JP2009205464A - Medical information processor, medical information processing method, and medical information processing program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a medical information processor for a modeling change corresponding to the level of the state or the condition of a disease or the detail sorting of the disease by automatically clustering and modeling large amounts of data with personal differences taken into sufficient consideration, and simultaneously achieving the more flexible modeling of time series, and the estimation of precise probability density. <P>SOLUTION: The medical information processor 10 classifies the group of medical information in which check items related with a plurality of candidates are prepared in time series into clusters based on the check items. The medical information processor 10 estimates multi-dimensional time series distribution in the group of a plurality of medical information related with a plurality of candidates as probability density functions in each classified cluster, and estimates a representative value as a model from the estimated probability density functions in the clusters. The medical information processor 10 moves the optimal time axis about all samples in each cluster by using the representative value as a reference, and converges the results. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a medical information processing apparatus, a medical information processing method, and a medical information processing program, and in particular, to extract statistical information regarding probabilistic business from a large amount of data by combining time series analysis and statistical methods. The present invention relates to a medical information processing apparatus, a medical information processing method, and a medical information processing program.

  From April 2008, due to the revision of the Health Insurance Law, medical insurers are required to undergo medical examinations and health guidance. In particular, health guidance is provided to medical insurers in line with prevention of metabolic syndrome. In providing diagnosis and health guidance to each individual, it is necessary to provide meticulous support and guidance tailored to the individual characteristics and circumstances.

Such detailed support and guidance required considerable costs. It was also difficult to understand and manage the situation for each individual.
In contrast, in recent years, attempts have been made to provide services for individuals by using information technology. In particular, there is an interest in medical information systems for supporting health insurance societies, public health nurses, doctors, and individuals for the implementation of specific medical examinations. There are also sites that provide health guidance tailored to each individual registered in advance via a Web system and e-mail.

  In these conventional technologies, how to efficiently input inspection data, how to display them in time series, or ease of use as a site is a major feature, and the examination results There was no mechanism for analyzing and interrogating data, calculating risks, and numerically managing and supporting health guidance courses.

  As such an approach, Non-Patent Document 1 discloses predicting the possibility of development of lifestyle-related diseases from personal checkup information. According to this, data from epidemiological surveys conducted in Kuyama City, Fukuoka Prefecture for about 40 years are used. From the disease risk calculation formula derived from this data and test data such as individual age, weight, blood pressure, exercise amount, electrocardiogram, cholesterol and blood sugar level, lifestyle-related diseases (cerebral infarction, ischemic heart disease) in the next 10 years , Diabetes, high blood pressure, etc.).

In addition, a system described in Patent Documents 1, 2, and 4 and a method for predicting the risk of developing lifestyle-related diseases described in Patent Document 3 have been proposed.
In the system of Patent Document 1, for clinical test numerical data items, there are different reference values for each laboratory and diagnostic decision values described in written documents, or test items predicted from side effects such as basic diseases, complications, treatments, etc. The normal range value and diagnosis decision value of each individual patient can be set arbitrarily. In the system, by inputting the clinical laboratory numerical data measured for the patient at the time of diagnosis, the above settings are used to obtain the clinical diagnostic evaluation value corresponding to the diagnostic severity based on the normal range value and the diagnostic decision value. Conversion processing is done.

  In the system of Patent Document 2, a living body model that reproduces the behavior of a living body by using a medical data input unit including patient test values and clinical findings, and a living body model that describes organs related to diabetes and the functions of the organs as mathematical models. A driving unit and a biological model unit that generates a patient-specific biological model by estimating a parameter set of the biological model based on the medical data. In addition, the system generates medical support information using a pathological condition analysis unit that analyzes the pathological condition of a patient's diabetes based on the parameter set of the generated model, and a diagnostic criterion that is determined for each analyzed pathological condition. A medical support information generating unit, and a medical support information output unit that outputs information obtained from the pathological condition analyzing unit and / or the medical support information generating unit are provided.

  Patent Document 3 proposes a method for estimating fluctuations in biochemical parameters selected from cholesterol, neutral lipids and blood glucose in blood by measuring the endotoxin concentration in the body fluid of healthy persons or healthy animals. A method for predicting the risk of developing lifestyle-related diseases has been proposed. Patent Document 3 discloses a biochemical parameter selected from cholesterol, neutral lipid, and blood glucose in blood by detecting or measuring the concentration of endotoxin derived from periodontal pathogens in the body fluid of a healthy person or healthy animal. A method for estimating the fluctuation of the lifespan is proposed, and a method for predicting the risk of developing lifestyle-related diseases by this method is proposed.

The diagnosis support system of Patent Document 4 includes a physiological data input unit, a diabetes disease risk analysis unit, a metabolic syndrome disease risk analysis unit, a diagnosis support information generation unit, a biological model generation unit, a disease state simulation unit, and a diagnosis support information output unit. Has functional blocks. The system analyzes the disease risk of diabetes and metabolic syndrome by the diabetes disease risk analysis unit and the metabolic syndrome disease risk analysis unit, respectively, and the diagnosis support information generation unit generates diagnosis support information from the analysis result. Has been.
JP 2004-227041 A Japanese Patent Laying-Open No. 2005-267042 Japanese Patent Laid-Open No. 2005-140618 JP 2006-304833 A Yutaka Kiyohara, “What is Hisayama Town Research”, [online], 2006, Hisayama Research Institute, Department of Environmental Medicine, Graduate School of Medicine, Kyushu University, [December 18, 2007], Internet <URL: http: // www.envmed.med.kyushu-u.ac.jp/about/index.html>

  However, in the above non-patent document 1, since the target number of people is about 2600 and the disease risk corresponding to individual lifestyle-related diseases is calculated, statistical information estimation for risk calculation is performed. As a whole, modeling according to the level of more detailed diseases and situations, and estimation considering individual differences have not yet been made. In addition, it is not yet common to use information system support in health guidance by public health nurses, registered dietitians, doctors, etc., and for detailed support corresponding to each individual's health condition and lifestyle, information Although it is recognized that there is a need to extract the optimal information to be prepared for the system, sufficient research results have not yet been obtained.

  In order to achieve detailed support and health guidance tailored to individual characteristics and circumstances, the time series such as test data is modeled more precisely, and statistical information obtained from a large amount of data is also created. Based on this, it is necessary to estimate the occurrence probability of the target event in more detail.

  The problem with the prior art in Patent Documents 1 to 4 is that the individual differences are not fully considered. When estimating the risk of a specific disease, the underlying model should be different depending on the symptom level and the detailed classification of the disease. It is not supported.

  The second is data variation processing. Examination data and data such as daily meals and exercise amount are highly dispersed and varied for each person's data. The prior art only estimates the overall probability density for these variations.

Thirdly, in the prior art, the temporal change model is simple such as ascending, descending, or average change pattern, and more detailed modeling is not possible.
Fourthly, even in the proposal of analyzing the inspection data and predicting the risk in the prior art, there is no proposal as to how to obtain the data that is the basis for such judgment. That is, the prior art has not proposed means for obtaining statistical knowledge for performing prediction from original data.

  The modeling method in the prior art is a model with coarse accuracy. Here, the coarse accuracy model (or probability estimation) is a test data that does not take into account temporal changes and individual differences. This is a rough modeling method for a small number of classes.

  In modeling by such a method or estimation of probability distribution (probability density function), although fluctuations due to various factors are included in the target data, it cannot be distinguished, so distribution It must be a broad estimation that estimates only the overall state of the.

  For example, if the changes over time are modeled without distinguishing samples at various stages from the state of health to the case of having a disease, only the overall distribution can be estimated. In addition, if all target samples are handled as a group without considering differences among individuals, it is necessary to estimate a probability distribution having a complicated shape, and thus cannot be expressed by a simple model. On the other hand, it is possible to estimate the probability distribution with a complicated model, but it is generally difficult to estimate with high accuracy from a model that contains fluctuations due to various factors unknown from the viewpoint of estimation accuracy. .

  The object of the present invention is to model changes in the degree of pathological conditions / symptoms and detailed classification of diseases by automatically modeling from a large amount of data by taking into account individual differences and modeling automatically. It is possible to provide a medical information processing method capable of simultaneously realizing more flexible time series modeling and precise probability density estimation.

  Further, since the object of the present invention is to perform clustering, the granularity setting enables more detailed modeling as compared with the prior art, and further, statistical analysis for predicting a subject's disease risk. It is to provide a medical information processing method capable of concretely showing how to obtain knowledge from original data.

  In addition, the second object of the present invention is to automatically change from a large amount of data into a model by clustering, taking into account individual differences, so that changes corresponding to the degree of disease state / symptom and the detailed classification of disease can be achieved. It is an object of the present invention to provide a medical information processing apparatus that can simultaneously model more flexible time-series modeling and precise probability density estimation.

  Further, since the object of the present invention is to perform clustering, the granularity setting enables more detailed modeling as compared with the prior art, and further, statistical analysis for predicting a subject's disease risk. It is to provide a medical information processing apparatus capable of specifically showing how to acquire knowledge from original data.

  The third object of the present invention is to model changes based on the degree of pathology / symptoms and detailed classification of diseases by automatically modeling from a large amount of data by taking into account individual differences. Another object of the present invention is to provide a medical information processing program that can simultaneously realize more flexible time-series modeling and precise probability density estimation.

  Further, since the object of the present invention is to perform clustering, the granularity setting enables more detailed modeling as compared with the prior art, and further, statistical analysis for predicting a subject's disease risk. It is to provide a medical information processing program capable of specifically showing how to acquire the knowledge from the original data.

  In order to achieve the above object, the invention according to claim 1 classifies a set of medical information to which data of a plurality of examination items relating to a plurality of subjects are given in time series into clusters according to the data of the examination items. A first step of estimating a multi-dimensional time-series distribution of a plurality of sets of medical information on a plurality of subjects as a probability density function in each classified cluster, and an estimation in the cluster A third step of determining a representative value to be a model from the obtained probability density function, and a fourth step of moving an optimal time axis for all the samples in each cluster based on the representative value. The gist of the medical information processing method is characterized by this.

  The invention of claim 2 is characterized in that, in claim 1, the result obtained in the fourth step is further converged by repeating in the first step to the fourth step.

  A third aspect of the present invention is the method according to the second aspect, wherein after the result obtained in the fourth step is converged, a step of assigning related information to the cluster and a new item in which inspection items are given in time series When the medical information is input, a step of calculating a certainty level indicating which cluster the test item of the new medical information belongs to, a certainty level regarding which cluster the new medical information belongs to, The method includes a step of outputting related information of a cluster to which the new medical information belongs.

  According to a fourth aspect of the present invention, in any one of the first to third aspects, the second step interpolates data having a missing portion based on the estimated probability density function. .

According to a fifth aspect of the present invention, in the fourth aspect, the second step is characterized in that when the medical information in the cluster has a time scale difference, the time scale is aligned.
The invention of claim 6 is a first means for classifying a set of medical information in which data of a plurality of examination items relating to a plurality of subjects are given in time series into clusters according to the data of the examination items, and each classified A second means for estimating, as a probability density function, a multi-dimensional time-series distribution in a plurality of sets of medical information related to a plurality of subjects in the cluster; and a model from the estimated probability density functions in the cluster; A medical information processing apparatus comprising: a third means for determining a representative value; and a fourth means for performing an optimal time axis movement for all samples in each cluster based on the representative value. It is a summary.

  The invention of claim 7 comprises the fifth means according to claim 6, further comprising a fifth means for repeatedly processing the result obtained by the fourth means by the first means to the fourth means for convergence. Features.

  The invention according to claim 8 is the novel medical information in which the related information giving means for giving the related information to the cluster converged by the fifth means and the test items are given in time series in claim 7 Is input, a certainty factor calculating means for calculating a certainty factor indicating to which cluster the test item of the new medical information belongs, and a certainty factor to which cluster the new medical information belongs, It is characterized by comprising output means for outputting related information of the cluster to which the new medical information belongs.

  The invention of claim 9 is characterized in that, in any one of claims 6 to 8, the second means interpolates data having a missing portion based on the estimated probability density function. .

The invention of claim 10 is characterized in that, in claim 9, the second means aligns the time scale when the medical information in the cluster has a time scale difference.
The invention according to claim 11 is a first means for classifying a set of medical information to which data of a plurality of examination items relating to a plurality of subjects are given in time series into clusters according to the examination item data; A second means for estimating, as a probability density function, a multidimensional time-series distribution in a plurality of sets of medical information relating to a plurality of subjects in each of the clusters, and among the estimated probability density functions in the cluster Medical information characterized by functioning as a fourth means for determining a representative value to be a model from the above and a fourth means for moving the optimal time axis for all samples in each cluster with reference to the representative value The gist of the processing program.

  According to a twelfth aspect of the invention, the computer according to the eleventh aspect causes the computer to function as a fifth means for causing the result obtained by the fourth means to be repeatedly processed by the first means to the fourth means to converge. It is characterized by.

  According to the invention of claim 1, by taking into account individual differences and modeling automatically from a large amount of data by clustering, a change corresponding to the degree of disease state / symptom and the detailed classification of the disease is modeled. It is possible to provide a medical information processing method that can simultaneously realize more flexible modeling of time series and precise estimation of probability density.

  Furthermore, according to the invention of claim 1, since clustering is performed, the granularity setting enables more detailed modeling than in the prior art, and furthermore, the risk of a subject's disease is predicted. It is possible to provide a medical information processing method that can specifically show how the statistical knowledge can be obtained from the original data.

  According to the invention of claim 2, since the result obtained in the fourth step is further repeated in the first step to the fourth step, the optimum classification result or the optimum probability density function is obtained. Therefore, the effect of claim 1 can be better realized.

  In conventional modeling or estimation of probability distribution (probability density function), fluctuations due to various factors are included in the target data and cannot be distinguished, so the overall state of the distribution It must be a broad estimate. However, according to the first and second aspects of the present invention, the variation can be absorbed by the process of matching the change with time.

  Further, according to the inventions of claim 1 and claim 2, it is possible to divide (classify) the problem of individual differences into clusters of an arbitrary number of categories. The distribution of samples within each cluster is made more compact.

This makes it possible to model the target phenomenon here more precisely and estimate the probability in more detail.
In the invention of claim 3, when new medical information is input, it is possible to easily obtain the certainty of which cluster the new medical information belongs and the related information of the cluster to which the new medical information belongs. Yes, it can be used as support when operating a health care program.

  In the prior art, it is inevitable that test values included in medical information have a missing value in time series, and only the data after cleansing excluding such data is targeted. According to this invention, it is possible to provide a medical information processing method capable of dealing with medical information having a missing value in time series.

  In the prior art, unified handling was difficult for a time series in which a year unit, a month unit, a week unit, a day unit, and a time unit were mixed. Even for medical information with different time scales, the difference in time scales can be absorbed and unified for time series in which year units, month units, week units, day units, and time units are mixed. Medical information processing methods that can be handled easily.

  According to the invention of claim 6, by taking into account individual differences and modeling automatically from a large amount of data by clustering, the degree of pathology / symptoms and changes corresponding to the detailed classification of the disease are modeled. It is possible to provide a medical information processing apparatus that can simultaneously realize more flexible time series modeling and precise probability density estimation. Further, according to the invention of claim 6, since clustering is performed, more detailed modeling is possible by setting the granularity as compared with the prior art, and further, the risk of the subject's disease is predicted. It is possible to provide a medical information processing apparatus that can specifically show how the statistical knowledge can be obtained from the original data.

  According to the invention of claim 7, since the result obtained by the fourth means is processed repeatedly by the first to fourth means, the optimum classification result or the optimum probability density function is obtained. Therefore, it is possible to realize the effect of claim 6 better.

  In the invention of claim 8, when new medical information is input, it is possible to easily obtain the certainty of which cluster the new medical information belongs and the related information of the cluster to which the new medical information belongs. It is possible to provide a medical information processing apparatus that can be used as support when operating a health care program.

  In the prior art, it is inevitable that test values included in medical information have a missing value in time series, and only the data after cleansing excluding such data is targeted. According to the invention, it is possible to provide a medical information processing apparatus that can cope with medical information having a missing value in time series.

  Further, according to the prior art, it has been difficult to uniformly handle a time series in which a year unit, a month unit, a week unit, a day unit, and a time unit are mixed. Can absorb the difference in time scale even for medical information with different time scales, and for time series where year units, month units, week units, day units, and time units are mixed, Medical information processing devices that can be handled in a unified manner can be provided.

According to the invention of claim 11, a medical information processing program capable of easily realizing the effect of the medical information processing apparatus of claim 5 can be provided.
According to the invention of claim 12, a medical information processing program capable of easily realizing the effect of the medical information processing apparatus of claim 6 can be provided.

Hereinafter, an embodiment of the present invention will be described with reference to FIGS.
FIG. 1 shows an overall schematic diagram of a medical information processing apparatus.
As shown in FIG. 1, the medical information processing apparatus 10 includes an input device 11 such as a keyboard and a data processing device 12 that operates according to a program. A storage device 13 for storing various data and an output device 14 are connected to the data processing device 12. The output device 14 includes a display and a printer, for example, and corresponds to output means.

  The data processing device 12 includes a computer 12c having a ROM 12a and a RAM 12b, and performs medical information processing using a medical information processing program stored in storage means such as the ROM 12a.

  The data processing device 12 includes a clustering means 21, a probability density function estimating means 22, a representative value determining means 23, a time axis moving means 24, a convergence means 25, a related information giving means 26, and a new data processing means 27. The clustering means 21 corresponds to a first means for classifying a set of medical information to which examination items relating to a plurality of subjects are given in time series into clusters according to the examination items.

  Here, as medical information, the following can be cited as representative examples, but is not limited thereto. These include test items in clinical tests and medical examinations.

  Values related to physical measurements (height, weight, BMI, waist circumference), blood pressure measurements, values related to blood chemistry tests (neutral fat, HDL cholesterol, LDL cholesterol), values related to liver function tests (AST (GOT), ALT (GPT) , Γ-GT (γ-GTP)), values related to blood glucose test (fasting blood glucose or HbA1c test), values related to urine test (urine sugar, etc.), electrocardiogram.

  In addition, medical information inspection items include inquiry items, physical inspection items performed by doctors, and the like. Each of these items has time information on the date and time of the inspection. This time information enables medical information (examination items) to have a time series distribution described later.

  The probability density function estimating means 22 corresponds to a second means for estimating, as a probability density function, a multidimensional time series distribution in a set of a plurality of medical information related to a plurality of subjects in each classified cluster. The representative value determining means 23 corresponds to a third means for determining a representative value to be a model from the estimated probability density functions in the cluster. The time axis moving means 24 corresponds to fourth means for moving the optimal time axis for all samples in each cluster with the representative value as a reference. The convergence means 25 corresponds to a fifth means for causing the first to fourth means to repeatedly process and converge. The related information giving means 26 is a means for giving relevant information such as disease, symptom, and medicine to each cluster. The new data processing means 27 is a means for performing various processes on new data.

Next, processing performed by the data processing device 12 according to the medical information processing program will be described with reference to FIG.
Prior to performing medical information processing by the data processing device 12, as preprocessing, for example, regarding various test items when determining that there is a disease or suspicion of a disease determined by the national or academic society based on medical information The threshold value Sh is stored in the storage device 13. Further, medical information of a plurality of subjects is stored in the storage device 13 from the input device 11 in advance.

(Step S10)
In S10, the clustering means 21 of the data processing device 12 reads the medical information of a plurality of subjects stored in advance in the storage device 13 from the input device 11, and selects an arbitrary according to the examination item from the set of the medical information. Cluster the number of categories. In this case, the larger the number of the plurality of subjects, the better.

  In the medical information of a plurality of subjects, the time information included in the medical information is t = 1, 2,..., T, the examination items are i = 1, 2,. Then, sample data, that is, medical information is represented by f (i, t, j).

When this clustering is performed, the data processing device 12 determines the granularity of the cluster in advance.
The cluster granularity is a numerical index representing the size of each cluster during clustering. Controlling this granularity is how many samples are assigned to each cluster when classifying the entire sample into clusters. In this embodiment, the number of samples belonging to the cluster is used as the cluster granularity.

Clustering includes, for example, the k-means method, but is not limited to the k-means method, and other clustering methods may be used. S10 corresponds to the first step.
(Step S20)
In S20, the probability density function estimation means 22 of the data processing apparatus 12 estimates the distribution of time series data in the samples in each cluster as a probability density function, but before estimating the probability dense time function, the time scale in each sample is estimated. If they are different, first set the time scale.

  A different time scale means that, for example, a specific test item for one sample is inspected every year, while the same specific test item for another sample is inspected monthly. It means that they are different. When the time scales are different as described above, the probability density function is estimated after aligning with the time scale set in advance for each inspection item by the input device 11. By aligning the time scale, the difference in the time scale of the sample can be absorbed, and a uniform treatment can be applied to a time series in which the units of year, month, week, day, and time are mixed. It will be possible.

If the time scales of the samples are the same, the probability density function is estimated as it is.
The estimation of the probability density distribution is usually performed by estimating a probability density function that gives a probability that an event occurs at each point in the multidimensional space. A parametric estimation method will be described in which a function representing the shape of the distribution is assumed as a function representing the distribution, and a parameter representing the characteristics of the distribution is estimated.

  Now, with respect to a certain cluster c, the inspection item values of M dimensions (i = 1, 2,..., M) at time t are obtained for N persons (j = 1, 2,..., N). To do. A probability density distribution in the M-dimensional space is estimated from the N samples. For example, assuming a multidimensional normal distribution as the shape of the distribution, the probability density function P (X) can be expressed by the mean vector μ and the covariance matrix Σ as follows.

Ｎ個のサンプルを用いて、平均ベクトルμと、共分散行列Σを推定すれば、確率密度関数が得られる。なお、平均ベクトルと共分散行列は、クラスターｃおよび、時間ｔごとに求める必要がある点に注意する必要がある。

A probability density function can be obtained by estimating the mean vector μ and the covariance matrix Σ using N samples. It should be noted that the mean vector and the covariance matrix need to be obtained for each cluster c and time t.

  Here, when an inspection item of an unknown sample Y is input with a keyboard or the like, the probability obtained by substituting Y into P (X) above belongs to cluster c at time t. Calculated as a numerical value representing the degree of risk when predicted.

  In addition, when a time series of inspection items in an unknown sample is given, time axis alignment is performed by the same procedure as S40 described later, and an optimal amount of time axis movement is obtained. The probability is obtained using the probability density function P (X).

  Here, an example of a multidimensional normal distribution is shown as the probability density function. However, when the number of samples that can be used for estimation is small, it is possible to use only a diagonal component instead of the covariance matrix. On the other hand, when the number of samples that can be estimated is large to some extent, a mixed normal distribution that is a weighted sum of a plurality of normal distributions can be used. These are parametric methods for expressing a function representing a distribution by parameters such as a mean vector and a covariance matrix.

The estimation of the probability density function is not limited to a parametric method, and may be performed by a nonparametric estimation method.
The nonparametric method does not particularly assume a functional form representing the shape of the distribution. A typical nonparametric probability density estimation method is a kernel density estimation method. Kernel density function estimation methods include multivariate kernel density estimation methods and binned kernel estimation methods.

  For example, in the binned kernel estimation method, the probability density function is estimated as follows.

ここで、Ｂはビンの数、Ｎはサンプル数、njはj番目のビンの度数、δはビン幅、Kh(x −jδ)は、バンド幅hのカーネル関数である。カーネル関数は、エパネックニコフ カーネル（Epanechnikov Kernel）等を用いることができる。

Here, B is the number of bins, N is the number of samples, nj is the frequency of the jth bin, δ is the bin width, and Kh (x−jδ) is a kernel function with a bandwidth h. As the kernel function, Epanechnikov Kernel or the like can be used.

Since the kernel density estimation method does not assume a specific function form for the entire distribution, more flexible modeling is possible.
(If there are missing values)
After estimating the probability density function, the probability density function estimating means 22 determines whether or not there is a missing value in the time series of the inspection items of the samples in each cluster. Based on the estimated probability density function, a missing value (that is, a missing portion) is interpolated. S20 corresponds to the second step.

(Step S30)
In S30, the representative value determining means 23 of the data processing device 12 determines a representative value in each cluster. Here, the representative value is determined in order to use the representative value (model) as a reference point when moving the sample in the cluster in S40 described later. The representative value may be a center value or an average value. The representative value may be determined based on information input from the outside. S30 corresponds to the third step.

(Step S40)
In S40, the time axis moving unit 24 of the data processing device 12 obtains a value (τ) that is the optimal time axis movement, and moves the data of the inspection item (that is, the inspection data). The method for obtaining the value (time difference) that results in the optimal movement of the time axis is determined in S30 by moving samples in the cluster for all values in a certain time width, for example, from -τ to τ. The value that minimizes the difference (sum of distances) from the representative value is the optimal time difference. This movement is performed for all inspection data.

As a result, the axis of change of the test value of the test item due to the disease is aligned.
FIG. 3A is a schematic diagram of medical information having certain examination items related to a plurality of subjects before clustering. In FIG. 3A, the vertical axis indicates the inspection value of the inspection item, and the horizontal axis indicates time. FIG. 3B is a schematic diagram when clustering is performed and the time axes are aligned. As shown in FIG. 3B, by performing clustering and aligning the time axis, in the inspection item, for example, in a specific cluster, the time-series inspection value is less than or equal to the threshold value Sh (or more than the threshold value). It is possible to determine that there is a possibility of illness in the case of having the above.

In FIG. 3A and FIG. 3B, the scale of the inspection value is different for convenience of explanation. S40 corresponds to the fourth step.
(Step S50)
In S50, the convergence means 25 of the data processing device 12 determines whether or not the data after moving in S40 has converged. That is, the convergence means 25 determines the convergence using a numerical index representing the goodness of the model in which the probability density function estimated as the goodness of classification into clusters represents the data.

  For example, the convergence unit 25 compares with an index representing the goodness of the model in the previous process (that is, the processes in S10 to S40), and determines that the convergence is achieved when improvement over a predetermined value cannot be obtained. The convergence means 25 uses a predetermined initial value index when there is no index representing the goodness of the model in the previous process (that is, the processes of S10 to S40), that is, when the determination of S50 is the first time. To do.

Here, the constant value includes an absolute value and a relative value (for example, the improvement converges at 1% or less of the previous time), but may be any.
As a numerical index, a probabilistic numerical value (likelihood) is used.

(Example of convergence judgment index)
Here, an example of a convergence determination index will be described.
Let D (j), j = 1, ..., N be the entire state where all sample data (X (j), j = 1, ..., N) are classified into a cluster. For example, when the third data (j = 3) is classified into cluster number 4, cluster numbers are assigned to all samples as D (3) = 4.

  The probability density function is estimated for each cluster. When the parametric method and the nonparametric method are expressed in common, the difference in the probability density function of each cluster is expressed as PM (X) using the model M. In the case of the parametric method, θ is a parameter for estimating the probability density function, and it is expressed as M (θ). For example, when the probability density is estimated with a normal distribution, the average value and the covariance matrix are parameters θ representing the probability density function PM (X).

  Given a classification D to a cluster and a model M of a probability density function, the probability generated from the classification D and the probability density function M can be calculated for each sample data. This probability is obtained for all samples, and the product is considered as the probability that all samples occur from the cluster classification D and the probability density function M. The probability that all these samples occur is used as an index for convergence determination. For example, when the log likelihood obtained by taking the logarithm of probability is used, the following formula is obtained.

ここで上記のモデルMは、クラスター番号D(j)に従って選択するものとする（ M=M(D(j)) ）。Πは積を表わす。

Here, the model M is selected according to the cluster number D (j) (M = M (D (j))). Π represents product.

  If the convergence means 25 determines in S50 that it has not converged, it returns to S10, and if it determines that it has converged, it proceeds to S60. By having S50, convergence is achieved by repeating the first to fourth steps.

(Step S60)
In S <b> 60, the related information adding unit 26 collects related information such as diseases, symptoms, and medicines by text mining from medical related text data stored in the storage device 13 in advance. In addition, the related information is aggregated from the medical-related text data, etc., and various information related to the disease name, symptoms, etc., such as instruction methods for diseases, drugs for improving or treating symptoms, etc. Contains information.

  Then, the related information giving unit 26 gives the related information to the cluster related to the aggregated related information. That is, based on the fact that the cluster includes information such as disease names and symptom levels that are frequently found in the cluster, the cluster relates to the disease name and symptom level from the aggregated information. Things are given as related information. As a result, the cluster is provided with information such as a guidance method for a disease and a drug for improving or treating symptoms as related information.

(Step S70)
In S <b> 70, the new data processing unit 27 waits for input of new data (that is, medical information), and when there is input of new data, calculates a certainty factor for the new data. The new data processing unit 27 corresponds to a certainty factor calculating unit.

There are two methods for expressing the certainty factor: a method for indicating a numerical value representing the certainty factor, and a method for setting a certain number of classes in the certainty factor in advance and indicating the class of the certainty factor.
When the certainty factor is represented by a numerical value, the new data processing means 27 calculates the probability that the new data belongs to each cluster for all clusters, and uses the probability as the certainty factor.

  As a method of displaying the certainty factor, there are a method of displaying all the clusters, a method of displaying a certain number of clusters in descending order of probability, and a method of displaying only clusters having a certain probability or higher.

  In the case of setting several classes, the new data processing means 27 determines the upper limit and the lower limit of the degree of certainty in advance using the above probability values. And about the new data used as the value in the meantime, only a class name (step value) is output as a certainty factor that the data processing device 12 outputs in S80 described later. In this way, for example, the certainty factor can be displayed in five stages.

  Further, since the new data is time series data, the new data processing means 27 estimates information about where the new data is on the time series of the model. This information is output to the data processing device 12 in S80 described later.

The information on the position on the time series is for indicating what stage the temporal change of the degree of the pathological condition / symptom represented by the model is.
In this case, estimation of the position on the time series by the new data processing means 27 is performed by moving the time axis with respect to the model of each cluster when obtaining the above probability from the input sample data. This is done by finding the amount of movement on the time axis that best matches the input data. Moreover, this method is performed similarly to S40.

  Also in this case, in addition to the numerical value for obtaining the movement amount of the time axis, several steps of movement amounts may be set in advance, and the optimum one may be selected from them. As a result, as in the case of the certainty factor, it is possible to indicate a movement amount having a width such as an upper limit and a lower limit.

  If the new data inspection item is different from the already obtained cluster in the time scale, the time axis is moved after the time scale of the new data is aligned with the time scale of the cluster.

(Step S80)
The data processing device 12 outputs the certainty factor relating to the new data obtained in S70, information about which position in the time series of the model of which cluster, and the threshold value Sh relating to various inspection items. Output (display and print) by the device 14. As a result, it is possible to know who the new data is in which position on the time series of which model of the cluster and the threshold value Sh for the various inspection items. Can be given to subjects.

The effects exhibited by this embodiment will be described below.
(1) The medical information processing method according to the present embodiment includes step S10 of classifying a set of medical information to which data of a plurality of examination items relating to a plurality of subjects are given in time series into clusters based on the examination item data. Prepare. The medical information processing method further includes a step S20 of estimating a multidimensional time-series distribution as a probability density function in a plurality of sets of medical information related to a plurality of subjects in each classified cluster. Furthermore, in the medical information processing method of this embodiment, step S30 for determining a representative value to be a model from the estimated probability density function in the cluster, and all samples in each cluster based on the representative value. Step S40 for performing an optimal time axis movement is provided.

  As a result, by taking into account individual differences and modeling automatically from a large amount of data by clustering, it is possible to model changes in the degree of pathological conditions / symptoms and detailed classification of diseases. Therefore, more flexible time series modeling and precise probability density estimation can be realized at the same time. Furthermore, according to the medical information processing method of this embodiment, if the granularity of clustering is set appropriately, more detailed modeling is possible as compared with the prior art, and the risk of the subject's disease can be predicted. It is possible to specifically show how statistical knowledge for performing can be obtained from the original data.

  Further, according to the medical information processing method of the present embodiment, since individual data is moved and matched on the time axis in consideration of the consistency with the model in S40, with respect to temporal changes, The variation can be absorbed by the process of matching it.

  (2) The medical information processing method of the present embodiment further converges the result obtained in step S40 by repeating it in S10 to S40 via S50. As a result, according to the medical information processing method of the present embodiment, an optimal classification result or an optimal probability density function can be obtained, and therefore the effect (1) can be realized better.

  In addition, modeling by a conventional method or estimation of a probability distribution (probability density function) includes fluctuations due to various factors in the target data and cannot be distinguished from each other. It must be an overall estimation that estimates only the state of the image.

  However, according to the medical information processing method of the present embodiment, when S10 to S40 are repeatedly processed, in S40, individual data is moved and matched on the time axis in consideration of consistency with the model. Therefore, the change can be absorbed by the process of matching the change with time.

  (3) In S60, the medical information processing method according to the present embodiment converges the result obtained in S40, and then gives related information to the cluster. In S70, the examination items are given in time series. When new medical information is input, a certainty factor indicating which cluster the test item of the new medical information belongs to is calculated. Then, in S80, the medical information processing method of the present embodiment outputs a certainty level to which cluster the new medical information belongs and related information of the cluster to which the new medical information belongs.

  As a result, in the medical information processing method of this embodiment, when new medical information is input, the certainty of which cluster the new medical information belongs and the related information of the cluster to which the new medical information belongs. Can be easily obtained.

  This is realized on the electronic medical record system and is supported when a doctor performs diagnosis and guidance for an individual patient, and is realized on a health management system so that a public health nurse and a nutritional manager can provide individual users. Can be provided as a support for health guidance. Furthermore, it can be used for epidemiological studies in the medical field. It can also be used as a support for the luck of health insurance associations and health management programs.

  (4) In step S20, the medical information processing method of the present embodiment interpolates data having a missing portion based on the estimated probability density function. As a result, in the prior art, it is inevitable that test values included in medical information have missing values in time series, and only the data after cleansing excluding such data is targeted, In the medical information processing method of the present embodiment, it is possible to cope with medical information having a missing value in time series.

  (5) In the medical information processing method, in step S20, when there is a time scale difference in the medical information in the cluster, the time scale is aligned. As a result, in the prior art, unified handling was difficult for time series in which year units, month units, week units, day units, and time units were mixed. Can absorb the difference of time scale even for medical information with time scale difference, unified for time series with mixed year unit, month unit, week unit, day unit and time unit Can be treated.

  (6) The medical information processing apparatus 10 includes clustering means 21 that classifies a set of medical information to which data of a plurality of examination items related to a plurality of subjects are given in time series into clusters according to the examination item data. Furthermore, the medical information processing apparatus 10 includes a probability density function estimating unit 22 that estimates, as a probability density function, a multidimensional time-series distribution in a plurality of sets of medical information related to a plurality of subjects in each classified cluster; A representative value determining unit 23 is provided for determining a representative value to be a model from the estimated probability density function in the cluster. Further, the medical information processing apparatus 10 includes a time axis moving unit 24 that performs an optimal time axis movement for all samples in each cluster with reference to the representative value. As a result, the medical information processing apparatus 10 of the present embodiment can be provided as an apparatus that can achieve the same effect as the above (1).

  (7) In the medical information processing apparatus 10, the result obtained by the time axis moving means 24 is further converted by the convergence means 25 into the clustering means 21, probability density function estimating means 22, representative value determining means 23, time axis moving means. The process is repeated at 24 to converge. As a result, it can provide as an apparatus which can have the same effect as the above (2).

  (8) When the medical information processing apparatus 10 receives the related information providing unit 26 that assigns the related information to the converged cluster and the new medical information in which the test items are given in time series, New data processing means 27 is provided for calculating a certainty factor indicating to which cluster the test item of the new medical information belongs. The medical information processing apparatus 10 includes an output device 14 that outputs the certainty of which cluster the new medical information belongs and the related information of the cluster to which the new medical information belongs.

As a result, the medical information processing apparatus 10 of the present embodiment can be provided as an apparatus that can achieve the same effect as the above (3).
(9) The probability density function estimation unit 22 of the medical information processing apparatus 10 interpolates data having a missing portion based on the estimated probability density function. As a result, the medical information processing apparatus 10 of the present embodiment can be provided as an apparatus that can achieve the same effect as the above (4).

  (10) The probability density function estimation means 22 of the medical information processing apparatus 10 aligns the time scale when there is a time scale difference in the medical information in the cluster. As a result, the medical information processing apparatus 10 of the present embodiment can be provided as an apparatus that can achieve the same effect as the above (5).

  (11) The medical information processing program of the present embodiment classifies a set of medical information to which data of a plurality of examination items relating to a plurality of subjects are given in time series into clusters according to the examination item data. It functions as the clustering means 21. In addition, the medical information processing program is a probability density function estimation unit that estimates a computer as a probability density function of a multi-dimensional time series distribution in a plurality of sets of medical information related to a plurality of subjects in each classified cluster. 22 and the representative value determining means 23 for determining a representative value to be a model from the estimated probability density function in the cluster. Further, the medical information processing program causes the computer to function as a time axis moving unit 24 that moves the optimal time axis for all samples in each cluster with the representative value as a reference. As a result, the program of the present embodiment can be provided as a program that exhibits the effect (1).

  (12) The medical information processing program of the present embodiment repeats the results obtained by the time axis moving means 24 by the clustering means 21, the probability density function estimating means 22, the representative value determining means 23, and the time axis moving means 24. The computer is caused to function as the convergence means 25 for processing and convergence. As a result, the program of the present embodiment can be provided as a program that exhibits the effect (2).

In addition, the said embodiment can also be changed and comprised as follows.
In S10, the cluster granularity is a method using the number of samples belonging to the cluster. However, in addition to this, a method using a numerical value representing the goodness of unity or a combination of both may be used. In addition, as a method of using the numerical value representing the goodness of the unit, a method using a sum of distances of each sample from a representative value in the cluster, or a diagonal component of a covariance matrix of a probability density function estimated as a normal distribution is used. Some use sums, etc., and the cluster granularity may be determined by these methods. Thus, the particle size of the cluster can be determined by various methods and is not limited.

1 is a schematic block diagram of a medical information processing apparatus 10. FIG. The flowchart of a medical information processing program. (A) is a schematic diagram of medical information having certain examination items related to a plurality of subjects before clustering, and (b) is a schematic diagram when clustering is performed and the time axes are aligned.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 ... Medical information processing apparatus, 11 ... Input device, 12 ... Data processing apparatus,
13 ... Storage device, 14 ... Output device (output means),
21 ... Clustering means (first means),
22 ... Probability density function estimation means (second means),
23 ... representative value determining means (third means), 24 ... time axis moving means (fourth means),
25 ... Convergence means (fifth means), 26 ... related information giving means,
27: New data processing means (confidence degree calculating means).

Claims (12)

A first step of classifying a set of medical information in which data of a plurality of examination items relating to a plurality of subjects are given in time series into clusters according to the data of the examination items;
A second step of estimating, as a probability density function, a multi-dimensional time-series distribution in a plurality of sets of medical information related to a plurality of subjects within each classified cluster;
A third step of determining a representative value as a model from the estimated probability density function in the cluster;
A medical information processing method comprising: a fourth step of performing an optimal time axis shift for all samples in each cluster with reference to the representative value.   The medical information processing method according to claim 1, wherein the result obtained in the fourth step is further converged by repeating the first step to the fourth step. Assigning relevant information to the cluster after converging the results obtained in the fourth step;
A step of calculating a certainty level indicating to which cluster the test item of the new medical information belongs when new medical information to which the test item is given in time series is input;
The medical information processing method according to claim 2, further comprising: a step of outputting a certainty level to which the new medical information belongs and a related information of the cluster to which the new medical information belongs.   The medical information processing method according to any one of claims 1 to 3, wherein the second step interpolates data having a missing portion based on the estimated probability density function.   5. The medical information processing method according to claim 4, wherein when the medical information in the cluster has a time scale difference, the second step aligns the time scale. A first means for classifying a set of medical information in which data of a plurality of examination items relating to a plurality of subjects are given in time series into clusters according to the data of the examination items;
A second means for estimating, as a probability density function, a multi-dimensional time-series distribution in a plurality of sets of medical information related to a plurality of subjects within each classified cluster;
A third means for determining a representative value as a model from the estimated probability density function in the cluster;
A medical information processing apparatus, comprising: a fourth means for performing an optimal time axis movement for all samples in each cluster with reference to the representative value.   The medical information processing according to claim 6, further comprising fifth means for causing the result obtained by the fourth means to be repeatedly processed by the first means to the fourth means to converge. apparatus. Related information giving means for giving related information to the cluster converged by the fifth means;
A certainty factor calculating means for calculating a certainty factor indicating to which cluster the test item of the new medical information belongs when the new medical information to which the test item is given in time series is input;
The medical information processing apparatus according to claim 7, further comprising: an output unit configured to output a certainty level to which cluster the new medical information belongs and related information of the cluster to which the new medical information belongs.   The medical information processing apparatus according to claim 6, wherein the second means interpolates data having a missing portion based on the estimated probability density function.   The medical information processing apparatus according to claim 9, wherein the second means aligns the time scale when there is a time scale difference in the medical information in the cluster. Computer
A first means for classifying a set of medical information in which data of a plurality of examination items relating to a plurality of subjects are given in time series into clusters according to the data of the examination items;
A second means for estimating, as a probability density function, a multi-dimensional time-series distribution in a plurality of sets of medical information related to a plurality of subjects within each classified cluster;
A third means for determining a representative value as a model from the estimated probability density function in the cluster;
A medical information processing program which functions as a fourth means for performing an optimal time axis movement for all samples in each cluster with the representative value as a reference. Computer
The medical information processing program according to claim 11, wherein the result obtained by the fourth means is caused to function as a fifth means for causing the first to fourth means to repeatedly process and converge.

Images