System and Method

Abstract

The system provides a method for identifying treatment options for patients corresponding to patient information. It also provides a method for identifying outcomes when a treatment option is implemented for a patient corresponding to patient information. [Solution] A system comprising a method identification means for identifying one or more treatment methods for a patient corresponding to acquired patient information, by requesting the execution of inference using a language model, with prompts including patient information and criteria for selecting a treatment method as input. A system comprising an outcome identification means for identifying the outcome when a treatment method is implemented for a patient corresponding to acquired patient information, by requesting the execution of inference using a language model, with prompts including patient information, a treatment method for a patient corresponding to the patient information as input, and a correspondence relationship between the patient information and the outcome when the treatment method is implemented for the patient corresponding to the patient information as input.

Description

Technical Field

[0001] The present disclosure relates to a program, apparatus, and system for assisting in formulating a treatment plan for a patient. More specifically, the present disclosure relates to a program, apparatus, and system that generate a treatment plan proposal using a correspondence table (matrix) or the like based on a new algorithm of treatment selection criteria such as treatment exclusion criteria and / or inclusion criteria in addition to existing treatment decision algorithms. Even more specifically, the present disclosure relates to a program, apparatus, and system that generate a treatment plan proposal using a large language model (LLM) and RAG (Retrieval-Augmented Generation) based on patient information. Furthermore, the present disclosure also relates to a program, apparatus, and system for assisting outcome prediction for assisting in formulating a treatment plan for a patient.

Background Art

[0002] The treatment of cancer (synonymous with malignant disease) is determined by the disease state (mainly the type and stage of cancer: stage) (existing treatment decision algorithm). The stage is classified according to the progression of the primary tumor, the degree of lymph node metastasis, and the presence or absence of distant metastasis, which is called the TNM classification. The treatment is mainly centered around the feasibility of surgery, and a general algorithm has been assembled to select from surgery, radiotherapy, and chemotherapy.

[0003] In actual clinical practice, the attending physician performs TNM classification in their head, and the treatment is selected according to the algorithm in the attending physician's memory. The treatment selected under limited knowledge and experience that varies from attending physician to attending physician is strongly influenced by the department of medicine specialized by the attending physician, and is not necessarily the optimal treatment for the patient. For example, if the attending physician is a medical oncologist who specializes in drug therapy, even if there is only one metastasis, it is often classified as having distant metastasis, and in reality, even if the metastasis is single and curative resection is possible, if the attending physician does not have that knowledge, the patient will miss the opportunity for cure.

[0004] Furthermore, in current clinical practice, patients must choose or accept treatments selected and presented by their attending physician without having any knowledge of those treatments themselves.

[0005] On the other hand, with advances in medical technology, the number of treatment options has increased, making treatment choices more complex, and the knowledge and experience of attending physicians are struggling to keep up. Disease conditions can now be subdivided, and not only have the types of treatments (major categories) increased, but the treatments themselves have also been subdivided (subcategories). For example, in cancer treatment, the major categories of treatment were surgery, radiation therapy, and systemic chemotherapy, but with the addition of new medical treatments such as immunotherapy, interventional radiology (IVR), photoimmunotherapy, and viral therapy, the number of categories has increased from three to seven. Regarding subcategories, for example, radiation therapy is classified into linear accelerators, more advanced high-precision X-rays, laser knives, brachytherapy, particle beam therapy (heavy ion therapy, proton therapy), BNCT, and theranostics, and it is expected that the types of treatments will continue to increase in the future.

[0006] Because the therapeutic effects and adverse events differ for each subdivided treatment, subdividing the disease state and selecting the appropriate treatment from among many subdivided treatments can improve survival rates, including complete cure, and reduce adverse events. Furthermore, complete cure (cure) has a positive impact on solving social problems and on the healthcare economy. Conversely, the current rough treatment selection based on the TNM classification by the attending physician misses opportunities for complete cure, resulting in social losses. This is true not only for cancer (malignant diseases) but also for benign diseases. For example, diabetic patients, who are numerous in number, are often treated by physicians other than specialists, and because the attending physician is unaware of the details of the treatment plan, the progression of the disease cannot be stopped, sometimes leading to dialysis being unavoidable.

[0007] Patent Document 1 discloses a treatment policy decision support device that assists in determining a treatment policy for a patient based on medical information indicating the content of medical procedures performed on each of multiple cancer patients, the device comprising a construction means for constructing a master database based on the medical information, each corresponding to one of the multiple patients and each containing multiple treatment records including cancer type and TNM classification value, the device detects one or more treatment records from the master database constructed by the construction means, each containing the cancer type and TNM classification value corresponding to the desired cancer type and TNM classification value, and returns the detected one or more treatment records to the issuer of the treatment record request.

[0008] However, it is difficult and unreasonable to expect all doctors to grasp the subdivided disease conditions, memorize the feasibility, effects, and adverse events of a vast number of treatments, and make accurate treatment selections. This is because, for doctors who are extremely busy with daily practice, it is difficult to catch up on knowledge outside of their own specialty in a departmentalized system, and it is also difficult to gain treatment experience. For example, it is difficult for an internist to correctly determine whether surgery is possible or to select the radiation therapy with the highest probability of curative treatment from among many radiation therapies.

[0009] To avoid such situations, clinical practice guidelines are created for each disease and condition, and optimal treatments are provided as guidelines for each more subdivided disease state. However, the amount of information in clinical practice guidelines that attending physicians must refer to is enormous, and there are limits to reading them all, memorizing the optimal treatments and their rationale for each disease and condition, and using them in treatment. For example, the clinical practice guidelines that a gastroenterologist must read include esophageal cancer practice guidelines, gastric cancer practice guidelines, liver cancer practice guidelines, pancreatic cancer practice guidelines, biliary tract cancer practice guidelines, colorectal cancer practice guidelines, small intestine cancer practice guidelines, peritoneal dissemination practice guidelines, and metastatic liver tumor practice guidelines. In addition, clinical practice guidelines are updated almost every year, and there are limits to keeping up with the latest treatment information.

[0010] This situation applies not only to cancer but also to chronic diseases in general, such as rheumatism, collagen diseases, diabetes, and dyslipidemia. Even with rheumatism and collagen diseases, the disease stages are subdivided, and guidelines for optimal treatment according to the stage of the disease are presented in clinical guidelines, etc. Furthermore, just like with cancer treatment, the guidelines themselves are updated daily.

[0011] The ideal treatment selection is based on the premise that medical treatment inevitably places a physical and mental burden on patients, and therefore, the principle is to provide treatment that is neither excessive nor insufficient. This involves presenting effective and feasible treatment options from a comprehensive list of treatment candidates that are updated daily in real time for each disease and condition, and then presenting the content, effects (treatment outcomes, etc.), and safety (adverse events, etc.) of each treatment, so that the patient understands the treatment and makes a choice in consultation with their doctor that aligns with their values. However, the current situation is as described in the background, and there is a significant gap between this ideal and reality. [Prior art documents] [Patent Documents]

[0012] [Patent Document 1] Japanese Patent Publication No. 2018-147274 [Overview of the project] [Problems that the invention aims to solve]

[0013] Examples of problems that the present invention aims to address include, but is not limited to, the following. The first problem that the present invention aims to address is to provide a system that can identify a treatment method for a patient corresponding to patient information. The second problem that the present invention aims to address is to provide a system that can identify the outcome when the treatment method is implemented for a patient corresponding to patient information. [Means for solving the problem]

[0014] Under the circumstances described above, the inventors have invented a new program, device, and system to support the formulation of treatment plans for patients by utilizing a large-scale language model (LLM), also known as generative AI, a real-time updated, up-to-date, and comprehensive treatment information database such as PubMed or clinical trial databases, and an AI / IT technology called RAG (Retrieval-Augmented Generation) that connects them.

[0015] This disclosure provides programs, devices, and systems to assist in the development of patient treatment plans. More specifically, this disclosure provides the following: [Embodiment 1] A program for assisting in the formulation of a patient's treatment plan, wherein one or more computers are instructed to perform the following steps: acquire patient information; generate a proposed treatment plan using a large-scale language model (LLM) based on the patient information, wherein the proposed treatment plan includes a list of one or more available treatment methods selected based on the patient information and treatment selection criteria (such as contraindication criteria and / or indication criteria) input to the LLM as prompts via RAG (Retrieval-Augmented Generation); and output the proposed treatment plan, wherein the proposed treatment plan further includes information including the advantages and disadvantages of the selected one or more available treatment methods, and / or predictive information on the probability of curative treatment and / or outcomes of the selected one or more available treatment methods. [Embodiment 2] A program for assisting in the formulation of a patient's treatment plan, which causes one or more computers to perform the steps of acquiring patient information, generating a proposed treatment plan based on the patient information, and outputting the proposed treatment plan. [Embodiment 3] The program according to Embodiment 2, wherein the step of generating a proposed treatment plan based on patient information includes the step of selecting one or more available treatment methods based on patient information and predefined treatment selection criteria. [Embodiment 4] The program according to any one of Embodiments 2 to 3, wherein the step of selecting one or more available treatment methods based on patient information and predefined treatment selection criteria is performed using a large-scale language model (LLM). [Embodiment 5] The program according to Embodiment 4, wherein treatment selection criteria are input to a large-scale language model (LLM) as prompts. [Embodiment 6] The program according to Embodiment 5, wherein treatment selection criteria are input into a large-scale word model (LLM) using RAG (Retrieval-Augmented Generation). [Embodiment 7] The program according to Embodiment 6, wherein treatment selection criteria are stored in a database. [Embodiment 8] The program according to Embodiment 7, wherein treatment selection criteria are stored in a database in the form of a correspondence table that associates patient information with available treatment methods. [Embodiment 9] A program according to any one of Embodiments 2 to 8, wherein the proposed treatment plan includes information on the advantages and disadvantages of one or more available treatment methods selected. [Embodiment 10] The proposed treatment plan includes information including predictive information on the probability of cure and / or outcome of one or more selected available treatment methods, as described in any one of the programs in Embodiments 2 to 9. [Embodiment 11] Patient information includes age, sex, disease name (disease name, primary site name, metastasis, recurrence, etc., complication name), left hepatic vein invasion, middle hepatic vein invasion, left hepatic vein invasion, vascular invasion, bile duct invasion, gastrointestinal invasion, Glisson's sheath invasion, metastasis to other organs, low liver function, S1, FLR (normal liver), FLR (impaired liver), massive hepatectomy (e.g., trisegmentectomy), massive hepatectomy (right hepatectomy), 3 segmental branches, left hepatic vein invasion, middle hepatic vein invasion, right hepatic vein invasion, portal vein invasion, distance from gastrointestinal tract (mm), blood vessels A program according to any one of Embodiments 1 to 10, comprising at least one piece of information selected from the group consisting of anatomical names and positional relationships such as invasion, bile duct invasion, gastrointestinal invasion, vascular proximity, bile duct proximity, gastrointestinal proximity, and Glisson's sheath proximity; stage of progression; histological and pathological findings; genetic information such as genetic abnormalities and their extent (including histological findings and gene expression levels); immunological findings such as the tumor microenvironment; and information regarding the degree of function. [Embodiment 12] A program according to any one of embodiments 1 to 11, wherein at least a portion of patient information is generated by automatic image interpretation. [Embodiment 13] Possible treatment methods include conventional surgery, advanced surgery, TSH stage 2 surgery, ALPPS, ALPTPS, venous combined surgery, robotic surgery, endoscopic surgery, radiotherapy, external radiotherapy, particle therapy, heavy particle therapy, proton therapy, high-precision X-ray therapy, IMRT, stereotactic radiotherapy, CyberKnife, brachytherapy, internal radiotherapy, small source therapy, Seranositics, BNCT boron neutron capture therapy, IVR, ablation, radiofrequency ablation, microwave ablation, cryotherapy, arterial infusion and embolization therapy, ethanol injection, HIFU high-intensity focused ultrasound, irreversible electroporation therapy (IRE), photoimmunotherapy, viral therapy, chemotherapy, hormonal therapy, molecularly targeted drugs, antibody-drug conjugates (ADC), immunotherapy, immune checkpoint inhibitors, fecal microbiota transplantation, cancer vaccines, CAR-T therapy, combined immunotherapy, gene therapy, stem cell transplantation, mRNA vaccines, CRISPR / Cas9, hyperthermia, TSH stage 2 surgery, ALPPS, venous combined surgery, conventional RFA, high-precision RFA, TACE, drug therapy, anticancer agents, immunosuppressants, investigational drugs, traditional Chinese medicine, transplantation (organ transplantation, bone marrow transplantation, etc.), nerve block, CART, dialysis, shunt (Denver shunt, etc.) nerve decompression, nerve transplantation, nerve repair, stem cell therapy, stem cell transplantation, inhaled drugs, intravenous drip, transdermal drugs, transdermal absorption preparations, transdermal patches, transdermal absorption preparations (TTS), eye drops, intracavitary administration, sustained administration, and new surgical procedures, psychotherapy (such as cognitive behavioral therapy CBT), physical therapy (electroconvulsive therapy: ECT, transcranial magnetic stimulation therapy: TMS, transcranial direct current stimulation therapy: tDCS, deep brain stimulation therapy (DBS), spinal cord stimulation therapy, etc.), diet therapy, exercise therapy, weight loss surgery, traditional Chinese medicine treatments such as acupuncture, and at least one treatment method selected from the group consisting of antioxidants, according to any one of Embodiments 1 to 12. [Embodiment 14] The step of generating a treatment plan based on patient information is performed using a machine learning model that outputs available treatment methods with patient information as input, according to the program of Embodiment 2.

[0016] Furthermore, the problem of the present invention is [1] A system comprising at least one computer device, the system comprising acquisition means for acquiring patient information, and method identification means for identifying one or more treatment methods for a patient corresponding to the acquired patient information based on the acquired patient information and selection criteria for treatment methods; [2] The system according to [1] above, wherein the method identification means identifies a surgical procedure, identifies a dosing regimen, identifies an immunotherapy or cell therapy regimen, or identifies a radiation therapy prescription; [3] The system according to [1] or [2] above, wherein the method identification means identifies one or more treatment methods for a patient corresponding to the acquired patient information by requesting execution of inference using a language model with the acquired patient information and a prompt including selection criteria for treatment methods as input; [4] The system according to any one of [1] to [3] above, wherein the method identification means identifies one or more treatment methods based on evaluation criteria for the patient's medical condition; [5] The system according to [3] or [4] above, wherein the method identification means identifies one or more treatment methods for a patient corresponding to the acquired patient information by requesting execution of inference using a language model with a prompt including selection criteria for treatment methods acquired from a knowledge base as input; [6] The system according to any one of [3] to [5] above, wherein the method identification means identifies one or more treatment methods for a patient corresponding to the acquired patient information by requesting execution of inference using a language model fine-tuned with information regarding selection criteria for treatment methods with the acquired patient information as input; [7] The system according to any one of [3] to [6] above, wherein the method identification means identifies one or more treatment methods for a patient corresponding to the acquired patient information by requesting execution of inference using a language model with a prompt including information regarding the patient's medical record as input and / or with information regarding an image of the patient's body taken as input; [8] A system according to any one of [3] to [7] above, wherein an input means for inputting a target outcome and a method identification means request the execution of inference using a language model, taking patient information and a prompt including the target outcome as input, to identify one or more treatment methods based on the target outcome for a patient corresponding to the acquired patient information; [9] A system according to any of [1] to [8] above, wherein the patient information includes image information relating to the body, the selection criteria can be determined from the image information relating to the body, and the method identification means identifies one or more treatment methods for a patient corresponding to the acquired patient information based on the acquired patient information and the selection criteria for the treatment method;

[10] A system according to any one of [1] to [9] above, wherein the patient information includes information that allows for the determination of the distance between the first and second organs in the patient's body, and the criteria for selecting a treatment method include information regarding the distance between the first and second organs as a criterion for enabling the selection of at least one treatment method;

[11] A system according to any of [1] to

[10] above, wherein patient information includes information that allows for the determination of the number of tumors in the patient's body, the size of the tumors, the stage of tumor progression, the anatomical location of the tumors, whether the site where the tumors are located is a common site, or whether the lesions are localized, and the criteria for selecting a treatment method include information that allows for the determination of the number of tumors in the patient's body, the size of the tumors, the stage of tumor progression, the anatomical location of the tumors, whether the site where the tumors are located is a common site, or whether the lesions are localized, as criteria for enabling the selection of at least one treatment method;

[12] A system comprising at least one computer device, comprising: an acquisition means for acquiring patient information; and a method identification means for identifying one or more treatment methods corresponding to the acquired patient information, based on a trained model that has been trained using patient information for machine learning as input data and treatment methods suitable for the patient information for machine learning as output data;

[13] The system according to

[12] , further comprising update means for updating a trained model by retraining based on new patient information for machine learning and a treatment method suitable for the new patient information for machine learning;

[14] A system comprising at least one computer device, comprising: an acquisition means for acquiring patient information; an input means for inputting target outcomes; and a method identification means for identifying one or more treatment methods corresponding to the acquired patient information and target outcomes, based on a trained model that has been trained using patient information for machine learning and outcomes as a result of performing the treatment method on patients corresponding to the patient information for machine learning as input data, and the treatment methods performed on patients corresponding to the patient information for machine learning as output data;

[15] The system according to

[14] , comprising new patient information for machine learning, outcomes as a result of performing the treatment method on patients corresponding to the new patient information for machine learning, and update means for updating a trained model by retraining based on the treatment method performed on patients corresponding to the patient information for machine learning;

[16] A system according to any one of [1] to

[15] , comprising a first information identification means for identifying information including the advantages and / or disadvantages of a treatment method based on a treatment method identified by a method identification means;

[17] A system according to any one of [1] to

[16] , comprising a second information identification means for identifying the survival rate, cure rate, recurrence rate, or local control rate resulting from the implementation of a treatment method identified by a method identification means;

[18] A system according to any one of [1] to

[17] , comprising a third information identification means for identifying a prediction of an outcome based on a treatment method identified by a method identification means;

[19] A system according to any one of [1] to

[18] , comprising an input means for inputting a patient's target outcome or the patient's risk tolerance, and a suitability determination means for identifying information that allows for understanding the degree to which a treatment method identified by a method determination means is appropriate to the input target outcome or risk tolerance;

[20] A system according to any of [1] to

[11] and

[16] to

[19] , wherein the criteria for selecting a treatment method are generated by requesting the execution of inference using a language model, with prompts containing text information about the treatment method as input;

[21] A system according to any of [1] to

[20] above, wherein the method identification means identifies only treatment methods covered by public health insurance;

[22] A system according to any one of [1] to

[21] above, wherein the method identification means can identify a treatment method for which public health insurance does not apply;

[23] A system according to any one of [1] to

[22] above, comprising a means for identifying missing information when there is insufficient information necessary to determine the appropriateness of a treatment method in the acquired patient information;

[24] A system comprising at least one computer device, comprising: acquisition means for acquiring patient information and a treatment method for a patient corresponding to said patient information; and outcome identification means for identifying a predicted outcome when the treatment method is performed on a patient corresponding to the acquired patient information, by requesting the execution of inference using a language model, taking as input a prompt that includes the acquired patient information, a treatment method for a patient corresponding to the patient information, and an outcome when the treatment method is performed on a patient corresponding to said patient information;

[25] A system comprising at least one computer device, comprising: acquisition means for acquiring patient information and a treatment method for a patient corresponding to said patient information; and outcome identification means for identifying an outcome when the treatment method is performed on a patient corresponding to the acquired patient information, by requesting the execution of inference using a language model finely tuned with information relating to the correspondence between patient information and the outcome when said treatment method is performed on a patient corresponding to said patient information, with the input being a prompt including the acquired patient information and a treatment method for a patient corresponding to the patient information; A system equipped with;

[26] A system comprising at least one computer device, comprising: acquisition means for acquiring patient information and treatment methods for patients corresponding to said patient information; and outcome identification means for identifying outcomes when said treatment methods are performed on patients corresponding to the acquired patient information, based on a trained model that has been trained using patient information for machine learning and treatment methods for patients corresponding to the patient information for machine learning as input data, and outcomes when said treatment methods are performed on patients corresponding to the patient information for machine learning as output data;

[27] New patient information for machine learning, a treatment method for patients corresponding to the patient information for machine learning, and an update means for updating a trained model by retraining based on the outcome when the treatment method is performed on patients corresponding to the new patient information for machine learning The system according to

[26] , comprising:

[28] A system according to any one of

[24] to

[27] , comprising a method identification means for identifying one or more treatment methods for a patient corresponding to the acquired patient information, based on acquired patient information and criteria for selecting a treatment method, wherein the outcome identification means identifies an outcome based on the identified treatment method;

[29] A system according to any one of

[24] to

[27] , comprising a method identification means that identifies one or more treatment methods corresponding to acquired patient information, based on a trained model that has been trained using patient information for machine learning as input data and treatment methods suitable for the patient information for machine learning as output data, wherein an outcome identification means identifies an outcome based on the identified treatment method;

[30] A system according to any of [1] to

[29] above, wherein patient information includes information relating to images taken inside the patient's body;

[31] A system according to any of [1] to

[30] above, wherein patient information includes text information;

[32] A method to be performed in a system comprising at least one computer device, comprising: an acquisition step of acquiring patient information; and a method identification step of identifying one or more treatment methods for a patient corresponding to the acquired patient information by requesting the execution of inference using a language model, with the acquired patient information and a prompt including criteria for selecting a treatment method as input;

[33] A method to be performed in a system comprising at least one computer device, comprising: an acquisition step of acquiring patient information; and a method identification step of identifying one or more treatment methods for a patient corresponding to the acquired patient information by taking a prompt containing the acquired patient information as input and requesting the execution of inference using a language model finely tuned with information on treatment method selection criteria;

[34] A method to be performed in a system comprising at least one computer device, comprising: an acquisition step of acquiring patient information; and a method identification step of identifying one or more treatment methods corresponding to the acquired patient information based on a trained model that has been trained using patient information for machine learning as input data and treatment methods suitable for the patient information for machine learning as output data;

[35] A method to be performed in a system comprising at least one computer device, comprising: an acquisition step of acquiring patient information; an input step of inputting a target outcome; and a method identification step of identifying one or more treatment methods corresponding to the acquired patient information and the target outcome, based on a trained model that has been trained using patient information for machine learning and outcomes as a result of performing the treatment method on patients corresponding to the patient information for machine learning as input data, and treatment methods performed on patients corresponding to the patient information for machine learning as output data;

[36] A method to be performed in a system comprising at least one computer device, comprising: an acquisition step of acquiring patient information and a treatment method for a patient corresponding to said patient information; and an outcome identification step of identifying an outcome for a patient corresponding to the acquired patient information by requesting the execution of inference using a language model, with input including the acquired patient information, a treatment method for a patient corresponding to the patient information, and a prompt including a correspondence between the patient information and the outcome when said treatment method is performed on the patient corresponding to said patient information;

[37] A method performed in a system comprising at least one computer device, comprising: an acquisition step of acquiring patient information and a treatment method for a patient corresponding to said patient information; and an outcome identification step of identifying an outcome for a patient corresponding to said patient information by taking a prompt including the acquired patient information and a treatment method for a patient corresponding to said patient information as input, and requesting the execution of inference using a language model finely tuned with information relating to the correspondence between the patient information and the outcome when said treatment method is performed on the patient corresponding to said patient information;

[38] A method to be performed in a system comprising at least one computer device, comprising: an acquisition step of acquiring patient information and a treatment method for a patient corresponding to said patient information; and an outcome identification step of identifying an outcome when the treatment method is performed on a patient corresponding to the acquired patient information, based on a trained model that has been trained using patient information for machine learning and a treatment method for a patient corresponding to said patient information as input data, and the outcome when the treatment method is performed on a patient corresponding to said patient information as output data; This can be resolved. [Effects of the Invention]

[0017] Examples of the effects of the present invention include, but are not limited to, the following. The first effect of the present invention is that it can provide a system that can identify a treatment method for a patient corresponding to patient information. The second effect of the present invention is that it can provide a system that can identify the outcome when the treatment method is implemented for a patient corresponding to patient information. [Brief explanation of the drawing]

[0018] [Figure 1] This shows a schematic configuration of the computer according to the embodiment. [Figure 2] This example shows how RAG can be used to formulate a patient's treatment plan according to the embodiment. [Figure 3] This document outlines the treatment selection and outcome prediction process using the treatment selection support program according to the embodiment. [Figure 4] This is a block diagram showing the configuration of the system according to the embodiment. [Figure 5] This figure shows a flowchart of the output processing according to the embodiment. [Figure 6] This figure shows a flowchart of the output processing according to the embodiment. [Modes for carrying out the invention]

[0019] The following describes embodiments of the present invention, but the present invention is not limited to the following embodiments unless it contradicts the spirit of the invention. The order of each process constituting the flowchart described below is not limited to any order that does not cause contradictions or inconsistencies in the processing content, and it is also possible to omit some of the processes constituting the flowchart or to add new processes to each process constituting the flowchart, as long as it does not contradict the spirit of the invention. Furthermore, the device that is the main entity that executes each process constituting the flowchart can be changed to another device, as long as it does not contradict the spirit of the invention. In this case, it is possible to change the processing content so as not to cause contradictions or inconsistencies in the processing content.

[0020] A program to support the development of treatment plans for patients. This disclosure relates, in one aspect, to a program for assisting in the formulation of treatment plans for patients. Patients' diseases may include, but are not limited to, malignant diseases such as cancer (synonymous with malignant diseases; classified into carcinoma, sarcoma, and hematological cancer), benign diseases such as rheumatism, collagen diseases, diabetes, and dyslipidemia, and a wide range of chronic diseases. Furthermore, it is not limited to patients already suffering from a disease, but also includes patients at high risk of developing a disease and patients who have completed treatment. The former includes, for example, intraductal papillary mucinous neoplasms (IPMN) in pancreatic cancer, hepatitis in liver cancer, and hyperglycemia in diabetes. The latter includes, for example, post-surgical treatment for cancer. Patients can be, for example, human cancer patients with any type of cancer, such as breast cancer, lung cancer, prostate cancer, colorectal cancer, pancreatic cancer, stomach cancer, hepatocellular carcinoma, leukemia, skin cancer, uterine cancer, sarcoma, soft tissue sarcoma, lymphoma, skull cancer, biliary tract cancer, brain tumor, esophageal cancer, cancer of unknown primary origin, prostate cancer, oral cancer, or ovarian cancer. Cancer patients may have metastases, and may be experiencing recurrent cancer as well as initial cancer.

[0021] In some embodiments, the program of this disclosure is a program for assisting in the formulation of a patient's treatment plan, and relates to a program that causes one or more computers to perform the steps of acquiring patient information, generating a proposed treatment plan based on the patient information, and outputting the proposed treatment plan. The one or more computers do not need to be physically located in the same place, and multiple computers with different roles (e.g., input / output terminals, LLM servers, database servers, etc.) may be connected via a network and function together.

[0022] As used herein, “treatment plan” includes the selection of treatment methods to be applied to a patient. For example, in the case of cancer patients, available treatment methods include conventional surgery, advanced surgery, TSH stage 2 surgery, ALPPS, ALPTPS, venous surgery, robotic surgery, laparoscopic surgery, radiotherapy, external beam radiotherapy, particle beam therapy, heavy ion beam therapy, proton beam therapy, high-precision X-ray therapy, IMRT, stereotactic radiotherapy, CyberKnife, internal radiotherapy, brachytherapy, brachytherapy, theranostics, BNCT (boron neutron capture therapy), IVR (interventional radiotherapy), ablation, radiofrequency ablation, microwave ablation, cryotherapy, intra-arterial embolization, ethanol injection, HIFU (high-intensity focused ultrasound), irreversible electroporation (IRE), photoimmunotherapy, viral therapy, chemotherapy, hormone therapy, molecular targeted drugs, antibody-drug conjugates (ADCs), immunotherapy, immune checkpoint inhibitors, fecal microbiota transplantation, cancer vaccines, CAR-T therapy, combination immunotherapy, gene therapy, stem cell transplantation, mRNA This includes, but is not limited to, vaccines, CRISPR / Cas9, hyperthermia, TSH stage 2 surgery, ALPPS, venous surgery, conventional RFA, high-precision RFA, TACE, drug therapy, anticancer drugs, immunosuppressants, investigational drugs, herbal medicines, transplantation (organ transplantation, bone marrow transplantation, etc.), nerve blocks, CART, dialysis, shunts (Denver shunt, etc.), nerve decompression, nerve transplantation, nerve repair, stem cell therapy, stem cell transplantation, inhaled drugs, intravenous infusions, transdermal drugs, transdermal formulations, transdermal patches, transdermal formulations (TTS), eye drops, intraoral administration, continuous administration, and new surgical procedures, psychotherapy (cognitive behavioral therapy, CBT, etc.), physical therapy (electroconvulsive therapy: ECT, transcranial magnetic stimulation: TMS, transcranial direct current stimulation: tDCS, deep brain stimulation (DBS), spinal cord stimulation, etc.), dietary therapy, exercise therapy, weight loss surgery, traditional Chinese medicine treatments such as acupuncture, and antioxidants. If multiple treatment methods are available, they may be combined, applied simultaneously, or sequentially. Where used herein, "treatment plan" may also include a plan for how to apply the available treatment methods.

[0023] As used herein, "patient information" includes several items relating to the patient's physical characteristics and medical condition. For example, in the case of a cancer patient, patient information includes, but is not limited to, age, sex, disease name (disease name, primary site name, metastasis, recurrence, etc., complication name), left hepatic vein invasion, middle hepatic vein invasion, left hepatic vein invasion, vascular invasion, bile duct invasion, gastrointestinal invasion, Glisson's sheath invasion, metastasis to other organs, low liver function, S1, FLR (normal liver), FLR (damaged liver), massive hepatectomy (e.g., trisegmentectomy), massive hepatectomy (right hepatectomy), 3 segmental branches, left hepatic vein invasion, middle hepatic vein invasion, right hepatic vein invasion, portal vein invasion, distance from gastrointestinal tract (mm), vascular invasion, bile duct invasion, gastrointestinal invasion, proximity to blood vessels, proximity to bile ducts, proximity to gastrointestinal tract, and proximity to Glisson's sheath, anatomical names and positional relationships, stage of progression, histological and pathological findings, genetic information such as genetic abnormalities and their extent (including histological findings and gene expression levels), immunological findings such as the tumor microenvironment, and information on the degree of function.

[0024] Patient information can be obtained, for example, through input from a keyboard, input via communication lines such as the internet, or input from storage devices such as hard disks. Patient information may be stored in a patient information database and retrieved as needed by accessing the database. At least a portion of the patient information may be obtained from, for example, electronic medical records. At least a portion of the patient information may also be obtained through character recognition of handwritten medical records. Furthermore, at least a portion of the patient information may be generated from examination images such as PET, CT, MRI, and X-ray images through automated image interpretation on a computer. AI-based image recognition technology may be used for automated image interpretation.

[0025] Generation of treatment plan In some embodiments, the step of generating a treatment plan based on patient information includes, for example, in the case of malignant diseases (cancer), the step of selecting one or more available treatment methods based on patient information and predefined treatment selection criteria (off-label criteria and / or applicable criteria and / or quasi-off-label criteria). The selected available treatment methods may be more than one, and a list of available treatment methods may be created and included in the treatment plan. The treatments displayed in the list may be presented in order of the outcome that is emphasized for each disease. For example, in malignant diseases (cancer), the outcome of curative treatment is important, and the treatment names may be presented in order of the probability of curative treatment. The probability of curative treatment may be calculated from the patient's disease condition information. The correspondence table may also use a correspondence table of progression and treatment established for each disease. For example, in malignant diseases, a correspondence table of stage (progression) and treatment may be used, or in the case of a benign disease such as rheumatoid arthritis, the phase (progression) is determined by the response (effect) to treatment obtained from the course of treatment, and treatment according to that phase is presented. Also, in the case of asthma, for example, a correspondence table is used that uses symptoms and the course of treatment as treatment selection criteria. Even for benign diseases, it is possible to create and use original correspondence tables that link subdivided disease symptoms to corresponding treatments.

[0026] As used herein, "off-label criteria" refer to criteria used to exclude a particular treatment method from treatment options based on one or more items of patient information. For example, the off-label criteria for pancreatic cancer surgery are when there is tumor contact or invasion of the SMA, CA, or CHA, or contact or invasion of less than 180 degrees or no occlusion of the SMV / PV, and when there is contact or invasion of 180 degrees or more of the SMA or CA, or contact or invasion of the proper hepatic artery or CA. Therefore, if the patient information indicates contact of 180 degrees or more of the SMA / PHA, pancreatic cancer surgery is not selected as an available treatment method.

[0027] Furthermore, for colorectal cancer patients with liver metastases, if the patient information indicates left hepatic vein invasion, standard surgery is not indicated, but advanced surgery (S1, etc.), TSH stage 2 surgery, and ALPPS surgery are considered available.

[0028] As used herein, “indication criteria” refers to criteria based on one or more items of patient information for including a particular treatment method as a treatment option.

[0029] As used herein, “off-label criteria” refers to criteria based on multiple items of patient information (for example, two or more, three or more, four or more, or five or more items) that allow for the exclusion of a particular treatment method from treatment options or the postponement of such a decision until certain conditions are met.

[0030] In this specification, “treatment selection criteria” refers to the patient’s condition or circumstances for administering treatment. In this specification, “treatment selection criteria” is intended to encompass “off-label criteria,” “indication criteria,” and “near-off-label criteria.”

[0031] In some embodiments, the step of selecting one or more available treatment methods based on patient information and predefined treatment selection criteria (off-label criteria and / or applicable criteria and / or quasi-off-label criteria) may be performed using a large-scale language model (LLM).

[0032] In the context of this disclosure, examples of large-scale language models (LLMs) include GPT (Generative Pre-trained Transformer), BERT (Bidirectional Encoder Representations from Transformers), T5 (Text-to-Text Transfer Transformer), and Llama (Large Language Model Meta AI). As used herein, “Large-scale language model (LLM)” refers to an advanced AI model that can learn from vast amounts of text data and understand and generate human language. Such models, also known as generative AI, are widely used in the field of natural language processing (NLP) and demonstrate high performance in a wide range of tasks, including text summarization, translation, question answering, text generation, and sentiment analysis. LLMs may run, for example, on a cloud server.

[0033] LLMs are typically trained using massive datasets containing billions of words collected from books, articles, and websites available on the internet. This training process employs a neural network architecture called a multi-layered transformer. Transformers are particularly adept at processing context-rich information and understanding interlingual dependencies. A key feature of this architecture is the self-attention mechanism, a process that assigns context-based weights to each word, considering its relevance to other words.

[0034] Furthermore, these models can be fine-tuned for specific tasks, making them adaptable to advanced language-related tasks suited to particular application areas and industries. For example, in the medical field, a fine-tuned LLM using specialized text data can be effectively used for complex tasks requiring expertise. Large-scale language models can serve as a foundational technology providing innovative solutions across a wide range of areas, from general-purpose natural language processing tasks to domain-specific applications.

[0035] In some embodiments, the large-scale language model (LLM) used in the program of this disclosure may be fine-tuned specifically for formulating treatment plans for patients. Fine-tuning may involve the use of treatment selection criteria and algorithms, such as off-label criteria, applicable criteria, and / or semi-off-label criteria, which are compiled and restructured from the latest treatment guidelines related to individual diseases, databases of past clinical cases, and literature databases such as PubMed. Existing, widely used disease stage / progression correspondence tables can be used, and these tables can be used as is, or they can all be restructured using treatment selection criteria such as applicable criteria and off-label / semi-off-label criteria. By retraining the model using such information (datasets), performance on the desired task of selecting available treatment methods can be significantly improved.

[0036] In other embodiments, the LLM used in the program of this disclosure may be a transfer-learned model for the task of formulating treatment plans for patients. Transfer learning utilizes the weights of a model previously trained on a broad dataset as initial values, and then retrains it to adapt it to a new task. Specifically, by fine-tuning a model already trained in general natural language processing or the medical field using data related to specific treatment plan formulation, rapid and efficient adaptation to new tasks is possible. Transfer learning is often considered part of fine-tuning to optimize a model for a specific task, and this method allows for rapid adaptation of the model to specialized and advanced tasks while making the most of already accumulated knowledge. As a result, the LLM used in the program of this disclosure can function as a powerful support tool for formulating optimal treatment plans tailored to the individual circumstances of patients.

[0037] As used herein, a “prompt” in a Large-Scale Language Model (LLM) refers to the text input provided to the model. This prompt serves as a starting point or instruction for the model to generate a response and is a crucial element in determining the content and format of the generated text. The role of a prompt is multifaceted; for example, it can instruct the language model to focus on a specific topic or task, providing appropriate direction. This allows the model to generate relevant and appropriate information according to the given context.

[0038] The input prompts determine the context of the generated text and define the scope of information the model refers to. Prompts can also instruct the model to generate text in a specific format or style. For example, prompts can be used to have the model generate answers in a question format or to create a summary of given text. Furthermore, prompt design is crucial for the model to generate more accurate responses. The content and structure of the prompts greatly influence what information the model prioritizes and in what format it generates its responses. Therefore, properly designing prompts can make the LLM output more specific and precise, improving the quality of the model's responses.

[0039] The program described herein can generate prompts to be input into a large-scale language model (LLM) based on patient information. These prompts may directly use the patient information entered by the user, or they may be modified appropriately. The technique of designing prompts to be input into the LLM so that the model produces desired outputs is known as "prompt engineering," and it is an important technique for improving the accuracy of the model's responses.

[0040] In some embodiments, prompts can include multiple task-related examples, allowing the model to make inferences based on them. This technique, called "fushot learning," can effectively improve model performance by presenting a small number of examples. Fushot learning is particularly effective in situations with limited data, enhancing the model's ability to quickly adapt to new tasks.

[0041] Furthermore, in some embodiments, prompts input to the LLM can be created using "RAG (Retrieval-Augmented Generation)". RAG is a method that searches and extracts external data (e.g., treatment selection criteria related to patient information (off-label criteria, applicable criteria, near-off-label criteria)) from the RAG database in real time, and combines the original question (prompt) with the search results to generate a more accurate and detailed response. This process enables the LLM to propose highly accurate treatment plans while incorporating the latest relevant information. When obtaining necessary information using RAG, for example, information highly relevant to queries such as patient information can be retrieved from the RAG database based on cosine similarity. By using RAG, real-time information updates for the LLM are possible, reducing the learning cost of the LLM and enabling the generation of responses that reflect the latest medical information. In addition, the fine-tuning described above can also be achieved in the RAG function, and further accuracy improvements can be expected by accumulating the information obtained during fine-tuning in the RAG database. Figure 2 shows an example of patient treatment plan formulation utilizing RAG in the program of this disclosure. By utilizing RAG, it becomes possible to acquire the latest external information in a timely manner and ask questions to LLMs based on that information, thereby enabling more advanced medical support. The RAG database may reside in a location physically separate from the computer used as the user terminal, for example, on a cloud server.

[0042] In the context of this disclosure, examples of prompts entered into the LLM include the following: (Prompt 1) "When performing surgery for pancreatic cancer, the criteria for contraindication are when the tumor has contact or infiltration into the SMA, CA, or CHA, or when there is no contact or infiltration of less than 180 degrees or occlusion of the SMV / PV, or when there is contact or infiltration of 180 degrees or more into the SMA or CA, or contact or infiltration into the proper hepatic artery or CA." (Prompt 2) The following is information regarding the patient's medical condition: • Infiltration into the duodenum • Infiltration into the lower part of the common bile duct • The superior mesenteric vein (SMV) is almost completely closed. • The superior mesenteric artery (SMA) and hepatic artery (PHA) are in contact at an angle of 180 degrees or more. • Infiltration into the horizontal portion of the duodenum. Can this patient undergo surgery?

[0043] The following are examples of LLM's responses to the above input. "The patient's condition is such that surgery is not indicated due to near-occlusion of the superior mesenteric vein (SMV), tumor contact of more than 180 degrees with the superior mesenteric artery (SMA) and hepatic artery (PHA), and infiltration into the duodenum and lower common bile duct. These conditions make surgery extremely high-risk and therefore not recommended. Non-surgical treatments such as chemotherapy and radiotherapy should be considered as alternative therapies."

[0044] In the example above, the user provides the off-label criteria as a prompt, but prompts may be generated using RAG (Retrieval-Augmented Generation) based on patient information, and treatment selection criteria such as off-label criteria may be trained in the LLM. That is, in some embodiments, treatment selection criteria (off-label criteria and / or applicable criteria) may be input to the Large-Scale Language Model (LLM) as prompts. Also, in some embodiments, treatment selection criteria (off-label criteria and / or applicable criteria and / or quasi-off-label criteria) may be input to the Large-Scale Language Model (LLM) by RAG (Retrieval-Augmented Generation). In some embodiments, treatment selection criteria (off-label criteria, off-label criteria, quasi-off-label criteria) are stored in a database. Treatment selection criteria (off-label criteria, off-label criteria, or quasi-off-label criteria) stored in the database may be used in the RAG. This is also true when using existing treatment decision algorithms stored in the database.

[0045] In some embodiments, treatment selection criteria (off-label criteria and / or applicable criteria and / or quasi-off-label criteria) are stored in the database in the form of a correspondence table (also called a correspondence table or matrix) that associates patient information with available treatment methods. The correspondence table may, if applicable, be automatically updated and / or expanded by LLM. In some embodiments, when treatment selection criteria (off-label criteria and / or applicable criteria and / or quasi-off-label criteria) are stored in the database in the form of a correspondence table that associates patient information with available treatment methods, available treatment methods may be selected by a standard matching algorithm (searching for correspondences between patient information and available treatment methods) without using LLM. For example, a treatment method can be selected by matching patient information, such as the patient's test results, with applicable or off-label treatment methods on the correspondence table and listing the applicable treatment methods. For example, a treatment method can be selected by excluding treatment methods that are unsuitable depending on the patient's condition, etc., from a list of all currently available treatment methods, and finally selecting the treatment method that remains in the list. Therefore, in some embodiments, the program relating to this disclosure is a program that performs the steps of acquiring patient (clinical) information, generating a proposed treatment plan based on the patient information, and outputting the proposed treatment plan, wherein the step of generating a proposed treatment plan based on the patient information may include the step of selecting one or more available treatment methods by referring to a correspondence table that associates patient information with available treatment methods. This is also true when using an existing treatment decision algorithm stored in a database.

[0046] Output of treatment plan In some embodiments, the step of outputting a proposed treatment plan may include, for example, outputting to a display directly or indirectly connected to the computer, outputting audio to a speaker, or outputting to a printer or storage device connected to the computer.

[0047] In some embodiments, the output treatment plan may include information on the advantages (e.g., effectiveness) and disadvantages (e.g., adverse events) of one or more selected available treatment methods. This information, including the advantages and disadvantages of individual treatment methods, may be stored in a database used by RAG or in a separate, independent database. This information may be retrieved by LLM or by a standard database query. If multiple available treatment methods are selected, they may be prioritized based on their advantages and disadvantages.

[0048] In some embodiments, the output treatment plan may include information including the cure probability of one or more selected available treatments. Information including the disease-specific outcome priority of individual treatments (such as the cure probability in cancer) may be stored in a database used by RAG or in a separate, independent database. Information including the cure probability of individual treatments may be obtained by LLM or by a regular database query. If multiple available treatments are selected, they may be prioritized (disease-specific) based on their cure probability.

[0049] In some embodiments, the proposed treatment plan may include predictive information on outcomes such as recurrence. The predictive information on outcomes may be automatically calculated by LLM by obtaining publication information from a database of publications such as PubMed and comparing it with patient information according to the algorithm in the publication information.

[0050] In some embodiments, the program for assisting in the formulation of a patient treatment plan according to the Disclosure may be a program that causes one or more computers to perform the following steps: acquiring patient information; generating a proposed treatment plan using a large-scale language model (LLM) based on the patient information, wherein the proposed treatment plan includes a list of one or more available treatment methods selected based on the patient information and treatment selection criteria (such as contraindication criteria and / or indication criteria) that are input to the LLM as prompts by RAG (Retrieval-Augmented Generation); and outputting the proposed treatment plan, wherein the proposed treatment plan further includes information including the merits and demerits of the selected one or more available treatment methods and / or information including predictive information on the curative probability and / or outcome of the selected one or more available treatment methods. In some embodiments, the program according to the Disclosure may further specify a step of pre-defining treatment names and treatment selection criteria for each treatment and storing them in a database. In some embodiments, the program according to the Disclosure may specify a step of collecting the pre-defined treatment names and treatment selection criteria for each treatment from a pre-created database or a sequentially updated database using RAG.

[0051] In some embodiments, the step of generating treatment plan proposals based on patient information can be performed using a suitable machine learning model (such as a neural network or random forest) that takes patient information as input and outputs usable treatment methods, in addition to a large-scale language model (LLM).

[0052] Figure 3 shows an overview of treatment selection and outcome prediction using the treatment selection support program related to this disclosure. A correspondence table (correspondence table) with patient information can be used in the algorithms for treatment selection and outcome / result prediction (such as recurrence prediction). In other words, patients can be stratified for each treatment using the correspondence table. Precise treatment selection becomes possible through outcome prediction. For example, the recurrence risk (recurrence score) after breast cancer surgery can be calculated using outcome prediction, and the degree of necessity for adjuvant therapy can be presented based on treatment information on the additional effect of adjuvant chemotherapy obtained from research papers and the recurrence score.

[0053] There may be two types of treatment selection tables. One is the traditional table of progression / phase and treatment, which can be an approach from the patient's "medical condition" (patient information) side. The other lists all treatments used for each disease, and associates the selection criteria for those treatments (off-label criteria, applicable criteria) with the patient's medical condition (patient information), allowing treatments that do not deviate from the off-label criteria or are applicable to the treatment to be considered as treatment options. Diseases can be broadly classified into malignant diseases (cancer) and benign diseases, but in malignant diseases, the feasibility of local therapy can be determined mainly by the off-label criteria. Outcome predictions can be calculated using patient information and a correspondence table or formula between factors / items related to outcomes proven in papers, etc., and the outcomes themselves.

[0054] 1) Creation of the original database (DB) (1) The original DB(1) contains, for example, a correspondence table for treatment selection, or a correspondence table and calculation formula for predicting treatment outcomes. This treatment selection or outcome prediction algorithm (correspondence table, calculation formula) may be created using RAG to access an external medical database, obtain the data necessary for algorithm creation, and then store in the original DB(1) using LLM or similar tools.

[0055] 2) Creation of the original database (DB) (2) External medical databases may include databases such as PubMed for obtaining the latest research paper information, databases containing treatment information that incorporates physician consensus such as guidelines, and other databases such as clinical trial information databases. The original database (2) may, for example, store treatment information (treatment content / explanation, treatment effect, adverse events, etc.) for each treatment, and treatment information may be obtained from external databases using RAG, organized with LLM, and stored in DB(2).

[0056] 3) Organization of patient information Patient information such as disease names (e.g., colorectal cancer, liver metastasis, lung metastasis), imaging information (interpretation findings from CT and MRI images, etc.), treatment progress information (date, treatment details, and findings), test information (blood tests, etc.), pathology information, genetic information, and symptoms / findings obtained from medical referral letters can be organized using LLM. Then, the patient information can be organized (adjusted) according to a correspondence table and converted into "disease condition data".

[0057] 4) Presentation of treatment options using comparison tables and calculation of outcomes / results By comparing the "medical condition data" (patient information organized in 3 above) with the comparison table obtained from DB(1) using RAG, using a checklist or LLM, treatment can be selected or outcome predictions can be performed.

[0058] 4a) In the case of cancer treatment options An example of an LLM prompt would be: "Access DB(1), retrieve the correspondence table from the disease name, compare the disease symptoms with the correspondence table, remove any treatment from the treatment options if any exclusion criteria are met, remove any treatments that do not meet the indication criteria, and display the remaining treatments in the same order as in the correspondence table."

[0059] 4b) In the case of outcome prediction (risk of recurrence) LLM prompts, when using a correspondence table, might include examples such as "Access DB(1), obtain the comparison table for the prediction you want, and calculate the predicted value based on the disease state and the comparison table" (for example, recurrence risk between stages 1 and 2 in stage 2 surgery for liver metastasis of colorectal cancer). When using a calculation formula, prompts might include expressions such as "Extract the items and their values ​​necessary for prediction from the disease state, substitute them into the calculation formula, and calculate the predicted value" (for example, the breast cancer recurrence rate calculation and cancer-related gene expression level weighting equation published in the NEJM and used in OncotypeDX®).

[0060] 5) Acquisition and display of treatment information Information about the treatment options can be retrieved from the original DB(2) using RAG and LLM. An example of an LLM prompt for retrieval might be, "Access DB(2) and retrieve and display treatment information for each presented treatment."

[0061] 6) Original DB (3): Patient Database Patient information and output results can be saved in a single file for each patient, and this file can then be stored in the original DB (3).

[0062] 7) Original DB (3): Utilization of patient databases The data can be completely anonymized and used as real-world data to improve the accuracy of the software related to this disclosure.

[0063] Those skilled in the art will understand that the features of the program described herein may also be applied to the devices and systems described below.

[0064] A device to support the formulation of treatment plans for patients. This disclosure relates, in one embodiment, to a device for assisting in the formulation of a patient's treatment plan.

[0065] In some embodiments, the present disclosure relates to an apparatus for assisting in the formulation of a patient's treatment plan, comprising a patient information acquisition unit for acquiring patient information, a treatment plan draft generation unit for generating a treatment plan draft based on the patient information, and a treatment plan draft output unit for outputting the treatment plan draft.

[0066] In some embodiments, the apparatus of the Disclosure may include a processor that performs various operations and memory connected to the processor. In some embodiments, the apparatus of the Disclosure may be realized by running the program of the Disclosure on a general-purpose computer. In some embodiments, the apparatus of the Disclosure is an apparatus that includes a processor connected to a storage device storing the computer program of the Disclosure and capable of executing instructions of the program.

[0067] In some embodiments, the apparatus relating to this disclosure includes means for inputting data. Examples of means for inputting data include a keyboard and a mouse.

[0068] In some embodiments, the apparatus according to the present disclosure includes a central processing unit (CPU) connected to a keyboard, mouse, etc. for data input, a hard disk, flash memory, etc. as a storage unit, and a memory (storage means) such as ROM, RAM, etc.

[0069] Means for outputting data including prediction results include, for example, monitors and printers. Alternatively, means for storing the data in storage devices such as hard disks, flash memory, ROM, and RAM can also be considered output methods.

[0070] The device relating to this disclosure may include means for storing a program for assisting in the formulation of a patient treatment plan relating to this disclosure. Examples of means for storing the program include hard disks and flash memory. Such storage means may be connected by a communication line. That is, the device relating to this disclosure may be part of a system obtained by connecting a device including the means for storing the program via a communication line.

[0071] Figure 1 is a schematic diagram showing an exemplary embodiment of the apparatus according to the present disclosure. In Figure 1, 100 is a computer, comprising a control unit 101, a storage unit 102, a peripheral device I / F unit 103, an input unit 104, a display unit 105, and a communication unit 106, which are connected by a bus 110. Note that this configuration is illustrative, and various configurations can be adopted as appropriate.

[0072] The control unit 101 consists of a CPM (Central Processing Unit), ROM (Read Only Memory), RAM (Random Access Memory), etc. The CPU calls programs stored in the memory unit 102, ROM, recording medium, etc., into the work memory area on RAM and executes them, drives and controls each device connected via the bus 110, and realizes the processing that the computer performs. ROM is a non-volatile memory that holds programs and data such as the computer 100's boot program and BIOS. RAM is a volatile memory that temporarily holds programs and data loaded from the memory unit 102, ROM, recording medium, etc., and also has a work area used by the control unit 101 when performing various processing. The memory unit 102 is, for example, an HDD (Hard Disk Drive) and stores programs executed by the control unit 101 and various other data.

[0073] The peripheral device interface (I / F) section 103 is a port for connecting the computer 100 to peripheral devices. The peripheral device interface section 103 consists of USB, Bluetooth, IEEE1394, RS-232C, etc. The connection method to peripheral devices can be wired or wireless. The input section 104 has pointing devices such as a keyboard or mouse, and input devices such as a numeric keypad, and provides operation instructions, action instructions, data input, etc. to the computer 100. The display section 105 is a logic circuit or device driver for displaying images and videos on a display device such as an LCD panel. The input section 104 and the display section 105 can also be configured as an integrated touch display.

[0074] The communication unit 106 has a communication control device, a communication port, etc., and is a wired or wireless communication interface that mediates communication with the communication network 120. The bus 110 is a communication path that mediates the exchange of control signals, data signals, etc. between each device. The communication network 120 may be further connected to an external server 130 or a database (or network storage) 140.

[0075] In some embodiments, the apparatus of the present disclosure may further include a prompt generation unit that generates prompts by RAG (Retrieval-Augmented Generation), an LLM input / output unit that inputs prompts to an LLM and obtains an output, an LLM operation unit that operates the LLM, a treatment method information acquisition unit that acquires information on available treatment methods, and a treatment method information storage unit that stores information on available treatment methods.

[0076] A system to support the formulation of treatment plans for patients. This disclosure relates, in one embodiment, to a system for assisting in the formulation of treatment plans for patients. The system relating to this disclosure may consist of one or more computers. The computers may be connected by a communication network such as the Internet, and some of the computers may be configured on a cloud server. An exemplary configuration is a system relating to this disclosure that consists of client terminals that primarily perform input / output and servers that primarily perform computational processing.

[0077] In some embodiments, the System of the Disclosure is a system for assisting in the formulation of a patient's treatment plan, comprising one or more computers, wherein the one or more computers comprises memory and a processor connected to the memory, and the processor performs the steps of acquiring patient information entered by a user, generating a proposed treatment plan based on the patient information, and outputting the proposed treatment plan. In some embodiments, the Program of the Disclosure is stored in the memory or storage device of a computer included in the System of the Disclosure, and when the Program is executed, each of the above steps is performed.

[0078] In some embodiments, the system of the present disclosure may further perform the steps of generating a prompt by RAG (Retrieval-Augmented Generation), inputting the prompt to an LLM and obtaining an output, operating the LLM, obtaining information about available treatment methods, and storing information about available treatment methods.

[0079] Non-temporary computer-readable recording medium This disclosure relates, in one embodiment, to a non-temporary computer-readable recording medium storing the program of this disclosure. Examples of computer-readable recording media include, but are not limited to, hard disk drives (HDDs), solid-state drives (SSDs), flash memory (such as USB memory sticks and SD cards), optical discs (such as CDs, DVDs, and Blu-ray® discs), magnetic tapes, floppy disks, and cloud storage.

[0080] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those generally understood by those skilled in the art to which the invention pertains. Any methods and materials similar or equivalent to those described herein may be used for carrying out or testing the invention, but several possible and preferred methods and materials are described herein. All publications referenced herein are incorporated herein by reference, and the methods and / or materials cited in relation to these publications are disclosed and described herein. In the event of any conflict, this disclosure shall prevail over the disclosure of the incorporated publication.

[0081] Where a range of values ​​is given, unless the context clearly indicates otherwise, it is understood that each intervening value between the upper and lower limits of that range, up to one-tenth of the lower limit unit, is also specifically disclosed. Each smaller range between any given value or intervening value within the given range and any other given value or intervening value within that given range is also included in this disclosure. The upper and lower limits of these smaller ranges may be included or excluded independently, and each range that includes either, either, or both of the limit values ​​in the smaller range is also included in the invention, but the limit values ​​specifically excluded in the given range are reserved. Where a given range includes one or both of the limit values, a range that excludes either or both of the included limit values ​​is also included in the invention. The term “about” with respect to numerical values ​​means within 5%.

[0082] The embodiments described herein are intended to be illustrative only, and those skilled in the art will be able to make numerous modifications and alterations without departing from the spirit of the invention. Certain modifications and alterations may yield satisfactory results, though not optimal. All such modifications and alterations are intended to fall within the scope of the invention as defined by the appended claims. Furthermore, any combination of the components disclosed herein, or the expressions of this disclosure in terms of methods, apparatus, systems, computer programs, data structures, recording media, etc., are also valid embodiments of this disclosure. Thus, the details described relating to the methods of this disclosure may be applied to systems, computer programs, data structures, recording media, etc.

[0083] This disclosure will be further understood by reference to the following embodiments. These embodiments are provided solely to illustrate the disclosure set forth in the claims, and the scope of this disclosure is not limited by the embodiments shown, which are intended only as examples of a single aspect of this disclosure. Any functionally equivalent method is included within the scope of this disclosure. In addition to those described herein, various modifications of this disclosure will be apparent to those skilled in the art from the foregoing description. Such modifications are intended to be within the scope of the appended claims.

[0084] <Embodiments 15 and 16> The matters described in Embodiments 15 and 16 can also be applied to Embodiments 1 to 14. Figure 4 is a block diagram showing the configuration of a system according to an embodiment. System 10 includes at least one computer device. System 10 may consist of, for example, one computer device (information processing device) (standalone type), one or more server devices and one or more terminal devices (client-server type), or multiple terminal devices (peer-to-peer type).

[0085] System 10 may include a user terminal 200 and a server device 100. The server device 100 corresponds to the computer 100 described above. Furthermore, the server device 100 may function in a distributed manner across multiple computer devices. For example, instead of the server device 100, a distributed ledger technology such as blockchain may be used.

[0086] The user terminal 200 and the server device 100 are connected to each other via the communication network 120 so that they can communicate with one another. The number of user terminals 200 is not particularly limited, and there may be one or more user terminals 200.

[0087] System 10 may include a language model server 300. The server device 100 and the language model server 300 are connected to each other via a communication network 120 so that they can communicate with each other.

[0088] (User terminal) User terminal 200 is a terminal operated by a user of system 10. Here, the user can be anyone who uses system 10. Examples of user terminal 200 include conventional mobile phones, tablet terminals, smartphones, and personal computers.

[0089] The user terminal 200 comprises a control unit, a storage unit, an input unit, a display unit, and a communication unit, each connected by a bus.

[0090] The control unit consists of a CPU. The control unit executes programs stored in the memory unit and controls the user terminal 200. The memory unit is a storage medium for storing programs and data. The memory unit also includes RAM. RAM is the work area of ​​the control unit. The control unit performs calculations based on the programs and data read from RAM, as well as the data input in the input unit. The program (program product) may be stored on a recording medium such as a CD-ROM.

[0091] The display unit has a display screen. The control unit outputs a video signal for displaying an image on the display screen according to the result of the calculation processing. Note that the display screen of the display unit may be a touch panel equipped with a touch sensor. In this case, the touch panel functions as an input unit.

[0092] The communication unit can connect to the communication network 120 wirelessly or via a wired connection, and can send and receive data with other computer devices via the communication network 120. Data received via the communication unit is loaded into RAM, and calculation processing is performed by the control unit.

[0093] (Server device) Figure 3 is a block diagram showing the hardware configuration of a server device according to an embodiment. The server device 100 comprises at least a control unit 101, a storage unit 102, and a communication unit 106, each connected by an internal bus.

[0094] The control unit 101 consists of a CPU and executes programs stored in the memory unit 102 to control the server device 100. The control unit 101 also has an internal timer for timing. The memory unit 102 includes RAM and main memory. RAM is the work area of ​​the control unit 101. Main memory is a storage area for saving programs and data. Main memory functions as a recording medium for storing programs. The control unit 101 reads the programs and data stored in the memory unit 102 and loads them into RAM, and performs program execution processing based on information received from the user terminal 200 or the language model server 300. The program (program product) may be stored on a recording medium such as a CD-ROM.

[0095] (Language model) The language model server 300 is equipped with a language model. The language model is a model that has learned statistical features obtained from a large amount of text data. Preferably, the language model is a large-scale language model that has been trained on a large dataset and has a vast number of parameters. The language model may also be a generative AI. A generative AI is an artificial intelligence that automatically generates new content such as text, images, audio, video, and program code based on input data. The generative AI may be trained on a large dataset using a neural network model, especially a deep learning model such as the Transformer architecture, and may generate content through probabilistic generation or pattern prediction.

[0096] The language model server 300 is equipped with an inference engine. The inference engine loads the model parameters into VRAM. When a prompt is input, the inference engine converts the prompt string into tokens, performs a forward computation of the Transformer, and sequentially generates tokens based on the resulting probability distribution. The generated output is output as an answer to the source of the prompt input.

[0097] Here, the language model is assumed to be located outside the server device 100, but it may also be located inside the server device 100.

[0098] (system) System 10 can identify one or more treatment methods for a patient corresponding to the patient information, based on patient information and information regarding the criteria for selecting treatment methods. System 10 can identify one or more treatment methods for a patient corresponding to the patient information by receiving prompts containing patient information and information regarding the criteria for selecting treatment methods as input and requesting the execution of inference using a language model. System 10 can also identify one or more treatment methods without using a language model.

[0099] The treatment methods identified by System 10 are not particularly limited. System 10 can identify surgical procedures as treatment methods. System 10 can also identify drug regimens as treatment methods. A drug regimen refers to a plan of medication, including the types of drugs to be administered to the patient, drug combinations, usage, dosage, route of administration, administration interval and / or number of cycles.

[0100] System 10 can identify immunotherapy or cell therapy regimens. An immunotherapy regimen refers to a plan of immunotherapy, including the type of drug, dosage, route of administration, administration interval, and / or number of cycles. A cell therapy regimen refers to a plan of cell therapy, including cell collection, preparation of cell products, pretreatment, administration, monitoring, or toxicology management. Furthermore, System 10 can identify radiotherapy prescriptions. A radiotherapy prescription refers to a plan of radiotherapy, including whether it is external or internal beam radiation, radiation quality, total dose, single dose, dose rate, number of fractions, irradiation frequency, and / or irradiation technique.

[0101] For each treatment method, selection criteria are set, which are the standards under which that treatment method can be used. If the patient information meets the selection criteria for a treatment method, that treatment method can be used; if the patient information does not meet the selection criteria for that treatment method, that treatment method cannot be used. By performing such a determination for each treatment method, it becomes possible to identify the usable treatment method from among multiple treatment methods. This identification of treatment methods can be performed in step S25 described later, without using a machine learning-trained model, or it can be performed in step S9 using a language model.

[0102] Alternatively, instead of the method described above, a treatment method or selection criterion similar to patient information can be identified based on the similarity between a patient information vector (a vectorized representation of patient information) and a treatment method vector (a vectorized representation of treatment methods) (or a selection criterion vector (a vectorized representation of selection criteria), or a vectorized representation of information consisting of treatment methods and selection criteria). The similarity is determined based on Euclidean distance or cosine similarity. Treatment methods with a similarity above a threshold can be identified as suitable for the patient information. By identifying suitable treatment methods based on the similarity between patient information and treatment methods and / or selection criteria, it becomes possible to identify which treatment method is more suitable for the patient. This identification of treatment methods can be performed in step S25 described later without using a machine learning-trained model, or it can be performed in step S9 using a language model. When using a language model, the prompt may include information about patient information, treatment methods and / or selection criteria, and information requesting the identification of treatment methods based on their similarity.

[0103] <Embodiment 15> The matters described in Embodiment 15 can also be applied to Embodiments 1 to 14 and 16.

[0104] The prompt is generated on the server device 100. The generation of the prompt is performed in step S7 or S12, which will be described later.

[0105] The patient information included in the prompt is not particularly limited as long as it pertains to the patient, but it is preferably information that influences the decision on treatment methods. Patient information may include, for example, basic attributes such as age and sex; information about the disease and condition such as disease name (e.g., hepatocellular carcinoma, colorectal cancer, diabetes mellitus), primary site (e.g., liver, pancreas, stomach), presence or absence of metastasis (e.g., lung metastasis, liver metastasis), presence or absence of recurrence, complications (e.g., diabetes mellitus, hypertension), and tumor invasion (e.g., left hepatic vein invasion, middle hepatic vein invasion, right hepatic vein invasion, vascular invasion, bile duct invasion, gastrointestinal invasion, Glisson's sheath invasion); information about anatomical relationships such as proximity to blood vessels, bile ducts, and gastrointestinal tracts; information about organs and bodily functions such as low liver function, FLR (e.g., normal liver / damaged liver), and feasibility of massive hepatectomy (e.g., trisegmentectomy, right hepatectomy); pathological and histological findings; genetic and molecular biological information such as genetic abnormalities and their degree and gene expression levels; and immunological findings such as the tumor microenvironment.

[0106] Furthermore, the patient information included in the prompt may include information about the distance between organ A in the patient's body and another organ B that is different from organ A. Also, the patient information in the image file format used for inference processing, as described later, may include information about the distance between organ A in the patient's body and another organ B that is different from organ A (information that allows us to determine the distance between organ A and organ B). For example, if the criteria for enabling the selection of treatment method X include a criterion regarding the distance between organ A and organ B (for example, a criterion that treatment method X can be used if the distance between organ A and organ B is ○ cm or more), it becomes possible to make an appropriate judgment as to whether or not to identify treatment method X as a treatment method for the patient corresponding to the patient information.

[0107] Furthermore, the patient information included in the prompt may include information regarding the number of tumors in the patient's body, the size of the tumors, the stage of tumor progression, the anatomical location of the tumors, whether the location of the tumors is a common site, or whether the lesions are localized. In addition, the patient information in the image file format used for inference processing, as described later, may include information regarding the number of tumors in the patient's body, the size of the tumors, the stage of tumor progression, the anatomical location of the tumors, whether the location of the tumors is a common site, or whether the lesions are localized (information that allows these pieces of information to be obtained). For example, if the criteria for enabling the selection of treatment method Y include criteria regarding the number of tumors, the size of the tumors, the stage of tumor progression, the anatomical location of the tumors, and whether the location of the tumors is a common site (for example, a criterion such as "treatment method Y can be used if the number of tumors is ○ or less"), it becomes possible to make an appropriate judgment on whether or not to identify treatment method Y as a treatment method for the patient corresponding to the patient information.

[0108] The patient information used in the inference processing by the language model may be transmitted from the user terminal 200 to the server device 100, or it may be registered in the patient information database in advance. The patient information database may be stored in the storage unit 102 of the server device 100, or it may be located outside the server device 100.

[0109] Patient information can be obtained from a patient information database. This patient information may be in text format or image format. Furthermore, patient information may also include text information obtained through image analysis using artificial intelligence capable of image analysis, or through multimodal AI.

[0110] The text information included in the prompt may be text information entered via a keyboard or the like, or it may be text information read by OCR from an electronic medical record or handwritten medical record. The text information may include, for example, findings observed with the naked eye or findings observed with a microscope. In addition to the text information entered as a prompt, the patient information used in the inference process may include data in electronic file format. Data in electronic file format may include image data in image file format or data in PDF file format. Image data may include images taken by tomographic imaging methods such as PET, CT, and MRI, as well as data related to X-ray images, ultrasound images, and images taken with an endoscope.

[0111] If the selection criteria can be determined from image information related to the body, then a language model (e.g., multimodal AI) can be used to perform inference using electronic files such as image data of the patient's body and prompts containing patient information as input, and one or more treatment methods can be identified based on the selection criteria for treatment methods.

[0112] The selection criteria for treatment methods included in a prompt are the conditions for safely and effectively implementing the treatment method. Preferably, the prompt includes selection criteria for each treatment method, which are the criteria for which that treatment method can be used. Furthermore, it is preferable that the prompt includes selection criteria for multiple treatment methods corresponding to the patient's disease or condition, as included in the patient information.

[0113] The criteria for selecting a treatment method include off-label criteria, applicable criteria, or quasi-off-label criteria. The treatment selection criteria (off-label criteria and / or applicable criteria and / or quasi-off-label criteria) may be stored in the treatment method database in the form of a correspondence table that associates patient information with treatment methods applicable to the patient corresponding to that patient information.

[0114] The criteria for selecting a treatment method may include consensus, diagnostic criteria, or treatable traits. Consensus refers to recommendations or opinions reached through deliberation among multiple experts, for example. Diagnostic criteria are defined items and thresholds for diagnosing and classifying a particular disease. Treatable traits are measurable, clinically significant, and treatable patient characteristics.

[0115] As mentioned above, the criteria for selecting a treatment method may include criteria regarding the distance between an organ in the patient's body and another organ different from that organ when the treatment method is being implemented. As mentioned above, the criteria for selecting a treatment method may also include criteria regarding the number of tumors in the patient's body, the size of the tumors, the extent of tumor progression, the anatomical location of the tumors, whether the site where the tumors are located is a common site, or whether the lesions are localized.

[0116] Based on patient information from numerous real patients, the treatment methods administered to these patients, and the resulting outcomes, it is possible to identify selection criteria corresponding to treatment methods through inference using language models. These identified selection criteria can then be registered in a treatment method database.

[0117] The selection criteria for treatment methods included in the prompt can utilize those pre-registered in the treatment method database. The treatment method database may be stored in the storage unit 102 of the server device 100, or it may be located outside the server device 100.

[0118] The selection criteria for treatment methods to be included in the prompt can be determined by searching the treatment method database for treatment methods that are highly similar to the patient's information. The treatment method database stores selection criteria for each treatment method, which are the conditions under which that treatment method can be used.

[0119] The search is performed by identifying treatment methods highly similar to the patient information based on the Euclidean distance or cosine similarity between a patient information vector, which is a vector representation of at least some of the patient information (e.g., information about the patient's disease or condition), and a treatment method vector, which is a vector representation of treatment methods (or information about the characteristics and details of treatment methods) stored in the treatment method database. This makes it possible to include in the prompt one or more selection criteria for treatment methods that correspond to the patient's disease or condition contained in the patient information.

[0120] Furthermore, the treatment method database may be tagged for each treatment method or selection criterion (for example, the name of the treatment method or keywords in the selection criteria may be assigned as tags). By following the tags based on keywords included in the patient information, it is possible to search for the selection criteria of the treatment method corresponding to the patient information. This makes it possible to include one or more selection criteria for treatment methods corresponding to the patient's disease or condition included in the patient information in the prompt. In addition, tags may be assigned for each treatment method or selection criterion using inference with a language model.

[0121] The server device 100 generates prompts that include, for example, patient information obtained from a patient information database or user terminal 200, and selection criteria for treatment methods obtained from a treatment method database. The generated prompts may include information requesting that the server output a treatment method appropriate for the patient corresponding to the patient information.

[0122] The selection criteria for treatment methods included in the prompt may be generated using a language model. Alternatively, the selection criteria generated using the language model may be stored in the treatment method database. The administrator or operator of system 10 may send text information about treatment methods from a management terminal to the server device 100, or from a user terminal 200 to the server device 100. The server device 100 generates a prompt containing text information about treatment methods, and based on this prompt, it can use the language model to output selection criteria for treatment methods as an answer. The output selection criteria for treatment methods are stored in the treatment method database.

[0123] Alternatively, the administrator or operator of system 10 may input selection criteria for treatment methods at a management terminal, and the input criteria may be stored in the treatment method database via the server device 100 (or without going through the server device 100). Alternatively, the user terminal 200 may input selection criteria for treatment methods, and the input criteria may be stored in the treatment method database via the server device 100 (or without going through the server device 100). In this way, users can customize the selection criteria for treatment methods.

[0124] Textual information regarding treatment methods is not particularly limited and can be selected as appropriate, but examples include research papers and guidelines that incorporate physician consensus.

[0125] The server device 100 requests the execution of inference using the language model by inputting the generated prompt to the language model server 300. The language model server 300's inference engine performs a forward calculation of the Transformer based on the input prompt and outputs an answer. The output answer is sent from the language model server 300 to the server device 100, and then from the server device 100 to the user terminal 200. The server device 100 may process a portion of the answer received from the language model server 300 and then send the processed answer to the user terminal 200.

[0126] The prompts generated by the server device 100 may include evaluation criteria for assessing the patient's condition. The evaluation criteria are not particularly limited as long as they allow for the assessment of the patient's condition and symptoms based on patient information, and may include, for example, criteria for assessing the severity of the condition according to the patient's condition and symptoms. In particular, for benign diseases, it is effective to identify the severity according to the patient's condition based on the evaluation criteria and then identify the treatment method according to this severity. In other words, the selection criteria may be such that different treatment methods can be selected depending on the evaluation results based on evaluation criteria such as severity.

[0127] While there are no particular limitations on the evaluation criteria, examples include the frequency and severity of symptoms (e.g., number of attacks, pain score, number of diarrhea episodes), the degree of impairment in daily living activities and physical function (e.g., walking ability, ADL, exercise limitations), objective laboratory values ​​and biomarkers (e.g., CRP, complement, BNP, CK levels), organ and tissue-specific clinical findings (e.g., joint swelling, skin hardening, endoscopic score), the patient's own subjective health assessment (e.g., VAS, QoL score, sense of control), and the presence or absence of complications and serious events (e.g., nephritis, neuropathy, fistula / abscess).

[0128] Evaluation criteria are established for each disease. For example, if the disease is asthma (ACT, GINA), the evaluation criteria include the frequency and severity of symptoms / subjective assessment (more specifically, the frequency of attacks, nocturnal symptoms, number of emergency inhaler use, limitations in daily life, and sense of control). For example, if the disease is rheumatoid arthritis (DAS28, CDAI), the evaluation criteria include organ-specific findings / biomarkers / subjective assessment (more specifically, the number of tender joints, the number of swollen joints, CRP / ESR, and patient VAS).

[0129] The evaluation criteria may be those pre-registered in the treatment method database, or they may be those pre-registered in a separate evaluation criteria database. The server device 100 generates a prompt that includes, for example, patient information obtained from the patient information database or user terminal 200, and treatment method selection criteria obtained from the treatment method database.

[0130] The evaluation criteria to be included in the prompt can be identified by searching for evaluation criteria with high similarity to the patient information from the evaluation criteria database. The search is performed by identifying evaluation criteria with high similarity to the patient information based on the Euclidean distance or cosine similarity between a patient information vector, which is a vector of at least some of the patient information (for example, information about the patient's disease or condition), and an evaluation criterion vector, which is a vector of information about the evaluation criteria stored in the evaluation criteria database. This makes it possible to include one or more evaluation criteria corresponding to the patient's disease or condition contained in the patient information in the prompt.

[0131] In the evaluation criteria database, tags may be assigned to each disease or each evaluation criterion (for example, the name of the disease or the evaluation items of the evaluation criteria may be assigned as tags). Based on keywords contained in the patient information, tags can be followed to search for evaluation criteria corresponding to the patient information. This makes it possible to include one or more evaluation criteria corresponding to the patient's disease or condition as contained in the patient information in the prompt. Furthermore, tags may be assigned to each disease or evaluation criterion using inference with a language model.

[0132] Diseases are broadly classified into two types: malignant diseases (cancer / tumors) and benign diseases. The prompt may include information that asks the user to identify whether the patient's disease is malignant (cancer / tumor) or benign, based on the patient's information. The prompt may also include information that asks the user to identify a treatment method based on selection criteria after identifying whether the patient's disease is malignant (cancer / tumor) or benign. Information on whether the patient's disease is malignant or benign, as determined by inference using a language model, may be displayed on the display screen of the user terminal 200.

[0133] Furthermore, treatment for malignant diseases can be classified into systemic therapy and local therapy. The key to selection is whether the number of lesion sites can be counted, which can be determined using visual data, i.e., images (CT, MRI, endoscopic findings, ultrasound findings, etc.). If the number of lesion sites cannot be counted, systemic therapy is selected; if the number of lesion sites can be counted, local therapy is selected. The prompt may include information that asks the user to identify whether systemic therapy or local therapy is appropriate for the malignant disease, based on patient information. The prompt may also include information that asks the user to identify whether systemic therapy or local therapy is appropriate for the malignant disease, based on patient information, and then to identify the treatment method according to the determination of which is appropriate. Furthermore, the prompt may include information that asks the user to determine whether the number of lesion sites is above a certain level (or whether the number of lesion sites can be counted), based on patient information, and then to identify whether systemic therapy or local therapy is appropriate. The information regarding which is appropriate, systemic therapy or local therapy, identified by inference using a language model, may be displayed on the display screen of the user terminal 200.

[0134] The prompt may include information requesting the user to create findings based on patient information (including image data taken inside the patient's body). Findings refer to facts and opinions that can be confirmed from the patient information. The prompt may also include information requesting the user to identify a treatment method based on the created findings. Findings generated by inference using a language model may be displayed on the display screen of the user terminal 200.

[0135] Furthermore, image data taken from inside the patient's body can be analyzed using artificial intelligence capable of analyzing images.

[0136] Furthermore, the prompts generated by the server device 100 may include not only patient information and criteria for selecting treatment methods, but also target outcomes. The generated prompts may also include information requesting that a treatment method based on the target outcome be output. By including target outcomes in the prompts, it is possible to identify one or more treatment methods based on the target outcome for the patient corresponding to the patient information.

[0137] An outcome refers to the symptoms, medical condition, health status, physical function / ability, or quality of life that a patient experiences as a result of treatment. Outcomes also include outcome indicators that show survival rates, mortality rates, recovery rates, or readmission rates after a specified period, such as a 5-year survival rate of X%. Alternatively, an outcome refers to a change in the patient's symptoms, medical condition, health status, physical function / ability, or quality of life as a result of treatment.

[0138] A target outcome is the desired outcome, and it is information about the treatment result that patients and medical professionals such as doctors desire. The content of a target outcome is not particularly limited, but examples include recovery of motor function, a change in cancer stage, or achieving a 5-year survival rate of X%.

[0139] The target outcome is input, for example, by a user terminal 200 and transmitted to the server device 100. The target outcome may also be input as text information.

[0140] The user can operate the user terminal 200 to select whether or not to limit the treatment methods received as answers to only those covered by public health insurance. If the user selects to limit the treatment methods received as answers to only those covered by public health insurance, information indicating that this selection has been made is sent from the user terminal 2 to the server device 100.

[0141] In this case, the prompt generated by the server device 100 may include information requesting that only treatment methods covered by public health insurance be identified. As a result, the language model server 300 identifies only treatment methods covered by public health insurance, and only the identified treatment methods covered by public health insurance are sent to the user terminal 200 via the server device 100.

[0142] The user can operate the user terminal 200 to select whether or not to include treatment methods not covered by public health insurance in the treatment methods obtained as a response. If the user selects whether or not to include treatment methods not covered by public health insurance in the treatment methods obtained as a response, information regarding the selection is sent from the user terminal 2 to the server device 100.

[0143] If the user chooses to include treatment methods not covered by public health insurance in the treatment methods obtained as answers, the prompt generated by the server device 100 may include information requesting the user to specify a treatment method while including treatment methods not covered by public health insurance as candidates. As a result, the language model server 300 will specify a treatment method while including treatment methods covered by public health insurance as candidates, and the specified treatment method will be sent to the user terminal 200 via the server device 100. On the other hand, if the user chooses not to include treatment methods not covered by public health insurance in the treatment methods obtained as answers, the prompt generated by the server device 100 may include information requesting the user to specify a treatment method without including treatment methods not covered by public health insurance as candidates.

[0144] System 10 can also determine whether there is insufficient patient information necessary to determine the appropriateness of a treatment method. Furthermore, if there is insufficient information necessary to determine the appropriateness of a treatment method, it can also identify the missing information. For example, the prompt generated by the server device 100 may include information requesting that the system determine whether there is insufficient information necessary to determine the appropriateness of a treatment method, along with the patient information. Moreover, the prompt generated by the server device 100 may include information requesting that, if there is insufficient information necessary to determine the appropriateness of a treatment method, the system output the missing information as an answer. This allows System 10 to determine whether there is insufficient information necessary to determine the appropriateness of a treatment method, and if there is insufficient information, to identify the missing information. The information output from the language model server 300, indicating that information is insufficient and which information is missing, is transmitted to the user terminal 200 via the server device 100.

[0145] In this example, the prompts generated by the server device 100 are assumed to include patient information and criteria for selecting a treatment method. However, for example, the language model may be fine-tuned using information regarding the criteria for selecting a treatment method. In this case, the prompts generated by the server device 100 only need to include patient information and do not need to include criteria for selecting a treatment method. Even in this case, the generated prompts may include information requesting the identification of a treatment method appropriate for the patient corresponding to the patient information.

[0146] The server device 100 can identify information regarding the benefits and / or drawbacks of performing a treatment method on a patient. This information regarding the benefits and / or drawbacks of performing a treatment method can be obtained from a treatment method database. The identified information, including the benefits and / or drawbacks, is transmitted from the server device 100 to the user terminal 200 and displayed on the user terminal 200.

[0147] The server device 100 can identify the patient's survival rate, the probability of complete cure of the disease or condition, the recurrence rate of the disease or condition, or the local control rate of the disease or condition as a result of implementing the identified treatment method. The patient's survival rate, the probability of complete cure of the disease or condition, the recurrence rate of the disease or condition, or the local control rate of the disease or condition when the treatment method is implemented may be obtained from the treatment method database or from the language model server 300. The patient's survival rate, the probability of complete cure of the disease or condition, the recurrence rate of the disease or condition, or the local control rate of the disease or condition when the treatment method is implemented is transmitted from the server device 100 to the user terminal 200 and displayed on the user terminal 200.

[0148] The server device 100 can identify the expected outcomes when a patient undergoes a specified treatment method. The expected outcomes when a patient undergoes a treatment method can be identified using information for identifying outcome predictions registered in the treatment method database. More specifically, the server device 100 can generate a prompt containing patient information, a treatment method, and information for identifying outcome predictions, and obtain outcome predictions from the language model server 300 that correspond to the patient information when that treatment method is performed.

[0149] Information for identifying outcome predictions can include, for example, a correspondence table between outcome predictions and outcome factors (proven in research papers, etc.), or a calculation formula for identifying outcome predictions. The calculation formula allows for the identification of outcome predictions by substituting items identified from the patient's condition. Alternatively, a correspondence table between outcome predictions and outcome factors, or a calculation formula for identifying outcome predictions, can be identified using language model inference based on patient information for a large number of real patients, the treatment methods performed on these patients, and the resulting outcome information.

[0150] System 10 can determine the degree of fit for a treatment method identified by the method identification means, which is the degree to which it is suitable for the target outcome or risk tolerance. The user can input the target outcome or risk tolerance by operating the user terminal 200. The target outcome or risk tolerance entered at the user terminal 200 is transmitted to the user terminal 200 via the server device 100 and displayed at the user terminal 200.

[0151] (Output processing flowchart) Next, the output processing will be explained. Figure 5 is a flowchart of the output processing according to the embodiment. The user operates the user terminal 200 to launch the application and log in (step S1). The application may be a web application or a native application. When logging in, the user enters a user ID (identification information that identifies the user).

[0152] Next, the user operates the user terminal 200 to input patient information (step S2). The input patient information is transmitted from the user terminal 200 to the server device 100 (step S3), and received by the server device 100 (step S4).

[0153] The server device 100 retrieves information about treatment methods related to the patient information from the treatment method database based on the patient information (step S5). This information about treatment methods includes information about the selection criteria for each treatment method. Next, the server device 100 retrieves information about evaluation criteria related to the patient information from the evaluation criteria database based on the patient information (step S6). This information about evaluation criteria includes information about evaluation criteria such as the severity of symptoms and conditions.

[0154] Next, the control unit of the server device 100 generates a prompt (step S7) that includes patient information received in step S4, information regarding treatment method selection criteria obtained in step S5, and information regarding evaluation criteria obtained in step S6. The prompt may also include information requesting that the severity of the patient corresponding to the patient information be identified according to the evaluation criteria, and that a treatment method suitable for the selection criteria be identified based on the patient information and severity. The generated prompt is transmitted from the server device 100 to the language model server 300 and received by the language model server 300 (step S8).

[0155] In the language model server 300, the inference engine uses the language model to perform inference based on the received prompt and generates a response (step S9). The response includes information on appropriate treatment methods for patients corresponding to the patient information received in step S4.

[0156] Next, the generated response (information on appropriate treatment methods for the patient) is transmitted from the language model server 300 to the server device 100 (step S10), and received by the server device 100 (step S11).

[0157] Next, the server device 100 generates a prompt that includes patient information received in step S4 and information regarding a treatment method suitable for the patient received in step S11 (step S12). The prompt may also include information requesting the expected outcome if the treatment method included in the prompt is implemented for the patient corresponding to the patient information. The generated prompt is transmitted from the server device 100 to the language model server 300 and received by the language model server 300 (step S13).

[0158] In the language model server 300, the inference engine uses the language model to perform inference based on the received prompt and generates a response (step S14). The response includes patient information received in step S4 and information about the expected outcome if a treatment method appropriate for the patient, received in step S11, is implemented.

[0159] Next, the generated response (information on the predicted outcome) is transmitted from the language model server 300 to the server device 100 (step S15), and received by the server device 100 (step S16). Next, the server device 100 transmits information on appropriate treatment methods for the patient received in step S11, and information on the predicted outcome received in step S16, to the user terminal 200 (step S17), and is received by the user terminal 200 (step S18). The information received by the user terminal 200 is displayed on the display screen of the user terminal 2 (step S19). The output processing is completed by steps S1 to S19.

[0160] <Embodiment 16> The matters described in Embodiment 16 can also be applied to Embodiments 1 to 15. The system 10 of Embodiment 16 can have the same configuration as the system 10 shown in Figure 4. However, the system 10 does not need to include a language model server 300. The system 10 includes at least one computer device. The system 10 may consist of, for example, one computer device (information processing device) (standalone type), one or more server devices and one or more terminal devices (client-server type), or multiple terminal devices (peer-to-peer type).

[0161] System 10 may include a user terminal 200 and a server device 100. The server device 100 corresponds to the computer 100 described above. Furthermore, the server device 100 may function in a distributed manner across multiple computer devices. For example, instead of the server device 100, a distributed ledger technology such as blockchain may be used.

[0162] The user terminal 200 and the server device 100 are connected to each other via the communication network 120 so that they can communicate with one another. The number of user terminals 200 is not particularly limited, and there may be one or more user terminals 200.

[0163] (User terminal) User terminal 200 is a terminal operated by a user of system 10. Here, the user can be anyone who uses system 10. Examples of user terminal 200 include conventional mobile phones, tablet terminals, smartphones, and personal computers.

[0164] The user terminal 200 comprises a control unit, a storage unit, an input unit, a display unit, and a communication unit, each connected by a bus. The functions and roles of the control unit, storage unit, input unit, display unit, and communication unit of the user terminal 200 are as described in Embodiment 15.

[0165] (Server device) Figure 3 is a block diagram showing the hardware configuration of a server device according to an embodiment. The server device 100 comprises at least a control unit 101, a storage unit 102, and a communication unit 106, each connected by an internal bus. The functions and roles of the control unit 101, storage unit 102, and communication unit 106 of the server device 100 are as described in Embodiment 15.

[0166] (system) System 10 can identify a treatment method corresponding to the patient information desired by the user, based on a trained model that uses patient information for machine learning as input data and treatment methods suitable for the patient for machine learning as output data. The trained model is constructed by machine learning on multiple datasets consisting of patient information and treatment methods suitable for that patient. This enables the identification of a treatment method suitable for the patient.

[0167] The server device 100 can also update the trained model by retraining it based on new patient information for machine learning and treatment methods suitable for the new patient information for machine learning.

[0168] Furthermore, System 10 can identify one or more treatment methods corresponding to patient information and target outcomes based on a trained model that has been trained using patient information for machine learning and outcomes as results of implementing treatment methods for patients corresponding to the patient information for machine learning as input data, and treatment methods implemented for patients corresponding to the patient information for machine learning as output data. The trained model is a model constructed by machine learning on multiple datasets consisting of patient information, outcomes as results of implementing treatment methods for patients corresponding to that patient information, and treatment methods implemented for patients corresponding to that patient information. This makes it possible to identify treatment methods that are suitable for the patient and that are compatible with the target outcome.

[0169] It is also possible to update the trained model by retraining it based on new patient information for machine learning, outcomes resulting from performing the treatment method on patients corresponding to the new patient information for machine learning, and the treatment method performed on patients corresponding to the patient information for machine learning.

[0170] The trained model is stored in the server device 100. When a user operates the user terminal 200 and inputs patient information, the input patient information is transmitted from the user terminal 200 to the server device 100. Based on the received patient information, the server device 100 can use the trained model to identify a treatment method corresponding to the received patient information. The identified treatment method is transmitted from the server device 100 to the user terminal 200 and displayed on the user terminal 200.

[0171] Furthermore, when a user operates the user terminal 200 to input patient information and target outcomes, the input patient information and target outcomes are transmitted from the user terminal 200 to the server device 100. Target outcomes are input as text information. Based on the received patient information and target outcomes, the server device 100 can use a trained model to identify a treatment method corresponding to the received patient information. The identified treatment method is transmitted from the server device 100 to the user terminal 200 and displayed on the user terminal 200.

[0172] Patient information for machine learning, or patient information of patients for whom treatment methods are to be identified, is not particularly limited as long as it is information about the patient, but it is preferable that it is information that influences the decision on treatment methods. Patient information may include, for example, basic attributes such as age and sex; information about the disease and condition such as disease name (e.g., hepatocellular carcinoma, colorectal cancer, diabetes mellitus), primary site (e.g., liver, pancreas, stomach), presence or absence of metastasis (e.g., lung metastasis, liver metastasis), presence or absence of recurrence, complications (e.g., diabetes mellitus, hypertension), and tumor invasion (e.g., left hepatic vein invasion, middle hepatic vein invasion, right hepatic vein invasion, vascular invasion, bile duct invasion, gastrointestinal invasion, Glisson's sheath invasion); information about anatomical relationships such as proximity to blood vessels, bile ducts, and gastrointestinal tracts; information about organs and bodily functions such as low liver function, FLR (e.g., normal liver / damaged liver), and feasibility of massive hepatectomy (e.g., trisegmentectomy, right hepatectomy); pathological and histological findings; genetic and molecular biological information such as genetic abnormalities and their degree and gene expression levels; and immunological findings such as the tumor microenvironment.

[0173] A treatment method appropriate to patient information is a treatment method appropriate for the patient corresponding to that patient information. A patient-appropriate treatment method used in machine learning may differ from the treatment method actually performed on the patient. For example, a patient-appropriate treatment method used in machine learning may be one that a specialist deduces to be appropriate based on patient information.

[0174] A target outcome is the desired outcome, and it is information about the treatment result that patients and medical professionals such as doctors desire. The content of a target outcome is not particularly limited, but examples include the recovery of motor function or a change in the stage of cancer.

[0175] System 10 uses patient information for machine learning and treatment methods for patients corresponding to that patient information as input data, and outputs the outcomes when those treatment methods are implemented for patients corresponding to that patient information as output data. Based on this trained model, it can identify the outcomes when a treatment method is implemented for patients corresponding to the acquired patient information. The trained model is constructed by machine learning on multiple datasets consisting of patient information, treatment methods for patients corresponding to that patient information, and outcomes when those treatment methods are actually implemented for patients corresponding to that patient information. This makes it possible to predict the outcomes when treatment is implemented for patients.

[0176] As described above, System 10 can identify a treatment method appropriate for a patient based on patient information, or based on patient information and target outcomes. Furthermore, System 10 can also predict outcomes based on this identified treatment method.

[0177] The machine learning algorithm used in the above-described trained model is not particularly limited and any known algorithm can be used, but it is preferable to use deep learning using a multilayer neural network. A multilayer neural network has an input layer, an output layer, and multiple hidden layers. Weights are assigned to the edges connecting the nodes in each layer. Each edge has a weight corresponding to each input to the node, and the weights corresponding to each input to the node are multiplied, and the value obtained by multiplying these weights is added to the bias. The resulting value is subjected to a nonlinear transformation using an activation function to calculate the activation value. The calculated activation value becomes the input value passed to the node of the next layer. The number of hidden layers can be designed as appropriate.

[0178] (Output processing flowchart) Next, the output processing will be explained. Figure 6 is a flowchart of the output processing according to the embodiment. The user operates the user terminal 200 to launch the application and log in (step S21). The application may be a web application or a native application. When logging in, the user enters a user ID (identification information that identifies the user).

[0179] Next, the user operates the user terminal 200 to input patient information (step S22). The input patient information is transmitted from the user terminal 200 to the server device 100 (step S23) and received by the server device 100 (step S24). Based on the input patient information, the server device 100 identifies one or more treatment methods corresponding to the acquired patient information (step S25). In step S25, the treatment method corresponding to the acquired patient information can be identified using a trained model that has been trained using patient information for machine learning as input data and treatment methods suitable for the patient information for machine learning as output data.

[0180] Next, based on the patient information received in step S24 and the treatment method identified in step S25, the predicted outcome when the treatment method identified in step S25 is performed on the patient corresponding to the patient information received in step S24 is identified (step S26). In step S26, the predicted outcome can be identified using a trained model that has been trained using patient information for machine learning and the treatment method for the patient corresponding to the patient information for machine learning as input data, and the outcome when the treatment method is performed on the patient corresponding to the patient information for machine learning as output data.

[0181] The treatment method identified in step S25 and the predicted outcome identified in step S26 are transmitted from the server device 100 to the user terminal 200 (step S27). The information received by the user terminal 200 is displayed on the display screen of the user terminal 2 (step S28). The output processing is completed by steps S21 to S28. [Explanation of Symbols]

[0182] 10 Systems 100 Computers (Server Devices) 101 Control Unit 102 Storage section 103 Peripheral Device I / F Section 104 Input section 105 Display section 106 Communications Department 110 Bus 120 Networks 130 External Servers 140 Databases 200 user terminals 300 language model servers

Claims

1. A system comprising at least one computer device, Means of obtaining patient information, A method and means for identifying one or more treatment methods for a patient corresponding to the acquired patient information, based on the acquired patient information and the criteria for selecting treatment methods. An outcome identification means that identifies the predicted outcome when the treatment method is performed on a patient corresponding to the acquired patient information, by requesting the execution of inference using a language model, taking as input the acquired patient information, the identified treatment method, and a prompt that includes the correspondence between the patient information and the outcome when the treatment method is performed on the patient corresponding to the patient information. A system equipped with these features.

2. The system according to claim 1, wherein the method-specific means specifies a surgical procedure, an immunotherapy or cell therapy regimen, or a radiotherapy prescription.

3. The method identification means, taking acquired patient information and prompts including criteria for selecting treatment methods as input, requests the execution of inference using a language model to identify one or more treatment methods for a patient corresponding to the acquired patient information. The system according to claim 1 or 2.

4. The system according to claim 1 or 2, wherein the method-specific means identifies one or more treatment methods based on evaluation criteria for the patient's medical condition.

5. The system according to claim 3, wherein the method identification means identifies one or more treatment methods for a patient corresponding to acquired patient information by taking a prompt including selection criteria for treatment methods obtained from a knowledge base as input and requesting the execution of inference using a language model.

6. The system according to claim 3, wherein the method identification means identifies one or more treatment methods for a patient corresponding to the acquired patient information by taking a prompt containing acquired patient information as input and requesting the execution of inference using a language model finely tuned with information on criteria for selecting treatment methods.

7. The system according to claim 3, wherein the method identification means identifies one or more treatment methods for a patient corresponding to acquired patient information by requesting the execution of inference using a language model, taking as input a prompt containing information about the patient's medical record and / or information about an image taken inside the patient's body.

8. An input method for inputting the target outcome, The system according to claim 3, wherein the method identification means identifies one or more treatment methods based on the target outcome for a patient corresponding to the acquired patient information by requesting the execution of inference using a language model, taking prompts including patient information and target outcomes as input.

9. Patient information includes image information related to the body, The selection criteria are those that can be determined from image information related to the body. The system according to claim 1 or 2, wherein the method identification means identifies one or more treatment methods for a patient corresponding to the acquired patient information, based on the acquired patient information and the criteria for selecting a treatment method.

10. The patient information includes information that allows us to determine the number of tumors in the patient's body, the size of the tumors, the stage of tumor progression, the anatomical location of the tumors, whether the location of the tumors is a common site, or whether the lesions are localized. The system according to claim 1 or 2, wherein the criteria for selecting a treatment method include information that allows for the selection of at least one treatment method, such as the number of tumors in the patient's body, the size of the tumors, the extent of tumor progression, the anatomical location of the tumors, whether the site where the tumors are located is a common site, or whether the lesions are localized.

11. A first information identification means that identifies information including the advantages and / or disadvantages of a treatment method based on the treatment method identified by the method identification means. The system according to claim 1 or 2, comprising:

12. A second information identification means that identifies the survival rate, cure rate, recurrence rate, or local control rate resulting from the implementation of the treatment method identified by the method identification means. The system according to claim 1 or 2, comprising:

13. A third information identification means that identifies the prediction of the outcome based on the treatment method identified by the method identification means. The system according to claim 1 or 2, comprising:

14. An input means for inputting the patient's target outcome or the patient's risk tolerance, A fit determination means that identifies information that allows for the determination of the degree to which a treatment method identified by a method determination means is appropriate for the input target outcome or risk tolerance. The system according to claim 1 or 2, comprising:

15. The system according to claim 1 or 2, wherein the criteria for selecting a treatment method are generated by requesting the execution of inference using a language model, with prompts containing text information about treatment methods as input.

16. The system according to claim 1 or 2, wherein the method-specific means specifies only treatment methods to which public health insurance applies.

17. The system according to claim 1 or 2, wherein the method-specific means can identify treatment methods that are not covered by public health insurance.

18. A means for identifying missing information when the acquired patient information lacks the necessary information to determine the appropriateness of a treatment method. The system according to claim 1 or 2, comprising:

19. A method and means for identifying one or more treatment methods for a patient corresponding to the acquired patient information, based on the acquired patient information and the criteria for selecting treatment methods. Equipped with, The system according to claim 1 or 2, wherein the acquisition means acquires a treatment method specified by the method-specific means.

20. The system according to claim 1 or 2, wherein patient information includes information relating to images taken inside the patient's body.

21. The system according to claim 1 or 2, wherein patient information includes text information.

22. The system according to claim 1 or 2, wherein the criteria for selecting a treatment method include off-label criteria, applicable criteria, or quasi-off-label criteria.

23. A method performed in a system comprising at least one computer device, Acquisition steps for obtaining patient information, A method for identifying one or more treatment methods for a patient corresponding to the acquired patient information, based on the acquired patient information and the criteria for selecting treatment methods. An outcome identification step involves requesting the execution of inference using a language model, taking as input the acquired patient information, the identified treatment method, and a prompt that includes the correspondence between the patient information and the outcome when the treatment method is performed on the patient corresponding to the patient information, thereby identifying the predicted outcome when the treatment method is performed on the patient corresponding to the acquired patient information. A method having

Images