How to assess frailty

Abstract

To provide an efficient and accurate frailty determination method which widely covers even such elderly persons who do not receive physical examinations.SOLUTION: There is provided a frailty determination method using a machine learning model. The machine learning model creates an objective variable using question items used for frailty determination as training data. An injury / disease name ICD10 code extracted from receipt data of a user is as a feature value, the feature value for the user is input to the machine learning model, and it is determined whether the user is in a frail state or not.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a technology for automatically determining whether an elderly person or the like is in a frail state by using existing medical receipt data as features of a machine learning model. [Background technology]

[0002] In Japan, where the population is aging, social security costs, especially medical and nursing care costs, are increasing, putting pressure on national finances and increasing the burden on individuals. The so-called baby boomer generation will be 75 years old or older in just a few years, and the aging rate is expected to reach 30% in 2025, and by 2033, one in four people will be elderly.

[0003] Currently, 80 to 90 percent of people in their late 70s and 80s have some kind of illness, and it is recognized that preventing chronic diseases from becoming severe is extremely important. The Ministry of Health, Labor and Welfare's "Guidelines for Health Care Activities Taking into Account the Characteristics of the Elderly (2018)" clearly states the importance of frailty prevention measures.

[0004] Frailty is a term proposed by the Japan Geriatrics Society as the Japanese translation of the word frailty, which means "a state in which the ability to recover from stress is reduced due to a decline in reserve capacity associated with aging." Frailty is considered a stage before the need for nursing care, but it is prone to multifaceted problems, including not only physical vulnerability but also mental and psychological vulnerability and social vulnerability, and is defined as "a high-risk state that is prone to health disorders, including disabilities and death."

[0005] Frailty is one of the factors that lead to the need for nursing care, but it has been shown to be related to various factors such as falls and fractures, illness onset and worsening, dementia, depression, admission to facilities, hospitalization, and death (chronic diseases cause frailty, but frailty can also be a factor in chronic diseases). Preventing frailty in the elderly with comorbid diseases can also prevent the condition from worsening.

[0006] On the other hand, frailty is reversible, and it has become clear that a certain percentage of people can return to a healthy state by taking necessary measures, so early detection of frailty and taking preventive measures are extremely important from the perspective of reducing medical and nursing care costs. Frailty is recognized as a main target of the high-risk approach in insurance programs for the elderly, and there is great significance in identifying those who qualify for this in promoting health programs.

[0007] The above-mentioned "Guidelines for Health Care Activities Taking into Account the Characteristics of the Elderly (2018)" clearly states the need to shift from measures against lifestyle-related diseases (specific health guidance, etc.) to measures that focus on frailty, such as malnutrition caused primarily by loss of weight and muscle mass, and declines in oral function, motor function, and cognitive function; however, it is necessary to determine (distinguish) which elderly people are in a frail state.

[0008] Currently, the CHS developed by Fried et al. is the most widely used test for frailty assessment, and in Japan, the partially improved Japanese version, the J-CHS, the Basic Checklist, and the Simple Frailty Check are all used. However, as these tests generate scores through questionnaires and physical function tests, administrative tasks such as measuring, tabulating, calculating scores, and assessing frailty require a lot of time and effort.

[0009] For example, at health checkups conducted by local governments, questionnaires are given to elderly people who visit the clinic, and frailty scores are calculated and assessed based on the results. However, this method is extremely inefficient, and the attendance rate for health checkups is very low to begin with, and many of those who do attend are healthy people who take care of their health on a daily basis, meaning that frailty assessments cannot be carried out on subjects who should be screened.

[0010] Taking this situation into consideration, for example, Patent Publication No. 2019-021146 (Patent Document 1), Patent Publication No. 2020-201579 (Patent Document 2), Patent Publication No. 2021-086466 (Patent Document 3), Patent No. 6815055 (Patent Document 4), and Patent No. 6943395 (Patent Document 5) disclose technologies for providing frailty assessment and prevention / treatment methods. However, all of these conventional technologies can only assess frailty in those who answer questions, limiting the subjects for whom frailty risk can be calculated, and require filling out checklists, answering evaluation items, obtaining biometric data, interviewing and measuring, etc., so they cannot widely cover elderly people who do not undergo health checkups, and cannot be said to be efficient and accurate assessment methods.

[0011] The importance of frailty assessment is not limited to Japan, but is similar in other countries as well. For example, the following British paper (Non-Patent Document 1) discloses a technology to automatically assess frailty using a machine learning model that uses medical receipt data. Specifically, ICD10 codes that can be used as markers for frailty assessment are prepared in advance, subjects are divided into several clusters by cluster analysis, and the cluster with the highest proportion of people with ICD10 codes corresponding to the above markers is determined to be the cluster corresponding to frailty. When comparing this cluster with other groups, 109 ICD10 codes that appear more than twice as frequently are used as explanatory variables, and logistic regression is performed with the objective variable being whether or not the subject corresponds to the cluster, and the weight is defined as 5 times the regression coefficient.

[0012] According to the method described in the British paper, although it can cover a wide range of elderly people, the three-digit ICD codes in the receipts are weighted, the total value is defined as a score, and frailty is judged based on the score. Since the determination of frailty itself is made using cluster analysis and the ICD10 code, there is an issue that the validity of the objective variable is low and the accuracy of the determination is also low. [Prior art documents] [Non-patent literature]

[0013] [Non-Patent Document 1] Gilbert, Thomas et al. “Development and validation of a Hospital Frailty Risk Score focusing on older people in acute care settings using electronic hospital records: an observational study.” Lancet(London,England)vol.391,10132(2018):1775-1782.doi:10.1016 / S0140-6736(18)30668-8.) [Patent documents]

[0014] [Patent Document 1] JP 2019-021146 A [Patent Document 2] JP 2020-201579 A [Patent Document 3] Patent Publication No. 2021-086466 [Patent Document 4] Patent No. 6815055 [Patent Document 5] Patent No. 6943395 Summary of the Invention [Problem to be solved by the invention]

[0015] Because frailty is the main target of the high-risk approach in health care programs for the elderly, there is a need for an efficient and accurate method of assessment that broadly covers elderly people, including those who do not undergo health checkups, and does not require the conventional interview and response processing work.The present invention aims to solve these problems. [Means for solving the problem]

[0016] In order to achieve the above-mentioned object, the present invention provides a method for determining frailty using a machine learning model, the machine learning model creating a response variable using questions used for frailty determination as training data, using the name of an injury or illness extracted from a user's prescription data or an ICD10 code corresponding to the name of an injury or illness as a feature, and inputting the feature of the user into the machine learning model to determine whether the user is frail or not. Effect of the Invention

[0017] According to the present invention, it is possible to cover a wide range of elderly people who have a history of visiting hospitals and who, although they should be screened, do not undergo health checkups, and it is possible to provide an efficient and accurate determination method that does not require the conventional interview and response processing work.By being able to quickly and accurately determine whether elderly people are frail while reducing the workload of local governments and other organizations, it is possible to contribute to maintaining the health of the elderly and reducing social security costs throughout Japan. [Brief description of the drawings]

[0018] [Figure 1] Comparison of the true positive rate and false positive rate of the ROC curve of the present model and the model in the British paper BEST MODE FOR CARRYING OUT THEINVENTION

[0019] Hereinafter, an embodiment of the present invention will be described in detail. In order to facilitate a comparison of accuracy between the assessment method according to the present invention (hereinafter referred to as the model of the present invention) and the assessment method described in the above-mentioned British paper (hereinafter referred to as the model of the British paper), the model of the present invention also uses logistic regression as its algorithm, just like the model of the British paper, and uses the same ICD10 code as a feature.

[0020] The input data used in the present invention model was prescription data from April 2020 to March 2021, and the output data was frailty questionnaire response data from April 2020 to March 2021. The total number of subjects was 2741 (breakdown: training data: 2055 cases, test data: 686 cases). The subjects of the analysis were elderly people aged 75 years or older who underwent a health checkup in fiscal year 2020.

[0021] The objective variable was dichotomized based on whether or not four or more of the following 12 items in the frailty questionnaire were met (the same as the existing criteria for determining whether one is frail). (a) Whether or not you eat three meals a day (b) Whether or not it is becoming more difficult to eat solid foods compared to six months ago (c) Whether or not you ever choke on tea, soup, etc. (d) Whether or not there was a weight loss of 2 to 3 kg or more in the past 6 months (e) Whether or not your walking speed has slowed compared to before (f) Have you fallen in the past year? (g) Whether or not you exercise, such as walking, at least once a week (h) Whether people around you have said that you have memory problems, such as "I always ask the same thing" (i) Whether there are times when you don't know what date it is (j) Do you go out at least once a week? (k) Do you usually have contact with family or friends? (l) Do you have someone nearby you can talk to when you are feeling unwell?

[0022] Tasks to which machine learning models can be applied include classification tasks, in which the appropriate option is selected from a set of options, and regression tasks, in which a numerical value is predicted; however, as many algorithms can be applied to both tasks, we will not limit this to one or the other. When applied as a classification task, the objective variable is whether or not four or more items, the criteria for frailty, and this results in a model that predicts whether or not a person is frail. When applied as a regression task, the objective variable is the number of applicable items on the questionnaire, and this results in a model that predicts the presence and degree of frailty.

[0023] The features used were the names of illnesses and injuries recorded in the medical receipt data, which were converted into three-digit ICD10 code classifications, and only the ICD10 codes that were the same as the model in the British paper were extracted from those and binarized. From there, the ICD10 codes were further limited to those held by more than 1% of the subjects in the total 2,741 data used this time. The reason for this is that, since the amount of data used for validation was small, if there were an ICD10 code with an extremely small number of applicable subjects, a bias would occur between the test data and training data, which would have a negative effect on the model. If there was more data, it would be possible to use the exact same features as the model in the British paper, and in that case it would clearly be possible to improve the accuracy, so the validity of this validation would not be lost.

[0024] Through the above process, the presence or absence of 33 ICD10 codes was set as a feature in the model of the present invention. The 33 codes are A09, D64, E83, E86, E87, F03, F32, G30, G40, H91, I63, I67, I69, K59, L08, M19, M25, M41, M48, M79, M81, N18, N19, N20, N28, N39, R00, R11, R31, R63, S00, S22, and S32.

[0025] The results of predictions made using the test data for the model of the present invention and the model from the British paper are shown in the ROC curve in Figure 1. The vertical axis of the ROC curve indicates the true positive rate, and the horizontal axis indicates the false positive rate. For evaluation values ​​between 0 and 1 given by the model, predictions are plotted for values ​​above the threshold as positive, and values ​​below the threshold as negative, with the threshold being varied. The further to the top left the plot is, the better the model is considered to be.

[0026] The predictions made by the model from the British paper are shown by point A, and the predictions made by the model of the present invention are shown by curve B. For the prepared test data, classification results using the method from the British paper showed a true positive rate of 0.36 and a false positive rate of 0.25, whereas the ROC curve for the model of the present invention was to the upper left of the results from the model from the British paper, demonstrating a prediction accuracy that exceeded those of the model from the British paper.

[0027] In addition, by changing the threshold value, the model of the present invention can change the balance between the true positive rate and the false positive rate, and can be adjusted to reduce missed samples depending on the application, or it can be narrowed down to a small number of subjects to determine only those who are definitely frail.

[0028] For example, in the model in the British paper, when the false positive rate is 0.25, the true positive rate is 0.36, meaning that 64% of those who are actually frail are mistakenly determined to be not frail, whereas in the model of the present invention, when the false positive rate is 0.25, the true positive rate is 0.43, meaning that 57% of those who are actually frail are mistakenly determined to be not frail. In other words, 7% more people can be correctly determined to be frail than in the model in the British paper.

[0029] In the above embodiment, the algorithm of the machine learning model uses logistic regression in accordance with the British paper model, but is not limited to this. Algorithms such as linear regression, decision tree, random forest, support vector machine, gradient boosting decision tree, neural network, and deep learning may also be used.

[0030] In the above embodiment, the feature amount used for the analysis is only the name of the injury or illness extracted from the medical receipt data, but the medical receipt data may further include age, sex, medical procedure, prescribed medicine, medical expenses, frequency of hospital visits, and frequency of hospitalization, and may further include household composition and place of residence in addition to the medical receipt data. By increasing the feature amount related to frailty, it becomes possible to perform a more accurate judgment.

[0031] In the above embodiment, the ICD10 code corresponding to the name of the injury or illness is used as input data, but the name of the injury or illness itself may also be used.

[0032] As described above, the frailty assessment method of the present invention makes it possible to identify people who may be frail but who do not undergo health checkups or answer conventional frailty questionnaires, thereby leaving their "frailty state" unattended and subsequently becoming dependent on care (in other words, frailty can be assessed based on medical receipt data alone, without the need for health checkups or answering questionnaires), and because there is no need to answer questionnaires, the workload of local government employees can be significantly reduced.

[0033] Frailty is reversible and can be returned to a healthy state through early detection and the implementation of appropriate measures. Therefore, by using this assessment method to detect frailty at a frequency of, for example, once every six months, it is possible to detect frailty early and return a person to a healthy state, which will ultimately lead to a reduction in social security costs. [Explanation of symbols]

[0034] A. Prediction based on the UK paper model B. Prediction by the present model

Claims

1. A method for determining frailty using a machine learning model, comprising: The machine learning model creates a response variable using questions used to assess frailty as training data, The ICD10 code corresponding to the name of the injury or illness extracted from the user's receipt data is used as a feature. By inputting the feature amount of the user into the machine learning model, it is determined whether the user is in a frail state; The ICD10 codes are A09, D64, E83, E86, E87, F03, F32, G30, G40, H91, I63, I67, I69, K59, L08, M19, M25, M41, M48, M79, M81, N18, N19, N20, N28, N39, R00, R11, R31, R63, S00, S22, and S32, which are 33 types. A method for determining frailty.

2. The frailty assessment method further includes, as features, age, sex, medical procedure, prescription drug, medical expenses, frequency of hospital visits, and frequency of hospitalization from the receipt data. The method for determining frailty according to claim 1 .

3. The frailty assessment method further includes, as features, household composition and place of residence in addition to the receipt data. The method for determining frailty according to claim 1 or 2.

4. The questions used to assess frailty are: (a) Are you eating three meals a day? (b) Whether or not you have difficulty eating solid foods compared to six months ago (c) Whether or not you ever choke on tea, soup, etc. (d) Whether or not there was a weight loss of 2 to 3 kg or more in the past 6 months (e) Whether or not the walking speed has become slower than before (f) Have you fallen in the past year? (g) Whether or not you exercise, such as walking, at least once a week (h) Have people around you said that you have forgetfulness, such as "I always ask the same thing?" (i) Whether there are times when you do not know what date it is (j) Do you go out at least once a week? (k) Whether or not you have regular contact with family and friends (l) Do you have someone nearby you can talk to when you are feeling unwell? These are the 12 items: The method for determining frailty according to claim 1 or 2.

5. The algorithm of the machine learning model is any one of logistic regression, linear regression, decision tree, random forest, support vector machine, gradient boosting decision tree, neural network, and deep learning; The method for determining frailty according to claim 1 or 2.

6. The period of the feature to be input to the machine learning model is any period of one month or more, and the period of the output data is the fiscal year included in the input feature or a period beyond that. The method for determining frailty according to claim 1 or 2.

Images