JP6548243B1 - Health checkup examination probability calculation method and medical checkup recommendation notice support system - Google Patents
Health checkup examination probability calculation method and medical checkup recommendation notice support system Download PDFInfo
- Publication number
- JP6548243B1 JP6548243B1 JP2018203585A JP2018203585A JP6548243B1 JP 6548243 B1 JP6548243 B1 JP 6548243B1 JP 2018203585 A JP2018203585 A JP 2018203585A JP 2018203585 A JP2018203585 A JP 2018203585A JP 6548243 B1 JP6548243 B1 JP 6548243B1
- Authority
- JP
- Japan
- Prior art keywords
- data
- checkup
- health
- probability
- insured
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Medical Treatment And Welfare Office Work (AREA)
Abstract
【課題】被保険者が健康診断を受診する確率を計算し、勧奨通知業務の効率化を図る。【解決手段】確率計算の基礎データとなる母集団を、複数又は単独の自治体、或いは複数又は単独の健保組合等が保有する過去ｎ年（ｎは３以上の整数）間において被保険者が受診した健康診断の健診データとし、処理プログラムは、第一ステップとして、全ての前記健診データから過去ｍ年（ｍは３以上の整数、ｍ≦ｎ）分のデータを抽出し、直近の１年分を除いたｍ−１年分のデータを、ｘ個の説明変数を持つデータに加工し、第二ステップとして、当該ｍ−１年分の前記健診データ及び直近１年分の健診受診の有無を教師データとして用いて学習し、第三ステップとして、前記学習済みモデルによって、特定の単独自治体又は特定の単独健保組合に加入する個々の被保険者の受診確率を算出する。【選択図】図２An object of the present invention is to calculate the probability that an insured person will receive a health checkup and to improve the efficiency of a recommendation notification work. SOLUTION: The insured is consulted in the past n years (n is an integer of 3 or more) held by multiple or single local governments or multiple or single health insurance associations, etc., as the basic data for probability calculation As a first step, the processing program extracts data for the past m years (m is an integer of 3 or more, m ≦ n) from all the medical examination data as the first step, and the latest 1 The data for m-1 year excluding the year is processed into data with x explanatory variables, and as the second step, the medical examination data for m-1 year and the medical examination for the most recent year The presence / absence of medical examination is learned as teacher data, and as a third step, the medical examination probability of each individual insured who joins a specific single local government or specific single health insurance association is calculated by the learned model. [Selection] Figure 2
Description
本発明は、医療保険機関が実施する健康診断の受診を促す被保険者への勧奨通知の効率化を支援するための健康診断受診確率計算方法及び健診勧奨通知支援システムに関する。 The present invention relates to a health checkup examination probability calculation method and a health checkup recommendation notice support system for supporting the streamlined notification of recommendation to an insured who promotes medical checkup conducted by a medical insurance organization.
我が国における医療費は毎年増加し続け、既に４０兆円を突破し、高齢化の進展に伴って今後も増え続けることが予想されている。中でも、国民の多数が加入する国民健康保険は、主として地方自治体（都道府県及び市区町村）が運営する公的保険であり、医療費の増加は、自治体の財政及び国家財政に対して大きな負担となっている。 Medical expenses in Japan continue to increase every year and have already surpassed 40 trillion yen, and it is expected that they will continue to increase as the population ages. Among them, the National Health Insurance, which a large number of people subscribe to, is a public insurance that is mainly operated by local governments (prefecture and municipality), and the increase in medical expenses is a big burden on local government finances and national finances. It has become.
増え続ける医療費の削減に向けて、各自治体は被保険者である住民の健康維持、病気の早期発見・早期治療を目的とした特定健診、特定保健指導を毎年実施しているものの、告知や通知によっても受診しない住民が多数存在する。 To reduce the ever-increasing cost of medical expenses, each municipality carries out health checkup of residents who are insured, specific medical checkup for the purpose of early detection and early treatment of diseases, and specific health guidance every year, but the notification There are a large number of residents who do not receive medical examinations even after notification.
こうした状況を踏まえ、各自治体は、受診を促す勧奨通知を郵送しているが、予算や人員の制約があるため、全ての被保険者に通知する余裕が無く、限られた予算及び人員の中で、最も効果的な勧奨通知を行うことが求められている。 Under such circumstances, each local government has sent a notice of recommendation to have a medical examination, but due to budget and personnel restrictions, there is no room to notify all insured persons, and it is among the limited budget and personnel. It is required to make the most effective recommendation notice.
従来、被保険者の健康状態に基づいて実施した保健事業の評価を支援するシステム（特許文献１）、複数の対象者の所定期間における健康診断の測定値又は受診履歴に基づいて対象者の新たな健診情報に基づく値とモデルとを用いてリスクの評価値を算出するシステム（特許文献２）、ある健康診断を受診可能な複数の団体に含まれるそれぞれの団体の検診の過去の受診率情報に応じて団体の組み合わせを決定し、該決定した組合せに含まれる団体を前記ある健康診断を受診する対象の団体として抽出するシステム（特許文献３）等が知られているが、効果的な健診勧奨通知を支援するシステムは存在しなかった。 Conventionally, a system for supporting the evaluation of a health business implemented based on the health status of the insured person (Patent Document 1), new values of the target person based on the measurement values or the medical examination history of the medical checkup in a predetermined period of a plurality of target people. System to calculate risk evaluation value using value and model based on medical checkup information (patent document 2), past visit rate of examination of each group included in multiple groups that can receive certain medical checkup There is known a system (Patent Document 3) etc. in which a combination of groups is determined according to the information and groups included in the determined combination are extracted as a group to be a target of the medical checkup. There was no system to support health check recommendation notifications.
本発明は、上記した課題を解決するため、過去の健診データを基にして、特定の国民健康保険又は特定の健保組合等に加入する被保険者の受診確率を計算し、勧奨通知をしなくても健康診断を受診する可能性の高い者、又は勧奨通知をしても健康診断を受診する可能性の低い者を特定して排除し、残った被保険者、即ち、勧奨通知をすることによって受診する可能性がある者を確度高く選定し、勧奨通知業務を効率化することを目的とする。 In order to solve the above-mentioned problems, the present invention calculates consultation probability of an insured person who joins a specific National Health Insurance or a specific health insurance association etc. based on past medical examination data, and sends a recommendation notification. Identify and exclude those who are not likely to receive a medical checkup, or who are unlikely to receive a medical checkup even after a recommendation notice, and the remaining insured persons, ie, a notification of recommendation The purpose is to select people who may receive a medical examination with high accuracy and to streamline the recommendation notification business.
上記目的を達成するため、本願の第一の発明は、被保険者の健康診断受診確率の計算方法であって、確率計算の基礎データとなる母集団を、複数又は単独の自治体、或いは複数又は単独の健保組合等が保有する過去ｎ年（ｎは３以上の整数）間において被保険者が受診した健康診断の健診データとし、前記健診データを基に、コンピューターが被保険者の受診行動を予測する処理プログラムを含み、前記コンピューターは、前記処理プログラムに従い、第一ステップとして、全ての前記健診データから過去ｍ年（ｍは３以上の整数、ｍ≦ｎ）分のデータを抽出し、直近の１年分を除いたｍ−１年分のデータを、ｘ個の説明変数を持つデータに加工し、第二ステップとして、当該ｍ−１年分の前記健診データ及び直近１年分の健診受診の有無を教師データとして用いて学習を行い、第三ステップとして、前記学習により構築したモデルによって、特定の単独自治体又は特定の単独健保組合に加入する個々の被保険者の受診確率を算出する、ことを特徴とする。 In order to achieve the above object, the first invention of the present application is a method of calculating a health checkup examination probability of an insured person, wherein a population serving as a basic data of probability calculation is a plurality of or a single municipality or a plurality of Medical checkup data of the medical checkup that the insured has consulted during the past n years (n is an integer of 3 or more) owned by a single health insurance association etc., and based on the health checkup data , the computer visits the insured According to the processing program, the computer extracts data for the past m years (m is an integer of 3 or more, m ≦ n) from all the medical checkup data according to the processing program. And processing m-1 year's worth of data excluding the last 1 year's worth into data with x explanatory variables, and as the second step, the medical checkup data for the m-1 year and the most recent 1 With or without medical checkup for the year Perform learning using as teacher data, wherein the third step, the by model constructed by learning, and calculates the consultation probabilities of individual insured to subscribe to a particular single municipality or specific single health insurance, that I assume.
また本願の第二の発明は、被保険者の健康診断受診確率の計算方法であって、確率計算の基礎データとなる母集団を、複数の自治体又は複数の健保組合等が保有する過去ｎ年（ｎは３以上の整数）間において被保険者が受診した健康診断の健診データとし、前記健診データを基に、コンピューターが被保険者の受診行動を予測する処理プログラムを含み、前記コンピューターは、前記処理プログラムに従い、第一ステップとして、複数の保険者の前記健診データから過去ｍ年（ｍは３以上の整数、ｍ≦ｎ）分のデータを抽出し、直近の１年分を除いたｍ−１年分のデータをｘ個の説明変数を持つデータに加工し、アルゴリズムはディープラーニングであり、当該ｍ−１年分のデータ及び直近１年での健診受診の有無を教師データとして用いることで、複数のレイヤーに対応する複数の重みＷｉを学習して第一特徴量を作成し、第二ステップとして、特定の単独自治体又は特定の単独健保組合等の前記健診データのみを抽出し、アルゴリズムにトランスファーラーニングを用いて前記第一ステップで作成した前記重みＷｉの一部を調整した第二特徴量を作成し、第三ステップとして、前記第二特徴量に基づいたモデルによって、前記特定の単独自治体又は前記特定の単独健保組合等に加入する個々の被保険者の受診確率を算出する、ことを特徴とする。 The second invention of the present application is a method of calculating a health checkup examination probability of an insured person, wherein a plurality of local governments or a plurality of health insurance associations etc. hold a population serving as basic data of probability calculation. (n is an integer of 3 or more) and medical examination of medical examination data insured was admitted between, on the basis of the medical examination data comprises a processing program computer to predict the visit behavior of the insured, the computer According to the processing program, as the first step, data for the past m years (m is an integer of 3 or more, m ≦ n) is extracted from the health checkup data of a plurality of insurers, The m-1 year's worth of data excluding it is processed into data with x explanatory variables, and the algorithm is deep learning, and the m-1 year's worth of data and the teacher for the health checkup visit in the most recent year are teachers Use as data And learn a plurality of weights Wi corresponding to a plurality of layers to create a first feature, and as a second step, extract only the health checkup data of a specific independent municipality or a specific independent health insurance association The second feature is generated by adjusting a part of the weight Wi created in the first step using transfer learning as an algorithm, and as the third step, the specification is performed using the model based on the second feature. Calculating a consultation probability of an individual insured who joins the independent local government or the specific independent health insurance association or the like.
また本願の第三の発明は、自治体又は健保組合等が実施する健康診断の受診を促す被保険者への勧奨通知を、受診確率計算に基づいて効率的に行う健診勧奨通知支援システムであって、複数又は単独の自治体、或いは複数又は単独の健保組合等が保有する被保険者が受診した過去ｎ年（ｎは３以上の整数）の健康診断の結果が蓄積された健診データベースと、前記健診データベースに蓄積されているデータを基に被保険者の受診行動を予測する処理プログラムが記録されているサーバーと、を備え、前記サーバーは、前記処理プログラムに従い、第一ステップとして、前記健診データベースから複数又は単独の自治体、或いは複数又は単独の健保組合等の過去ｍ年（ｍは３以上の整数、ｍ≦ｎ）分のデータを抽出し、直近の１年分を除いたｍ−１年分のデータを、ｘ個の説明変数を持つデータに加工し、第二ステップとして、当該ｍ−１年分のデータ及び直近１年での健診受診の有無を教師データとして用いて学習を行い、第三ステップとして、前記学習により構築したモデルによって、特定の自治体又は特定の健保組合等に加入する個々の被保険者の受診確率を算出し、第四ステップとして、前記特定の自治体又は前記特定の健保組合等が定めた所定の高確率値以上及び所定の低確率値以下を除く中間確率値の被保険者に対してのみ、又は予め定めた所定数の中間確率値の被保険者に対してのみ、勧奨通知対象リストを作成する、ことを特徴とする。 The third invention of the present application is a medical examination recommendation notification support system that efficiently performs, based on the reception probability calculation, recommendation notification to an insured person who promotes medical checkup conducted by a local government or a health insurance association or the like. Medical checkup database in which the results of the medical checkup of the past n years (n is an integer of 3 or more) visited by the insured persons owned by multiple or independent local governments or multiple or independent health insurance associations etc. A server on which a processing program for predicting a medical checkup behavior of the insured person is recorded based on the data stored in the medical checkup database, the server, according to the processing program, as the first step Data for the past m years (m is an integer of 3 or more, m n n) of multiple or single municipalities or multiple or single health insurance associations etc. are extracted from the medical checkup database and m excluding the last 1 year The data for one year, and processed into data having x number of explanatory variables, using as a second step, a medical examination whether visits in the m-1_Nenbun'nodetaoyobichokkin 1 year as teacher data learning As a third step, according to the model constructed by the learning , the consultation probability of an individual insured who joins a specific municipality or a specific health insurance association etc. is calculated, and as the fourth step, the specific municipality or The insured person of the intermediate probability value excluding the predetermined high probability value or more and the predetermined low probability value or less defined by the specific health insurance association etc., or the insured person of the predetermined number of intermediate probability values determined in advance The recommendation notification target list is created only for.
また本願の第四の発明は、自治体又は健保組合等が実施する健康診断の受診を促す被保険者への勧奨通知を、受診確率計算に基づいて効率的に行う健診勧奨通知支援システムであって、複数の自治体又は複数の健保組合等が保有する被保険者が受診した過去ｎ年（ｎは３以上の整数）の健康診断の結果が蓄積された健診データベースと、前記健診データベースに蓄積されているデータを基に被保険者の受診行動を予測する処理プログラムが記録されているサーバーと、を備え、前記サーバーは、前記処理プログラムに従い、第一ステップとして、前記健診データベースから過去ｍ年（ｍは３以上の整数、ｍ≦ｎ）分のデータを抽出し、直近の１年分を除いたｍ−１年分のデータの中からｘ個の説明変数を持つデータに加工し、アルゴリズムはディープラーニングであり、当該ｍ−１年分のデータ及び直近１年での健診受診の有無を教師データとして用いることで、複数のレイヤーに対応する複数の重みＷｉを学習して第一特徴量を作成し、第二ステップとして、特定の自治体又は特定の健保組合等のデータのみを抽出し、アルゴリズムにトランスファーラーニングを用いて前記第一ステップで作成した前記重みＷｉの一部を調整した第二特徴量を作成し、第三ステップとして、前記第二特徴量に基づいたモデルによって、前記特定の自治体又は前記特定の健保組合等に加入する個々の被保険者の受診確率を算出し、第四ステップとして、前記特定の自治体又は前記特定の健保組合等が定めた所定の高確率値以上及び所定の低確率値以下を除く中間確率値の被保険者に対してのみ、又は予め定めた所定数の中間確率値の被保険者に対してのみ、勧奨通知対象リストを作成する、ことを特徴とする。
The fourth invention of the present application is a medical checkup recommendation notification support system that efficiently performs, based on consultation probability calculation, recommendation notification to an insured person who encourages a medical checkup visit conducted by a local government or a health insurance association etc. Medical checkup database in which the results of the medical checkup of the past n years (n is an integer of 3 or more) visited by the insured persons owned by multiple municipalities or multiple health insurance associations etc. And a server in which a processing program for predicting an insured person's examination behavior is recorded based on accumulated data, the server, according to the processing program, as a first step, from the medical checkup database Extract data for m years (m is an integer of 3 or more, m n n), and process it into data with x explanatory variables from data for m-1 years excluding the last 1 year , The algorithm is The e-learning is used to learn multiple weights Wi corresponding to multiple layers by using the m-1 year's worth of data and the presence or absence of a health checkup visit in the most recent year to learn the first feature amount A second feature of creating and extracting only data of a specific municipality or a specific health insurance association as a second step, and adjusting a part of the weight Wi created in the first step using transfer learning as an algorithm The third step is to calculate the visit probability of the specific municipality or the individual insured who joins the specific health insurance association etc. by the model based on the second feature amount as the third step, and the fourth step As for the insured person of intermediate probability value except for the specific high probability value more than the specific local government or the specific health insurance association etc. and the predetermined low probability value or less, or Only for a predetermined insured of a predetermined number of intermediate probability values, creating a encouraged notified list, characterized in that.
本願の第一発明、及び第三発明によれば、被保険者が健康診断を受診する確率計算の基礎データとなる母集団を、複数又は単独の自治体、或いは複数又は単独の健保組合等が保有する過去ｎ年（ｎは３以上の整数）間において被保険者が受診した健康診断の健診データとし、高い性能の受診確率計算モデルを作成するのに十分な健診データ数を有する大都市（又は大きな健保組合等）は、地方小都市（又は小さな健保組合等）のデータがノイズとならないように当該大都市（又は大きな健保組合等）のみのデータを用いて確率モデルを作成し、高い性能の受診確率計算モデルを作成するのに十分な健診データ数が無い地方小都市（又は小さな健保組合等）は、全国各地の複数の自治体（又は健保組合）のデータを用いて確率計算モデルを作成することにより、特定の単独自治体又は特定の単独健保組合等に加入する個々の被保険者の受診確率を精度よく計算（推測）することができる。その結果、勧奨通知を発送しても健康診断を受診しない可能性の高い者の被保険者グループと、勧奨通知を発送しなくても健康診断を受診する可能性の高い者の被保険者グループを高い確度で特定することができ、これらの２つのグループ以外に属する被保険者グループ、即ち、勧奨通知を発送することによって受診可能性が高くなる被保険者グループを特定することで、勧奨通知に係る事務作業を効率化し、予算及び人員の適正化を図ることができる。また、結果として健康診断を受診する被保険者が増えることで、医療費の削減にも繋げることができる。 According to the first invention and the third invention of the present application, a plurality of or a single municipality, or a plurality or a single health insurance association or the like possess a population serving as the basic data of probability calculation for the insured to receive a medical examination. In the past n years (n is an integer of 3 or more) as a health checkup data of a health checkup that the insured has received, and a large city with a sufficient number of health checkup data to create a high performance consultation probability calculation model (Or large health insurance association etc.) create a probabilistic model using data only for the big city (or large health insurance association etc) so that the data of small local cities (or small health insurance association etc) do not become noise, and high Local small cities (or small health insurance associations etc.) that do not have enough medical examination data to create a consultation probability calculation model for performance use probability calculation models using data from multiple local governments (or health insurance associations) across the country. Create Rukoto, it is possible to identify a single municipality or specific single health insurance such visits probability of individual insured that subscribe to accurately calculate (guess). As a result, the insured person group of those who are likely not to receive a medical checkup even if the recommendation notice is sent and the insured group of those who are likely to receive a medical checkup without sending the recommendation notice Can be identified with a high degree of certainty, and by identifying an insured group that belongs to other than these two groups, ie, an insured group that is likely to receive a medical examination by sending a recommendation notice, Administrative work pertaining to can be streamlined, and budgets and personnel can be optimized. In addition, as a result, by increasing the number of insured persons who receive a medical examination, it can be connected to the reduction of medical expenses.
また本願の第二発明、及び第四発明によれば、被保険者が健康診断を受診する確率計算の基礎データとなる母集団を、複数の自治体又は複数の健保組合等が保有する過去ｎ年（ｎは３以上の整数）間において被保険者が受診した健康診断の健診データとし、第一ステップとして、複数の自治体又は複数の健保組合等が有する全てのデータを用いて第一特徴量を作成し、第二ステップとして、特定の単独自治体又は特定の単独健保組合等の前記健診データのみを用いて第一ステップで作成した特徴量の一部を調整した第二特徴量を作成し、第三ステップとして、当該第二特徴量に基づいた確率計算モデルによって、特定の単独自治体又は特定の単独健保組合等に加入する個々の被保険者の受診確率を計算することにより、複数の自治体又は複数の健保組合等の大量なデータに基づいて基本形となる第一特徴量を作成する一方、計算対象となる特定の単独自治体又は特定の単独健保組合等の被保険者に特有な行動パターンを加味することができる。その結果、勧奨通知を発送しても健康診断を受診しない可能性の高い者の被保険者グループと、勧奨通知を発送しなくても健康診断を受診する可能性の高い者の被保険者グループを高い確度で特定することができ、これらの２つのグループ以外に属する被保険者グループ、即ち、勧奨通知を発送することによって受診可能性が高くなる被保険者グループを特定することで、勧奨通知に係る事務作業を効率化し、予算及び人員の適正化を図ることができる。 Further, according to the second and fourth inventions of the present application, the past n years in which a plurality of local governments, a plurality of health insurance associations, etc. possess a population serving as basic data of probability calculation for the insured to receive a medical examination. (N is an integer greater than or equal to 3) health checkup data of the health checkup that the insured person consulted among, as the first step, using all the data possessed by multiple municipalities or multiple health insurance associations etc. As a second step, a second feature amount is generated by adjusting a part of the feature amounts created in the first step using only the health checkup data of a specific independent local government or a specific independent health insurance association etc. As a third step, a plurality of local governments are calculated by calculating the consultation probability of each insured who joins a specific local government or a specific health insurance association by a probability calculation model based on the second feature amount. Or multiple While creating a first feature quantity to be a basic form based on a large amount of data from health insurance associations etc., take into consideration a behavior pattern specific to a specific self-governing body to be calculated or a insured person such as a specific health insurance union etc. Can. As a result, the insured person group of those who are likely not to receive a medical checkup even if the recommendation notice is sent and the insured group of those who are likely to receive a medical checkup without sending the recommendation notice Can be identified with a high degree of certainty, and by identifying an insured group that belongs to other than these two groups, ie, an insured group that is likely to receive a medical examination by sending a recommendation notice, Administrative work pertaining to can be streamlined, and budgets and personnel can be optimized.
上記のとおり本願の各発明は、勧奨通知をしなくても健康診断を受診する可能性の高い被保険者、又は勧奨通知をしても健康診断を受診する可能性の低い被保険者を特定して排除し、残った被保険者、即ち、勧奨通知をすることによって受診する可能性がある者を、確度高く抽出するものであるが、実際のデータを用いて分析すると、各自治体の特徴又は各健保組合等の特徴によって実測値との乖離が異なるため、それぞれの自治体や健保組合等の特徴を生かす形で、どの手法を用いるかを適宜に選択することができる。 As described above, each invention of the present application identifies an insured person who is likely to receive a medical examination without a recommendation notice, or an insured person who is unlikely to receive a medical examination after a recommendation notice. And the remaining insured persons, ie, those who may receive a medical examination by making a recommendation notice, are extracted with high accuracy, but when it is analyzed using actual data, the characteristics of each municipality Alternatively, since the difference from the actual measurement value differs depending on the characteristics of each health insurance association etc., it is possible to appropriately select which method to use in the form of taking advantage of the characteristics of each local government, health insurance association etc.
なお、本発明で言う「健保組合等」とは、自治体が運営する国民健康保険以外の各種健康保険組合（各企業の健康保険組合、全国健康保険協会管掌健康保険（いわゆる協会けんぽ）、共済組合など）を意味するものである。 In addition, "health insurance association etc." referred to in the present invention means various health insurance associations other than the national health insurance operated by the local government (health insurance association of each company, national health insurance association governing health insurance (so-called association kenpo), mutual aid association Etc.) is meant.
図１〜１２を用いて、本発明の実施形態について詳細に説明する。なお各実施形態は本願発明の範囲を限定的に解釈するためのものではなく、特許請求の範囲に記載された内容と発明の趣旨に基づいて、適宜に実施して良いことは言うまでもない。以下説明する本実施形態は、一例として特定の自治体向けにサービスを行うことを想定したシステムとして説明する。 Embodiments of the present invention will be described in detail with reference to FIGS. It is needless to say that each embodiment is not intended to interpret the scope of the present invention in a limited manner, and that the present invention may be appropriately implemented based on the contents described in the claims and the spirit of the invention. The present embodiment described below will be described as a system assuming that a service is provided to a specific local government as an example.
図１は、本願発明の基本となるシステム構成図である。本発明に係るシステムは、第一実施形態、第二実施形態ともに共通であり、複数の自治体が有する過去の健康診断受診結果を蓄積した健診データベース１と、当該健診データベース１に蓄積されているデータを基に被保険者の受診行動を予測する処理プログラム２が記録されているサーバー３とからなる。健診データベース１は、サーバー３内に構築しても良い。以下説明する各実施形態は、処理プログラム２の処理方法が相違することで、被保険者の受診確率の計算方法が異なっている。 FIG. 1 is a system configuration diagram that is the basis of the present invention. The system according to the present invention is common to both the first embodiment and the second embodiment, and is stored in the medical checkup database 1 in which the results of past medical checkup visits owned by a plurality of local governments are accumulated and the medical checkup database 1 It comprises the server 3 in which the processing program 2 which predicts the medical checkup behavior of the insured person based on the existing data is recorded. The medical checkup database 1 may be built in the server 3. Each embodiment described below is different in the processing method of the processing program 2 and in the calculation method of the consultation probability of the insured person.
まず、本発明の第一実施形態に係る処理プログラム２について説明する。
処理プログラム２は、データ数の多い大都市の自治体（都道府県単位であっても良い）とデータ数の少ない小都市の自治体（都道府県単位であっても良い）とで、確率計算の基礎となるデータの抽出方法が異なる。まず小都市の被保険者の受診確率を求める方法について説明する。
First, the processing program 2 according to the first embodiment of the present invention will be described.
Processing program 2 is the basis of probability calculation between the municipality of a large city with a large number of data (may be a prefectural unit) and the municipality of a small city with a small number of data (a prefectural unit may be). Data extraction methods differ. First, a method of obtaining the consultation probability of an insured person in a small city will be described.
小都市の場合、データ数が少なく、高い性能の受診確率計算モデルを構築することが難しいため、健診データベース１に蓄積されている大都市を含む他の自治体のデータを合わせて活用する。 In the case of a small city, the number of data is small, and it is difficult to construct a high performance consultation probability calculation model, so data of other local governments including large cities stored in the health checkup database 1 are used together.
図２は、処理プログラムによる処理フローの概要を示す説明図である。
処理プログラム２は、第一ステップとして、健診データベース１に蓄積されている複数自治体のデータを対象として、この中から過去６年分のデータを、教師データセットを作成する目的で抽出し、直近の１年分のデータを、健診受診の有無（正解ラベルデータ）に加工し、正解ラベルデータ以外の５年分のデータを、図４に示すような４４個の説明変数を持つデータに加工する。
FIG. 2 is an explanatory view showing an outline of a processing flow by the processing program.
As a first step, the processing program 2 extracts data for the past six years from among the data of a plurality of local governments accumulated in the medical checkup database 1 for the purpose of creating a teacher data set. Data for 1 year is processed to check for medical checkup (correct label data), and data for 5 years other than correct label data is processed to data with 44 explanatory variables as shown in Figure 4 Do.
図４は、説明変数のリストであり、過去に健診を受診した被保険者個人の年齢、性別、身長、体重、ＢＭＩ、腹囲、収縮期血圧、拡張期血圧、中性脂肪等の４４項目から構成される。図５は、教師データの構成を示す図である。 FIG. 4 is a list of explanatory variables, and 44 items such as age, sex, height, weight, BMI, abdominal circumference, systolic blood pressure, diastolic blood pressure, neutral fat, etc. of the insured individual who had a medical examination in the past It consists of FIG. 5 is a diagram showing the configuration of teacher data.
処理プログラム２は、第二ステップとして、当該４４個の項目を説明変数とし、受診確率を目的変数とするモデルを作成する。本実施形態では、図６に示すとおりランダムフォレストを用いてモデルを構築する。複数の自治体の健診データの中から、１５００個のサンプルをランダムに抽出し、４４個の説明変数から６個の説明変数をランダムに抽出して１００個の決定木を作成する。当該５年分の健診データ及び直近１年での健診受診の有無を教師データに用いて学習し、モデルとなる決定木を確定する。 As a second step, the processing program 2 creates a model with the 44 items as an explanatory variable and the examination probability as an objective variable. In this embodiment, a model is constructed using a random forest as shown in FIG. From health checkup data of a plurality of municipalities, 1500 samples are randomly extracted, and 6 explanatory variables are randomly extracted from 44 explanatory variables to create 100 decision trees. The health checkup data for the five years and the presence or absence of the health checkup visit in the most recent year are used as training data to learn, and a decision tree to be a model is determined.
処理プログラム２は、第三ステップとして、確定した決定木を用いて、特定小都市の住人である全ての被保険者の受診確率を計算する。全ての被保険者とは、実態上の住人全員ではなく、過去に健康診断を受診した記録が健診データベース１に存在する者の全員を意味する。 As the third step, the processing program 2 uses the decided decision tree to calculate the check probabilities of all the insured persons who are residents of a specific small city. All insured persons are not all the residents in reality, but all the persons who have a record of medical examination in the past exist in the medical examination database 1.
処理プログラム２は、第四ステップとして、算定された個々の被保険者の受診確率をアウトプットとしてリスト化する。算定した個々の被保険者の受診確率が、例えば８０％を超えていた場合には、当該者は高い確度で健康診断を受診する可能性があり、逆に２０％より低い者は、高い確度で健康診断を受診しない可能性がある。即ち、２０％以上８０％以下の者は、勧奨通知を行うことで、健康診断を受診する可能性が高くなることが分かる。 As a fourth step, the processing program 2 lists the calculated individual insured's consultation probability as an output. If, for example, the consultation probability of each insured person calculated exceeds 80%, the person concerned may receive a medical examination with high accuracy, and conversely, the person who is lower than 20% has high accuracy May not receive a medical checkup. That is, it is understood that the notification of the recommendation increases the possibility that the person of 20% or more and 80% or less receives the medical examination.
上記結果を基に、特定小都市は、例えば全被保険者の中から確率２０％以上８０％以下の者のみを特定し、特定した者に対して勧奨通知を行うことで、限られた予算、限られた人員の中で、効率よく勧奨通知業務を実施することができる。 Based on the above results, a specified small city, for example, identifies only those with a probability of 20% or more and 80% or less among all insured persons, and sends a recommendation notification to those who have identified the limited budget. And, within the limited number of personnel, can be efficiently implemented the notification work recommendation.
また勧奨通知の発送数が予め予算化されている場合、発送数に合うように発送対象とする確率閾値を、適宜に調整変更すれば良い。或いは、発送数を予め１０００通としている場合、高確率値・低確率値の人数を均等に排除した中間の人数が１０００人となるように選択すれば良い。 In addition, when the number of shipments of the recommendation notice is budgeted in advance, the probability threshold value to be shipped may be adjusted and changed appropriately so as to match the number of shipments. Alternatively, in the case where the number of shipments is set to 1000 in advance, it may be selected so that the number of intermediate persons who equally exclude the number of persons of high probability value and low probability value is 1,000.
第四ステップは、最終的に特定の自治体が定めた所定の高確率値以上及び所定の低確率値以下を除く中間確率値の被保険者に対してのみ、又は予め定めた所定数の中間確率値の被保険者に対してのみ、勧奨通知対象リストを作成することとなる。 The fourth step is performed only for the insured person of intermediate probability value excluding a predetermined high probability value finally determined by a specific local government and not more than a predetermined low probability value, or a predetermined number of intermediate probabilities determined in advance Only for the valued insured, the recommendation notification target list will be created.
次に、大都市自治体の被保険者の受診確率を求める方法について説明する。大都市の場合、小都市とは逆にデータ数が多く、性能の高いモデルを構築するのに十分なサンプルを有しており、かつ特に地方の小都市住民の行動パターンがノイズとなる可能性がある。そのため、大都市自治体の被保険者の受診確率を求める際の基礎データは、当該自治体のみのデータとする。 Next, a method of obtaining the consultation probability of an insured person in a metropolitan area will be described. In the case of a large city, contrary to a small city, the number of data is large, and it has sufficient samples to build a high-performance model, and in particular, the behavior pattern of small city residents in rural areas may become noise. There is. Therefore, the basic data when obtaining the consultation probability of the insured person of the metropolitan area is the data only for the local government.
処理プログラム２は、第一ステップとして、健診データベース１に蓄積されている当該大都市自治体のデータのみを対象として、上記小都市の場合と同様に、この中から過去６年分のデータを、教師データセットを作成する目的で抽出し、直近の１年分のデータを、健診受診の有無（正解ラベルデータ）に加工し、正解ラベルデータ以外の５年分のデータを、図４に示すような４４個の説明変数を持つデータに加工する。その後の処理は、上記小都市の場合と同じであるため、説明は省略する。 As the first step, the processing program 2 targets only the data of the large city municipality accumulated in the health checkup database 1, and, as in the case of the above-mentioned small city, the data for the past six years from this Extracted for the purpose of creating a teacher data set, the data for the last 1 year is processed into the presence or absence of medical checkup (correct label data), and data for 5 years other than the correct label data is shown in FIG. 4 Process it into data with such 44 explanatory variables. The subsequent processing is the same as in the case of the above-mentioned small city, so the description will be omitted.
なお、上記した第一実施形態では、過去６年分の健診データを用いて計算したが、３年以上のデータがあれば十分予測可能である。また、ランダムフォレストのデータ抽出数、決定木数等の各種パラメータは、確度を高める目的の範囲内で適宜に決定しても良い。 In addition, although it calculated using the medical examination data for the past six years in 1st embodiment mentioned above, if there exist data of three years or more, it can fully predict. In addition, various parameters such as the number of data extractions of the random forest and the number of decision trees may be appropriately determined within the range intended to improve the accuracy.
さらに、上記した第一実施形態では、モデルを作成するアルゴリズムとしてランダムフォレストを利用した例を示したが、これに限らずディープラーニングであっても良い。その場合、誤差関数に交差エントロピーを用い、最適化学習にＡｄａｍを用いて個々の被保険者の受診確率を算出するのが好ましいが、交差エントロピー以外の損失関数（例えば、二乗誤差やヒンジ損失関数等）、Ａｄａｍ以外の他の勾配降下法等の最適化アルゴリズム（例えば、Ｎｅｓｔｅｒｏｖ ａｃｃｅｌｅｒａｔｅｄ ｇｒａｄｉｅｎｔ、ＡｄａｇｒａｄやＡｄａｄｅｌｔａ等）を用いても良く、特に限定されない。 Furthermore, although the example which used the random forest as an algorithm which produces a model was shown in above-mentioned 1st embodiment, not only this but deep learning may be sufficient. In that case, it is preferable to use the cross entropy for the error function and Adam for the optimization learning to calculate the visit probability of each insured person, but loss functions other than the cross entropy (for example, square errors and hinge loss functions) Other optimization algorithms such as gradient descent other than Adam, etc. (eg, Nesterov accelerated gradient, Adagrad, Adadelta, etc.) may be used without particular limitation.
次に、本発明の第二実施形態に係る処理プログラム２について説明する。
処理プログラム２は、データ数の多い大都市の自治体もデータ数の少ない小都市の自治体も、同じ方法で被保険者の受診確率を求めるものである。
Next, the processing program 2 according to the second embodiment of the present invention will be described.
The processing program 2 is for the local government of a big city with a large number of data and the local government of a small city with a small number of data to obtain the consultation probability of the insured person in the same way.
図７は、その処理フローの概要を示す説明図である。全体の概略構成は、まず複数自治体のデータを用いたモデルを作成し、確率計算を必要とする特定の自治体のデータを用いて当該モデルの一部を修正して、特定自治体の被保険者全ての計算に用いる最終モデルを確定するものである。 FIG. 7 is an explanatory view showing an outline of the process flow. The whole outline configuration first creates a model using data of multiple municipalities, corrects a part of the model using data of a specific municipality that requires probability calculation, and all insured persons of a specific municipality The final model to be used for the calculation of
処理プログラム２は、第一ステップとして、健診データベース１に蓄積されている複数自治体のデータを対象として、この中から過去６年分のデータを、教師データセットを作成する目的で抽出し、直近の１年分のデータを、健診受診の有無（正解ラベルデータ）に加工し、正解ラベルデータ以外の５年分のデータを、図４に示すような４４個の説明変数を持つデータに加工する。 As a first step, the processing program 2 extracts data for the past six years from among the data of a plurality of local governments accumulated in the medical checkup database 1 for the purpose of creating a teacher data set. Data for 1 year is processed to check for medical checkup (correct label data), and data for 5 years other than correct label data is processed to data with 44 explanatory variables as shown in Figure 4 Do.
図８に示すとおり、第一ステップでのアルゴリズムはディープラーニングであり、誤差関数に交差エントロピーを、最適化学習にＡｄａｍを用い、教師データセットに基づいて、複数のレイヤーに対応する複数の重みＷ１、Ｗ２、Ｗ３を学習させて第一特徴量を作成する。 As shown in FIG. 8, the algorithm in the first step is deep learning, and using the cross entropy for the error function, using Adam for the optimization learning, and based on the teacher data set, a plurality of weights W1 corresponding to a plurality of layers , W2 and W3 to learn the first feature.
図９に示すとおり、処理プログラム２は、第二ステップとして、確率計算する特定の自治体のデータのみを抽出し、アルゴリズムにトランスファーラーニング（転移学習）を用いて第一ステップで作成した重みＷ１、Ｗ２はそのままとし、Ｗ３のみを調整した第二特徴量を作成する。トランスファーラーニングは、誤差関数に交差エントロピーを用い、最適化学習にＡｄａｇｒａｄを用いている。なお、ディープラーニングやトランスファーラーニングにおいて、交差エントロピー以外の損失関数（例えば、二乗誤差やヒンジ損失関数等）、Ａｄａｍ、Ａｄａｇｒａｄ以外の他の勾配降下法等の最適化アルゴリズム（例えば、Ｎｅｓｔｅｒｏｖ ａｃｃｅｌｅｒａｔｅｄ ｇｒａｄｉｅｎｔやＡｄａｄｅｌｔａ等）を用いても良く、特に限定されない。 As shown in FIG. 9, as the second step, the processing program 2 extracts only data of a specific municipality to be subjected to probability calculation, and weights W1 and W2 created in the first step using transfer learning (transfer learning) as an algorithm. Is left as it is, and a second feature amount in which only W3 is adjusted is created. Transfer learning uses cross entropy for the error function and Adagrad for optimization learning. In deep learning and transfer learning, optimization algorithms such as loss functions other than cross entropy (for example, square error, hinge loss function, etc.) and gradient descent methods other than Adam and Adagrad (for example, Nesterov accelerated gradient or Adadelta Etc.) may be used, and it is not particularly limited.
処理プログラム２は、第三ステップとして、上記の第二特徴量に基づいたモデルに基づいて、確率計算する特定の自治体の被保険者全員の受診確率を計算する。 As a third step, the processing program 2 calculates the visit probability of all the insured persons of a specific local government who performs probability calculation based on the model based on the above-mentioned second feature value.
処理プログラム２は、第四ステップとして、第一実施形態と同様の方法により、特定の自治体が定めた所定の高確率値以上及び所定の低確率値以下を除く中間確率値の被保険者に対してのみ、又は予め定めた所定数の中間確率値の被保険者に対してのみ、勧奨通知対象リストを作成する。 As the fourth step, the processing program 2 applies to the insured person of intermediate probability value excluding the predetermined high probability value or more and the predetermined low probability value or less determined by a specific local government by the same method as the first embodiment. The recommendation notification target list is created only for the insured person who has only a predetermined number of intermediate probability values or a predetermined number of intermediate probability values.
なお、上記した第二実施形態では、過去６年分の健診データを用いて計算したが、３年以上のデータがあれば十分予測可能である。また、ディープラーニングにおけるレイヤー数等の各種パラメータは、確度を高める目的の範囲内で適宜に決定しても良い。 In addition, although it calculated using the medical examination data for the past six years in 2nd embodiment mentioned above, if there are data of three years or more, it can fully predict. In addition, various parameters such as the number of layers in deep learning may be appropriately determined within the range intended to improve the accuracy.
以上、本願発明の各実施形態について説明したが、出願人による多くの自治体及び健保組合等を対象にした実証実験によれば、予測値と実測値との乖離は、データ数の多い大都市又は規模の大きい健保組合等では、第二実施形態のシステム≒第一実施形態のシステム（単独自治体、単独健保組合等のデータを使用）≧第一実施形態のシステム（全ての自治体、又は全ての健保組合等のデータを使用）の傾向が強く、データ数の少ない小都市又は規模の小さい健保組合等では、第二実施形態のシステム≧第一実施形態のシステム（全ての自治体、又は全ての健保組合等のデータを使用）＞第一実施形態のシステム（単独自治体、単独健保組合等のデータを使用）となる傾向が強かった。 As mentioned above, although each embodiment of this invention was described, according to the proof experiment for many local governments, a health insurance association, etc. by applicants, the gap between a predicted value and a measured value is a big city with a large number of data or In a large scale health insurance association etc., the system of the second embodiment ≒ the system of the first embodiment (using data of an independent municipality, an independent health insurance association etc.) ≧ the system of the first embodiment (all municipalities or all health insurance The system of the second embodiment 健 the system of the first embodiment (all municipalities or all the health insurance unions) in a small city or small-sized health insurance association where the tendency of the data of unions and the like is strong and the number of data is small Etc.)> The tendency of the system of the first embodiment (using data of an independent local government, an independent health insurance association, etc.) was strong.
上記した実証結果は、あくまで全体としての傾向であって、それぞれの自治体やそれぞれの健保組合が有する特異性に基づき、実証実験の積重ねにより、いずれのシステムを利用するかを適宜に選択決定すれば良い。 The above-mentioned demonstration result is a tendency as a whole to the last, and based on the peculiarities of each local government and each health insurance association, by selecting and deciding which system to use appropriately by stacking the demonstration experiments good.
図１０は、都道府県ベースで実施した実証実験の結果を示す図であり、データ数が数百レベルの保険者（自治体Ａ、自治体Ｂ、自治体Ｃ、自治体Ｄ、自治体Ｅ）について実施した結果である。 FIG. 10 is a diagram showing the results of a demonstration experiment conducted on a prefectural basis, and the results of the experiments conducted on insurers with several hundreds of data (local government A, local government B, local government C, local government D, local government E) is there.
各県それぞれ３つの棒グラフで示されている受診率は、左の棒グラフが勧奨通知を行った者の実際の受診率であり、中央の棒グラフが当該県のデータのみを用いたモデルで確率計算した予測値であり、右の棒グラフが複数保険者のデータを用いたモデルに更に当該保険者のデータを用いて転移学習させたモデルで確率計算した予測値である。 The reception rates shown in the three bar graphs for each prefecture are the actual consultation rates for those who gave recommendations on the left bar graph, and the central bar graph was probability calculated with a model using only the data for that prefecture It is a predicted value, and the bar graph on the right is a predicted value calculated by probability with a model in which transfer learning is performed using data of the insurer on the model using data on multiple insurers.
図１０から明らかなとおり、データ数の少ない５つの保険者全てにおいて、転移学習したモデルで確率計算した予測値の方が実測値に近い値となっていることから、当該方法によって選定した被保険者に対してのみ勧奨通知を行えばよく、自治体の予算削減、人員削減に貢献することができる。 As apparent from FIG. 10, in all five insurers with a small number of data, the predicted value calculated by probability calculation using the transfer learning model is closer to the actual measurement value, so the insured selected by the method is It is sufficient to notify the recommendation only to those who are responsible for reducing the budget of the local government and reducing the workforce.
図１１は、データ数が数万レベルの３つの保険者（自治体Ｘ、自治体Ｙ、自治体Ｚ）について実施した結果である。各保険者それぞれ３つの棒グラフで示されている受診率は、図１０と同じである。 FIG. 11 shows the results of implementation for three insurers (local government X, local government Y, local government Z) having tens of thousands of data numbers. The consultation rates shown in the three bar graphs for each insurer are the same as in FIG.
図１１から明らかなとおり、データ数の多い３つの保険者においても転移学習したモデルで確率計算した予測値の方が実測値に近い値となっているが、Ｘ保険者及びＹ保険者では、単独データを用いた場合と大きな差はないことが理解できる。 As is clear from FIG. 11, in the three insurers with a large number of data, the predicted value calculated by probability calculation using the transfer learning model is closer to the actual measurement value, but for the X insurer and the Y insurer, It can be understood that there is no significant difference from the case of using single data.
図１２は、ある保険者における予測対象者の受診率予測値と、受診勧奨後の実際の受診率の差を示す図である。当該図より明らかなとおり、受診率予測値が３０〜７０％の対象者に対する勧奨効果は明らかに高くなり、予測対象者の平均勧奨効果は１２．６％であった。 FIG. 12 is a diagram showing the difference between the predicted visit rate of the person to be predicted and the actual visit rate after the recommendation for a visit by a certain insurer. As is clear from the figure, the recommendation effect for subjects with a predicted visit rate of 30 to 70% was clearly higher, and the average recommendation effect for those to be predicted was 12.6%.
上記した各実施形態は、国民健康保険の運営主体である自治体に対するサービスとして実施するものを説明したが、健康保険組合等に対するサービスとしても同様に実施可能である。その場合、複数の健保組合のデータを健診データベース１に蓄積し、データ数の多い大規模健保組合とデータ数の少ない小規模健保組合にそれぞれ応じた処理を行えば良い。 Although each embodiment mentioned above demonstrated what was implemented as a service to the local government which is a main body of management of National Health Insurance, it can implement similarly as a service to a health insurance association etc. In that case, data of a plurality of health insurance associations may be stored in the health checkup database 1, and processing may be performed according to the large scale health insurance association with a large number of data and the small scale health insurance association with a small number of data.
以上のとおり、本願発明によれば、勧奨通知を発送しても健康診断を受診しない可能性の高い者の被保険者グループと、勧奨通知を発送しなくても健康診断を受診する可能性の高い者の被保険者グループを高い確度で特定することができ、これらの２つのグループ以外に属する者、即ち、勧奨通知を発送することによって受診可能性が高くなる被保険者グループを特定することで、勧奨通知に係る事務作業を効率化し、予算及び人員の適正化を図ることができる。また、結果として健康診断を受診する被保険者が増えることで、医療費の削減にも繋げることができる。 As described above, according to the present invention, the insured person group of a person who is likely not to receive a medical examination even if the recommendation notice is sent, and the possibility to receive the medical examination without sending the recommendation notice. Identify high-income insured groups with a high degree of certainty, and identify those who belong to other than these two groups, ie, those insured groups who are more likely to receive a medical examination by sending out a notification of recommendation In this way, it is possible to streamline office work related to the recommendation notice and to optimize budgets and personnel. In addition, as a result, by increasing the number of insured persons who receive a medical examination, it can be connected to the reduction of medical expenses.
１ 健診データベース
２ 処理プログラム
３ サーバー
1 medical checkup database 2 processing program 3 server
Claims (14)
確率計算の基礎データとなる母集団を、複数又は単独の自治体、或いは複数又は単独の健保組合等が保有する過去ｎ年（ｎは３以上の整数）間において被保険者が受診した健康診断の健診データとし、
前記健診データを基に、コンピューターが被保険者の受診行動を予測する処理プログラムを含み、
前記コンピューターは、前記処理プログラムに従い、
第一ステップとして、
全ての前記健診データから過去ｍ年（ｍは３以上の整数、ｍ≦ｎ）分のデータを抽出し、
直近の１年分を除いたｍ−１年分のデータを、ｘ個の説明変数を持つデータに加工し、
第二ステップとして、
当該ｍ−１年分の前記健診データ及び直近１年分の健診受診の有無を教師データとして用いて学習を行い、
第三ステップとして、
前記学習により構築したモデルによって、特定の単独自治体又は特定の単独健保組合に加入する個々の被保険者の受診確率を算出する、
ことを特徴とする被保険者の健康診断受診確率計算方法。 It is a calculation method of health checkup consultation probability of the insured person,
Health checkup of the insured at a medical checkup in the past n years (n is an integer of 3 or more) held by multiple or single local governments or multiple or single health insurance associations etc. As medical examination data,
Including a processing program in which a computer predicts the health checkup of the insured based on the health checkup data;
The computer follows the processing program,
As a first step
Extract data for the past m years (m is an integer of 3 or more, m ≦ n) from all the medical checkup data,
Process the m-1 year's worth of data excluding the most recent 1 year's worth into data with x explanatory variables,
As a second step
Conduct learning using the health checkup data for the m-1 years and the presence or absence of health checkup visits for the last 1 year as teacher data,
As the third step
According to the model constructed by the learning, the consultation probability of an individual insured who joins a specific independent municipality or a specific independent health insurance association is calculated.
Calculation method for health checkup consultation probability of the insured person characterized in that.
ことを特徴とする請求項１に記載の被保険者の健康診断受診確率計算方法。 The algorithm of the processing program is a random forest, and a samples are randomly extracted from the medical checkup data held by multiple or single local governments, or multiple or single health insurance associations, etc., and x descriptions Randomly extract b explanatory variables from the variables to create y decision trees, and calculate the visit probability of the individual insured by an ensemble of all y decision trees.
The method according to claim 1, wherein the insured person's health checkup probability is calculated.
ことを特徴とする請求項１に記載の被保険者の健康診断受診確率計算方法。 The algorithm of the processing program is deep learning, and based on the health checkup data held by multiple or single local governments, or multiple or single health insurance associations etc., using cross entropy as an error function, Adam for optimization learning Calculate the consultation probability of the individual insured using
The method according to claim 1, wherein the insured person's health checkup probability is calculated.
ことを特徴とする請求項１ないし３のいずれか１項に記載の被保険者の健康診断受診確率計算方法。 The x explanatory variables are 43 or more items created by processing the date of birth, sex, height, weight, various inquiry results, various test values, etc.
The method according to any one of claims 1 to 3, wherein the insured person's health examination consultation probability calculation method.
確率計算の基礎データとなる母集団を、複数の自治体又は複数の健保組合等が保有する過去ｎ年（ｎは３以上の整数）間において被保険者が受診した健康診断の健診データとし、
前記健診データを基に、コンピューターが被保険者の受診行動を予測する処理プログラムを含み、
前記コンピューターは、前記処理プログラムに従い、
第一ステップとして、
複数の保険者の前記健診データから過去ｍ年（ｍは３以上の整数、ｍ≦ｎ）分のデータを抽出し、
直近の１年分を除いたｍ−１年分のデータをｘ個の説明変数を持つデータに加工し、
アルゴリズムはディープラーニングであり、
当該ｍ−１年分のデータ及び直近１年での健診受診の有無を教師データとして用いることで、複数のレイヤーに対応する複数の重みＷｉを学習して第一特徴量を作成し、
第二ステップとして、
特定の単独自治体又は特定の単独健保組合等の前記健診データのみを抽出し、アルゴリズムにトランスファーラーニングを用いて前記第一ステップで作成した前記重みＷｉの一部を調整した第二特徴量を作成し、
第三ステップとして、
前記第二特徴量に基づいたモデルによって、前記特定の単独自治体又は前記特定の単独健保組合等に加入する個々の被保険者の受診確率を算出する、
ことを特徴とすることを特徴とする被保険者の健康診断受診確率計算方法。 It is a calculation method of health checkup consultation probability of the insured person,
We assume population serving as basic data of probability calculation as medical examination data of medical examination which insured person consulted in the past n years (n is integer more than 3) which plural municipalities or plural health insurance associations hold,
Including a processing program in which a computer predicts the health checkup of the insured based on the health checkup data;
The computer follows the processing program,
As a first step
Extracting data for the past m years (m is an integer of 3 or more, m ≦ n) from the health checkup data of a plurality of insurers,
Process the m-1 year's worth of data excluding the most recent 1 year's worth into data with x explanatory variables,
The algorithm is deep learning,
By using the data for m-1 years and the presence or absence of the medical checkup in the most recent year as teacher data, a plurality of weights Wi corresponding to a plurality of layers are learned to create a first feature value.
As a second step
Extract only the medical checkup data of a specific independent municipality or a specific independent health insurance association, etc., and create a second feature that adjusts part of the weight Wi created in the first step using transfer learning as an algorithm And
As the third step
According to the model based on the second feature amount, the consultation probability of each insured person who joins the specific independent local government or the specific independent health insurance association is calculated.
A method of calculating a health checkup examination probability of an insured person characterized in that
前記トランスファーラーニングは、誤差関数に交差エントロピーを用い、最適化学習にＡｄａｇｒａｄを用いている、
ことを特徴とする請求項５に記載の被保険者の健康診断受診確率計算方法。 The deep learning uses cross entropy for the error function and Adam for optimization learning.
The transfer learning uses cross entropy for the error function and Adagrad for optimization learning.
The method according to claim 5, wherein the insured person's health checkup probability is calculated.
ことを特徴とする請求項５又は６に記載の被保険者の健康診断受診確率計算方法。 The x explanatory variables are 43 or more items created by processing the date of birth, sex, height, weight, various inquiry results, various test values, etc.
7. The method according to claim 5 or 6, wherein the health checkup probability of the insured person.
複数又は単独の自治体、或いは複数又は単独の健保組合等が保有する被保険者が受診した過去ｎ年（ｎは３以上の整数）の健康診断の結果が蓄積された健診データベースと、
前記健診データベースに蓄積されているデータを基に被保険者の受診行動を予測する処理プログラムが記録されているサーバーと、を備え、
前記サーバーは、前記処理プログラムに従い、
第一ステップとして、
前記健診データベースから複数又は単独の自治体、或いは複数又は単独の健保組合等の過去ｍ年（ｍは３以上の整数、ｍ≦ｎ）分のデータを抽出し、
直近の１年分を除いたｍ−１年分のデータを、ｘ個の説明変数を持つデータに加工し、
第二ステップとして、
当該ｍ−１年分のデータ及び直近１年での健診受診の有無を教師データとして用いて学習を行い、
第三ステップとして、
前記学習により構築したモデルによって、特定の自治体又は特定の健保組合等に加入する個々の被保険者の受診確率を算出し、
第四ステップとして、
前記特定の自治体又は前記特定の健保組合等が定めた所定の高確率値以上及び所定の低確率値以下を除く中間確率値の被保険者に対してのみ、又は予め定めた所定数の中間確率値の被保険者に対してのみ、勧奨通知対象リストを作成する、
ことを特徴とする健診勧奨通知支援システム。 A health check recommendation notification support system that efficiently performs, based on consultation probability calculation, recommendation notifications to an insured person who encourages a medical checkup visit conducted by a local government or a health insurance association etc.
A medical checkup database in which results of medical examinations of the past n years (n is an integer of 3 or more) visited by multiple insured persons owned by multiple or single local governments or multiple or single health insurance associations etc.
A server in which a processing program for predicting an insured's examination behavior is recorded based on the data stored in the medical checkup database;
The server follows the processing program
As a first step
Extract data for the past m years (m is an integer of 3 or more, m ≦ n), such as multiple or single local governments, or multiple or single health insurance associations, from the health checkup database,
Process the m-1 year's worth of data excluding the most recent 1 year's worth into data with x explanatory variables,
As a second step
We perform learning using the data for the m-1 year and the presence or absence of medical examination consultation in the latest one year as teacher data,
As the third step
Calculate the consultation probability of each insured who joins a specific municipality or a specific health insurance association etc. by the model constructed by the learning,
As the fourth step
Only for the insured person of intermediate probability value excluding the predetermined high probability value or more and the predetermined low probability value or less defined by the specific local government or the specific health insurance association etc., or a predetermined number of intermediate probabilities determined in advance Create a list of recommendations for coverage only for value insureds,
Health check recommendation notification support system characterized by.
ことを特徴とする請求項８に記載の健診勧奨通知支援システム。 The algorithm of the processing program is a random forest, and a samples are randomly extracted from data stored in the medical checkup database such as multiple or single local governments, or multiple or single health insurance associations, Randomly extract b explanatory variables from x explanatory variables to create y decision trees, and calculate a consultation probability of the individual insured by an ensemble of all y decision trees.
The health check recommendation notification support system according to claim 8, wherein
ことを特徴とする請求項８に記載の健診勧奨通知支援システム。 The algorithm of the processing program is deep learning, and based on data stored in the medical checkup database such as one or more local governments or multiple or single health insurance associations stored in the medical checkup database, Use the cross-entropy for the error function and Adam for the optimization learning to calculate the visit probability of the individual insured
The health check recommendation notification support system according to claim 8, wherein
ことを特徴とする請求項８ないし１０のいずれか１項に記載の健診勧奨通知支援システム。 The x explanatory variables are 43 or more items created by processing the date of birth, sex, height, weight, various inquiry results, various test values, etc.
The health checkup recommendation notification support system according to any one of claims 8 to 10, characterized in that
複数の自治体又は複数の健保組合等が保有する被保険者が受診した過去ｎ年（ｎは３以上の整数）の健康診断の結果が蓄積された健診データベースと、
前記健診データベースに蓄積されているデータを基に被保険者の受診行動を予測する処理プログラムが記録されているサーバーと、を備え、
前記サーバーは、前記処理プログラムに従い、
第一ステップとして、
前記健診データベースから過去ｍ年（ｍは３以上の整数、ｍ≦ｎ）分のデータを抽出し、
直近の１年分を除いたｍ−１年分のデータの中からｘ個の説明変数を持つデータに加工し、
アルゴリズムはディープラーニングであり、
当該ｍ−１年分のデータ及び直近１年での健診受診の有無を教師データとして用いることで、複数のレイヤーに対応する複数の重みＷｉを学習して第一特徴量を作成し、
第二ステップとして、
特定の自治体又は特定の健保組合等のデータのみを抽出し、アルゴリズムにトランスファーラーニングを用いて前記第一ステップで作成した前記重みＷｉの一部を調整した第二特徴量を作成し、
第三ステップとして、
前記第二特徴量に基づいたモデルによって、前記特定の自治体又は前記特定の健保組合等に加入する個々の被保険者の受診確率を算出し、
第四ステップとして、
前記特定の自治体又は前記特定の健保組合等が定めた所定の高確率値以上及び所定の低確率値以下を除く中間確率値の被保険者に対してのみ、又は予め定めた所定数の中間確率値の被保険者に対してのみ、勧奨通知対象リストを作成する、
ことを特徴とする健診勧奨通知支援システム。 A health check recommendation notification support system that efficiently performs, based on consultation probability calculation, recommendation notifications to an insured person who encourages a medical checkup visit conducted by a local government or a health insurance association etc.
A medical checkup database in which results of medical examinations of the past n years (n is an integer of 3 or more) visited by an insured person owned by a plurality of local governments or a plurality of health insurance associations etc.
A server in which a processing program for predicting an insured's examination behavior is recorded based on the data stored in the medical checkup database;
The server follows the processing program
As a first step
Data for the past m years (m is an integer of 3 or more, m ≦ n) is extracted from the medical checkup database,
Process into data with x explanatory variables out of m-1 year's worth of data excluding the last 1 year's worth,
The algorithm is deep learning,
By using the data for m-1 years and the presence or absence of the medical checkup in the most recent year as teacher data, a plurality of weights Wi corresponding to a plurality of layers are learned to create a first feature value.
As a second step
Only data of a specific municipality or a specific health insurance association etc. is extracted, and a second feature amount is generated by adjusting a part of the weight Wi created in the first step using transfer learning as an algorithm,
As the third step
According to the model based on the second feature amount, the consultation probability of the individual insured who joins the specific municipality or the specific health insurance association is calculated.
As the fourth step
Only for the insured person of intermediate probability value excluding the predetermined high probability value or more and the predetermined low probability value or less defined by the specific local government or the specific health insurance association etc., or a predetermined number of intermediate probabilities determined in advance Create a list of recommendations for coverage only for value insureds,
Health check recommendation notification support system characterized by.
前記トランスファーラーニングは、誤差関数に交差エントロピーを用い、最適化学習にＡｄａｇｒａｄを用いている、
ことを特徴とする請求項１２に記載の健診勧奨通知支援システム。 The deep learning uses cross entropy for the error function and Adam for optimization learning.
The transfer learning uses cross entropy for the error function and Adagrad for optimization learning.
The medical checkup recommendation notification support system according to claim 12, characterized in that:
ことを特徴とする請求項１２又は１３に記載の健診勧奨通知支援システム。 The x explanatory variables are 43 or more items created by processing the date of birth, sex, height, weight, various inquiry results, various test values, etc.
The medical examination recommendation notice support system according to claim 12 or 13, characterized in that:
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| JP2018203585A JP6548243B1 (en) | 2018-10-30 | 2018-10-30 | Health checkup examination probability calculation method and medical checkup recommendation notice support system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| JP2018203585A JP6548243B1 (en) | 2018-10-30 | 2018-10-30 | Health checkup examination probability calculation method and medical checkup recommendation notice support system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| JP6548243B1 true JP6548243B1 (en) | 2019-07-24 |
| JP2020071562A JP2020071562A (en) | 2020-05-07 |
Family
ID=67390351
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| JP2018203585A Active JP6548243B1 (en) | 2018-10-30 | 2018-10-30 | Health checkup examination probability calculation method and medical checkup recommendation notice support system |
Country Status (1)
| Country | Link |
|---|---|
| JP (1) | JP6548243B1 (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2021082009A (en) * | 2019-11-19 | 2021-05-27 | 株式会社日立製作所 | Information processing device, and information processing method |
Family Cites Families (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130159023A1 (en) * | 2011-12-16 | 2013-06-20 | Neela SRINIVAS | System and method for evidence based differential analysis and incentives based heal thcare policy |
| JP5888063B2 (en) * | 2012-03-29 | 2016-03-16 | 日本電気株式会社 | Health guidance target person selection support device |
| SG11201406704RA (en) * | 2012-04-26 | 2014-11-27 | Nec Corp | Health guidance receiver selection condition generation support device |
| AU2014239852A1 (en) * | 2013-03-15 | 2015-11-05 | The Cleveland Clinic Foundation | Self-evolving predictive model |
| WO2016207862A1 (en) * | 2015-06-25 | 2016-12-29 | Timothy Burton | Methods and systems using mathematical analysis and machine learning to diagnose disease |
| US20190156240A1 (en) * | 2016-04-18 | 2019-05-23 | Nippon Telegraph And Telephone Corporation | Learning apparatus, learning method, and recording medium |
-
2018
- 2018-10-30 JP JP2018203585A patent/JP6548243B1/en active Active
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2021082009A (en) * | 2019-11-19 | 2021-05-27 | 株式会社日立製作所 | Information processing device, and information processing method |
| JP7053557B2 (en) | 2019-11-19 | 2022-04-12 | 株式会社日立製作所 | Information processing equipment, information processing method |
Also Published As
| Publication number | Publication date |
|---|---|
| JP2020071562A (en) | 2020-05-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Wylie et al. | Patterns of chlamydia and gonorrhea infection in sexual networks in Manitoba, Canada | |
| US10943676B2 (en) | Healthcare information technology system for predicting or preventing readmissions | |
| US20200315518A1 (en) | Apparatus for processing data for predicting dementia through machine learning, method thereof, and recording medium storing the same | |
| Wiens et al. | Learning data-driven patient risk stratification models for Clostridium difficile | |
| KR20170061222A (en) | The method for prediction health data value through generation of health data pattern and the apparatus thereof | |
| Zmiri et al. | Classification of patients by severity grades during triage in the emergency department using data mining methods | |
| US20210174968A1 (en) | Visualization of Social Determinants of Health | |
| CN111696673A (en) | Mental health management model system and method | |
| Jeffery et al. | Predicting 1-year disability and mortality of injured older adults | |
| JP6548243B1 (en) | Health checkup examination probability calculation method and medical checkup recommendation notice support system | |
| Chand et al. | Learning from machines to close the gap between funding and expenditure in the Australian National Disability Insurance Scheme | |
| US11537888B2 (en) | Systems and methods for predicting pain level | |
| Cheng et al. | Optimal vaccination strategies for COVID-19 based on dynamical social networks with real-time updating | |
| Kaushik et al. | Disease management: clustering–based disease prediction | |
| KR20180061551A (en) | Method and apparatus of similar case clustering for future health trends prediction | |
| Farooq et al. | A novel cardiovascular decision support framework for effective clinical risk assessment | |
| Daniel et al. | Recurrent Neural Networks for Feature Extraction from Dengue Fever | |
| Tekieh et al. | Analysing healthcare coverage with data mining techniques | |
| Kumar et al. | A Clinical Data Analysis Based Diagnostic Systems for Heart Disease Prediction Using Ensemble Method | |
| Avati et al. | Predicting inpatient discharge prioritization with electronic health records | |
| Fatma et al. | Outpatient Diversion using Real-time Length-of-Stay Predictions | |
| Shabbeer et al. | Prediction of Sudden Health Crises Owing to Congestive Heart Failure with Deep Learning Models. | |
| Kapoor et al. | Development of prediction model using machine learning for vector borne diseases in India | |
| Kapoor et al. | Machine Learning Based Classification Algorithm for Classification of Dengue (Dengue Fever-DF, Dengue Harmonic Fever-DHF, Serve Dengue-SD) | |
| Intia et al. | Prediction of Agoraphobia Disease Based on Machine Learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| A621 | Written request for application examination |
Free format text: JAPANESE INTERMEDIATE CODE: A621 Effective date: 20181030 |
|
| A871 | Explanation of circumstances concerning accelerated examination |
Free format text: JAPANESE INTERMEDIATE CODE: A871 Effective date: 20181030 |
|
| A521 | Request for written amendment filed |
Free format text: JAPANESE INTERMEDIATE CODE: A523 Effective date: 20181030 |
|
| A975 | Report on accelerated examination |
Free format text: JAPANESE INTERMEDIATE CODE: A971005 Effective date: 20181115 |
|
| A977 | Report on retrieval |
Free format text: JAPANESE INTERMEDIATE CODE: A971007 Effective date: 20190314 |
|
| A131 | Notification of reasons for refusal |
Free format text: JAPANESE INTERMEDIATE CODE: A131 Effective date: 20190325 |
|
| A521 | Request for written amendment filed |
Free format text: JAPANESE INTERMEDIATE CODE: A523 Effective date: 20190401 |
|
| A131 | Notification of reasons for refusal |
Free format text: JAPANESE INTERMEDIATE CODE: A131 Effective date: 20190530 |
|
| A521 | Request for written amendment filed |
Free format text: JAPANESE INTERMEDIATE CODE: A523 Effective date: 20190530 |
|
| TRDD | Decision of grant or rejection written | ||
| A01 | Written decision to grant a patent or to grant a registration (utility model) |
Free format text: JAPANESE INTERMEDIATE CODE: A01 Effective date: 20190617 |
|
| A61 | First payment of annual fees (during grant procedure) |
Free format text: JAPANESE INTERMEDIATE CODE: A61 Effective date: 20190619 |
|
| R150 | Certificate of patent or registration of utility model |
Ref document number: 6548243 Country of ref document: JP Free format text: JAPANESE INTERMEDIATE CODE: R150 |
|
| S531 | Written request for registration of change of domicile |
Free format text: JAPANESE INTERMEDIATE CODE: R313531 |
|
| R350 | Written notification of registration of transfer |
Free format text: JAPANESE INTERMEDIATE CODE: R350 |
|
| R250 | Receipt of annual fees |
Free format text: JAPANESE INTERMEDIATE CODE: R250 |