Therapeutic effect prediction system and method

Abstract

The present invention provides a therapeutic effect prediction system and method that make it possible to improve user-friendliness. A therapeutic effect prediction system 1 comprises: a machine learning unit 132 that outputs a prescribed parameter from patient data which has been input; and a mathematical model unit 131 that calculates a therapeutic effect from prescription data and the prescribed parameter which has been output by the machine learning unit. The machine learning unit is trained to reduce the difference between treatment history data and the therapeutic effect which has been calculated by the mathematical model unit. Thus, it is possible to predict a therapeutic effect by making combined use of the machine learning unit and the mathematical model unit without preparing a data set which is a correct answer.

Description

Treatment effect prediction system and method

[0001] The present invention relates to a system and method for predicting therapeutic effect.

[0002] There is a demand for personalized medicine and precision medicine, which can understand the state of tumors and immune function within patients by analyzing patient data obtained before and during treatment (e.g., age, gender, smoking history, gene mutations, gene expression in the tumor environment, various immune ratios, amounts of various metabolites, CT image information, etc.), and provide medicines and treatments that are appropriate for each patient.

[0003] As shown in Non-Patent Document 1, in the field of cancer treatment, machine learning models have attracted attention as a method for deriving correlations between the response rates of therapeutic drugs such as immune checkpoint inhibitors and molecular targeted drugs and various patient data. However, machine learning models are generally not good at extrapolation, and there is a possibility that prediction accuracy will decrease for patients with characteristics that differ significantly from those of the group used for learning.

[0004] One challenge posed by such machine learning models is that, for example, in the field of radiation therapy, it is difficult to use machine learning models to explore new irradiation methods (e.g., irradiation dose, number of fractions, combination with drugs, etc.) that differ from conventional standard treatments in order to improve therapeutic efficacy. Furthermore, because machine learning models are black boxes, they lack the ability to explain results, and it is expected to be difficult to present appropriate alternatives when a treatment is determined to be ineffective.

[0005] To address these challenges with machine learning models, Non-Patent Document 2 discloses a method for predicting response rates using a mathematical model. The mathematical model is composed of a group of mathematical formulas describing human body mechanisms, such as changes in the number of tumor cells in the body and drug accumulation, which have been revealed through various basic human research. It predicts treatment prognosis by numerically calculating the temporal changes in the pathological condition during treatment. Therefore, it is possible to analyze the reasons why a treatment is effective (or ineffective) from the calculation process, and even if a treatment is determined to be ineffective, it is expected that appropriate alternatives can be considered. Furthermore, by basing the model on human body mechanisms, it is possible to make relatively valid predictions even for prescriptions that have never been used before.

[0006] The mathematical model contains multiple parameters that represent the biological diversity of patients, such as the proliferation ability of cancer cells and the strength of immune activity, and highly accurate predictions are only possible by setting an appropriate parameter set for each patient. Such biological diversity is thought to be indirectly reflected in patient data through various biological mechanisms. However, there are still many unknowns about the human body, making it difficult to construct mathematical formulas that link patient data with parameters.

[0007] To address this issue, Patent Document 1 discloses a response rate prediction model that is a hybrid of a mathematical model and a machine learning model. Patent Document 1, based on the hypothesis that a patient's biological diversity is also reflected in CT image information, obtains a parameter set for the mathematical model by inputting the patient's CT image information into a machine learning model. At this time, a correct parameter set is required for training the machine learning model. In Patent Document 1, the correct parameter set is determined using the least squares method or the like so that the mathematical model best reproduces the temporal transition of tumor cell count obtained by observation. The temporal transition of tumor cell count is estimated from CT image information taken daily.

[0008] Tan P, et al. , Deep learning predictors immune checkpoint inhibitor-related pneumonitis from pretreatment computed tomography images. Front Physiol. 2022 Jul 25;13:978222Friedrich T, Scholz M, Durante M. A Predictive Biophysical Model of the Combined Action of Radiation Therapy and Immunotherapy of Cancer. Int J Radiat Oncol Biol Phys. 2022 Jul 15;113(4):872-884.

[0009] WO2022 / 220300

[0010] The method of observing the temporal progression of a physical quantity, such as tumor cell count, and determining a correct parameter set so that a mathematical model best reproduces the observed value may be difficult to implement at some medical institutions. This is because such physical quantities cannot necessarily be obtained daily due to invasiveness and throughput considerations. For example, while tumor cell count is generally estimated from CT image information, as in Patent Document 1, CT image information for the same patient is not necessarily obtained daily. Furthermore, in some cases, it may be difficult to determine changes in tumor volume during treatment from CT image information.

[0011] Furthermore, when reproducing the transition of a physical quantity such as the number of tumor cells in a mathematical model, it may not be possible to determine a unique set of parameters that is the correct answer. For example, consider a simple mathematical model that expresses the number of tumor cells remaining in the body, n_tumor, as a single differential equation shown in the following equation 1.

[0012] Here, a and b are parameters that represent the proliferation rate of cancer cells and the effectiveness of anticancer drugs, respectively. When a-b = constant, it is clear that there are countless possible combinations of the parameter set (a, b). This can be interpreted as meaning that even in patients for whom anticancer drugs are ineffective, treatment can be effective if the cancer proliferation rate is slow, and conversely, even in cases where the cancer proliferation rate is fast, treatment can be effective in patients for whom anticancer drugs are effective.

[0013] The present invention has been made in view of the above-mentioned problems, and an object of the present invention is to provide a system and method for predicting therapeutic effect that can improve ease of use for users.

[0014] In order to solve the above problem, a treatment effect prediction system according to one aspect of the present invention comprises a machine learning unit that outputs predetermined parameters from input patient data, and a mathematical model unit that calculates the treatment effect from prescription data and the predetermined parameters output by the machine learning unit, and the machine learning unit is trained to reduce the difference between the treatment history data and the treatment effect calculated by the mathematical model unit.

[0015] According to the present invention, it is possible to predict the therapeutic effect by using a machine learning unit and a mathematical model unit in combination, without preparing a correct data set.

[0016] 1 is a configuration diagram of a treatment selection support system as a "treatment effect prediction system" according to a first embodiment; FIG. 2 is an explanatory diagram showing the functions and processing flow of the treatment selection support system; FIG. 3 is an example of a setting screen; FIG. 4 is an example of a screen for displaying a response rate; FIG. 5 is an example of a screen for displaying parameters; FIG. 6 is an example of a screen for displaying data on similar cases; FIG. 7 is an outline diagram of a hybrid prediction model; FIG. 8 is an explanatory diagram of a treatment selection support system according to a second embodiment; FIG. 9 is a flowchart of a process for outputting a predicted response rate and parameters according to a third embodiment; FIG. 10 is a flowchart of a process for outputting a predicted response rate and parameters according to a fourth embodiment; and FIG. 11 is a flowchart of a process for outputting a predicted response rate and parameters according to a fifth embodiment.

[0017] The following describes an embodiment of the present invention with reference to the drawings. As will be described later, the treatment selection support system according to this embodiment uses a hybrid prediction model in which a mathematical model unit having an arithmetic expression such as simultaneous differential equations and a machine learning unit (e.g., a neural network) are connected in series.

[0018] The machine learning unit learns the difference between the actual treatment effect (treatment history data) and the output result of the mathematical model unit as a loss function. The machine learning unit receives patient data as input and outputs predetermined parameters adjusted for each patient to the mathematical model unit. The mathematical model unit calculates a predetermined arithmetic formula using the predetermined parameters adjusted for each patient by the machine learning unit and the prescription data, and predicts and outputs a response rate indicating the effect of treatment prognosis.

[0019] Furthermore, the treatment selection support system according to this embodiment displays the predetermined parameters output from the machine learning unit on a screen in association with past treatment outcomes. This allows the user of the treatment selection support system to determine what factors determine the response rate predicted by the mathematical model unit. This allows the user to propose other treatments (alternative treatments) for patients with poor prognosis (patients who are not seeing any therapeutic benefit).

[0020] As a result, according to the treatment selection support system of this embodiment, the machine learning unit adjusts the parameters used by the mathematical model unit and provides them to the mathematical model unit, thereby enabling highly accurate prognosis prediction for each patient. As described above, the treatment selection support system provides interpretability to the prediction results, allowing the user to propose appropriate alternative treatments for patients determined to have a poor prognosis.

[0021] For example, the mathematical model unit includes predetermined parameters that can be adjusted for each patient, such as the cancer growth rate and the cytotoxicity of immune cells. The treatment selection support system according to this embodiment predicts treatment prognosis by appropriately adjusting these predetermined parameters. The machine learning unit learns the difference between the actual treatment outcome and the output result of the mathematical model unit as a loss function, and provides predetermined parameters adjusted for each patient using patient data (patient test data) as input to the mathematical model unit. This makes it possible to predict treatment prognosis based on patient data such as gene expression levels or blood metabolites, thereby supporting the provision of appropriate medical care for each patient.

[0022] The treatment selection support system according to this embodiment is, for example, a treatment selection support system that outputs predicted results of treatment prognosis, and includes a mathematical model unit that calculates the response rate using prescription data and predetermined parameters as input, and a machine learning unit that outputs the predetermined parameters using patient data, which is data obtained by examining a patient, as input, and the machine learning unit is trained to minimize the difference between the actual treatment outcome and the response rate calculated by the mathematical model unit.

[0023] The treatment selection support system according to this embodiment predicts treatment effects using a hybrid prediction model that combines a mathematical model unit and a machine learning unit, and provides the difference between the predicted result and the actual treatment effect (treatment outcome) as a loss function to the machine learning model used by the machine learning unit. Therefore, the treatment selection support system according to this embodiment does not require specific correct parameters for training the machine learning model. Furthermore, the treatment selection support system according to this embodiment presents information to the user about the specific parameters used in the mathematical model unit, thereby enabling the treatment effect prediction to be explained more clearly.

[0024] For example, even when patient data such as CT (Computed Tomography) image information cannot be obtained daily, the treatment selection support system according to this embodiment can predict treatment effects and further enhance the interpretability of the predictions, improving user usability. According to this embodiment, a hybrid prediction model is used, so that consideration of appropriate alternatives can be supported for patients who are determined to have a poor treatment effect (non-response). Furthermore, according to this embodiment, even for prescriptions that have not previously been proven, prognosis predictions with a relatively high degree of validity can be made based on the human body mechanisms in the mathematical model section.

[0025] Example 1 will be described using Figures 1 to 7. Figure 1 shows an example of the configuration of a treatment selection support system 1 as a "treatment effect prediction system." The treatment selection support system 1 can also be called, for example, a treatment policy formulation support system. The treatment selection support system 1 is configured with a device capable of various information processing, for example, an information processing device such as a computer, and includes a calculation device 110, an input device 140, an output device 150, a memory 130, a communication device 120, and a database 2. The calculation device 110 is connected to the input device 140, the output device 150, the memory 130, and the communication device 120. This connection method is not particularly limited, and may be a connection method via a LAN (Local Area Network) or a WAN (Wide Area Network) such as the Internet.

[0026] The arithmetic device 110 is a processor such as a CPU (Central Processing Unit), a GPU (Graphic Processing Unit), or an FPGA (Field-Programmable Gate Array), and constitutes a control unit that controls the entire treatment selection support system 1.

[0027] The input device 140 is a device that receives various information from an operator who operates the treatment selection support system 1, and is, for example, a mouse and a keyboard.

[0028] The output device 150 is a device that outputs various information such as the predicted results of treatment effects output from the treatment selection support system 1. The output device 150 is, for example, a monitor display, a printer, etc. Here, the output device 150 will be described mainly as an example of display output.

[0029] The memory 130 is configured from the same or a different recording medium, and records a program (computer program) that defines the operation of the arithmetic device 110, and various information used and generated by the arithmetic device 110. Examples of the various information include a mathematical model 131, a machine learning model 132, patient data, parameters used in the mathematical model, etc. These will be described later.

[0030] The recording medium may be, for example, a magnetic storage medium such as a hard disk drive (HDD), a semiconductor storage medium such as a random access memory (RAM), a read-only memory (ROM) and a solid state drive (SSD), or a combination of an optical disk and an optical disk drive such as a digital versatile disk (DVD). When the treatment selection support system 1 starts operating (for example, when the power is turned on), the arithmetic unit 110 reads a program from the memory 130 and executes the read program to perform various processes, thereby controlling the entire treatment selection support system 1.

[0031] The communication device 120 is a communication interface that can communicate with external devices, and is connected to, for example, a database 2. The database 2 stores, for example, patient data (examination data) 21 acquired during past treatments, prescription data, and treatment results (shown as prescription data 22 in FIG. 1 ). Details of patient data and prescriptions as examples of "treatment history data" will be described later. Treatment results include survival time, progression-free survival time, and the occurrence or non-occurrence of various side effects. The database 2 also stores, for past treatments, parameters of the mathematical model 131 output from the machine learning model 132 and prediction results of the response rate by the mathematical model 131.

[0032] A computer program for realizing the functions of the treatment selection support system 1 can be stored in a storage medium MM and installed in another computer (not shown). Conversely, the computer program stored in the storage medium MM can be transferred to and stored in the treatment selection support system 1. The storage medium MM can be any device that can non-temporarily store a computer program, such as an HDD or SSD. Note that instead of the storage medium MM, the computer program can be transferred via a communication network.

[0033] Figure 2 shows an example of a series of procedures and functions for a new cancer patient who is scheduled to receive radiation therapy and immune checkpoint inhibitor therapy in treatment selection using the treatment selection support system 1. Figures 3 to 6 show examples of the operation screen 3 displayed on the output device 150.

[0034] The entity (function) performing each process S1 to S5 is described within each process. The patient data acquisition unit F1, parameter calculation unit F2, prescription data acquisition unit F3, predicted response rate calculation unit F4, and calculation result output unit F5 are realized by the calculation device 110 executing a predetermined computer program. The patient data acquisition unit F1 acquires patient data. The parameter calculation unit F2 uses the machine learning model 132 to calculate and output predetermined parameters (hereinafter sometimes abbreviated as parameters) used in the mathematical formula model 131. The prescription data acquisition unit F3 acquires prescription data. The predicted response rate calculation unit F4 calculates the predicted response rate using the mathematical formula model 131. The calculation result output unit F5 displays and outputs the calculated predicted response rate in association with the parameters used to calculate the predicted response rate.

[0035] The machine learning model 132 and the parameter calculation unit 20 are examples of a "machine learning model unit." The predicted response rate calculation unit 40 and the mathematical model 131 are examples of a "mathematical model unit." The output device 150 and the calculation result output unit 50 are examples of a "display output unit."

[0036] First, the operator presses the patient data input button 311 in the setting tab 31 of the operation screen 3 shown in Fig. 3 to input patient data of a new patient acquired by any means into the treatment selection support system 1. The patient data is saved in the memory 130 (step S1).

[0037] Patient data includes, for example, age, gender, smoking history, the presence or absence of various gene mutations, the expression levels of various genes in the tumor environment, the expression levels of various proteins on the cell surface, the proportion of various immune cells, and the amounts of various metabolites. These values ​​are obtained by analyzing tumor samples and blood collected during a prior surgery using specialized equipment and reagents. However, this is not limited to this, as long as it represents the patient's biological characteristics. For example, CT image information, MR images, and PET images also qualify as patient data. Furthermore, patient data may include not only information acquired before the start of treatment, but also information acquired during treatment. Although the present embodiment describes patient data as being input by an operator, a testing device (not shown) for acquiring patient data and the treatment selection support system 1 may be connected via a communication device 120, and patient data may be automatically input from the testing device by inputting the patient's name or an identification number such as an ID instead.

[0038] The treatment selection support system 1 inputs patient data into the machine learning model 132 stored as a computer program in the memory 130, and outputs parameters 23 for the mathematical formula model 131 (step S2). Details of the machine learning model 132 and the learning method will be described later. The output parameters 23 are stored in the memory 130.

[0039] The operator presses the prescription input button 312 and inputs a prescription created by any means into the treatment selection support system 1 as prescription data 22. Hereinafter, this may be referred to as a prescription 23. The input prescription data 22 is saved in the memory 130 (step S3). Note that step S3 may be performed before step S1 or step S2. Herein, the prescription is information that describes the type of treatment to be administered to the patient, and includes, for example, the type of therapeutic agent, the daily dosage of each therapeutic agent, the number of radiation fractions, and the daily irradiation dose. Information on the radiation dose may be stored as an average value within a specific region of interest (ROI), or may be stored as a dose distribution, i.e., the value of each voxel on a CT image, or a dose-volume histogram for each ROI. Generally, the dose distribution is determined by dose calculation using treatment planning software (not shown), but the treatment selection support system 1 may also predict the dose distribution from past treatment results based on the ROI information on the input CT image information.

[0040] The operator presses the calculation start button 313 to instruct calculation of the patient's predicted response rate. The treatment selection support system 1 inputs the above-mentioned parameters 23 and prescription 22, also stored in the memory 130, into the mathematical model 131, which is stored as a program in the memory 130, and calculates the patient's predicted response rate (step S4).

[0041] An example of a method for calculating the predicted response rate using the mathematical model 131 is shown below. In this example, it is assumed that the number of tumor cells remaining in the body follows Poisson statistics. When the expected number of tumor cells remaining in the body at a time T_end preset by the operator is n_tumor(T_end), the response rate, i.e., the probability P of reducing the number of tumor cells to zero, is expressed by the following formula:

[0042] In the mathematical model of this embodiment, the differential equation of the expected value n_tumor of the number of tumor cells at time t is expressed as follows:

[0043] C_1 is a parameter representing the cancer growth rate per unit time, and is output from the machine learning model. λ_RT is the survival rate of cancer cells following radiation exposure at time t, and is expressed, for example, as the following equation using a linear-quadratic curve model:

[0044] α and β are parameters that represent the radiosensitivity of cancer cells. d represents the radiation dose to the tumor. λ_IMM is the survival rate of cancer cells related to immune cytocidal action. Based on the commonly known finding that the cytocidal effect of cancer immunity depends on the number of cytotoxic T-lymphocytes (CTLs) that infiltrate the tumor, λ_IMM is expressed by the following formula:

[0045] C_2 is a parameter representing the cytotoxic effect per unit CTL number and is output from the machine learning model. n_CTL(t) is the expected number of CTLs infiltrating into the tumor at time t, and the differential equation is expressed as follows:

[0046] C_3 is a parameter representing the number of CTLs induced by antigens released from cancer cells and is output from the machine learning model. μ is a parameter representing the inactivation rate of CTLs due to immunosuppressive mechanisms and is expressed by the following formula:

[0047] C_4 and C_5 are parameters that represent the CTL inactivation rates before and after administration of an immune checkpoint inhibitor, respectively, and are output from the machine learning model like other parameters.

[0048] λ_RT^((CTL)) is the survival rate of CTLs following radiation exposure at time t, and is expressed using a linear-quadratic curve model, for example, as shown in Equation 4. Here, α and β are parameters representing the radiosensitivity of cancer cells. In this case, the input dose d may be the radiation dose to the tumor, or may be the dose to organs and tissues, such as lymph nodes, aorta, and heart, where immune-related cells such as CTLs are abundant.

[0049] In this example, an example of a mathematical model was shown using cancer treatment with immune checkpoint inhibitors as an example, but the effects of the present invention can also be obtained with mathematical models constructed for other treatment modalities such as cytotoxic anticancer drugs, molecular targeted drugs, and radiation therapy.

[0050] The treatment selection support system 1 calculates the predicted response rate for new patients by numerically solving Equations 2 to 6 from t = 0 to t = T_end using the prescription and parameters obtained from the machine learning model as input. While the initial values ​​n_tumor (t = 0) and n_CTL (t = 0) are not specifically described in this example, they may be fixed values ​​or may be treated as parameters output from machine learning, similar to C_1 to C_5. In this example, α and β shown in Equation 4 are also described as fixed values ​​determined in advance through cell irradiation experiments, etc., but they may also be treated as parameters output from machine learning, similar to C_1 to C_5.

[0051] The calculated predicted response rate is stored in the memory 130 and displayed in the predicted response rate display section 321 of the response rate display 32, as shown in FIG. 4 (step S5). At this time, the time change of any variable included in the mathematical model (in this embodiment, the expected value n_tumor of the tumor cell count and the expected value n_CTL count) can be displayed in the mathematical model calculation result display section 322. The type of variable to be displayed can be switched using the variable selection button 323. Furthermore, the treatment selection support system 1 can connect to the database 2 via the communication device 120 to calculate the actual response rate from the treatment results of the same case (e.g., lung cancer) in the past and display it in the actual response rate display section 324. The treatment selection support system 1 can also display statistical information of the patient group used to calculate the actual response rate in the statistical information display section 325, based on the patient data 21 stored in the database 2.

[0052] By checking the predicted response rate for the new patient displayed on the output device 150 along with the actual response rate, the operator can determine the validity of the prescription that the operator himself / herself previously input into the treatment selection support system 1.

[0053] If the predicted response rate is lower than the actual result, it is possible to input a different prescription into the treatment selection support system 1 to see if the predicted response rate improves, or to consider other treatments. The parameters, predicted response rate, prescription, and patient data stored in the memory 130 are also stored in the database 2 via the communication device 120. In this embodiment, this can be used in conjunction with an optimization function that determines a prescription by repeated calculations so that the predicted response rate exceeds a threshold value preset by the operator.

[0054] As shown in FIG. 5 , the treatment selection support system 1 can display parameters output from the machine learning model, as well as the predicted response rate, in the parameter display section 331 of the parameter display tab 33. In the example of FIG. 5 , information about two parameters is displayed. When no particular distinction is made, the parameter display sections 331(1) and 331(2) can be referred to as the parameter display section 331. Similarly, the parameter display sections 331(1) and 331(2) can be written as the parameter display section 331. Similarly, the parameter selection buttons 334(1) and 334(2), which will be described later, can be referred to as the parameter selection buttons 334.

[0055] The displayed parameters can be switched using the parameter selection button 332. At this time, the treatment selection support system 1 can obtain parameters calculated for patients who have previously undergone treatment from the database 2 via the communication device 120 and display them superimposed on the parameter display section 331.

[0056] Parameters can also be displayed in two dimensions along the X and Y axes, as in the two-axis parameter display section 333. In this case, the displayed parameters can be switched using the X-axis and Y-axis parameter selection buttons 334. In this example, the parameters are displayed as a two-dimensional scatter plot, but three parameters can also be selected and displayed as a three-dimensional scatter plot. Furthermore, the treatment selection support system 1 may determine which patients responded and which did not respond based on the treatment results stored in the database 2, and display the parameters by distinguishing them by color, shape, etc.

[0057] As mentioned above, the parameters of the mathematical model represent the biological diversity of patients, such as the proliferation ability of cancer cells and the strength of immune activity. Therefore, even if the operator does not have specialized knowledge of, for example, molecular biology or the expression pathways of each gene, the operator can intuitively evaluate the pathological characteristics of a new patient by analyzing the parameters displayed on the output device 150 and comparing them with those of past patients. As a result, if the predicted response rate output from the treatment selection support system 1 is low, it is possible to appropriately consider changing the prescription or alternative treatment based on the patient's pathology.

[0058] 6, the treatment selection support system 1 can extract a group of patients having parameter sets similar to those of the patient for whom treatment is currently being selected, and display them in a similar case tab 34. The similarity is expressed as a similarity score. In this embodiment, the sum of squares of the differences between the parameters is assumed to be the similarity score, but the calculation method is not necessarily limited to this.

[0059] The treatment selection support system 1 can sort the patient list in descending order of treatment outcome. In this embodiment, survival time is shown as an example of treatment outcome, but other numerical values ​​may be used. By pressing the details button 341, the operator can check detailed treatment outcomes and patient data for each patient stored in the database 2. By pressing the display button 342, the operator can display the prescription for each patient stored in the database 2. This allows for efficient optimization of prescriptions and treatment selection.

[0060] 7 is a schematic diagram of a treatment effect prediction algorithm composed of a machine learning model and a mathematical model. Details of the machine learning model that outputs the parameters of the mathematical model and its learning procedure in the treatment selection support system 1 will be described below using FIG. 7. In this embodiment, a neural network will be used as an example of the machine learning model, but other machine learning models may also be used.

[0061] First, when the operator presses the model learning button 314 shown in FIG. 3 via the input device 140, the treatment selection support system 1 starts learning the machine learning model. The treatment selection support system 1 acquires patient data, prescriptions, and treatment outcomes of previously treated patients from the database 2 and stores them in the memory 130. Furthermore, it starts learning the machine learning model based on the following formula. A series of algorithms is stored in the memory 130 as a computer program. The target to be optimized by learning is the weight of each layer of the machine learning model. If the weight matrix of the i-th layer is w_i, the weight adjustment amount Δw_i can be expressed as the average value of the values ​​obtained from each learning data by the following formula 8.

[0062] Here, N is the number of patient data used for learning. Δw_i^((n)) is the adjustment amount of the weight matrix obtained from the nth data, and is a matrix of j(i) × j(i-1) where j(i) is the number of neurons in the i-th layer. However, if bias neurons are included, the matrix becomes j(i) × {j(i-1) + 1}. As with the general procedure shown below, Δw_i^((n)) can be found by minimizing the loss function E^((n)) using the steepest descent method.

[0063] Here, η is a hyperparameter known as the learning rate, which is set in advance by the operator.

[0064] In general neural network learning, the loss function E^((n)) is expressed using the squared difference between the output value from the machine learning model and the correct data. However, in this embodiment, the output value from the machine learning model is the parameter of the mathematical model, and the correct parameter is unknown. Therefore, as shown in Equation 10, the treatment selection support system 1 expresses the loss function E^((n)) using the squared difference between the predicted response rate P calculated by the mathematical model using the parameters output from the machine learning model during learning as input, and the treatment result P_meas.

[0065] Here, P_meas^((n)) indicates the treatment result of the nth data, and is defined as 1 for patients who responded and 0 for patients who did not. P(x_I^((n))) is the predicted response rate obtained from the mathematical model and Equation 2. x_I^((n)) is a simulation parameter output from a neural network with a total number of layers I, and is a j(I)-dimensional vector. From Equations 2 to 6, in this embodiment, x_I^((n)) = (C_1, C_2, C_3, C_4, C_5). Here, Equation 8 can be transformed into the following Equation 11.

[0066] The input vector x_i^((n)) to the i layer is expressed by the following equation 12 using the weight matrix w_i of the i layer and the output vector z_(i-1)^((n)) from the activation function of the i-1 layer.

[0067] As a result, Eq 7 showing the weight adjustment amount Δw_i is expressed by Equation 13.

[0068] z_(i-1) is a j(i-1) x N matrix with element z_(i-1)^((n)) in the row direction. E_i is a j(i) x N matrix with element ε_i^((n)) = -∂E^((n)) / ∂x_i^((n)) in the row direction. ε_(i=I)^((n)) of the final layer is expressed by the following equation 14.

[0069] Since ∂P(x_I^((n))) / ∂x_I^((n)) is a partial differential that cannot be solved analytically, the treatment selection support system 1 calculates it numerically by slightly changing the input parameters to the mathematical model around x_I^((n)). On the other hand, ε_i^((n)) in the intermediate layer can be calculated from E_(i+1) in the next layer and the weight matrix w_(i+1) by the following transformation:

[0070] ∂z_i^((n)) / ∂x_i^((n)) is the differential of the activation function of the hidden layer i. In this embodiment, a sigmoid function is used as the activation function, which is expressed by the following equation 16.

[0071] Combining Equation 14 and Equation 15, E_i of the intermediate layer is expressed by Equation 17 below.

[0072] The treatment selection support system 1 determines the weights w_i of the machine learning model by sequential calculation based on Equations 7 to 16, and stores them in the memory 130. When learning is completed using the above procedure, the treatment selection support system 1 calculates the parameters of the mathematical model using the weights w_i stored in the memory 130 when patient data is input to the machine learning model.

[0073] The weight w_i can be recalculated every time treatment for a patient is completed by updating the database 2. Note that, although a sigmoid function is used as the activation function for the intermediate layer in this embodiment, the same effect can be obtained even if another function such as the ReLU function is used.

[0074] The treatment selection support system 1 configured in this manner predicts treatment efficacy using a hybrid prediction model that combines a mathematical model and a machine learning model. This eliminates the need for a set of correct parameters required for training the machine learning model. Therefore, even when patient data such as CT image information cannot be obtained daily for various reasons, highly interpretable treatment efficacy predictions are possible using the hybrid prediction model. Furthermore, the use of the hybrid prediction model is expected to enable the consideration of appropriate alternatives even for patients who are determined to be non-responsive. Furthermore, by basing the model on the mechanisms of the human body, relatively valid prognosis predictions are possible even for prescriptions that have not previously been used.

[0075] Example 2 will be described using FIG. 8 . In the following examples, including this example, differences from Example 1 will be mainly described. Example 1 described a method for estimating all parameters of a mathematical formula model using one machine learning model. In contrast, in this example, as shown in FIG. 8 , multiple machine learning models 132(1), 132(2), and 132(3) are used to individually estimate the parameters 23. For example, the first machine learning model 132(1) estimates the first parameter 231(1) of the parameters 23, the second machine learning model 132(2) estimates the second parameter 231(2) of the parameters 23, and the third machine learning model 132(3) estimates the third parameter 231(3) of the parameters 23.

[0076] The patient data 211, 212, and 213 items, such as age, gender, and gene expression information, input to each of the machine learning models 132(1), 132(2), and 132(3) may overlap. In the example of Figure 8, certain data 214 included in the patient data 21 is used by both the first machine learning model 132(1) and the second machine learning model 132(2).

[0077] Each machine learning model 132(1), 132(2), 132(3) may output two or more parameters.

[0078] In this example, because the relationship between patient data and the parameters of the mathematical model is unknown, a method has been described in which all patient data is input and all parameters are output from a single machine learning model. However, there are cases in which the relationship between a specific item of data and a specific parameter is well known in advance (for example, when a correlation between the characteristics of cancer cells and a certain group of genes has been confirmed, but it is clear that there is no correlation with other genes). In such cases, separating the machine learning model and the data used for estimation for each parameter enables more accurate predictions.

[0079] This embodiment configured in this manner can also achieve the same effects as those of embodiment 1. Furthermore, in this embodiment, multiple machine learning models 132(1) to 132(3) are used to share and estimate the parameter 23, thereby further improving prediction accuracy. Furthermore, it is also possible to use the method in which some of the parameters are set to given values ​​and other parameters are estimated by machine learning models.

[0080] Example 3 will be described with reference to Fig. 9. Fig. 9 is a flowchart showing the process of displaying and outputting the predicted response rate and parameters.

[0081] When displaying the calculated predicted response rate on the screen (S51), the treatment selection support system 1 of this embodiment also displays parameters designated by the user on the screen (S52). That is, among the parameters used to calculate the predicted response rate, the parameters designated by the user are displayed on the screen together with the predicted response rate.

[0082] In this embodiment configured as above, parameters designated by the user among the parameters used in calculating the predicted response rate can be displayed on the screen, thereby improving usability for the user.

[0083] Example 4 will be described with reference to Figure 10. Figure 10 is a flowchart showing a process for displaying and outputting the predicted response rate and parameters. In this example, parameters used in calculating the predicted response rate that fall outside a predetermined range are displayed on the screen together with the predicted response rate (S52A). Parameters that fall outside the predetermined range include, for example, parameters that fall outside an average range or parameters that fall outside a range obtained from past cases.

[0084] This embodiment configured in this manner also achieves the same effects as those of Example 1. Furthermore, in this embodiment, it is possible to output parameters that are out of a predetermined range among the parameters used in calculating the predicted response rate, and therefore, for example, parameters that indicate abnormal values ​​can be displayed on the screen to draw the user's attention, thereby improving usability.

[0085] Example 5 will be described with reference to Fig. 11. Fig. 11 shows a process of displaying and outputting parameters that are associated in advance with the case related to the prediction, among the parameters used in calculating the predicted response rate (S52B).

[0086] The treatment selection support system 1 of this embodiment includes a case-parameter correspondence table T1 that associates parameters to be displayed for each case. This table T1 can be stored, for example, in the memory 130. When the treatment selection support system 1 displays the calculated predicted response rate (S51), it also displays the parameters used in calculating the predicted response rate that are associated with the case (S52B).

[0087] This embodiment configured as described above also achieves the same effects as those of Example 1. Furthermore, in this embodiment, parameters associated with cases related to the predicted response rate are displayed on the screen together with the predicted response rate, so that related parameters can be displayed on the screen to enhance explainability, thereby improving usability for the user.

[0088] It should be noted that the present invention is not limited to the above-described embodiments. Those skilled in the art can make various additions and modifications within the scope of the present invention. The above-described embodiments are not limited to the configuration examples shown in the accompanying drawings. The configurations and processing methods of the embodiments can be modified as appropriate within the scope of achieving the object of the present invention.

[0089] For example, it is clear that the present disclosure includes configurations expressed as follows:

[0090] (Representation 1) A treatment effect prediction system that predicts the results of treatment, comprising: a machine learning unit that outputs predetermined parameters from input patient data; and a mathematical model unit that calculates treatment effects from prescription data and the predetermined parameters output by the machine learning unit, wherein the machine learning unit learns to reduce the difference between treatment history data and the treatment effects calculated by the mathematical model unit.

[0091] (Representation 2) A therapeutic effect prediction system according to Representation 1, wherein the machine learning unit is composed of a neural network.

[0092] (Representation 3) A therapeutic effect prediction system according to Representation 1 or 2, wherein in the training of the neural network, the amount of adjustment of the weight of the final layer is calculated by making small changes to the parameters input to the mathematical model unit.

[0093] (Representation 4) A therapeutic effect prediction system according to any one of Representations 1 to 3, further comprising a display output unit that displays the predetermined parameters output from the machine learning unit and parameters obtained from treatment history data.

[0094] (Representation 5) A therapeutic effect prediction system described in any one of Representations 1 to 4, wherein the display output unit further displays and outputs the therapeutic effect calculated by the mathematical model unit and the therapeutic effect obtained from the treatment history data.

[0095] (Representation 6) A treatment effect prediction system described in any one of Representations 1 to 5, wherein the display output unit further extracts and displays predetermined treatment history data having the predetermined parameters and parameters within a predetermined range from the treatment history data based on the predetermined parameters output from the machine learning unit.

[0096] (Representation 7) A therapeutic effect prediction system described in any one of Representations 1 to 6, wherein the display output unit further rearranges the extracted specified treatment history data in order of treatment effect and displays and outputs prescription data corresponding to the specified treatment history data.

[0097] (Representation 8) A treatment effect prediction system described in any one of Representations 1 to 7, in which multiple machine learning units are provided, and the multiple machine learning units share the responsibility of outputting the specified parameters based on the patient data.

[0098] (Representation 9) A therapeutic effect prediction system described in any one of Representations 1 to 8, wherein the display output unit displays and outputs specified parameters from among the predetermined parameters output from the machine learning unit.

[0099] (Representation 10) A therapeutic effect prediction system described in any one of Representations 1 to 9, wherein the display output unit displays and outputs parameters out of the specified parameters output from the machine learning unit that fall outside a specified range.

[0100] (Representation 11) A therapeutic effect prediction system described in any one of Representations 1 to 10, in which the display output unit displays and outputs parameters corresponding to the case of the target patient among the specified parameters output from the machine learning unit.

[0101] (Representation 12) A method for predicting the effectiveness of a treatment using a treatment effect prediction system that predicts the outcome of a treatment, wherein the treatment effect prediction system acquires patient data, causes a machine learning unit to output predetermined parameters from the patient data, and causes a mathematical model unit to calculate the treatment effect from prescription data and the predetermined parameters output by the machine learning unit, and the machine learning unit learns to reduce the difference between treatment history data and the treatment effect calculated by the mathematical model unit.

[0102] 1: Treatment selection support system, 2: Database, 3: Screen, 21: Patient data, 22: Prescription data, 131: Mathematical model, 132: Machine learning model

Claims

1. A treatment effect prediction system for predicting the outcome of treatment, comprising: a machine learning unit that outputs predetermined parameters from input patient data; and a mathematical model unit that calculates treatment effects from prescription data and the predetermined parameters output by the machine learning unit, wherein the machine learning unit is trained to reduce the difference between treatment history data and the treatment effects calculated by the mathematical model unit.

2. The therapeutic effect prediction system according to claim 1, wherein the machine learning unit is configured as a neural network.

3. The therapeutic effect prediction system according to claim 2, wherein in the training of the neural network, the amount of adjustment of the weights in the final layer is calculated by slightly changing the parameters input to the mathematical model section.

4. The treatment effect prediction system according to claim 1, further comprising a display output unit that displays the predetermined parameters output from the machine learning unit and parameters obtained from treatment history data.

5. The therapeutic effect prediction system according to claim 4, wherein the display output unit further displays the therapeutic effect calculated by the mathematical model unit and the therapeutic effect obtained from the treatment history data.

6. The treatment effect prediction system described in claim 4, wherein the display output unit further extracts, based on the predetermined parameters output from the machine learning unit, predetermined treatment history data having the predetermined parameters and parameters within a predetermined range from the treatment history data, and displays and outputs the extracted data.

7. The therapeutic effect prediction system according to claim 6, wherein the display output unit further sorts the extracted predetermined treatment history data in order of treatment effect, and displays and outputs prescription data corresponding to the predetermined treatment history data.

8. The therapeutic effect prediction system according to claim 1, wherein the machine learning unit is provided in plurality, and the plurality of machine learning units share the responsibility of outputting the predetermined parameters based on the patient data.

9. The therapeutic effect prediction system according to claim 4, wherein the display output unit displays and outputs designated parameters from among the predetermined parameters output from the machine learning unit.

10. The therapeutic effect prediction system according to claim 4, wherein the display output unit displays and outputs parameters that are outside a predetermined range among the predetermined parameters output from the machine learning unit.

11. The therapeutic effect prediction system according to claim 4, wherein the display output unit displays and outputs parameters associated with the case of the target patient from among the predetermined parameters output from the machine learning unit.

12. A method for predicting the effectiveness of a treatment using a treatment effect prediction system that predicts the outcome of a treatment, wherein the treatment effect prediction system acquires patient data, causes a machine learning unit to output predetermined parameters from the patient data, and causes a mathematical model unit to calculate the treatment effect from prescription data and the predetermined parameters output by the machine learning unit, and the machine learning unit is trained to reduce the difference between treatment history data and the treatment effect calculated by the mathematical model unit.

Images