Information processing device, information processing method, program

Abstract

The predictive model's accuracy in predicting the probability of the target state is insufficient. [Solution] The information processing device of this disclosure includes an acquisition unit that acquires training samples consisting of the characteristics and state of an object at each time point, and training data which is a collection of training samples, and a learning unit that uses the training data to machine-learn a prediction model that predicts the probability of a predetermined state of an object at a predetermined time. The learning unit machine-learns the prediction model using a loss function that includes a first loss based on the predicted value by the prediction model for the training samples, and a second loss based on the value of a hazard function obtained by automatically differentiating the prediction model for the training samples. This makes it possible to support physician decision-making by predicting the patient's state using the prediction model.

Description

Technical Field

[0001] The present disclosure relates to an information processing apparatus, an information processing method, and a program.

Background Art

[0002] As described in Patent Document 1, predicting the state of a patient's illness using a prediction model has been performed. Specifically, Patent Document 1 describes that a prediction model for predicting a state transition probability of a patient is machine-learned using medical characteristics and states of a patient at a predetermined time, and the probability of the state of the patient is predicted using such a prediction model. At this time, in Patent Document 1, machine learning of the prediction model is performed using a loss function including a likelihood term that imposes a penalty on the fact that the state predicted by the prediction model does not match the dataset.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, in the technique described in Patent Document 1 above, an approximation is performed in which a time-direction interval is discretized as a set of small intervals. As a result, there arises a problem that the accuracy of predicting the probability of the state of the object by the prediction model due to the approximation is insufficient.

[0005] For this reason, one object of the present disclosure is to solve the above-described problem that the accuracy of predicting the probability of the state of the object by the prediction model is insufficient.

Means for Solving the Problems

[0006] An information processing apparatus according to one aspect of the present disclosure is An acquisition unit that acquires training samples consisting of the characteristics and state of the target at each time point, and training data which is a collection of said training samples, A learning unit that uses the aforementioned training data to machine-learn a predictive model that predicts the probability of a predetermined state at a predetermined time for the subject, Equipped with, The learning unit performs machine learning on the prediction model using a loss function that includes a first loss based on the predicted value by the prediction model for the training samples and a second loss based on the value of the hazard function obtained by automatically differentiating the prediction model for the training samples. This is the structure it takes. Furthermore, the information processing method, which is one form of this disclosure, Information processing device, We acquire training samples consisting of the characteristics and state of the target at each time point, and training data which is a collection of said training samples. When machine learning a predictive model to predict the probability of a predetermined state at a predetermined time for the target using the aforementioned training data, the predictive model is machine-learned using a loss function that includes a first loss based on the predicted value by the predictive model for the training samples and a second loss based on the value of the hazard function obtained by automatically differentiating the predictive model for the training samples. This is the structure it takes. Furthermore, one form of this disclosure is a program, In an information processing device, We acquire training samples consisting of the characteristics and state of the target at each time point, and training data which is a collection of said training samples. When machine learning a predictive model to predict the probability of a predetermined state at a predetermined time for the target using the aforementioned training data, the predictive model is machine-learned using a loss function that includes a first loss based on the predicted value by the predictive model for the training samples and a second loss based on the value of the hazard function obtained by automatically differentiating the predictive model for the training samples. To execute the process This is the structure it takes. [Effects of the Invention]

[0007] This disclosure, configured as described above, can improve the accuracy of predicting the probability of a given state using a predictive model. [Brief explanation of the drawing]

[0008] [Figure 1] This figure shows an example of the state transitions of the predicted target in this disclosure. [Figure 2] This is a block diagram showing an example of the configuration of the information processing device related to this disclosure. [Figure 3] This flowchart shows an example of the processing operation of the information processing device related to this disclosure. [Figure 4] This flowchart shows an example of the processing operation of the information processing device related to this disclosure. [Figure 5] This flowchart shows an example of the processing operation of the information processing device related to this disclosure. [Figure 6] This block diagram shows an example of the hardware configuration of the information processing device related to this disclosure. [Figure 7] This is a block diagram showing an example of the configuration of the information processing device related to this disclosure. [Modes for carrying out the invention]

[0009] <First Embodiment> A first embodiment of this disclosure will be described with reference to the drawings. The drawings may be relevant to any embodiment.

[0010] [composition] As an example, the information processing device 10 of this disclosure is used to machine-learn a predictive model that predicts the probability of a person's state transitions. In addition, the information processing device 10 of this disclosure is used to predict the probability of a person's state transitions using the machine-learned predictive model.

[0011] Here, in the present embodiment, the state of the person who predicts the probability is assumed to be the state of the patient's illness. As an example, as shown in FIG. 1, there are three states: state 1: healthy, state 2: mild illness, and state 3: illness. That is, in the present embodiment, a prediction model is generated to predict the probability that the patient transitions from a predetermined state to another state among the above three states as shown in FIG. 1, and the state of the patient is predicted using such a prediction model. Thereby, it is possible to support the decision-making by medical staff such as a doctor who examines the patient.

[0012] However, in the present disclosure, the state for predicting the probability is not limited to the state such as the patient's illness, and the target for predicting the state is also not limited to a person. For example, the target for prediction may be equipment such as industrial machinery installed in a factory, and the state for predicting the probability may be a state such as a failure of the equipment. Further, the state for predicting the probability is not limited to the three states as described above, and may be a larger number of multi-states.

[0013] The information processing apparatus 10 in the present embodiment is composed of one or more information processing apparatuses including an arithmetic unit and a storage unit. As shown in FIG. 1, the information processing apparatus 10 includes an acquisition unit 11, a learning unit 12, and a prediction unit 13. Each function of the acquisition unit 11, the learning unit 12, and the prediction unit 13 can be realized by the arithmetic unit executing a program for realizing each function stored in the storage unit. Further, the information processing apparatus 10 includes a data storage unit 16 and a model storage unit 17 composed of a storage unit. Each configuration will be described in detail below.

[0014] The acquisition unit 11 acquires a sample composed of the characteristics and state of a target patient at each time point and data that is a set of such samples. Note that the characteristics include not only those that change over time but also those that do not change over time, as exemplified below. Specifically, the acquisition unit 11 acquires, as the characteristics of the patient, personal data such as gender and age, and vital data such as blood pressure values and heart rates at a predetermined time point. As the state of the patient, the acquisition unit 11 acquires which of the three above-described states the patient is in at a predetermined time point, or the probability (degree) of the three states. Then, the acquisition unit 11 stores the data in which the acquired characteristics and state of the patient are associated with time in the data storage unit 16. At this time, as will be described later, the acquired data is divided into learning data used when machine-learning a prediction model and verification data used when verifying the learning result.

[0015] Also, as will be described later, after machine-learning a prediction model, when predicting the state of a predetermined patient using such a prediction model, the acquisition unit 11 acquires prediction data that is the characteristics (such as gender, age, blood pressure value, heart rate, etc.) of the target patient at a predetermined time point.

[0016] The learning unit 12 machine-learns a prediction model that predicts the probability of a predetermined state of a patient at a predetermined time point using the above-described data. That is, the learning unit 12 machine-learns a prediction model that takes time and the characteristics of the patient (such as gender, age, blood pressure value, heart rate, etc.) as inputs and outputs the state transition probability of the patient at that time. At this time, the output by the prediction model is configured to output the probabilities of the three above-described states. Here, as shown in Equation 1, the prediction model is expressed as a vector of the state transition probability P at time t. Thus, since the output of the prediction model is a state transition probability, it is necessary to design it to satisfy the constraint conditions as shown in Equation 2. As an example of a specific design, it is conceivable to use the softmax function.

Equation

Equation

[0017] Specifically, the learning unit 12 first initializes the weight parameters set for the prediction model constructed using a neural network, etc., and also initializes the evaluation values ​​that serve as the basis for evaluating the prediction model. Then, the learning unit 12 divides the data stored in the data storage unit 16 into training data and validation data. For example, the learning unit 12 divides the data into training data and validation data in a predetermined proportion randomly.

[0018] The learning unit 12 then inputs the time of a predetermined learning sample from the learning data and the patient's characteristics (gender, age, blood pressure, heart rate, etc.) into the prediction model. The learning unit 12 sets a loss function L that includes the error between the predicted probability of each state output by the prediction model and the state of the corresponding learning sample (which state it is, or the probability of each state). The learning unit then calculates the weight parameters of the prediction model using backpropagation to minimize this loss function L. At this time, the loss function L, as shown in Equation 3 below, is a loss corresponding to the prediction error described above, and includes a first loss L1 corresponding to the first constraint on the state transition probability and a second loss L2 corresponding to the second constraint on the state transition probability. The loss function L is set by the sum of these two losses.

number

[0019] Here, the first loss L1 is expressed as a value based on the negative log-likelihood, calculated using the state transition probabilities, which are the output of the prediction model. However, the first loss L1 can be any value as long as it is based on the output of the prediction model for the training data.

[0020] Furthermore, the second loss L2 is a loss based on the hazard function λ obtained by automatically differentiating the prediction model, and is set to a value that satisfies the constraints of the hazard function λ, as shown in Equation 4. Specifically, the second loss L2 is set to be larger when the value of the hazard function λ for the training samples is negative, in order to impose a penalty when the hazard function λ < 0. Note that the hazard function from state i to j can be calculated by automatically differentiating the state transition function P, as shown in Equation 5.

number

number

[0021] By setting the loss function L as described above and using machine learning to minimize it, the constraints on state transition probabilities mentioned above can be satisfied. As a result, the degree of freedom in the structure of the prediction model increases, and accuracy can be improved. Furthermore, it can handle complex constraints and appropriately manage situations where there are multiple states to predict.

[0022] Here, regarding the second loss L2 of the loss function L described above, it is not limited to setting it for all training samples in the training data during machine learning, but may be set for only some of the training samples in the training data. For example, training samples from a few time points within the target period in each training cycle may be extracted, and a hazard function λ may be calculated only for the training samples at the extracted time points, and the second loss L2 may be set to impose a penalty if λ < 0. In other words, in this case, for the training samples at the extracted time points, a loss function L consisting of the sum of the first loss L1 and the second loss L2 will be set, and for the training samples at time points that were not extracted, a loss function L consisting only of the first loss L1 will be set.

[0023] Furthermore, the learning unit 12 may add a third loss L3 to the loss function L described above and train the prediction model to minimize this loss function L. In this case, the value of the third loss L3 is set according to the time relationship of the training samples and the magnitude relationship of the predicted values ​​by the prediction model for the training samples. For example, there are first and second training samples that occur one after the other. If the time of the first training sample is earlier, it is assumed that the predicted value of the probability of a certain state (e.g., illness) by the prediction model for the first training sample will be greater than the predicted value for the second training sample, which occurs later. In situations where the above-mentioned relationship does not exist, that is, when the predicted value for the first training sample is smaller than the predicted value for the second training sample, a third loss L3 is added to the loss function L, which has a larger value in such cases. By performing machine learning in this way, it is possible to obtain a prediction model that can make more robust predictions.

[0024] Furthermore, the correspondence between the time-series relationships of the training samples and the magnitude of the predicted values ​​by the prediction model for those training samples is not necessarily limited to the relationship described above. For example, depending on the state, the correspondence between the time-series relationships and the magnitude of the predicted values ​​may be set in the opposite direction to the above, and if such a correspondence does not exist, a third loss L3 with a larger value may be set.

[0025] As described above, the learning unit 12 calculates and updates the weight parameters of the prediction model using machine learning, and then evaluates the prediction model with these updated weight parameters. Specifically, the learning unit 12 calculates an evaluation value for the prediction model with updated weight parameters using validation data, and if the evaluation value on the validation data is better than the baseline evaluation value, it saves the updated weight parameters to be adopted into the prediction model. In addition, the learning unit 12 also saves the improved evaluation value so that it becomes the previous evaluation value for the next evaluation and is used as the baseline evaluation value. At this time, the evaluation value is calculated based on the prediction error of the prediction model on the validation data, and as an example, the smaller the error, the higher the evaluation value calculated.

[0026] Then, the learning unit 12 terminates machine learning of the predictive model when learning converges, such as when there is no further improvement in the evaluation value, or when a predetermined number of learning iterations have been completed. The learning unit 12 saves the predictive model with the weight parameters set at the time machine learning is terminated to the model storage unit 17.

[0027] The prediction unit 13 takes the patient's characteristics (gender, age, blood pressure, heart rate, etc.) at a predetermined time as prediction data and inputs them to the prediction model stored as described above, and obtains a predicted value of the patient's state transition probability, which is the output from the prediction model. The prediction unit 13 then outputs the predicted value to a display device or the like.

[0028] [Operation] Next, the processing operations of the information processing device 10 will be explained. First, the information processing device 10 collects data from multiple patients, consisting of training samples with characteristics and states at each time point, and a collection of such training samples (Step S1 in Figure 3). Then, the information processing device 10 uses the data to machine-learn a predictive model (Step S2 in Figure 3). Here, the operation of the information processing device 10 in machine-learning a predictive model will be explained with reference to Figure 4.

[0029] First, the information processing device 10 initializes the weight parameters set for the prediction model constructed using a neural network or the like, and initializes the evaluation values ​​used to evaluate the prediction model. The information processing device 10 also divides the collected data into training data and validation data (step S11 in Figure 4). Then, using the training data, the information processing device 10 repeats the processing of Loop 1 shown in steps S12 to S19 in Figure 4 until learning converges, such as when there is no improvement in the evaluation values ​​described later, or until a predetermined number of training iterations are completed.

[0030] Then, within Loop 1, the information processing device 10 repeats the processing of Loop 2, steps S13 to S17 in Figure 4, for the number of training samples included in the training data. Specifically, the information processing device 10 takes the time of the training sample and the patient's characteristics (gender, age, blood pressure, heart rate, etc.) as input, and performs machine learning on a prediction model that outputs the patient's state transition probability at that time, and updates the weight parameters of the prediction model. At this time, the information processing device 10 sets a loss function L which is the sum of a first loss L1 corresponding to the prediction error and the first constraint condition of the state transition probability, and a second loss L2 corresponding to the second constraint condition of the state transition probability (steps S14, S15 in Figure 4), and updates the weight parameters of the prediction model to minimize this loss function L (step S16 in Figure 4).

[0031] Here, since the output of the prediction model is the state transition probability, it is designed to satisfy the constraints shown in equation 2 above, and the first loss L1 is expressed as a loss value based on the negative log-likelihood corresponding to the prediction error calculated using the state transition probability. The second loss L2 is a loss based on the hazard function λ obtained by autodifferentiating the prediction model, and is set to a value that satisfies the constraints of the hazard function λ, as shown in equation 4 above. Specifically, the second loss L2 is set to be larger when the value of the hazard function λ for the training data is negative, in order to impose a penalty when the hazard function λ < 0. Note that the second loss L2 is not limited to being set during machine learning for all training samples, but may be set only for some of the training samples in the training data.

[0032] Furthermore, the information processing device 10 may add a third loss L3 to the loss function L described above and train the prediction model to minimize this loss function L. In this case, a third loss L3 is set, which will have a larger value, when there is no predetermined correspondence between the time series relationships of the training samples and the magnitude relationships of the predicted values ​​by the prediction model for the training samples.

[0033] Next, when the information processing device 10 finishes processing Loop 2, which is the repetition of machine learning for the training samples (steps S13-S17 in Figure 4), it evaluates the prediction model with updated weight parameters (step S18 in Figure 4). At this time, the information processing device 10 calculates an evaluation value of the prediction model with updated weight parameters using the validation data. If the evaluation value on the validation data is an improvement over the previous evaluation value, it adopts and saves the updated weight parameters in the prediction model, and also saves the improved evaluation value. When the learning described above converges, or when a predetermined number of learning iterations are completed, the information processing device 10 terminates the processing of Loop 1 (step S19 in Figure 4). The information processing device 10 then stores the prediction model with the weight parameters set at the time the machine learning of the prediction model is completed (step S3 in Figure 3).

[0034] Subsequently, the information processing device 10 predicts the patient's condition using a machine learning-prepared predictive model. At this time, the information processing device 10 reads the stored machine learning-prepared predictive model (step S21 in Figure 5) and obtains predictive data from the patient consisting of the patient's characteristics at a predetermined time (gender, age, blood pressure, heart rate, etc.) (step S22 in Figure 5). Then, the information processing device 10 inputs the patient's characteristics, which are the acquired predictive data, into the predictive model and predicts the predicted value of the patient's state transition probability, which is the output from the predictive model (step S23 in Figure 5). The information processing device 10 outputs the prediction results, such as the predicted value of the patient's state transition probability, from a display device or the like, as needed (step S24 in Figure 5).

[0035] As described above, the information processing device 10 in this disclosure can satisfy the constraints on state transition probabilities by performing machine learning on a loss function L that includes the first loss L1 and the second loss L2 described above, and minimizing it. As a result, the degree of freedom in the structure of the prediction model is increased, and accuracy can be improved. Furthermore, it can handle complex constraints and can respond even when there are multiple states to predict.

[0036] In the above example, we used machine learning to create a predictive model that predicts a patient's condition based on their characteristics, but as mentioned above, it can also be applied to predicting the condition of equipment based on its characteristics. In this case, the equipment characteristics are, for example, operating time, temperature, current value, etc., and the equipment condition is, normal, minor malfunction, malfunction, etc. However, the object whose condition is to be predicted is not limited to people or equipment, but can be any object, and the condition for which the probability is predicted can be any condition depending on the object.

[0037] <Second Embodiment> Next, a second embodiment of the present disclosure will be described with reference to the drawings. This embodiment shows an outline of the information processing device, etc., described in the above-described embodiment. Note that the drawings may be relevant to any embodiment.

[0038] First, the hardware configuration of the information processing device 100 in this disclosure will be described. The information processing device 100 is composed of a general information processing device, and as an example, it is equipped with the following hardware configuration as shown in Figure 6. ·CPU(Central Processing Unit)101(Arithmetic unit) ROM (Read Only Memory) 102 (Storage Device) • RAM (Random Access Memory) 103 (Storage Device) • Program group 104 loaded into RAM 103 • Storage device 105 for storing the program group 104 • Drive device 106 for reading and writing to external storage medium 110 of the information processing device. • Communication interface 107 connecting to a communication network 111 outside the information processing device. • Input / output interface 108 for data input and output. • Bus 109 connecting each component

[0039] Figure 6 shows an example of the hardware configuration of the information processing device 100, and the hardware configuration of the information processing device is not limited to the case described above. For example, the information processing device may consist of only a part of the configuration described above, such as not having a drive device 106. In addition, the information processing device may use a GPU (Graphic Processing Unit), DSP (Digital Signal Processor), MPU (Micro Processing Unit), FPU (Floating point number Processing Unit), PPU (Physics Processing Unit), TPU (Tensor Processing Unit), quantum processor, microcontroller, or a combination thereof instead of the CPU described above.

[0040] The information processing device 100 can be equipped with the acquisition unit 121 and learning unit 122 shown in Figure 7 by having the CPU 101 acquire the program group 104 and execute it. The program group 104 is, for example, stored in advance in the storage device 105 or ROM 102, and the CPU 101 loads it into the RAM 103 and executes it as needed. The program group 104 may also be supplied to the CPU 101 via the communication network 111, or it may be stored in advance in the storage medium 110, and the drive device 106 reads the program and supplies it to the CPU 101. However, the acquisition unit 121 and learning unit 122 described above may be constructed with dedicated electronic circuits to realize such means.

[0041] The acquisition unit 121 acquires training samples consisting of the characteristics and state of the target at each time point, and training data which is a collection of said training samples. The learning unit 122 uses the training data to machine-learn a prediction model that predicts the probability of the target being in a predetermined state at a predetermined time point. Furthermore, the learning unit 122 machine-learns the prediction model using a loss function that includes a first loss based on the predicted value by the prediction model for the training samples, and a second loss based on the value of the hazard function obtained by automatically differentiating the prediction model for the training samples.

[0042] This disclosure, by being configured as described above, can satisfy the constraints on state transition probabilities by setting and minimizing a loss function that includes the first and second losses and performing machine learning. As a result, the degree of freedom in the structure of the prediction model is increased, and accuracy can be improved. Furthermore, it can handle complex constraints and can accommodate situations where there are multiple states to predict.

[0043] Furthermore, at least one of the functions of the acquisition unit 121 and the learning unit 122 described above may be performed on an information processing device installed and connected to any location on the network, that is, it may be performed using so-called cloud computing.

[0044] Furthermore, the aforementioned programs can be stored and supplied to a computer using various types of non-transitory computer-readable media. Non-transitory computer-readable media include various types of tangible storage media. Examples of non-transitory computer-readable media include magnetic recording media (e.g., flexible disks, magnetic tapes, hard disk drives), magneto-optical recording media (e.g., magneto-optical disks), CD-ROMs (Read Only Memory), CD-Rs, CD-R / Ws, and semiconductor memory (e.g., mask ROMs, PROMs (Programmable ROMs), EPROMs (Erasable PROMs), flash ROMs, and RAMs (Random Access Memory)). Programs may also be supplied to a computer using various types of transient computer-readable media. Examples of transient computer-readable media include electrical signals, optical signals, and electromagnetic waves. Transitory computer-readable media can be supplied to a computer via wired communication channels such as electric wires and optical fibers, or via wireless communication channels.

[0045] Although the present disclosure has been described above with reference to embodiments, the present disclosure is not limited to the embodiments described above. Various modifications to the structure and details of the present disclosure are possible, as can be understood by those skilled in the art within the scope of the present disclosure. Furthermore, each of the embodiments described above can be combined with other embodiments as appropriate.

[0046] <Note> Some or all of the above embodiments may also be described as follows. The general configuration of the information processing apparatus, information processing method, and program in this disclosure is described below. However, this disclosure is not limited to the configuration described in the following notes. Furthermore, some or all of the configurations and functions described in Appendices 2 to 8, which are dependent on Appendice 1 below, may also be dependent on Appendices 9 and 10 in the same way as Appendices 2 to 8. Moreover, not limited to Appendices 1, 9, and 10, some or all of the configurations and functions described as appendices may also be dependent on similar hardware, software, various recording means for recording software, or systems, without departing from the embodiments described above. (Note 1) An acquisition unit that acquires training samples consisting of the characteristics and state of the target at each time point, and training data which is a collection of said training samples, A learning unit that uses the aforementioned training data to machine-learn a predictive model that predicts the probability of a predetermined state at a predetermined time for the subject, Equipped with, The learning unit performs machine learning on the prediction model using a loss function that includes a first loss based on the predicted value by the prediction model for the training samples and a second loss based on the value of the hazard function obtained by automatically differentiating the prediction model for the training samples. Information processing device. (Note 2) The information processing device described in Appendix 1, The learning unit performs machine learning on the prediction model to minimize the loss function, which includes the sum of the first loss and the second loss. Information processing device. (Note 3) The information processing device described in Appendix 2, The learning unit performs machine learning on the prediction model to minimize the loss function, which includes a second loss whose value increases when the value of the hazard function for the training samples is negative. Information processing device. (Note 4) The information processing device described in Appendix 1, The learning unit extracts training samples from the training data at multiple time points, and uses the loss function, which includes the second loss based on the value of the hazard function for each of the extracted training samples, to machine-learn the prediction model. Information processing device. (Note 5) The information processing device described in Appendix 2, The learning unit performs machine learning on the prediction model to minimize the loss function, which is further calculated by adding a third loss that corresponds to the temporal relationship of the training samples and the relative magnitudes of the predicted values ​​by the prediction model for the training samples. Information processing device. (Note 6) The information processing device described in Appendix 5, The learning unit performs machine learning on the prediction model to minimize the loss function, which is further increased by adding a third loss, when the temporal relationship between the learning samples and the magnitude relationship between the predicted values ​​of the prediction model for the learning samples does not conform to a predetermined relationship. Information processing device. (Note 7) The information processing device described in Appendix 1, The learning unit performs machine learning on the prediction model that predicts the probabilities of three or more states. Information processing device. (Note 8) The information processing device described in Appendix 1, The acquisition unit acquires the characteristics of the target at a predetermined time as prediction data. The system includes a prediction unit that predicts the probability of a predetermined state of the target based on the prediction model for the prediction data, Information processing device. (Note 9) Information processing device, We acquire training samples consisting of the characteristics and state of the target at each time point, and training data which is a collection of said training samples. When machine learning a predictive model to predict the probability of a predetermined state at a predetermined time for the target using the aforementioned training data, the predictive model is machine-learned using a loss function that includes a first loss based on the predicted value by the predictive model for the training samples and a second loss based on the value of the hazard function obtained by automatically differentiating the predictive model for the training samples. Information processing methods. (Note 9.1) The information processing method described in Appendix 9, The aforementioned information processing device The prediction model is trained to minimize the loss function, which includes the sum of the first loss and the second loss. Information processing methods. (Appendix 9.2) The information processing method described in Appendix 9.1, The aforementioned information processing device The predictive model is trained to minimize the loss function, which includes a second loss that is larger when the value of the hazard function for the training samples is negative. Information processing methods. (Appendix 9.3) The information processing method described in Appendix 9, The aforementioned information processing device The predictive model is trained using machine learning by extracting training samples from the training data at multiple time points, and using the loss function that includes the second loss based on the value of the hazard function for each of the extracted training samples. Information processing methods. (Appendix 9.4) The information processing method described in Appendix 9.1, The aforementioned information processing device The prediction model is trained to minimize the loss function, which is further calculated by adding a third loss that depends on the temporal relationship of the training samples and the relative magnitudes of the predicted values ​​by the prediction model for the training samples. Information processing methods. (Note 9.5) The information processing method described in Appendix 9.1, The aforementioned information processing device The characteristics of the aforementioned target at a predetermined time are acquired as prediction data. The prediction model predicts the probability of a predetermined state for the target based on the prediction data. Information processing methods. (Note 10) In an information processing device, We acquire training samples consisting of the characteristics and state of the target at each time point, and training data which is a collection of said training samples. When machine learning a predictive model to predict the probability of a predetermined state at a predetermined time for the target using the aforementioned training data, the predictive model is machine-learned using a loss function that includes a first loss based on the predicted value by the predictive model for the training samples and a second loss based on the value of the hazard function obtained by automatically differentiating the predictive model for the training samples. A program that executes a process. [Explanation of Symbols]

[0047] 10 Information Processing Devices 11 Acquisition Department 12. Learning Department 13 Prediction Section 16 Data Storage Unit 17 Model Memory Unit 100 Information Processing Devices 101 CPU 102 ROM 103 RAM 104 Program Groups 105 Storage device 106 Drive unit 107 Communication Interface 108 Input / Output Interfaces 109 Bus 110 Storage medium 111 Communication Network 121 Acquisition Department 122 Learning Department

Claims

1. An acquisition unit that acquires training samples consisting of the characteristics and state of the target at each time point, and training data which is a collection of said training samples, A learning unit that uses the aforementioned training data to machine-learn a predictive model that predicts the probability of a predetermined state at a predetermined time for the subject, Equipped with, The learning unit performs machine learning on the prediction model using a loss function that includes a first loss based on the predicted value by the prediction model for the training samples and a second loss based on the value of the hazard function obtained by automatically differentiating the prediction model for the training samples. Information processing device.

2. An information processing apparatus according to claim 1, The learning unit performs machine learning on the prediction model to minimize the loss function, which includes the sum of the first loss and the second loss. Information processing device.

3. An information processing apparatus according to claim 2, The learning unit performs machine learning on the prediction model to minimize the loss function, which includes a second loss whose value increases when the value of the hazard function for the training samples is negative. Information processing device.

4. An information processing apparatus according to claim 1, The learning unit extracts training samples from the training data at multiple time points, and uses the loss function, which includes the second loss based on the value of the hazard function for each of the extracted training samples, to machine-learn the prediction model. Information processing device.

5. An information processing apparatus according to claim 2, The learning unit performs machine learning on the prediction model to minimize the loss function, which is further calculated by adding a third loss that corresponds to the temporal relationship of the training samples and the relative magnitudes of the predicted values ​​by the prediction model for the training samples. Information processing device.

6. An information processing device according to claim 5, The learning unit performs machine learning on the prediction model to minimize the loss function, which is further increased by adding a third loss, when the temporal relationship between the learning samples and the magnitude relationship between the predicted values ​​of the prediction model for the learning samples does not conform to a predetermined relationship. Information processing device.

7. An information processing apparatus according to claim 1, The learning unit performs machine learning on the prediction model that predicts the probabilities of three or more states. Information processing device.

8. An information processing apparatus according to claim 1, The acquisition unit acquires the characteristics of the target at a predetermined time as prediction data. The system includes a prediction unit that predicts the probability of a predetermined state of the target based on the prediction model for the prediction data, Information processing device.

9. Information processing device, We acquire training samples consisting of the characteristics and state of the target at each time point, and training data which is a collection of said training samples. When machine learning a predictive model to predict the probability of a predetermined state at a predetermined time for the target using the aforementioned training data, the predictive model is machine-learned using a loss function that includes a first loss based on the predicted value by the predictive model for the training samples and a second loss based on the value of the hazard function obtained by automatically differentiating the predictive model for the training samples. Information processing methods.

10. In an information processing device, We acquire training samples consisting of the characteristics and state of the target at each time point, and training data which is a collection of said training samples. When machine learning a predictive model to predict the probability of a predetermined state at a predetermined time for the target using the aforementioned training data, the predictive model is machine-learned using a loss function that includes a first loss based on the predicted value by the predictive model for the training samples and a second loss based on the value of the hazard function obtained by automatically differentiating the predictive model for the training samples. A program that executes a process.

Images