supplied from the data acquisition unit 171, and extracts the contribution of each patient data to each class and the feature amount for each record item. Specifically, the soft clustering unit 172 calculates a basis matrix indicating the clustering result and a coefficient matrix indicating the feature amount extraction result.

[0101] The soft clustering unit 172 supplies the feature extraction results to the class integration unit 173 and supplies the clustering results to the labeling unit 174 .

[0102] The class integrating unit 173 integrates the classes into which each piece of data is classified by the soft clustering unit 172 until the number of classes reaches the optimal number of classes determined by the server 101. Specifically, the class integrating unit 173 integrates classes until the number of classes reaches the optimal number of classes determined by the server 101, for example by adjusting the threshold value of the correlation coefficient while performing processing in accordance with the flow of processing for determining the optimal number of classes by the server 101.

[0103] In the information processing device 102, the threshold value of the correlation coefficient is set so that it does not fall below a certain value (for example, 0.5). If the number of classes does not become the optimal number even when a combination of classes with a correlation coefficient of 0.5 or more is integrated, re-learning is performed with that number of classes.

[0104] The class integrating unit 173 notifies the learning unit 175 of the optimal number of classes. The class integrating unit 173 also supplies the labeling unit 174 with flag information indicating which class each basis vector (each class before integration) has been integrated into.

[0105] The labeling unit 174 assigns an initial class label to each piece of patient data included in the dataset supplied from the data acquisition unit 171, based on the clustering result by the soft clustering unit 172 and the flag information supplied from the class integration unit 173. The labeling unit 174 supplies the learning unit 175 with the dataset including the patient data to which the initial class label has been assigned.

[0106] The learning unit 175 receives as input the optimal number of classes notified by the class integration unit 173 and the dataset supplied from the labeling unit 174, and re-learns the prediction model supplied from the communication unit 176. In other words, the learning unit 175 re-learns the prediction model using multiple patient data classified into the optimal number of classes. Specifically, the learning unit 175 re-learns while updating the class labels assigned to each patient data set so as to reduce the difference between the predicted value and the true value. The learning unit 175 supplies the re-learned prediction model to the communication unit 176.

[0107] The communication unit 176 receives the prediction model and information indicating the optimal number of classes transmitted from the server 101. The communication unit 176 supplies the prediction model to the learning unit 175 and supplies the information indicating the optimal number of classes to the class integration unit 173. The communication unit 176 also transmits the prediction model supplied from the learning unit 175 to the server 101.

[0108] Server Operation Next, the processing performed by the server 101 will be described with reference to the flowchart of FIG.

[0109] In step S1, the soft clustering unit 152 applies soft clustering to each cell culture data included in the data set. By soft clustering, clustering of each sample (extraction of the contribution of each cell culture data to each class) and extraction of feature amounts for each record item are performed.

[0110] In step S2, the optimal class number determination unit 153 performs an optimal class number determination process based on the feature extraction results by the soft clustering unit 152. By the optimal class number determination process, a combination of classes with high feature similarity is integrated into one class, and the optimal number of classes for the associative learning is determined. Details of the optimal class number determination process will be described later with reference to FIG. 23.

[0111] In step S3, the labeling unit 154 performs an initial class labeling process based on the clustering result by the soft clustering unit 152 and the class integration result by the optimal class number determination unit 153. By the initial class labeling process, an initial class label is assigned to each piece of cell culture data. Details of the initial class labeling process will be described later with reference to FIG. 24.

[0112] In step S4, the learning unit 155 performs learning using the optimal number of classes determined by the optimal class number determination unit 153 and the cell culture data to which initial class labels have been assigned by the labeling unit 154 as input, thereby generating prediction models for the optimal number of classes.

[0113] In step S5, the communication unit 156 transmits the prediction model to the information processing device 102 of each medical institution.

[0114] In step S6, the communication unit 156 receives the prediction model transmitted from the information processing device 102 of each medical institution.

[0115] In step S7, the integrating unit 157 integrates the prediction models re-trained by the information processing devices 102 of the medical institutions to generate a new prediction model. Then, the process returns to step S5, and the subsequent steps are repeated. For example, the prediction model is repeatedly updated until the prediction accuracy of the prediction model exceeds a predetermined threshold.

[0116] The process of determining the optimum number of classes performed in step S2 of FIG. 22 will be described with reference to the flowchart of FIG.

[0117] In step S21, the optimum class number determination unit 153 extracts the feature amount for each class for the prediction item from the feature amount extraction result by the soft clustering unit 152.

[0118] In step S22, the optimum class number determination unit 153 calculates the correlation coefficient of the temporal change in the feature amount between classes.

[0119] In step S23, the optimum class number determination unit 153 determines whether or not there is a combination of classes whose correlation coefficient is equal to or greater than a threshold value.

[0120] If it is determined in step S23 that there is a combination of classes whose correlation coefficient is equal to or greater than the threshold, in step S24, the optimal class number determination unit 153 combines the combination of classes whose correlation coefficient is equal to or greater than the threshold into one class.

[0121] In step S25, the optimal class number determination unit 153 averages the feature quantities of each class that has been integrated into one class for each day. After that, the process returns to step S22, and the subsequent processes are performed.

[0122] On the other hand, if it is determined in step S23 that there is no combination of classes whose correlation coefficient is equal to or greater than the threshold, then in step S26, the optimum class number determination unit 153 determines the number of classes at this time as the optimum number of classes.

[0123] Thereafter, the process returns to step S2 in FIG. 22, and the subsequent processes are carried out.

[0124] The initial class labeling process performed in step S3 of FIG. 22 will be described with reference to the flowchart of FIG.

[0125] In step S41, the labeling unit 154 extracts the class with the highest contribution (z score) for each sample based on the clustering result by the soft clustering unit 152.

[0126] In step S42, the labeling unit 154 combines the classes that have been integrated in the process of determining the optimal number of classes into one class.

[0127] In step S43, the labeling unit 154 assigns an initial class label to the cell culture data of each sample. Specifically, the labeling unit 154 assigns the class label of the integrated class with the highest contribution for each sample to the cell culture data of that sample.

[0128] Thereafter, the process returns to step S3 in FIG. 22, and the subsequent processes are carried out.

[0129] Operation of Information Processing Device Next, the processing performed by the information processing device 102 will be described with reference to the flowchart in Fig. 25. The processing in Fig. 25 is started, for example, when the information processing device 102 receives information indicating a prediction model and an optimal number of classes transmitted from the server 101.

[0130] In step S61, the soft clustering unit 172 applies soft clustering to each piece of cell culture data (patient data) included in the dataset. By soft clustering, clustering of each sample (extraction of the contribution of each patient data to each class) and extraction of feature amounts for each record item are performed.

[0131] In step S62, the class integrating unit 173 performs class integrating processing based on the feature extraction results obtained by the soft clustering unit 172. Through the class integrating processing, combinations of classes with high feature similarity are integrated into one class until the number of classes reaches an optimal number. Details of the class integrating processing will be described later with reference to FIG. 26 .

[0132] In step S63, the labeling unit 174 performs an initial class labeling process based on the clustering result by the soft clustering unit 172 and the class integration result by the class integration unit 173. By the initial class labeling process, an initial class label is assigned to each patient data. Details of the initial class labeling process will be described later with reference to FIG. 27 .

[0133] In step S64, the learning unit 175 receives as input the optimal number of classes determined by the server 101 and the data to which the initial class labels have been assigned by the labeling unit 174, and re-learns the prediction model.

[0134] In step S65, the communication unit 176 transmits the re-learned prediction model to the server 101.

[0135] The class integration process performed in step S62 of FIG. 25 will be described with reference to the flowchart of FIG.

[0136] In step S81, the class integration unit 173 extracts feature amounts for each class of the predicted item from the feature amount extraction results obtained by the soft clustering unit 172.

[0137] In step S82, the class integration unit 173 determines whether the number of classes matches the optimum number of classes.

[0138] If it is determined in step S82 that the number of classes does not match the optimum number of classes, then in step S83, the class integration unit 173 calculates the correlation coefficient of the temporal changes in the feature amounts between the classes.

[0139] In step S84, the class integrating unit 173 integrates into one class a combination of classes whose correlation coefficient is equal to or greater than a threshold value.

[0140] In step S85, the class integration unit 173 averages the feature quantities of each class integrated into one class for each day. After that, the process returns to step S82, and the subsequent processes are performed.

[0141] If it is determined in step S82 that the number of classes matches the optimum number of classes, the process returns to step S62 in FIG. 25, and the subsequent processes are carried out.

[0142] The initial class labeling process performed in step S63 of FIG. 25 will be described with reference to the flowchart of FIG.

[0143] In step S101 , the labeling unit 174 extracts the class with the highest contribution (z score) for each sample based on the clustering result by the soft clustering unit 172 .

[0144] In step S102, the labeling unit 174 combines the classes combined in the class combining process into one class.

[0145] In step S103, the labeling unit 174 assigns an initial class label to the patient data of each sample. Specifically, the labeling unit 174 assigns the class label of the integrated class with the highest contribution for each sample to the patient data of that sample.

[0146] Thereafter, the process returns to step S63 in FIG. 25, and the subsequent processes are carried out.

[0147] As described above, in the learning system of the present technology, soft clustering is applied to multiple data used to train a predictive model, feature quantities for each class of prediction items of the predictive model are extracted, and the optimal number of classes for training the predictive model (optimal number of classes) is determined based on the feature quantities. Furthermore, in the learning system of the present technology, soft clustering is applied to multiple data used to train a predictive model, feature quantities for each class of prediction items are extracted, as well as the contribution of each of the multiple data to each class, and a class label is assigned to each of the multiple data based on the contribution quantities and the results of class integration performed when determining the optimal number of classes.

[0148] Before federated learning, the number of classes into which each piece of data used in training (generating) or relearning a predictive model is classified and the initial class labels assigned to each piece of data are optimized, thereby making it possible to improve the predictive accuracy of the predictive model constructed by federated learning.

[0149] <3. Modification> When associative learning is performed using data that is hierarchically clustered, it is desirable to appropriately classify each piece of data into groups and assign an optimal group label to each piece of data before learning a predictive model. A group is a lower level of a class, and data classified into the same group is also classified into the same class. A group label indicates which group the data assigned the group label is classified into.

[0150] The server 101 applies soft clustering to each piece of data to assign an optimal group label to each piece of data.

[0151] FIG. 28 is a diagram showing another example of processing including soft clustering performed by the server 101.

[0152] Before generating a prediction model, the server 101 performs soft clustering, determines the number of initial classes, and assigns initial class labels, as described with reference to Fig. 4. As shown in #51 in Fig. 28, the server 101 assigns group labels to each data item during the initial class labeling process.

[0153] FIG. 29 is a diagram for explaining the details of the initial data labeling and group labeling processes.

[0154] When assigning an initial class label and a group label to each data, a matrix W1 is generated based on the basis matrix W, and a matrix W2 is generated based on the matrix W1, as in the case described with reference to FIG.

[0155] Next, as indicated by arrow Ar11 in Fig. 29, the class label of the class with the largest z-score for each sample in matrix W2 is assigned to the cell culture data of that sample. For example, sample 1 is assigned the group label of class 5 (group 5).

[0156] 29, the classes that were combined into one class in determining the optimal number of classes are also combined into one class in matrix W2 to generate matrix W3, and in matrix W3, the class label of the class with the largest z-score for each sample is assigned to the cell culture data of that sample. For example, the class label of class 1 is assigned to the cell culture data of sample 1.

[0157] In this way, before combining the classes integrated in determining the optimal number of classes into one class, the group label of the class (group) with the greatest contribution for each sample is assigned to the cell culture data of that sample. After combining the classes integrated in determining the optimal number of classes into one class, the class label of the class with the greatest contribution for each sample is assigned to that sample.

[0158] As described above, by applying soft clustering to multiple cell culture data, an initial class label from Class 1 to Class 4 was assigned to each cell culture data, and a group label from Group 1 to Group 7 was assigned to each cell culture data.

[0159] The present applicant extracted nine data (data on the final day of culture) from multiple cell culture data sets as test data, and used the remaining data (data on the samples other than the final day of culture used as test data and data on all days of the other samples) as training data to have the server 101 perform learning, thereby determining prediction accuracy. This learning was performed by inputting multiple cell culture data sets to which initial class labels had been assigned by the initial class labeling process and to which group labels had been assigned by the group labeling process, and the optimal number of classes (=4) determined by the optimal number of classes determination process.

[0160] The present applicant created a linear regression model using four cumulative dose stimuli (the cumulative doses of OKT3 and IC3 / 28 as antibody stimuli, and the cumulative doses of IL2 and IL7 as cytokine stimuli) as explanatory variables and the CD8+ cell count as the target variable, ultimately obtaining four prediction models. The present applicant used these prediction models to predict the CD8+ cell counts of test data and evaluated them by comparing them with the true values ​​of the test data. For comparison purposes, the present applicant also prepared a prediction model in which group labels were randomly assigned to each cell culture data set (all conditions except for the group labels were the same), and evaluated this prediction model in the same way.

[0161] FIG. 30 is a diagram showing an example of the prediction accuracy of a prediction model when group labels based on the clustering results are assigned and when random group labels are assigned.

[0162] In Fig. 30, the error rate between the predicted value and the true value calculated for each of the nine test data is shown in a box plot. In the box plot in Fig. 30, the horizontal axis indicates the group label assigned to each cell culture data, and the vertical axis indicates the error rate.

[0163] When comparing the prediction accuracy of a prediction model when group labels based on clustering results (optimal group labels) are assigned with the prediction accuracy of a prediction model when random group labels (non-optimal group labels) are assigned, the overall error rate is smaller and the prediction accuracy of the prediction model is higher when optimal group labels are assigned.

[0164] In this way, by assigning optimal group labels to the data used to train (generate) the predictive model before federated learning, it is possible to improve the predictive accuracy of the predictive model constructed by federated learning.

[0165] <Example of Computer Configuration> The above-described series of processes can be executed by hardware or software. When the series of processes is executed by software, the program constituting the software is installed from a program recording medium into a computer incorporated in dedicated hardware, or into a general-purpose personal computer, etc.

[0166] FIG. 31 is a block diagram showing an example of the hardware configuration of a computer that executes the above-described series of processes by a program.

[0167] A CPU (Central Processing Unit) 501 , a ROM (Read Only Memory) 502 , and a RAM (Random Access Memory) 503 are interconnected by a bus 504 .

[0168] An input / output interface 505 is also connected to the bus 504. An input unit 506 including a keyboard, a mouse, etc., and an output unit 507 including a display, a speaker, etc. are connected to the input / output interface 505. Also connected to the input / output interface 505 are a storage unit 508 including a hard disk, a nonvolatile memory, etc., a communication unit 509 including a network interface, etc., and a drive 510 that drives removable media 511.

[0169] In a computer configured as described above, the CPU 501 performs the above-described series of processes by, for example, loading a program stored in the storage unit 508 into the RAM 503 via the input / output interface 505 and the bus 504 and executing it.

[0170] The program executed by the CPU 501 is installed in the storage unit 508 by being recorded on, for example, a removable medium 511 or provided via a wired or wireless transmission medium such as a local area network, the Internet, or digital broadcasting.

[0171] The program executed by the computer may be a program that processes in chronological order according to the order described in this specification, or may be a program that processes in parallel or at the required timing, such as when called.

[0172] In this specification, a system refers to a collection of multiple components (devices, modules (components), etc.), regardless of whether all of the components are housed in the same housing. Therefore, multiple devices housed in separate housings and connected via a network, and a single device housed in a single housing with multiple modules, are both systems.

[0173] The effects described in this specification are merely examples and are not limiting, and other effects may also be present.

[0174] The embodiments of the present technology are not limited to the above-described embodiments, and various modifications are possible without departing from the spirit of the present technology.

[0175] For example, the present technology can be configured as a cloud computing system in which a single function is shared and processed collaboratively by a plurality of devices via a network.

[0176] Furthermore, each step described in the above flowchart can be executed by one device, or can be shared and executed by a plurality of devices.

[0177] Furthermore, when one step includes multiple processes, the multiple processes included in that one step can be executed by one device or can be shared and executed by multiple devices.

[0178] <Examples of Combinations of Configurations> The present technology can also have the following configurations.

[0179] (1) An information processing method comprising: applying soft clustering to a plurality of data used for training a predictive model to extract features for each class for prediction items of the predictive model; and determining an optimal number of classes for training the predictive model based on the features. (2) The information processing method described in (1), wherein the optimal number of classes is determined by merging a plurality of classes into one class based on similarities between the features between the classes. (3) The information processing method described in (2), wherein a combination of classes for which a correlation coefficient of temporal changes in the features between the classes is equal to or greater than a threshold is merged into one class. (4) The information processing method described in (3), further comprising: applying soft clustering to a plurality of the data to extract the features and the contribution of each of the plurality of data to each class; and assigning a class label indicating the class to each of the plurality of data based on the class integration result and the contribution. (5) The information processing method described in (4), further comprising assigning a group label indicating a group that is a lower level of the class to each of the plurality of data based on the contribution. (6) The information processing method according to any one of (3) to (5), further comprising presenting to a user flag information indicating into which class the class before integration has been integrated. (7) The information processing method according to any one of (1) to (6), wherein the predictive model is trained by associative learning. (8) The information processing method according to (7), wherein the feature amounts are extracted by applying the soft clustering to the plurality of data used to generate the predictive model. (9) The information processing method according to (8), wherein a first information processing device extracts the feature amounts, determines the optimal number of classes, and generates the predictive model by learning using the plurality of data, and wherein a second information processing device re-trains the predictive model using the plurality of data classified into the optimal number of classes determined by the first information processing device.(10) The information processing method according to (9), further comprising: the first information processing device integrating the predictive models that have been re-trained by a plurality of the second information processing devices. (11) The information processing method according to any one of (1) to (10), wherein the soft clustering includes jNMF. (12) An information processing system comprising: a soft clustering unit that applies soft clustering to a plurality of data used for training a predictive model to extract features for each class for prediction items of the predictive model; and an optimal class number determination unit that determines an optimal number of classes for training the predictive model based on the features.

[0180] REFERENCE SIGNS LIST 101 Server, 102 Information processing device, 151 Data acquisition unit, 152 Soft clustering unit, 153 Optimal class number determination unit, 154 Labeling unit, 155 Learning unit, 156 Communication unit, 157 Integration unit, 171 Data acquisition unit, 172 Soft clustering unit, 173 Class integration unit, 174 Labeling unit, 175 Learning unit, 176 Communication unit

Claims

1. An information processing method comprising: applying soft clustering to a plurality of data used to train a predictive model, extracting features for each class regarding the prediction items of the predictive model; and determining the optimal number of classes for training the predictive model based on the features.

2. The information processing method according to claim 1, wherein the optimal number of classes is determined by merging multiple classes into one class based on the similarity of the feature amounts between classes.

3. The information processing method according to claim 2, wherein a combination of classes in which the correlation coefficient of the temporal change in the feature between the classes is equal to or greater than a threshold is integrated into one class.

4. The information processing method according to claim 3, further comprising: applying the soft clustering to the plurality of data to extract the feature values ​​and the contribution of each of the plurality of data to each class; and assigning a class label indicating the class to each of the plurality of data based on the class integration result and the contribution.

5. The information processing method according to claim 4, further comprising: assigning a group label indicating a group that is a lower level of the class to each of the plurality of data based on the contribution degree.

6. The information processing method according to claim 3, further comprising: presenting to the user flag information indicating into which class the class before integration has been integrated.

7. The information processing method according to claim 1, wherein the predictive model is trained by associative learning.

8. The information processing method according to claim 7, wherein the feature amounts are extracted by applying the soft clustering to the plurality of data used to generate the prediction model.

9. The information processing method according to claim 8, wherein a first information processing device extracts the features, determines the optimal number of classes, and generates the predictive model by learning using a plurality of the data, and a second information processing device re-learns the predictive model using the plurality of data classified into the optimal number of classes determined by the first information processing device.

10. The information processing method according to claim 9, further comprising the first information processing device integrating the predictive models that have been retrained by a plurality of the second information processing devices.

11. The information processing method according to claim 1, wherein the soft clustering includes jNMF.

12. An information processing system comprising: a soft clustering unit that applies soft clustering to multiple data used to train a predictive model and extracts features for each class regarding the prediction items of the predictive model; and an optimal class number determination unit that determines the optimal number of classes for training the predictive model based on the features.

Images