Estimation support device and trained model, care support device and outcome prediction device equipped with said estimation support device, and outcome prediction device equipped with said care support device
The estimation support device and learned model effectively estimate cancer patients' symptoms, aiding in formulating care plans and predicting outcomes, addressing the challenge of subjective symptom evaluation in cancer patients.
Patent Information
- Application Number
- JP2020215401
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2019-12-25
- Filing Date
- 2020-12-24
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2040-12-24
AI Technical Summary
Evaluating the subjective symptoms of cancer patients, particularly in their final stages, is challenging and often difficult even for experienced medical personnel.
An estimation support device and a learned model that estimates cancer patients' symptoms using patient information and a pre-generated trained model, which can consider attribute information and initial evaluation information, including a decision tree configuration to output symptom estimation results.
Enables accurate estimation of cancer patients' symptoms, facilitating the formulation of optimal care plans and predicting outcomes, thereby improving care planning and decision-making for cancer patients.
Smart Images

Figure 0007589981000001 
Figure 0007589981000002 
Figure 0007589981000003
Abstract
Description
[Technical field]
[0001] The present disclosure relates to an estimation support device and a learned model, a care support device and an outcome prediction device including the estimation support device, and an outcome prediction device including the care support device. [Background technology]
[0002] For example, Patent Document 1 discloses a system for objectively grasping the subjective symptoms of a disease patient. [Prior art documents] [Patent documents]
[0003] [Patent Document 1] JP 2002-336259 A Summary of the Invention [Problem to be solved by the invention]
[0004] In the terminal stage of cancer patients, it is often difficult to communicate with them and to estimate (assess) their symptoms. In particular, even experienced medical professionals have great difficulty in assessing the patient's subjective symptoms.
[0005] An object of the present disclosure is to provide an estimation support device and a trained model that can be used to estimate the condition of a cancer patient, a care support device and an outcome prediction device equipped with the estimation support device, and an outcome prediction device equipped with the care support device. [Means for solving the problem]
[0006] An estimation support device according to one aspect of the present disclosure includes an input unit into which patient information including attribute information of a cancer patient and initial assessment information regarding plant system functional disorders is input, an estimation unit that estimates symptoms of a cancer patient using the patient information input into the input unit and a trained model generated in advance, and an output unit that outputs the estimation results of the estimation unit, where the trained model is a trained model generated using teacher data so as to output symptom estimation information of a cancer patient when patient information is input.
[0007] According to the above estimation support device, the symptoms of a cancer patient are estimated using the patient information and a trained model generated in advance, and the estimation result is output. Therefore, the estimation support device can be used to estimate the symptoms of a cancer patient.
[0008] The attribute information may include at least one of information regarding the medical setting, age, sex, type of cancer, treatment stage, Eastern Cooperative Oncology Group Performance Status (ECOG-PS), and requester. The symptoms of the cancer patient can be estimated by taking these items into consideration.
[0009] The initial evaluation information may include information regarding at least one of loss of appetite, nausea and vomiting, abdominal distension, constipation, edema, and insomnia. By taking these items into consideration, the symptoms of the cancer patient can be estimated.
[0010] Symptoms of cancer patients may include at least one of the following symptoms caused by animal system dysfunction: pain, dyspnea, cough, phlegm, fatigue, drowsiness, anxiety, depression, emotional discomfort, delirium, lack of informed consent, and spiritual pain. This allows us to estimate symptoms that are difficult to evaluate objectively (subjective symptoms).
[0011] The trained model may be a decision tree. In this manner, a decision tree can be used as the trained model.
[0012] A care support device according to one aspect of the present disclosure includes the above-mentioned estimation support device, and a generation unit that generates a proposed care plan for a cancer patient using at least one of patient information, estimation results, and observation information on the estimation results, as well as a second trained model generated in advance, wherein the second trained model is generated using teacher data different from the trained model, and the output unit outputs the proposed care plan.
[0013] According to the above-mentioned care support device, a care plan based on the symptom estimation information of the cancer patient can be obtained by using at least one of the estimation result by the estimation support device and the observation information on the estimation result. This allows general medical practitioners to refer to the care plan when formulating a care plan for the cancer patient. Therefore, the care support device can be used to formulate the optimal care plan for the cancer patient.
[0014] An outcome prediction device according to one aspect of the present disclosure includes the care support device, and a prediction unit that generates an outcome prediction for a cancer patient over a specified period of time using the above-mentioned care support device, at least one of patient information, estimated results and observation information, at least one of a proposed care plan and a planned care plan, and a third trained model generated in advance, wherein the third trained model is generated using teacher data different from the trained model and the second trained model, and the output unit outputs the outcome prediction.
[0015] An outcome prediction device according to one aspect of the present disclosure includes the above-mentioned estimation support device, and a prediction unit that generates an outcome prediction for a cancer patient over a specified period of time using patient information, estimation results, and / or observation information on the estimation results, a planned care plan input via an input unit, and a third trained model generated in advance, wherein the third trained model is generated using teacher data different from the trained model, and an output unit outputs the outcome prediction.
[0016] According to the above-mentioned outcome prediction device, by using at least one of the estimation result and the observation information by the estimation support device, and at least one of the proposed care plan and the planned care plan, etc., a cancer patient's outcome prediction for a predetermined period of time can be obtained. This allows, for example, a medical professional to objectively confirm the appropriateness of the planned care plan for the cancer patient. Therefore, the outcome prediction device can be used to improve the accuracy of the optimal care plan. In addition, for example, a cancer patient can objectively view examples of his or her future condition transition by checking the outcome prediction. This makes it easier for the cancer patient to make decisions regarding his or her wishes and choices regarding medical care. Therefore, the outcome prediction device can be used to plan Advance Care Planning (ACP).
[0017] A trained model according to one aspect of the present disclosure includes an input node, an output node, and an intermediate node provided between the input node and the output node, and is generated using training data that matches patient information, including attribute information of cancer patients and information on disorders of plant system functions, with symptom estimation information of cancer patients, and causes a computer to function such that when patient information is input to the input node, the output node outputs the symptom estimation information of the cancer patient.
[0018] According to the trained model, symptom estimation information of a cancer patient can be obtained from patient information. Therefore, the trained model can be used to estimate the symptoms of a cancer patient.
[0019] The trained model may be a decision tree in which the input node is a root node, the output node is a leaf node, and the intermediate node branches the path between the root node and the leaf node. In this way, a decision tree can be used as the trained model. Effect of the Invention
[0020] According to the present disclosure, there is provided an estimation support device and a trained model that can be used to estimate the condition of a cancer patient, a care support device and an outcome prediction device equipped with the estimation support device, and an outcome prediction device equipped with the care support device. [Brief description of the drawings]
[0021] [Figure 1] FIG. 1 is a diagram showing an outline of the estimation support device. [Diagram 2] FIG. 2 is a diagram showing an outline of the estimation support device. [Diagram 3] FIG. 3 is a diagram showing an outline of the estimation support device. [Figure 4] FIGS. 4(a), (b), and (c) are diagrams showing examples of functional blocks of the estimation support device. [Diagram 5] FIG. 5 is a diagram illustrating an example of a decision tree configuration. [Figure 6] FIG. 6 is a diagram illustrating an example of input data in the training data. [Figure 7] FIG. 7 is a diagram illustrating an example of output data in the teacher data. [Figure 8] FIG. 8 is a diagram showing the occurrence frequency of feature amounts at upstream nodes of a decision tree. [Figure 9] FIG. 9 is a diagram showing the occurrence frequency of feature amounts at upstream nodes of a decision tree. [Figure 10] FIG. 10 is a diagram showing the occurrence frequency of feature amounts at upstream nodes of a decision tree. [Figure 11] FIG. 11 is a diagram showing the occurrence frequency of feature amounts at upstream nodes of a decision tree. [Figure 12] FIG. 12 is a diagram showing the occurrence frequency of feature amounts at upstream nodes of a decision tree. [Figure 13] FIG. 13 is a diagram showing the occurrence frequency of feature amounts at upstream nodes of a decision tree. [Figure 14] FIG. 14 is a diagram showing the occurrence frequency of feature amounts in upstream nodes of a decision tree. [Figure 15]FIG. 15 is a diagram showing the occurrence frequency of feature amounts at upstream nodes of a decision tree. [Figure 16] FIG. 16 is a diagram illustrating an example of functional blocks of the care support device. [Figure 17] FIG. 17 shows an example of output data in the training data of the second trained model. [Figure 18] FIG. 18 is a graph generated using the evaluation data results. [Figure 19] FIG. 19 is a graph generated using the evaluation data results. [Figure 20] FIG. 20 is a graph generated using the evaluation data results. [Figure 21] FIG. 21 is a diagram illustrating an example of functional blocks of an outcome prediction device. [Figure 22] FIG. 22 shows an example of output data for the third trained model. [Figure 23] FIG. 23 is a graph formed using the validation results. [Figure 24] FIG. 24 is a graph formed using the validation results. [Diagram 25] FIG. 25 is a graph formed using the validation results. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0022] Hereinafter, an embodiment of the present disclosure will be described with reference to the drawings. In the drawings, the same elements are designated by the same reference numerals, and redundant description will not be repeated.
[0023] (First embodiment) An outline of the estimation support device according to the first embodiment will be described with reference to FIGS. 1 to 3. FIG. 1 illustrates an estimation support device 1 and a user U using the estimation support device 1. In this example, the estimation support device 1 is a tablet terminal, and the user U operates the estimation support device 1 using a display 1a (touch panel display) of the estimation support device 1. The estimation support device 1 is not limited to a tablet terminal, and a smartphone, a PC, or the like may be used. The user U is a person related to a patient (not shown), such as a medical professional. The user U may be a specialist who has been consulted by a medical professional about palliative care, or a medical professional who has support needs related to palliative care. When information about a cancer patient (patient information) is input, the estimation support device 1 estimates the symptoms of the cancer patient and outputs the estimation result. Such an estimation support device 1 can be used to estimate the symptoms of a cancer patient.
[0024] FIG. 2 shows an example of an input screen of the estimation support device. In this example, the display 1a of the estimation support device 1 displays "medical setting," "age," "gender," "type of cancer," "treatment stage," "ECOG-PS," "client," "loss of appetite," "nausea and vomiting," "abdominal distension / ascites," "constipation," "edema," and "insomnia" as patient information. "Medical setting" is information indicating the setting in which the cancer patient is receiving medical treatment, and check boxes are provided so that either hospitalization or outpatient treatment can be selected and input. "Age" is information indicating the age of the cancer patient, and a corresponding numerical value can be directly input. "Gender" is information indicating the gender of the cancer patient, and check boxes are provided so that either male or female can be selected and input. "Type of cancer" is information indicating the type of cancer the cancer patient has, and a pull-down menu is provided so that the corresponding type of cancer (for example, lung cancer, etc.) can be selected and input. "Treatment stage" is information indicating the treatment stage (time) in which the cancer patient is in treatment, and check boxes are provided so that either no anti-cancer treatment is planned, treatment is planned (during detailed examination), or anti-cancer treatment is being performed is selected and input. "ECOG-PS" is information indicating the overall condition of a cancer patient (Eastern Cooperative Oncology Group Performance Status), and the corresponding numerical value is entered directly. "Requester" is information indicating the person who requested palliative care for a cancer patient, and check boxes are provided so that either a doctor or a nurse can be selected and entered. "Loss of appetite" is information indicating whether the cancer patient has lost their appetite, and check boxes are provided so that either yes or no can be selected and entered. The same applies to "Nausea and vomiting," "Abdominal distension / ascites," "Constipation," "Edema," and "Insomnia."
[0025] FIG. 3 shows an example of the output screen of the estimation support device. In this example, the display 1a of the estimation support device 1 displays the probability of "pain," "dyspnea, cough, phlegm," "fatigue," "sleepiness," "anxiety, depression, emotional distress," "delirium," "spiritual pain," and "lack of informed consent" as percentages as estimation results. For example, the value corresponding to "pain" indicates the estimation result of the probability that a cancer patient has pain (pain sensitivity) or the probability that he or she does not have pain (pain specificity). Both sensitivity and specificity may be displayed. The same applies to other symptoms. Incidentally, "spiritual pain" refers to the suffering related to death that a cancer patient may face, and refers to the suffering of meaninglessness, worthlessness, emptiness, etc. of life. "Lack of informed consent" refers to insufficient understanding and selection of diagnosis and treatment, anxiety and worries about side effects of anti-cancer treatment, and circumstances that make it difficult to communicate with medical professionals. Furthermore, general-purpose assessment scales for "spiritual pain" and "lack of informed consent" have not yet been established in countries such as Japan. "Spiritual pain" and "lack of informed consent" are often first recognized during examinations by members of the palliative care team (specialists, etc.) in clinical practice. For this reason, it is not uncommon for "spiritual pain" and "lack of informed consent" to be discovered late.
[0026] The user U can estimate (evaluate) the symptoms of a cancer patient by using the above-described inference support device 1. For example, the user U can improve the accuracy of estimating the symptoms of a cancer patient by considering not only the result of his / her own judgment but also the inference result of the inference support device 1.
[0027] 4(a) is a diagram showing an example of functional blocks of the estimation support device 1. The estimation support device 1 includes, as its functional blocks, an input unit 10, an estimation unit 20, a trained model 30, and an output unit 40.
[0028] The input unit 10 is a part (input means) into which patient information is input. The input unit 10 accepts, for example, the patient information described above with reference to FIG. 2. More specifically, the patient information includes attribute information of the cancer patient. The attribute information is unique information that the cancer patient possesses, and is information on the medical treatment setting, age, sex, cancer type, treatment stage, ECOG-PS, or requester, or any combination of these. The patient information also includes initial evaluation information of the cancer patient. The initial evaluation information is information on symptoms that are easy to evaluate by observation by a specialist (medical professional, etc.). Symptoms that are easy to evaluate include symptoms caused by plant system dysfunction (disorders of nutrition, metabolic endocrine system, blood circulation system, autonomic nervous system, etc.) in the cancer patient. Examples of symptoms caused by plant system dysfunction are loss of appetite, nausea and vomiting, abdominal distension, constipation, edema, and insomnia. The initial evaluation information may be information on loss of appetite, nausea and vomiting, abdominal distension, constipation, edema, or insomnia, or any combination of these. In this specification, "vegetative function" refers to a physiological function that is directly involved in maintaining life. Vegetative functions include, for example, nutrition, metabolism, endocrinology, blood circulation, and autonomic nerves. Physiological functions that are directly involved in maintaining life are often regulated unconsciously. The state of vegetative function is easy to evaluate by observation by others, and in a pathological state, it can be recognized by others as an objective symptom.
[0029] The estimation unit 20 is a part (estimation means) that estimates the symptoms of a cancer patient using the patient information input to the input unit 10 and the trained model 30. The estimated symptoms of a cancer patient are symptoms that are difficult to evaluate even by observation by a specialist (medical professional, etc.). Examples of symptoms that are difficult to evaluate include symptoms caused by animal system dysfunction (motor system (skeletal muscles and nerves), respiratory and circulatory system, and higher brain function) of a cancer patient. Examples of symptoms caused by animal system dysfunction are pain, dyspnea, cough, phlegm, fatigue, drowsiness, anxiety, depression, emotional distress, delirium, lack of informed consent, and spiritual pain. The symptoms of a cancer patient may be at least one of these or any combination. The estimation unit 20 estimates the symptoms of a cancer patient by inputting patient information into the trained model 30 and obtaining symptom estimation information of the cancer patient from the trained model 30. The symptom estimation information of a cancer patient is an estimation result of sensitivity and specificity of "pain" and the like, as described above with reference to FIG. 3. In this specification, "animal function" refers to physiological functions related to perception, sensation, nervous mechanisms, and movement. Animal function is, for example, a function represented by the sensory organs, skeletal muscles, and nervous system. In many cases, animal function is regulated unconsciously. The state of animal function tends to be difficult to evaluate by observation from others, and in a pathological condition, it is recognized by the person (patient) as a subjective symptom.
[0030] The trained model 30 is a trained model generated using teacher data so as to output symptom estimation information of a cancer patient when patient information is input. The trained model may include an input node (input layer) that accepts input data (patient information in this example), an output node (output layer) that outputs output data (symptom estimation information in this example), and an intermediate node (intermediate layer) provided between the input node and the output node. As such a trained model, a decision tree, a neural network, or various other known models may be used. The neural network includes an explanation of one generated by deep learning.
[0031] FIG. 5 is a diagram showing an example of the configuration of a decision tree. The decision tree 30a illustrated in FIG. 5 is configured to include a plurality of nodes N. For convenience, each node is given a distinguishable code such as "N1". The node N1 is a root node (input node) to which patient information is input, and is a first branch. The nodes N2-1 and N2-2 are nodes branched from the node N1, and are a second branch. The nodes N3-1 and N3-2 are nodes branched from the node N2-1, and are a third branch. The nodes N3-3 and N3-4 are nodes branched from the node N2-2, and are a third branch. In this way, branching is repeated from upstream to downstream, leading to the n-th branch, which is the nodes Nn-p and Nn-q. The nodes Nn-p and Nn-q are not branched, and are leaf nodes (output nodes) that output symptom estimation information of cancer patients. Nodes N2-1, N2-2, N3-1, N3-2, N3-3 and N3-4, which are provided between node N1 and nodes Nn-p and Nn-q, are intermediate nodes that branch between node N1 and nodes Nn-p and Nn-q.
[0032] The decision tree 30a is generated by machine learning using teacher data. A decision tree having a configuration according to the teacher data is automatically generated by machine learning. The configuration of the decision tree is the number of nodes included in the decision tree, the connection relationships between each node, the branching conditions at the input node and the intermediate node, the disease condition estimation information output by the output node, and the like. The branching conditions at each node are determined using any of the items of the patient information. An example of the item is "age" and the like shown in FIG. 2 described above and FIG. 6 described later. An example of the branching condition is a condition such as whether or not the age is equal to or less than a predetermined age value. An example of the disease condition estimation information to be output is the estimation result of the sensitivity and specificity of "pain" and the like shown in FIG. 3 described above and FIG. 7 described later.
[0033] Returning to (a) of FIG. 4, the output unit 40 is a part (output means) that outputs the estimation result of the estimation unit 20. The output unit 40 presents, for example, the estimation result described above with reference to FIG. 3. The manner of output by the output unit 40 is not particularly limited. For example, items may be rearranged and displayed (in ascending order, descending order, etc.) according to the value of the estimation result. Only a part of the estimation result may be displayed. The estimation result may be displayed using various graphs. The estimation result may be output as sound (audio, etc.).
[0034] 4A, the estimation unit 20 and the trained model 30 are provided inside the estimation support device 1. However, the estimation unit 20 and the trained model 30 may be provided outside the estimation support device 1.
[0035] The estimation support device 1A illustrated in (b) of FIG. 4 differs from the estimation support device 1 in that the trained model 30 is provided in an external server 2A. The estimation support device 1 also includes a communication unit 50. The server 2A includes the trained model 30 and a communication unit 60. In the estimation support device 1A, the estimation unit 20 uses the trained model 30 in the server 2A via the communication unit 50 of the estimation support device 1A and the communication unit 60 of the server 2A. A system (estimation support system) thus realized by the cooperation of the estimation support device 1A and the server 2A is also one aspect of the estimation support device.
[0036] The estimation support device 1B illustrated in FIG. 4(c) is different from the estimation support device 1A in that an estimation unit 20 is provided in the server 2B. In the estimation support device 1B, patient information input to the input unit 10 is sent to the estimation unit 20 in the server 2B via the communication unit 50 of the estimation support device 1B and the communication unit 60 of the server 2B. The estimation unit 20 estimates the symptoms of the cancer patient using the trained model 30. The estimation result of the estimation unit 20 is sent to the estimation support device 1B via the communication unit 60 of the server 2B and the communication unit 50 of the estimation support device 1B, and is output by the output unit 40. The "system (estimation support system)" thus realized by the cooperation of the estimation support device 1B and the server 2B is also one aspect of the estimation support device.
[0037] The hardware configuration of the estimation support device 1 will be described. The estimation support device 1 may be configured by a computer device including a processor (CPU, etc.), a main storage device (memory such as ROM, RAM, etc.), and an auxiliary storage device. The auxiliary storage device may be a computer readable medium, and may be a non-transitory recording medium. Examples of the auxiliary storage device include, but are not limited to, various known storage devices (for example, CD-ROM, DVD, hard disk drive, USB memory, etc.). The auxiliary storage device stores a program for causing the computer device to function as the estimation support device. This program is read into the main storage device, and the processor operates the computer device according to the program. The trained model 30 may be stored in the auxiliary storage device (recorded in a recording medium) as a program or something equivalent to a program. The trained model 30 causes the computer device to function so that, when patient information is input to the input node, the output node outputs symptom estimation information. The computer device may also include an input interface for receiving an operation by the user U, and an output interface for presenting the estimation result to the user U. 1, the display 1a of the estimation support device 1 is one embodiment of the input interface and the output interface. Note that the servers 2A and 2B ((b) and (c) in FIG. 4) may also have the same hardware configuration as the estimation support device 1.
[0038] An example of generating the trained model 30 will be further described. Teacher data, more specifically, a data set (training data) composed of a plurality of teacher data, is used to generate the trained model 30. In this embodiment, the teacher data is data in which patient information corresponding to input data is associated with symptom estimation information corresponding to output data.
[0039] The method of collecting the data set (patient information and symptom estimation information) is not particularly limited. The attribute information (medical treatment setting, etc.) of the patient information can be easily obtained, so the method of collecting it will not be described here. The initial evaluation information (loss of appetite, etc.) of the patient information is obtained by examination by a specialist, etc. The symptom estimation information is obtained by actually prescribing something and determining whether or not the expected effect was obtained. For example, if it is confirmed that the symptoms of a cancer patient are alleviated by prescribing pain, it is presumed that the symptom of the cancer patient was pain. In general, in hospital facilities that have palliative care wards, palliative care team activity records are created in a specified format, so that the records may be used to collect the data set. An example of the format is the Standard Format for Reporting Palliative Care Team Activities 1.0 (SF-PCTA 1.0), but is not limited to this.
[0040] Hereinafter, an example of generating the trained model 30 will be described. In this example, a data set based on 213 cancer cases, that is, 213 pieces of training data, was prepared, and machine learning was performed using the data set to generate the trained model 30.
[0041] 6 shows an example of input data in the training data. In this example, the patient information corresponding to the input data is represented by 32 items. For convenience, each item is associated with a code (X1 to X32).
[0042] Items X1 to X26 are attribute information of the cancer patient. Specifically, item X1 "Hospitalization" is information indicating whether the cancer patient is hospitalized or not. An example of data corresponding to this information is "1" if the patient is hospitalized, and "0" if not. Item X2 "Outpatient" is information indicating whether the cancer patient is an outpatient or not. An example of data is "1" if the patient is an outpatient, and "0" if not. Item X3 is information indicating the age of the cancer patient. An example of data is the value of age (continuous variable). Instead of a continuous variable, binary data may be used. For example, if the age is higher than a predetermined age, it may be "1", and if not, it may be "0". An example of the predetermined age is the median, average, etc. of all teacher data. Item X4 "Male" is information indicating whether the cancer patient is male or not. An example of data is "1" if the patient is male, and "0" if not. The same applies to item X5 "Female". Item X6 "Pancreatic cancer" is information indicating whether the cancer patient has pancreatic cancer or not. An example of the data is "1" if it is pancreatic cancer, and "0" if it is not. The same is true for item X7 "cancer of unknown primary origin", item X8 "lung cancer", item X9 "breast cancer", item X10 "head and neck cancer", item X11 "biliary system cancer", item X12 "colorectal and rectal cancer", item X13 "prostate cancer", item X15 "kidney and bladder cancer", item X16 "esophageal cancer", item X17 "uterine and ovarian cancer", item X18 "liver cancer", item X19 "stomach cancer", and item X20 "lymphatic and blood cancer". Item X14 "cancer under examination" is information indicating whether the type of cancer has not been clarified and is in the middle of examination. An example of the data is "1" if it is under examination, and "0" if it is not. Item X21 "plan not to perform anti-cancer treatment" is information indicating whether the cancer patient plans not to perform anti-cancer treatment. An example of the data is "1" if anti-cancer treatment is not planned, and "0" if not. Item X22 "Planned treatment" is information indicating whether or not a cancer patient plans to undergo treatment. An example of the data is "1" if treatment is planned, and "0" if not. Item X23 "Currently undergoing anti-cancer treatment" is information indicating whether or not a cancer patient is undergoing anti-cancer treatment. An example of the data is "1" if anti-cancer treatment is planned, and "0" if not.Item X24 "ECOG-PS" is information indicating the ECO-PS value of a cancer patient. An example of the data is the ECOG-PS value (continuous variable). Instead of a continuous variable, it may be binary data. For example, if it is higher than a specified value, it may be set to "1", and if not, it may be set to "0". Examples of specified values include the median and average of all teaching data. Item X25 "Requester is a physician" is information indicating whether the person who requested palliative care for a cancer patient is a physician. An example of the data is "1" if the person is a physician, and "0" if not. The same applies to item X26 "Requester is a nurse".
[0043] Items X27 to X32 are initial evaluation information for cancer patients. Specifically, item X27 "Loss of appetite / difficulty in oral intake" is information indicating whether or not the cancer patient has symptoms of loss of appetite / difficulty in oral intake. An example of data is "1" if loss of appetite / difficulty in oral intake is observed, and "0" if not. The same is true for item X28 "Nausea / vomiting", item X29 "Abdominal distension / ascites", item X30 "Constipation", item X31 "Edema / lymphedema" and "Insomnia".
[0044] 7 shows an example of output data in the training data. In this example, the symptom estimation information corresponding to the output data is represented by eight items. For convenience, each item is associated with a code (Y1 to Y8).
[0045] Item Y1 "Pain" is information indicating that pain is presumed to exist in cancer patients, or that it is presumed not to exist. Examples of data are numbers indicating the degree of possibility of existence / nonexistence, ranging from 0 to 1, or this number converted into a percentage ranging from 0 to 100. The same is true for other items Y2 "Difficulty breathing, cough, phlegm," item Y3 "Fatigue," item Y4 "Sleepiness," item Y5 "Anxiety, depression, emotional distress," item Y6 "Delirium," item Y7 "Spiritual pain," and item Y8 "Lack of informed consent."
[0046] Various known methods may be used to generate the model. In this embodiment, the model was generated by machine learning using RapidMinder, which is provided as a machine learning platform. A decision tree was used as the trained model.
[0047] The cross validation method was used to validate the trained model. The cross validation method itself is well known, so it will be briefly explained here. First, 1 / 10 of the data in the dataset is left as test data. This test data is not used in machine learning. The remaining 9 / 10 of the data is divided into 10 groups, and one arbitrary group is used as validation data, and the remaining 9 groups are used as training data. A trained model is generated using the training data, and the validation data is used to check the operation, thereby obtaining a validation result for the trained model. Similarly, the test data is used to check the operation, thereby obtaining another validation result for the trained model. Next, a group other than the group used for the previous validation data is used as new validation data, and the remaining 9 groups are used as training data. The training data is used to generate another trained model, and the validation data is used to check the operation, thereby obtaining a validation result for the trained model. Similarly, the test data is used to check the operation, thereby obtaining another validation result for the trained model. By repeating this process, validation results for each of the 10 trained models are obtained. Then, the average of these verification results is calculated. Note that, although the number of divisions is set to 10 in this embodiment, the number of divisions may be any integer (for example, the number of divisions may be 5, etc.).
[0048] A confusion matrix and a receiver operating characteristic curve (ROC curve) were used as indicators of the verification results. Confusion matrices and ROC curves are publicly known, so they will be briefly explained here without detailed explanation. That is, a confusion matrix describes the results of machine learning class classification, and if a class that is actually true (positive) is estimated to be that class (positive), it is called TP (True Positive). If a class that is not actually true (negative) is estimated to be not that class (negative), it is called TN (True Negative). If a class that actually is (positive) is presumed to be the other (negative) class, it is called FN (False Negative). If a class that actually is not (negative) is presumed to be the other (positive) class, it is called FP (False Positive). Using these, accuracy = (TP+TN) / (TP+FP+TN+FN), sensitivity = TP / (TP+FN), specificity = TN / (FP+TN) and precision = TP / (TP+FP). The ROC curve is a curve with sensitivity (TP / TP+FN) on the vertical axis and 1-specificity (FP / (FP+TN)) on the horizontal axis while changing the parameters of the learning model, and the area under the ROC curve, AUC, indicates the performance (ability) of class classification using the curve.
[0049] The following results were obtained as the average of the validation results for each of the 10 trained models. Accuracy, sensitivity, specificity, and precision are in percent (%). The numbers after ± indicate standard deviation. AUC is dimensionless (maximum 1.0).
[0050] Regarding the symptom "pain" in cancer patients, the results of the verification using the verification data are as follows. Accuracy = 63.3± 2.95. Sensitivity = 77.7 ± 3.40. Specificity = 29.9± 5.46. Precision rate = 72.0± 2.28. AUC = 0.541 ± 0.0250. The verification results using the test data are as follows. Accuracy = 68.5± 14.2. Sensitivity = 84.9 ± 7.05. Specificity = 24.1± 19.6. Precision rate = 72.9± 16.4. AUC = 0.582 ± 0.151. As described above, relatively high values were obtained for accuracy, sensitivity, and precision. At least in this respect, the method is considered to be useful for predicting the symptom "pain" in cancer patients.
[0051] Regarding the symptom "dyspnea" in cancer patients, the verification results using the verification data are as follows. Accuracy = 62.6± 3.46. Sensitivity = 18.1 ± 6.91. Specificity = 80.0± 6.00. Precision rate = 26.4± 6.36. AUC = 0.486 ± 0.0337. The verification results using the test data are as follows. Accuracy = 59.5± 11.4. Sensitivity = 27.9 ± 28.1. Specificity = 70.6± 13.8. Precision rate = 19.6± 20.3. AUC = 0.482 ± 0.175. As described above, relatively high values were obtained for accuracy and specificity. At least in this respect, it is considered to be useful for predicting the symptom "dyspnea" in cancer patients. High specificity can be useful, for example, for screening and review.
[0052] Regarding the symptom "fatigue" in cancer patients, the results of the verification using the verification data are as follows. Accuracy = 70.5± 2.75. Sensitivity = 21.0 ± 7.23. Specificity = 85.7± 3.42. Precision rate = 30.8± 7.81. AUC = 0.552 ± 0.0484. The verification results using the test data are as follows. Accuracy = 73.5± 6.26. Sensitivity = 34.5 ± 30.8. Specificity = 88.0± 12.0. Precision rate = 48.0± 38.9. AUC = 0.706 ± 0.146. As shown above, relatively high values were obtained for AUC, sensitivity, and specificity. At least in this respect, the method is considered to be useful for predicting the symptom of "fatigue" in cancer patients.
[0053] Regarding the symptom "sleepiness" in cancer patients, the results of the verification using the verification data are as follows. Accuracy = 88.0± 1.71. Sensitivity = 12.9 ± 5.74. Specificity = 95.7± 1.75. Precision rate = 25.6± 12.6. AUC = 0.531 ± 0.123. The validation results using the test data are as follows. Note that there were two sets of split data for which the ACU was not determined (AUC=NA), so the average validation results for the remaining eight sets are shown here. Accuracy = 85.0± 5.98. Sensitivity = 4.17 ± 11.8. Specificity = 95.8± 3.95. Precision rate = 12.5± 35.4. AUC = 0.450 ± 0.207. As shown above, relatively high values were obtained for sensitivity and specificity. At least in this respect, this method is considered to be useful for predicting the symptom of "sleepiness" in cancer patients.
[0054] Regarding the symptom "anxiety" in cancer patients, the results of the verification using the verification data are as follows. Accuracy = 58.6± 2.52. Sensitivity = 61.1 ± 5.05. Specificity = 55.9± 5.11. Precision rate = 59.6± 2.96. AUC = 0.549 ± 0.0402. The verification results using the test data are as follows. Accuracy = 56.0± 8.10. Sensitivity = 67.3 ± 12.5. Specificity = 41.8± 21.6. Precision rate = 56.8± 15.8. AUC = 0.533 ± 0.162. As can be seen from the above, a relatively high sensitivity was obtained. At least in this respect, it is believed to be useful for estimating the symptom "anxiety" in cancer patients.
[0055] Regarding the symptom "delirium" in cancer patients, the results of verification using verification data are as follows. Accuracy = 70.8± 2.35. Sensitivity = 31.2 ± 5.04. Specificity = 83.2± 2.42. Precision rate = 36.9± 4.63. AUC = 0.612 ± 0.0301. The verification results using the test data are as follows. Accuracy = 71.0± 10.2. Sensitivity = 29.8 ± 20.9. Specificity = 85.7± 11.2. Precision rate = 40.2± 27.4. AUC = 0.654 ± 0.147. As can be seen from the above, relatively high values were obtained for accuracy and specificity. At least in this respect, the method is considered to be useful for predicting the symptom of "delirium" in cancer patients.
[0056] Regarding the symptom of "lack of informed consent" in cancer patients, the results of verification using verification data are as follows. Accuracy = 60.1± 2.83. Sensitivity = 41.8 ± 5.73. Specificity = 70.8± 4.58. Precision rate = 46.3± 3.47. AUC = 0.556 ± 0.0383. The verification results using the test data are as follows. Accuracy = 55.5± 14.6. Sensitivity = 28.0 ± 16.2. Specificity = 71.2± 9.98. Precision rate = 36.6± 17.8. AUC = 0.460 ± 0.134. As shown above, a relatively high value was obtained in terms of specificity. At least in this respect, it is considered to be useful in predicting the symptom of "lack of informed consent" in cancer patients.
[0057] Regarding the symptom "spiritual pain" in cancer patients, the results of verification using verification data are as follows. Accuracy = 72.3± 3.79. Sensitivity = 17.9 ± 5.68. Specificity = 86.4± 3.84. Precision rate = 26.0± 7.44. AUC = 0.548 ± 0.0412. The verification results using the test data are as follows. Accuracy = 74.0± 10.7. Sensitivity = 21.7 ± 35.2. Specificity = 90.8± 8.40. Precision rate = 25.0± 35.4. AUC = 0.558 ± 0.235. As described above, relatively high values were obtained for accuracy and specificity. At least in this respect, it is considered to be useful for predicting the symptom "spiritual pain" in cancer patients.
[0058] From the above, it can be seen that the generated trained model is useful for estimating symptoms. It is expected that higher estimation accuracy can be obtained by increasing the number of training data or appropriately selecting the type of training model. The machine learning method may be changed depending on the number of training data. For example, when the number of training data is about 11 to 100, the Leave One Out method may be used. When the number of training data is about 101 to 1000, the Leave One Out method may be used in addition to the above-mentioned Cross Validation method. When the number of training data is about 1001 to 10000, the Cross Validation method or the Hold Out method may be used. When the number of training data is about 10001 to 100000, the Hold Out method may be used. When the number of training data is further increased to 100001 or more, a more appropriate method may be considered by a specialized institution or the like. Regarding the type of training model, the verification result of estimating the symptom "dyspnea" when deep learning is applied to a neural network using the above-mentioned dataset is shown below. This result was obtained using test data. Accuracy = 73.6± 13.6. Sensitivity = 0.00 ± 0.00. Specificity = 100 ± 0.00. Precision rate = 0.00± 0.00. The neural network described above had the advantage of being more accurate and specific than the decision tree. As shown above, different learning models (decision tree or neural network) can provide different characteristics, which shows that various learning models can be used and that the estimation accuracy can be improved by appropriately selecting the type of learning model.
[0059] When a decision tree is adopted as a learning model, there are further advantages as follows. That is, a person can confirm what kind of structure the decision tree generated by machine learning has. By examining the structure, more specifically, the branching process, it is expected that knowledge leading to new pathophysiology and diagnostic techniques can be obtained. For example, items appearing in upstream branches of a decision tree are considered to be particularly important features in symptom estimation. From this perspective, the results of examining the structure of the decision tree created in the above embodiment will be described with reference to Figs. 8 to 15.
[0060] FIG. 8 is a graph showing items that frequently appear in upstream branches of a decision tree when estimating the symptom "pain". The horizontal axis of the graph shows the branch number, and the vertical axis shows the frequency of appearance (%) of the items in each branch. The items are items X1 to X32 shown in FIG. 6 described above. For example, the frequency of appearance of item X3 "age" in branch 1 is 80%, which means that of the 10 generated decision trees, 8 decision trees have a node N in branch 1 that judges item X3 "age". From the graph shown in FIG. 8, it can be seen that item X3 "age", item X4 "male" (gender), item X11 "biliary tract cancer", item X24 "ECOG-PS", and item X27 "loss of appetite / difficulty in oral intake" are important features when estimating the symptom "pain". In other words, it can be seen that there may be some causal relationship between these items and the symptom "pain".
[0061] Figure 9 is a graph showing the items that frequently appear in the upstream branches of the decision tree when inferring the symptom "dyspnea." From the graph shown in Figure 9, it can be seen that item X3 "age," item X7 "cancer of unknown primary source," item X9 "breast cancer," item X13 "prostate cancer," item X24 "ECOG-PS," item X29 "abdominal distension / ascites," and item X31 "edema / lymphedema" are important features when inferring the symptom "dyspnea."
[0062] Figure 10 is a graph showing items that frequently appear in the upstream branches of the decision tree when estimating the symptom "fatigue." From the graph shown in Figure 10, it can be seen that item X3 "age," item X7 "cancer of unknown primary source," item X9 "breast cancer," item X13 "prostate cancer," item X31 "edema / lymphedema," and item X32 "insomnia" are important features when estimating the symptom "fatigue."
[0063] Figure 11 is a graph showing items that frequently appear in the upstream branches of the decision tree when estimating the symptom "drowsiness." From the graph shown in Figure 11, it can be seen that item X3 "age," item X30 "constipation," and item X32 "insomnia" are important features when estimating the symptom "drowsiness."
[0064] Figure 12 is a graph showing the items that frequently appear in the upstream branches of the decision tree when inferring the symptom "anxiety." From the graph shown in Figure 12, we can see that item X3 "age," item X20 "lymphatic / blood cancer," and item X31 "edema / lymphedema" are important features when inferring the symptom "anxiety."
[0065] Figure 13 is a graph showing the items that frequently appear in the upstream branches of the decision tree when inferring the symptom "delirium." From the graph shown in Figure 13, we can see that item X3 "age," item X7 "cancer of unknown primary," item X24 "ECOG-PS," and item X31 "edema / lymphedema" are important features when inferring the symptom "delirium."
[0066] Figure 14 is a graph showing the items that frequently appear in the upstream branches of the decision tree when inferring the symptom "lack of informed consent." From the graph shown in Figure 14, we can see that item X1 "hospitalization," item X3 "age," item X7 "cancer of unknown primary source," item X8 "lung cancer," item X9 "breast cancer," item X14 "cancer under examination," item X17 "uterine / ovarian cancer," item X24 "ECOG-PS," item X25 "requester is a doctor," item X29 "abdominal distension / ascites," item X30 "constipation," and item X31 "edema / lymphedema" are important features.
[0067] Figure 15 is a graph showing the items that frequently appear in the upper branches of the decision tree when predicting the symptom "spiritual pain". From the graph shown in Figure 15, we can see that item X1 "hospitalization", item X3 "age", item X6 "pancreatic cancer", item X8 "lung cancer", item X11 "biliary system cancer", item X13 "prostate cancer", item X17 "uterine / ovarian cancer", item X18 "liver cancer", item X19 "stomach cancer", item X20 "lymphatic / blood cancer", item X21 "no plans to perform anti-cancer treatment in the future", item X24 "ECOG-PS", item X29 "abdominal distension / ascites" and item X32 "insomnia" are important features.
[0068] As described above, by examining the structure of the decision tree generated by machine learning, we can obtain clues about the features that are important for estimating each symptom.
[0069] The estimation support device 1 described above (including the estimation support device 1A and the estimation support device 1B) is specified, for example, as follows.
[0070] The estimation support device 1 includes an input unit 10, an estimation unit 20, and an output unit 40. Patient information including attribute information of a cancer patient and initial evaluation information regarding plant system functional disorders is input to the input unit 10. The estimation unit 20 estimates symptoms of a cancer patient using the patient information input to the input unit 10 and a trained model 30 generated in advance. The trained model 30 is a trained model generated using teacher data so as to output symptom estimation information of a cancer patient when patient information is input.
[0071] According to the estimation support device 1, symptoms of a cancer patient are estimated and the estimation result is output using patient information and a trained model 30 generated in advance. Therefore, the estimation support device 1 can be used to estimate symptoms of a cancer patient.
[0072] The attribute information may include at least one of information regarding the medical setting, age, sex, type of cancer, treatment stage, ECOG-PS, and requester. The symptoms of the cancer patient can be estimated by taking these items into consideration.
[0073] The attribute information may be at least one of the above items or any combination thereof. A preferred example of the combination is age, cancer type, and ECOG-PS.
[0074] The initial evaluation information may include information regarding at least one of loss of appetite, nausea and vomiting, abdominal distension, constipation, edema, and insomnia. By taking these items into consideration, the symptoms of the cancer patient can be estimated.
[0075] The initial evaluation information may be at least one or any combination of the above items. A preferred example of the combination is abdominal distension, constipation, edema, and insomnia.
[0076] Symptoms of cancer patients may include at least one of the following symptoms caused by animal system dysfunction: pain, dyspnea, cough, phlegm, fatigue, drowsiness, anxiety, depression, emotional discomfort, delirium, lack of informed consent, and spiritual pain. This allows us to estimate symptoms that are difficult to evaluate objectively (subjective symptoms).
[0077] In particular, in the terminal stage of cancer patients, it is often difficult to communicate with them and to estimate their symptoms, and in such cases, the benefits of using the estimation support device 1 become more apparent. For example, the estimation support device 1 can be used to estimate (point out) symptoms that are difficult to evaluate as described above, leading to a diagnosis by a specialist, thereby making up for the shortage of specialists. As of April 1, 2019, there were 244 specialists certified by the Japanese Society for Palliative Medicine and 518 doctors certified with a simpler certification than specialists, but in 2017, there were more than 370,000 deaths from cancer nationwide. Palliative care is required to be provided not only in the terminal stage but also from the time of cancer diagnosis, so there is a reality that palliative care cannot be provided in local medical care due to a shortage of specialists, including patients newly diagnosed with cancer.
[0078] The trained model 30 may be a decision tree 30a. In this manner, the decision tree 30a can be used as the trained model 30. As described above with reference to Figs. 8 to 15, by examining the structure (classification process) of the decision tree 30a, it is possible to know feature quantities (items of patient information) that are important in estimating symptoms.
[0079] The trained model 30 is identified, for example, as follows:
[0080] The trained model 30 includes an input node, an output node, and an intermediate node. The intermediate node is provided between the input node and the output node. The trained model 30 is generated using training data that associates patient information, including attribute information of a cancer patient and information on a disorder of plant system functions, with symptom estimation information of the cancer patient, and causes a computer to function so that when patient information is input to the input node, the output node outputs the symptom estimation information of the cancer patient.
[0081] According to the trained model 30, symptom estimation information of a cancer patient can be obtained from the patient information. Therefore, the trained model 30 can be used to estimate the symptoms of a cancer patient.
[0082] The trained model 30 may be a decision tree 30a. In this case, the input node is a root node (node N1). The output nodes are leaf nodes (nodes Nn-p, Nn-q, etc.). The intermediate nodes are nodes that branch the path between the root node and the leaf nodes (nodes N2-1, N2-2, N3-1, N3-2, N3-3, N3-4, etc.). In this way, a decision tree can be used as the trained model. Also, by examining the structure (classification process) of the decision tree, it is possible to know features (items of patient information) that are important in estimating symptoms.
[0083] Second embodiment In the following, an overview of a care support device according to a second embodiment will be described as one application example of the estimation support device 1 according to the first embodiment. In the following second embodiment, the description of the points that overlap with the first embodiment will be omitted. Therefore, in the following, the points that differ from the first embodiment will be mainly described.
[0084] The care support device according to the second embodiment is a device that proposes a care plan for a cancer patient whose symptoms have been estimated by the estimation support device 1, and includes the estimation support device 1. For this reason, the care support device according to the second embodiment is a tablet terminal or the like, similar to the estimation support device 1. Users of the care support device are medical professionals who provide medical treatment, nursing, etc. to the patient, and experts who receive consultations regarding palliative care from the medical professionals. The hardware configuration of the care support device is the same as that of the estimation support device 1.
[0085] Fig. 16 is a diagram illustrating an example of functional blocks of a care support device 100. As illustrated in Fig. 16, the care support device 100 includes, as its functional blocks, an input unit 10A, an estimation unit 20, a trained model 30, a generation unit 31, a second trained model 32, and an output unit 40A.
[0086] The input unit 10A is a part to which observation information on the estimation result estimated by the estimation unit 20 is input in addition to the patient information. The input unit 10A has the same configuration as the input unit 10 of the first embodiment. The observation information is information on the subjective symptoms of the cancer patient that can be determined by observation by a general medical professional who is not a palliative care specialist. For example, the general medical professional extracts the observation information through verbal and non-verbal communication with the patient with reference to the estimation result, and then inputs the extracted observation information to the care support device 100 via the input unit 10A. In this embodiment, the general medical professional inputs the observation information based on the observation result of the cancer patient by himself and the estimation result by the estimation support device 1, but is not limited to this. The information on the subjective symptoms of the cancer patient is, for example, the judgment result of the general medical professional on the presence or absence of each symptom (i.e., "pain", "dyspnea, cough, phlegm", "fatigue", "sleepiness", "anxiety, depression, emotional distress", "delirium", "spiritual pain", and "lack of informed consent"). A general medical professional may input the probability (sensitivity) of the above subjective symptoms as information on the symptoms of cancer patients. Note that the observation information may be the same as the estimation result.
[0087] The generating unit 31 is a part (a care plan generating part) that generates a care plan proposal for a cancer patient whose symptoms are estimated by the estimation support device 1. The generating unit 31 generates a care plan proposal using the patient information input from the input unit 10A, at least one of the estimation result of the estimation unit 20 and the observation information on the estimation result, and the second trained model 32. In this embodiment, when the observation information is different from the estimation result, the generating unit 31 generates a care plan proposal using the patient information, the observation information, and the second trained model 32. When the observation information is the same as the estimation result, the generating unit 31 generates a care plan proposal using the patient information, at least one of the estimation result and the observation information, and the second trained model 32. However, even when the observation information is different from the estimation result, the generating unit 31 may generate a care plan proposal using the patient information, the estimation result and the observation information, and the second trained model 32.
[0088] The proposed care plan includes, for example, at least one of the following major items: "comprehensive assessment of the patient," "care regarding the patient's physical pain," "mental care of the patient," "care regarding the patient's decision-making," "care regarding the selection and transition of the place of care," "care of the patient receiving home care," "care of the family," "ethical issues," "assessment and referral of the need for specialist intervention," "support of medical staff," "coordination within the palliative care team," "medical treatment and examination," and "pharmacotherapy." That is, the proposed care plan includes at least one or any combination of the above major items. The generation unit 31 obtains the proposed care plan from the second trained model 32 by inputting the patient information, the estimation result, and the observation information into the second trained model 32. The above major items include one or more minor items. For example, the major item: "care regarding the patient's physical pain" may include a minor item such as "use of medical narcotics." Therefore, the proposed care plan may correspond to at least one or any combination of the above minor items.
[0089] The second trained model 32 is a trained model generated using teacher data so as to output a care plan proposal when the patient information and at least one of the estimation result of the estimation unit 20 and the observation information on the estimation result are input. The second trained model 32 is generated using teacher data different from that of the trained model 30. The teacher data of the second trained model 32 is, for example, data in which patient information, symptom estimation information, and observation information on the symptom estimation information are associated with each other. The second trained model 32 includes input nodes, output nodes, and intermediate nodes different from those of the trained model 30. As the second trained model, a decision tree, a neural network, or various other known models may be used, as in the trained model 30. The generation of the decision tree in the second trained model 32 is the same as the generation of the decision tree 30a in the trained model. The branching condition at each node is determined using any one of the items of the patient information, the estimation result, and at least one of the observation information on the estimation result. Examples of items are "age" included in the patient information, and "pain" included in the estimation result and the observation information. An example of a branching condition is a condition such as the presence or absence of pain. Note that the second trained model 32 and the trained model 30 may have the same hardware configuration.
[0090] The output unit 40A is a part that outputs the care plan proposal in addition to the estimation result of the estimation unit 20. The output unit 40A has the same configuration as the output unit 40 of the first embodiment. The output unit 40A presents, for example, one or more major items included in the above-mentioned care plan proposal. At this time, minor items may be presented at the same time. The output unit 40A can present the care plan proposal in the same manner as the estimation result. When the output unit 40A presents the major items and / or minor items included in the care plan proposal in text and / or sound, the presented expression may be adjusted according to the sensitivity or specificity of the presented major item and minor item. For example, in the case of the first minor item, when its sensitivity is dominant over its specificity, the output unit 40A adjusts to a first expression (for example, "Please consider the first minor item first") that encourages the first minor item to be considered preferentially. On the other hand, when the specificity is more important than the sensitivity in the second subitem, the output unit 40A adjusts the second expression (for example, "please consider the second subitem as appropriate") to encourage the medical practitioner to consider the second subitem so as not to overlook / forget it. By adjusting the expression for each major item and minor item in this way, the medical practitioner can easily grasp the items that should be considered first, the items that should not be overlooked, the items that should not be forgotten, etc. This makes it easier for the medical practitioner to plan a more appropriate care plan. Since the sensitivity and the specificity are in a trade-off relationship, the closer the sensitivity is to 100% (i.e., the closer the specificity is to 0%), the more the sensitivity is important than the specificity, and the closer the specificity is to 100% (i.e., the closer the sensitivity is to 0%), the more the specificity is important than the sensitivity, but this is not limited to this. For example, when the sensitivity is 0.9 times or more, 0.8 times or more, or 0.75 times or more of the specificity in a specific subitem, the sensitivity may be set to be important than the specificity. In other words, in a given subitem, even if the sensitivity is slightly lower than the specificity, the sensitivity may be set to be more important than the specificity. Alternatively, in a given subitem, if the specificity is 0.9 times or more, 0.8 times or more, or 0.75 times or more of the sensitivity, the specificity may be set to be more important than the sensitivity. In other words, in a given subitem, even if the sensitivity is slightly lower than the specificity, the specificity may be set to be more important than the sensitivity.
[0091] The care support device according to the second embodiment may include the trained model 30 and the second trained model 32 in an external server, as in the above-described first embodiment. Alternatively, the care support device may include the estimation unit 20, the trained model 30, the second trained model 32, and the generation unit 31 in a server. In other words, a "system (care support system)" realized by the cooperation of the care support device and the server is also one aspect of the care support device.
[0092] The care support device 100 described above is specified as follows, for example. The care support device 100 includes an input unit 10A, an estimation unit 20, a generation unit 31, and an output unit 40A. In addition to patient information, observation information on the estimation result of the estimation unit 20 is input to the input unit 10A. The generation unit 31 generates a care plan proposal using the patient information, at least one of the estimation result and the observation information, and a second trained model 32 generated in advance. According to this care support device 100, a care plan proposal based on the symptom estimation information of a cancer patient is obtained by using at least one of the estimation result by the estimation support device 1 and the observation information on the estimation result. As a result, when a medical professional plans a care plan for the cancer patient, the medical professional can refer to the care plan proposal. Therefore, the care support device 100 can be used to plan an optimal care plan for the cancer patient.
[0093] As described above, the number of palliative care specialists certified by the Japanese Society for Palliative Medicine is extremely small compared to the number of cancer deaths nationwide. Therefore, by using the care support device 100 according to the second embodiment, for example, by general medical practitioners, it becomes easier to create appropriate palliative care plans even in areas where there is a shortage of specialists.
[0094] An example of a verification result of the second trained model 32 in the second embodiment will be described below. Here, a second dataset was used in which observation information for the output data (symptom estimation information) of the dataset used in the first embodiment was added to the dataset. Examples of input data in the teacher data of the second trained model 32 are patient information including attribute information and initial evaluation information of cancer patients, and observation information determined by general medical professionals, etc. In this example, the patient information is the same as in the first embodiment and is represented by 32 items, and the observation information is represented by 8 items. In addition, each item of the observation information is represented by binary data. For example, the observation information item "pain" is information indicating the presence or absence of pain, and is "1" when it is determined that there is pain, and is "0" when it is determined that there is no pain. Other items of the observation information are also set to "1" or "0" in the same manner.
[0095] FIG. 17 shows an example of output data in the second trained model 32. In this example, the proposed care plan for the output data is represented by 13 major items. Each major item includes one or more minor items. For convenience, a code (Z1 to Z100) is associated with each minor item, and minor items Z1 to Z4 are included in the major item "comprehensive assessment of the patient," minor items Z5 to Z15 are included in the major item "care related to the patient's physical pain," minor items Z16 to Z30 are included in the major item "mental care of the patient," major items Z31 to Z33 are included in "care related to the patient's decision-making," minor items Z34 to Z38 are included in the major item "care related to selection and transition of treatment location," and minor item Z39 is included in the major item "care of patients receiving home care." "A" is included in the main category, sub-items Z40 to Z47 are included in the main category "Family care", sub-item Z48 is included in the main category "Ethical issues", sub-items Z49 to Z53 are included in the main category "Assessment of the need for specialist intervention and referral", sub-items Z54 to Z64 are included in the main category "Support for medical staff", sub-items Z65 to Z66 are included in the main category "Coordination within the palliative care team", sub-items Z67 to Z71 are included in the main category "Medical procedures and testing", and sub-items Z72 to Z100 are included in the main category "Drug therapy".
[0096] The major item "Comprehensive Patient Assessment" captures the patient's distress from four perspectives: physical, psychological, social, and spiritual, and provides information indicating whether specialized treatment is necessary without overlooking the distress, or whether such treatment is unnecessary. An example of data is a number indicating the degree of possibility of necessity / unnecessity, which is a number between 0 and 1, or a percentage between 0 and 100. Such numbers are output for each of the minor items Z1 to Z4. The same applies to the other minor items Z5 to Z100.
[0097] In this example, a model was generated by machine learning using RapidMinder, as in the first embodiment. A decision tree was used for the second trained model. The second trained model was verified by the same verification method as in the first embodiment. As a result, a care plan including at least one of the verified subitems was obtained. Next, evaluation data results, which are the results of the specialist's evaluation of the care plan, were obtained. The evaluation data results also include Zn(+), which indicates the frequency of care plan implementation among the evaluation data results, which are the results of the specialist's evaluation, for each of the subitems Z1 to Z100. The units of accuracy, sensitivity, specificity, and Zn(+) are percentages (%). The lowest accuracy among the subitems Z1 to Z100 was 46%.
[0098] Each of Figs. 18 to 20 is a graph formed using the evaluation data results. In Fig. 18, the horizontal axis indicates Zn(+), and the vertical axis indicates accuracy. In Fig. 19, the horizontal axis indicates Zn(+), and the vertical axis indicates sensitivity. In Fig. 20, the horizontal axis indicates Zn(+), and the vertical axis indicates specificity. Figs. 18 to 20 show approximation curves AC1 to AC3, respectively. As shown by the approximation curve AC1 in Fig. 18, it can be seen that the smaller items in which Zn(+) is closer to 0% or 100% tend to show higher accuracy. As shown by the approximation curve AC2 in Fig. 19, it can be seen that the smaller items in which Zn(+) is closer to 100% tend to show higher sensitivity. As shown by the approximation curve AC3 in Fig. 20, it can be seen that the smaller items in which Zn(+) is closer to 100% tend to show lower specificity. From these graphs, it can be seen that the smaller items with Zn(+) closer to 0% or 100% tend to be easier to choose whether to include in the planned care plan, and the smaller items with Zn(+) closer to 50% tend to be more difficult to choose in the planned care plan. By obtaining such a tendency, it becomes easier to plan an appropriate palliative care plan even in areas where there is a shortage of specialists. In addition, by applying the above-mentioned first expression (an expression encouraging the consideration of small items as a priority) to small items with Zn(+) closer to 100%, it is easy to understand the matters that should be considered first when planning a palliative care plan. In addition, by applying the above-mentioned second expression (an expression encouraging the consideration of small items so as not to overlook them) to small items with Zn(+) closer to 0%, it is easy to understand the matters that should not be overlooked and the matters that must not be forgotten when planning a palliative care plan. In addition, in any of Figures 18 to 20, stratification by major items of the care plan was not observed.
[0099] Third embodiment In the following, an overview of an outcome prediction device according to a third embodiment will be described as another application example of the estimation support device 1 according to the first embodiment. In the following third embodiment, descriptions of points that overlap with the first and second embodiments will be omitted. Therefore, in the following, differences from the first and second embodiments will be mainly described.
[0100] The outcome prediction device according to the third embodiment is a device that objectively evaluates the validity of a planned care plan for a cancer patient whose symptoms are estimated by the estimation support device 1, and includes the estimation support device 1. For this reason, the outcome prediction device according to the third embodiment is a tablet terminal or the like, similar to the estimation support device 1. The user of the outcome prediction device is the same as the care support device 100 according to the second embodiment. The hardware configuration of the outcome prediction device is, for example, the same as the hardware configuration of the estimation support device 1. In the third embodiment, the planned care plan is planned based on a care plan proposal generated by the care support device 100. For this reason, the outcome prediction device includes the care support device 100.
[0101] Fig. 21 is a diagram showing an example of functional blocks of an outcome prediction device. As shown in Fig. 21, an outcome prediction device 200 includes, as its functional blocks, an input unit 10B, an estimation unit 20, a trained model 30, a generation unit 31, a second trained model 32, a prediction unit 33, a third trained model 34, and an output unit 40B.
[0102] The input unit 10B is a part where a planned care plan for the proposed care plan generated by the generating unit 31 is input in addition to the patient information and observation information. The input unit 10B has the same configuration as the input unit 10 of the first embodiment. The planned care plan is a treatment method, a recuperation method, etc. for a cancer patient that is planned by a palliative care specialist or a general medical practitioner. The planned care plan corresponds to, for example, one or a combination of a plurality of sub-items included in the proposed care plan. In this embodiment, the medical practitioner inputs the planned care plan based on his / her own observation results of the cancer patient and the proposed care plan generated by the generating unit 31, but is not limited to this. The planned care plan may be the same as the proposed care plan.
[0103] The prediction unit 33 is a part (outcome prediction part) that predicts the outcome of a cancer patient whose symptoms have been estimated by the estimation support device 1 in a predetermined period. The prediction unit 33 generates an outcome prediction using patient information input from the input unit 10A, at least one of the estimation result of the estimation unit 20 and observation information on the estimation result, at least one of the proposed care plan and the planned care plan, and the third trained model 34. In this embodiment, when the planned care plan is different from the planned care plan, the prediction unit 33 generates an outcome prediction using the patient information, at least one of the estimation result and observation information on the estimation result, the planned care plan, and the third trained model 34. Also, when the planned care plan is the same as the planned care plan, the prediction unit 33 generates an outcome prediction using the patient information, at least one of the estimation result and observation information on the estimation result, at least one of the proposed care plan and the planned care plan, and the third trained model 34. However, even if the planned care plan is different from the proposed care plan, the prediction unit 33 may generate an outcome prediction using patient information, at least one of the estimation result and observation information on the estimation result, at least one of the proposed care plan and the planned care plan, and the third trained model 34.
[0104] The outcome prediction includes, for example, at least one of the following items: "end of observation period," "death," "discharge to home," "discharge / transfer to other facility," "transfer to hospice / palliative care ward," and "problem resolved / disappeared." The specified period may be divided into multiple phases. In this case, the prediction unit 33 performs outcome prediction in each phase. Then, the predicted outcome in each phase is set. In this embodiment, the prediction unit 33 sets the specified period to one month and each phase to seven days (one week), but is not limited to this.
[0105] The third trained model 34 is a trained model generated using teacher data so as to output an outcome prediction for a predetermined period when patient information, at least one of the estimation result of the estimation unit 20 and the observation information on the estimation result, and at least one of the planned care plan and the proposed care plan are input. The third trained model 34 is generated using teacher data different from the trained model 30 and the second trained model 32. The teacher data of the third trained model 34 is, for example, data in which patient information, symptom estimation information, observation information on the symptom estimation information, the proposed care plan, and the planned care plan are associated with each other. The third trained model 34 includes input nodes, output nodes, and intermediate nodes different from the trained model 30 and the second trained model 32. The input data in the third trained model 34 is, for example, patient information, at least one of the estimation result and the observation information on the estimation result, and at least one of the planned care plan and the proposed care plan. The output data in the third trained model 34 is, for example, the sensitivity and specificity of each item. As the third trained model, a decision tree, a neural network, or various other known models may be used, as in the trained model 30. The generation of the decision tree in the third trained model 34 is the same as the generation of the decision tree 30a in the trained model. The branching condition at each node is determined using any of the items of the patient information, at least one of the estimation result and the observation information for the estimation result, and at least one of the planned care plan and the proposed care plan. Examples of the items are "age" included in the patient information, "pain" included in the estimation result and the observation information, and "use of medical narcotics" included in at least one of the planned care plan and the proposed care plan. The third trained model 34 and the trained model 30 may have the same hardware configuration.
[0106] The output unit 40B is a part that outputs the outcome prediction for a predetermined period in addition to the estimation result of the estimation unit 20 and the proposed care plan. The output unit 40B has, for example, the same configuration as the output unit 40 of the first embodiment. The output unit 40B presents, for example, the items included in the outcome prediction described above for each phase. At this time, all phases and the items corresponding to the phases may be presented at the same time.
[0107] The outcome prediction device according to the third embodiment may include the trained model 30, the second trained model 32, and the third trained model 34 in an external server, as in the above-described second embodiment. Alternatively, the outcome prediction device may include the estimation unit 20, the trained model 30, the generation unit 31, the second trained model 32, the prediction unit 33, and the third trained model 34 in a server. In other words, a "system (outcome prediction system)" realized by the cooperation of the outcome prediction device and the server is also one aspect of the outcome prediction device.
[0108] The outcome prediction device 200 described above is specified as follows, for example. The outcome prediction device 200 includes an input unit 10B, an estimation unit 20, a generation unit 31, a prediction unit 33, and an output unit 40B. In addition to patient information, a planned care plan for the proposed care plan is input to the input unit 10B. The prediction unit 33 generates an outcome prediction for a predetermined period using patient information, at least one of an estimation result and observation information, at least one of a planned care plan and a proposed care plan, and a third trained model 34 generated in advance. According to this outcome prediction device 200, an outcome prediction for a predetermined period of a cancer patient is obtained by using at least one of an estimation result and observation information by the estimation support device 1, at least one of a proposed care plan and a planned care plan, etc. As a result, for example, a medical professional can objectively confirm the validity of the planned care plan for the cancer patient. Therefore, the outcome prediction device 200 can be used to improve the accuracy of optimal care plan planning. In addition, for example, a cancer patient can objectively view examples of his / her condition transition over a given period by checking his / her outcome prediction. This makes it easier for the cancer patient to make decisions regarding his / her medical treatment wishes and choices. Therefore, the outcome prediction device 200 can be used for planning advance care planning (ACP).
[0109] As described above, the number of palliative care specialists certified by the Japanese Society for Palliative Medicine is extremely small compared to the number of cancer deaths nationwide. For this reason, it is difficult to objectively evaluate the appropriateness of a care plan proposal using the care support device 100 according to the second embodiment and a planned care plan based on the proposed care plan, particularly in areas where there is a shortage of specialists. In contrast, in the third embodiment, a medical professional can easily objectively evaluate the appropriateness of a planned care plan by checking the outcome prediction by the prediction unit 33. Therefore, by using the outcome prediction device 200, a medical professional can easily plan a more appropriate palliative care plan (and ACP). In particular, it is important to plan and document an ACP at a time when the cancer patient can express his or her wishes. This makes it possible to respond in accordance with the wishes and choices of the cancer patient even if the cancer patient falls into a critical situation where he or she cannot make medical decisions. Therefore, even if the patient falls into the above-mentioned critical situation, it is not necessary for the patient's family or the like to make decisions on his or her behalf.
[0110] An example of a verification result of the third trained model 34 in the third embodiment will be described below. Here, a second data set was used in which observation information and a planned care plan were added to the data set used in the first embodiment. Examples of input data in the teacher data of the third trained model 34 are patient information, observation information, and a planned care plan. In this example, the patient information is the same as in the first embodiment and is represented by 32 items, and the observation information is represented by 8 items. The planned care plan corresponds to the small items described in the second embodiment and is represented by 100 items. Each item of the planned care plan is represented by binary data. For example, the item "use of medical narcotics" of the planned care plan is information indicating whether or not medical narcotics are needed, and is "1" if it is determined that medical narcotics are needed, and is "0" if it is determined that medical narcotics are not needed. Other items of the planned care plan are also set to "1" or "0" in the same manner.
[0111] FIG. 22 shows an example of output data in the third trained model 34. In this example, outcome prediction regarding the output data is represented by six major items. Each major item includes a number of sub-items corresponding to the number of phases included in the specified period. In the following, each major item includes four sub-items by including four phases (first phase to fourth phase) in the specified period. The subitems W1, W7, W13, and W19 are included in the major item "End of observation period", the subitems W2, W8, W14, and W20 are included in the major item "Death", the subitems W3, W9, W15, and W21 are included in the major item "Discharge to home", the subitems W4, W10, W16, and W22 are included in the major item "Discharge / transfer to other facilities", the subitems W5, W11, W17, and W23 are included in the major item "Transfer to hospice / palliative care ward", and the subitems W6, W12, W18, and W24 are included in the major item "Problem resolution / disappearance". In addition, the subitems W1 to W6 are used in the first phase, the subitems W7 to 12 are used in the second phase, the subitems W13 to W18 are used in the third phase, and the subitems W19 to W24 are used in the fourth phase.
[0112] The major item "End of observation period" is information indicating that further medical treatment is no longer necessary due to reasons such as the patient being discharged from hospital. Examples of data are numbers indicating the degree of possibility of necessity / no need, ranging from 0 to 1, or 0 to 100 calculated as a percentage. Such numbers are output for each of the minor items W1, W7, W13, and W19. The same applies to minor items contained in other major items.
[0113] In this example, a model was generated by machine learning using RapidMinder, as in the first embodiment. A decision tree was used for the third trained model. The third trained model was verified by the same verification method as in the first embodiment. The average of the verification results by the third trained model is shown in Figs. 23 to 25. Each of Figs. 23 to 25 is a graph formed using the verification results. Fig. 23 shows the accuracy for the verified subitems, Fig. 24 shows the sensitivity for the verified subitems, and Fig. 25 shows the specificity for the verified subitems. In Figs. 23 to 25, the subitems W1, W4, W7, W13, and W18 were omitted because they were not output. As shown in Figs. 23 to 25, the accuracy and specificity of outcome prediction are approximately 80% or more. It is considered to be useful as outcome prediction at least in these respects. In particular, since the specificity of each subitem is generally high, the outcome prediction device 200 may issue a warning depending on the outcome prediction result. For example, when the outcome prediction of the subitem "death" is made, it may be necessary to prepare for ACP response, and when the outcome prediction of the subitem "discharge to home" is made, it may be necessary to prepare a system for cooperation with home care staff. Therefore, by the outcome prediction device 200 issuing a warning according to the outcome prediction result, general medical personnel can easily determine whether or not a response is necessary according to the outcome prediction, which may lead to a reduction in the burden on medical sites. Examples of warnings include displaying messages such as "Please be careful just in case" and "Consider whether ACP response preparation is necessary," and audio display.
[0114] In the third embodiment, the outcome prediction device 200 includes the care support device 100, but is not limited thereto. For example, the outcome prediction device may not include the care support device 100 (particularly, the generation unit 31 and the second trained model 32). That is, the outcome prediction device may include the estimation support device 1 instead of the care support device 100. In this case, the outcome prediction device may generate an outcome prediction using patient information, at least one of the estimation result of the cancer patient obtained from the estimation support device 1 and the observation information on the estimation result, the planned care plan input via the input unit, and the third learning model. Even in this case, the same action and effect as in the third embodiment can be achieved.
[0115] For example, a specialist in palliative care plans may not need a device capable of supporting the estimation of symptoms of a cancer patient (estimation support device) and a device capable of supporting the creation of a care plan for a cancer patient (care support device), but may need to objectively confirm how the outcome of a cancer patient can be predicted by a care plan that he or she has created. In order to meet such needs, an outcome prediction device used by a specialist in palliative care plans may not include both a care support device and an estimation support device (particularly, the estimation unit 20, the trained model 30, the generation unit 31, and the second trained model 32). In other words, the outcome prediction device may be a device that includes at least an input unit, a prediction unit, a third trained model, and an output unit. Such an outcome prediction device may generate an outcome prediction for a cancer patient in a predetermined period of time using, for example, patient information (patient attribute information and initial evaluation information) input by a specialist or the like via the input unit, the subjective symptoms of the cancer patient determined by the specialist or the like, and the planned care plan created by the specialist or the like, in addition to the third trained model. This allows the expert or other person to objectively grasp the validity of the care plan that he or she has created, and the outcome prediction device can be used to consider whether or not the care plan needs to be revised. In addition, the use of the outcome prediction device can also provide the same effects as those of the third embodiment. [Explanation of symbols]
[0116] 1, 1A, 1B...estimation support device, 2A, 2B...server, 1a...display, U...user, 10...input unit, 20...estimation unit, 30...trained model, 31...generation unit, 32...second trained model, 33...prediction unit, 34...third trained model, 30a...decision tree, 40...output unit, 50...communication unit, 60...communication unit, 100...care support device, 200...outcome prediction device.
Claims
1. an input unit for inputting patient information including attribute information of a cancer patient and initial evaluation information regarding a functional disorder of the plant system; An estimation unit that estimates a subjective symptom of the cancer patient using patient information input to the input unit and a trained model generated in advance; an output unit that outputs an estimation result of the subjective symptom by the estimation unit; Equipped with The trained model is a trained model generated using training data so as to output symptom estimation information of the cancer patient when the patient information is input. Estimation support device.
2. The attribute information includes at least one of information regarding medical treatment setting, age, sex, cancer type, treatment period, ECOG-PS, and requester; The estimation support device according to claim 1 .
3. The initial evaluation information includes information regarding at least one of loss of appetite, nausea and vomiting, abdominal distension, constipation, edema, and insomnia; The estimation support device according to claim 1 .
4. The subjective symptoms include symptoms resulting from animal system dysfunction; The estimation support device according to any one of claims 1 to 3.
5. The subjective symptoms are pain, dyspnea, cough, phlegm, fatigue, drowsiness, anxiety, depression, and mood swings. Distress, delirium, lack of informed consent, and spiritual pain At least one of The estimation support device according to any one of claims 1 to 4.
6. The trained model is a decision tree. The estimation support device according to any one of claims 1 to 5.
7. An estimation support device according to any one of claims 1 to 6, A generation unit that generates a care plan proposal for the cancer patient using the patient information, the estimation result, or at least one of observation information on the estimation result, and a second trained model generated in advance; Equipped with The second trained model is generated using teacher data different from the trained model, The output unit outputs the care plan proposal. Care support devices.
8. The care support device according to claim 7 ; A prediction unit that generates an outcome prediction for the cancer patient in a predetermined period of time using the patient information, at least one of the estimation result and the observation information, at least one of the proposed care plan and the planned care plan, and a third trained model generated in advance; Equipped with The third trained model is generated using teacher data different from the trained model and the second trained model; The output unit outputs the outcome prediction. Outcome predictor.
9. An estimation support device according to any one of claims 1 to 6, A prediction unit that generates an outcome prediction for the cancer patient in a predetermined period of time using the patient information, at least one of the estimation result and observation information on the estimation result, a planned care plan input via the input unit, and a third trained model generated in advance; and Equipped with The third trained model is generated using teacher data different from the trained model, The output unit outputs the outcome prediction. Outcome predictor.
10. The present invention includes an input node, an output node, and an intermediate node provided between the input node and the output node, The cancer patient information is generated using teacher data that corresponds to the attribute information of the cancer patient and the patient information including the disorder information of the plant system function, and the symptom estimation information used to estimate the subjective symptoms of the cancer patient; A trained model for causing a computer to function such that when the patient information is input to the input node, the symptom estimation information is output by the output node.
11. the input node is a root node, the output node is a leaf node, the intermediate nodes branch paths between the root node and the leaf nodes; The trained model of claim 10, which is a decision tree.
Citation Information
Patent Citations
Decision system for patient'S subjective symptom
JP2002336259A
Technique for classifying data
JP2009122851A
Patient risk assessment based on data from multiple sources in a healthcare facility
JP2019207684A
Distributed machine learning system, apparatus, and method
JP2019526851A
