JP2006202235A - Time-based phenomenon occurrence analysis apparatus and time-based phenomenon occurrence analysis method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a new survival time analysis apparatus and survival time analysis method that can analyze the probability of the occurrence of a predetermined phenomenon in an analysis object (single sample) at a predetermined time point by associating and analyzing feature amount data on the analysis object and the occurrence of the predetermined phenomenon. <P>SOLUTION: The survival time analysis apparatus 100 comprises an input part 10 for inputting feature amount data acquired from an analysis object, and a probability calculation part 20 for calculating predetermined survival rates of the analysis object according to gene expression profile data input from the input part 10, and the probability calculation part 20 has a plurality of estimators 21 for respective predetermined time points. The associated analysis of the gene expression profile and survival rate of the analysis object can analyze the survival rates of the analysis object (single sample) at the predetermined time points. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a time-dependent phenomenon occurrence analysis apparatus and a time-related phenomenon occurrence analysis method for analyzing the lifespan of a living organism such as a human being or the life of an industrial product, and in particular, the survival rate and the industrial product in the treatment prognosis of a disease. The present invention relates to a temporal phenomenon occurrence analysis apparatus and a temporal phenomenon occurrence analysis method for estimating and analyzing an occurrence rate of a failure of the machine.

  Living organisms always die after a certain period of time, and they break down if they are industrial products such as machinery. Survival analysis (also known as survival analysis) is known as a method of analyzing the relationship between various factors such as the lifespan of an individual with respect to irreversible changes such as the death of a living organism or mechanical failure. Yes.

  Survival time analysis is a statistical method with time as an objective variable. In the case of survival time analysis, the objective variable is the time until death, but it can be applied to other cases when considered as the time until a certain phenomenon occurs. For example, not only survival and death of living organisms, but also metastasis of cancer, morbidity, occurrence of failures in industrial products, and the like can be objects of survival time analysis.

  In addition to survival and death of these organisms, in addition to cancer metastasis, disease morbidity, and the occurrence of failures in industrial products, the time until a given phenomenon occurs in a given analysis target is statistical In this specification, the analysis method is particularly referred to as “time-lapse phenomenon generation analysis”. Further, in this specification, in the analysis of occurrence of phenomena over time, the case of analyzing life / death information of a living organism is particularly distinguished and referred to as “survival time analysis”. Hereinafter, for convenience of explanation, a survival time analysis will be described as an example.

  Survival analysis in medical statistics shows that for each case, the survival time to death is obtained for the death case, and as long as the observed survival time is obtained for the survival case. It can be understood as a study.

  For example, what are the factors that affect the prognosis of patients who have undergone surgery for lung cancer? Also, is there a long-lived effect on patients with new therapies? To answer these questions, it is first necessary to follow the patient for a long time and collect data on subsequent changes (death, severity, etc.). The life-and-death status of each case should be examined and the effect of each variable on the patient's final state should be examined. For this purpose, an appropriate statistical method for survival analysis is indispensable.

  As a representative method of survival time analysis in such medical statistics, for example, a method of “drawing a survival curve for each layer after stratifying the target data by feature amount” and “features that make continuous values” The technique of expressing the regression relationship between the quantity and the survival rate by the regression coefficient can be cited as a standard one. Here, “stratification” means that each example is classified into two or more groups according to the feature amount. The “survival curve” refers to a curve in which the horizontal axis represents time and the vertical axis represents the ratio of the number of surviving cases at each time point, and the ratio of all cases alive is plotted on the time axis.

  A typical example of the former method is Kaplan-Meier analysis, and a typical example of the latter method is Cox proportional hazard analysis (Non-Patent Document 1). , 2).

In addition, techniques for predicting human health, lifespan, and the like using medical statistics as described above have been developed. For example, Patent Document 1 uses an index that comprehensively represents an individual's health status, and contributes to the quantitative grasp of daily health status, the effects of performing health management guidance, and optimization of guidance content. , Using life expectancy prediction data, which is the basic data for life expectancy prediction for various health checkup results, to calculate the predicted value of life expectancy for each individual from the health checkup results entered in the health check result input procedure However, a technique for displaying on a display unit or printing on a printing unit is disclosed.
JP 2003-167959 A (publication date: June 13, 2003) Haruo Yanai and Satoshi Takagi, edited by Handbook of Multivariate Analysis, Contemporary Mathematics Company, 1986. Cox, DR: Regression models and life-tables, Journal of the Royal Statistical Society, Ser. B 34: 187-220, 1972.

  As described above, representative methods of survival time analysis include “Kaplan-Meier analysis” and “Cox proportional hazard analysis”.

  However, in order to perform a more accurate survival analysis, it is preferable to target the case where there are a wide variety of features and fine stratification is required, but in Kaplan-Meier analysis, In such a case, there is a problem that the number of cases in each layer is reduced and the reliability of the Kaplan-Meier survival curve is lowered.

  In addition, in order to perform a reliable survival time analysis, it is necessary to target cases where various time scales are mixed, such as factors related to mortality risk at an early stage after diagnosis and factors related to mortality risk after delay. However, there is a problem that these cannot be handled simultaneously in the “Cox proportional hazard analysis”.

  Furthermore, the life expectancy prediction data generation device disclosed in Patent Document 1 is intended for lifestyle habits and health checkup results as input data, and it is necessary to stratify and provide input amounts such as lifestyle habits. It becomes complicated. In addition, the output result of the health expectancy prediction data generation apparatus selects and outputs a curve calculated in advance for each layer from the database. For this reason, the output curve has only simple information that can be reduced to one parameter of “healthy life expectancy” and does not reflect the risk change in each period after diagnosis. For this reason, when considering application to tailor-made (custom-made) medicine, which is expected to be further developed in the future, such a conventional apparatus is not sufficient.

  In addition, the life science technology in recent years has been remarkably advanced. For example, a large amount of feature amount can be obtained for the medical data of each case by a gene expression analysis technology or the like. For this reason, development of a new survival time analysis method for analyzing such a large amount of information in association with the survival time is strongly desired.

  Furthermore, as described above, the survival time analysis method developed in this way is not limited to the survival and death of living organisms, but is also applied to predetermined methods such as cancer metastasis at a predetermined point in time, morbidity of disease, occurrence of failures in industrial products, etc. The present invention can also be applied to analysis of the occurrence of phenomena over time, which analyzes the occurrence of phenomena.

  The present invention has been made in view of the above-described problems, and an object of the present invention is to analyze feature quantity data to be analyzed in association with occurrence of a predetermined phenomenon, and to analyze an object at a predetermined time (single example). It is an object of the present invention to provide a novel temporal phenomenon occurrence analysis apparatus and temporal phenomenon occurrence analysis method capable of analyzing the probability of occurrence of a predetermined phenomenon.

  As a result of intensive studies to solve the above problems, the present inventors used, as learning data, a gene expression profile and a survival time at a predetermined time point after treatment for the prognosis of a patient after cancer treatment. By learning the correlation of these data by supervised machine learning of binary classification with probability output, survival time analysis device (time-dependent phenomenon occurrence analysis device) that can analyze survival probability for each patient at a predetermined time point It was found that can be produced. And, by inputting a gene expression profile for a predetermined time point to this survival time analyzer, the survival rate of a single patient at a predetermined time point can be calculated, and a single case survival curve can be drawn New technology was established and the present invention was completed. The present invention has been completed based on such novel findings, and includes the following inventions.

  (1) input means for inputting feature quantity data obtained from the analysis target; probability calculation means for calculating an occurrence probability of a predetermined phenomenon for the analysis target based on the feature quantity data input by the input means; The probability calculation means includes, as learning data, whether or not the predetermined phenomenon has occurred at a predetermined elapsed time from the time when the feature amount data and the individual who acquired the feature amount data are earned. An estimator obtained by supervising the correlation between the feature data and the phenomenon information using a plurality of sets of phenomenon information regarding the learning data, and at any time other than the individual used for the learning data When the feature amount data of an arbitrary individual at is input, when the learning data has a correlation with the arbitrary feature amount data and the learning data from the time point The estimator that predicts whether or not the predetermined phenomenon occurs in the individual that acquired the feature amount data and outputs the probability, and the probability calculating means corresponds to each of a plurality of predetermined elapsed times. A temporal phenomenon occurrence analysis apparatus that includes a plurality of estimators and calculates the probability of occurrence of a predetermined phenomenon for the analysis target at a plurality of elapsed time points corresponding to the estimators.

  (2) Further, by using the occurrence probability values at a plurality of elapsed time points calculated by the probability calculating means, a phenomenon occurrence probability curve with time from the time of acquiring feature amount data in an arbitrary analysis target is created. The temporal phenomenon occurrence analysis device according to (1), comprising a curve creation means.

  (3) The temporal phenomenon according to (1) or (2), further comprising expected value calculation means for calculating an expected value of a time that elapses until the predetermined phenomenon occurs using the phenomenon occurrence probability curve. Generation analysis device.

(4) The measurement object is a living organism, and the feature data is biological data.
The temporal phenomenon occurrence analysis apparatus according to any one of (1) to (3), wherein the occurrence of the predetermined phenomenon is death of a living organism to be analyzed, morbidity of a disease, or metastasis of a cell proliferative disease.

  (5) The biological phenomenon generation analysis device according to (4), wherein the biological data is biological pathological diagnosis data.

  (6) The biological phenomenon generation analysis apparatus according to (4) or (5), wherein the biological data is a gene expression profile of an organism.

  (7) The biological data generation analyzer according to any one of (4) to (6), wherein the biological data relates to a prognosis of a cell proliferative disease.

  (8) The estimator is a discriminant function whose estimation accuracy is increased by the learning data, and a predetermined phenomenon occurs using a discriminant function that receives the feature quantity data to be analyzed and outputs a real value. A discriminant function processing unit that calculates whether or not, and an fP conversion processing unit that calculates the probability of occurrence of a predetermined phenomenon by performing fP conversion processing on the output value f from the discriminant function processing unit. The temporal phenomenon occurrence analysis device according to any one of (1) to (7).

  (9) The discriminant function in the discriminant function processing unit uses one-dimensional linear discriminant analysis and a weighted voting method, and the fP conversion processing in the fP conversion processing unit uses logistic regression. The temporal phenomenon occurrence analysis device according to (8).

  (10) An input step of inputting feature amount data obtained from the analysis target; a probability calculation step of calculating a probability of occurrence of a predetermined phenomenon for the analysis target based on the feature amount data input by the input unit; The probability calculation step includes, as learning data, whether or not the predetermined phenomenon has occurred at a predetermined elapsed time from the time when the characteristic amount data and the individual who acquired the characteristic amount data are earned. An estimator obtained by supervising the correlation between the feature data and the phenomenon information using a plurality of sets of phenomenon information regarding the learning data, and at any time other than the individual used for the learning data When the feature amount data of an arbitrary individual at is input, a predetermined progress in the learning data from the time point correlated with the arbitrary feature amount data A step of using an estimator that predicts whether or not the predetermined phenomenon occurs in the individual that acquired the feature amount data and outputs the probability, and the probability calculation step includes a plurality of predetermined elapsed time points. A temporal phenomenon occurrence analysis method, which is a step of calculating a probability of occurrence of a predetermined phenomenon with respect to the analysis target at a plurality of elapsed time points corresponding to the estimator using a plurality of corresponding estimators.

  (11) Further, by using the value of the occurrence probability of the phenomenon at a plurality of elapsed time points calculated by the probability calculation step, a time-dependent phenomenon occurrence probability curve from the time of acquiring the feature amount data in an arbitrary analysis target is created. The method of analyzing occurrence of a phenomenon with time according to (10), including a curve creation step.

  (12) The temporal phenomenon according to (10) or (11), further including an expected value calculating step of calculating an expected value of a time elapsed until the predetermined phenomenon occurs using the phenomenon occurrence probability curve. Occurrence analysis method.

  The temporal phenomenon occurrence analysis apparatus or the temporal phenomenon occurrence analysis method may be realized by a computer. In this case, the survival time analysis apparatus is operated by the computer by operating the computer as each means. The control program of the temporal phenomenon occurrence analysis apparatus to be realized and the computer-readable recording medium on which the control program is recorded also fall within the scope of the present invention.

  According to the temporal phenomenon occurrence analysis apparatus and the temporal phenomenon occurrence analysis method according to the present invention, it is possible to analyze the feature quantity data to be analyzed and the occurrence of the predetermined phenomenon in association with each other at a predetermined time point. At this point, it is possible to analyze the probability of occurrence of a predetermined phenomenon for a single analysis target example.

  For example, the feature amount data can include life-related data such as pathological diagnosis data, and the occurrence of a predetermined phenomenon includes life and death of a living organism, disease (including cancer metastasis), product failure, and the like. Examples of the probability of occurrence of a predetermined phenomenon include survival rate, disease incidence (including cancer metastasis rate), product failure rate, and the like.

  Furthermore, a phenomenon occurrence probability curve (for example, a survival rate curve or the like) representing the probability of occurrence of a predetermined phenomenon at each time point can be drawn. In this case, it is possible to accurately and easily determine the probability that a predetermined phenomenon will occur even when the estimator is not provided (time point). Thus, a unique phenomenon occurrence probability curve can be obtained for each analysis target (for each case) without stratification as in a conventional apparatus for performing survival time analysis.

  The temporal phenomenon occurrence analysis apparatus and the temporal phenomenon occurrence analysis method according to the present invention will be described in detail below. Since the description of the temporal phenomenon occurrence analysis method overlaps with the explanation of the processing steps in the temporal phenomenon occurrence analysis apparatus, the following explanation will be given by taking the temporal phenomenon occurrence analysis apparatus as an example, and only the method. Do not do.

  The temporal phenomenon occurrence analysis apparatus according to the present invention calculates an occurrence probability of a predetermined phenomenon for an analysis target at a predetermined time point using an estimator that performs binary classification processing with probability output by supervised machine learning. It is. The term “occurrence of a predetermined phenomenon” in this specification includes not only survival and death of living organisms, but also metastasis of cell proliferative diseases (cancer, tumor), morbidity, and failure of industrial products. included.

  In other words, the temporal phenomenon occurrence analysis apparatus according to the present invention calculates the probability of occurrence of a predetermined phenomenon with respect to the analysis target at a predetermined time point based on various feature amount data of the analysis target. Analysis can be performed. In particular, the present invention proposes a new method for analyzing the occurrence of a phenomenon over time, which has not been seen so far, in that the occurrence rate of a phenomenon at each time point can be calculated for a single analysis target example (one sample).

  Furthermore, the temporal phenomenon occurrence analysis apparatus according to the present invention can create and display a phenomenon occurrence probability curve using the calculated occurrence rates of a plurality of phenomena. Here, the phrase “phenomenon occurrence probability curve” is a graph in which the probability of occurrence of a predetermined phenomenon at each time point is displayed in the form of a graph with the vertical axis representing the occurrence rate of the phenomenon and the horizontal axis representing time. For example, a survival rate curve can be given as an example of the phenomenon occurrence probability curve. In particular, according to the temporal phenomenon occurrence analysis apparatus according to the present invention, a phenomenon occurrence probability curve for a single example can be drawn.

  The basic principle of the processing method of the temporal phenomenon occurrence analysis apparatus according to the present invention will be briefly described. First, the correlation between the feature value data representing the features of each example as multivariate and the phenomenon information regarding whether or not a predetermined phenomenon has occurred at a predetermined time point is learned by supervised machine learning, and the probability output An estimator that can perform the binary classification process is created. Then, feature amount data to be analyzed is input to these estimators, and a probability that a predetermined phenomenon occurs in the analysis target at a predetermined time point is calculated. A plurality of such estimators are prepared for each predetermined time point, and the process of calculating the occurrence rate of the phenomenon is performed at a plurality of time points. Then, from the obtained result, a single phenomenon occurrence probability curve is created and displayed.

  That is, the present invention discloses a technique of arranging a plurality of estimators (also referred to as predictors and classifiers) in parallel for risk evaluation in a plurality of hours. Each individual estimator adopts a machine learning method capable of handling multivariate feature quantity data, and is configured to output a phenomenon occurrence rate at a certain point in time. And the output result of several estimators is expressed in the form of a single example survival rate curve.

  The single occurrence probability curve obtained in this way shows not only survival and death of living organisms, but also metastasis of cell proliferative diseases (cancer, tumor, etc.), morbidity, industrial product failures, etc. Very useful in survival analysis. For example, when pathological diagnosis data or medical data is used as the feature amount data, it is possible to calculate the survival rate of organisms and the incidence of disease. Moreover, when the data regarding the lifetime of an industrial product is used as feature-value data, the failure rate of an industrial product can also be calculated. Such analysis results are particularly necessary for making detailed judgments (diagnostics) for each example at the analysis site (clinical site for medical applications, quality control site for industrial products, etc.). Judgment (diagnosis) criteria can be given.

  Such a time-dependent phenomenon occurrence analyzing apparatus according to the present invention will be specifically described below by taking as an example the case of analyzing the survival and death of a living organism. That is, the survival time analysis apparatus will be described below as an embodiment of the temporal phenomenon occurrence analysis apparatus. In the present embodiment, for convenience of explanation and ease of understanding, biological data, particularly gene expression profile data is used as the feature amount data, and cancer treatment (cancer surgery) is performed as an analysis target. A case where a patient (human) is an analysis target and the prognostic survival rate of a patient who has been treated for cancer is calculated as an occurrence rate of a predetermined phenomenon will be described as an example. It should be noted that the present invention can be modified as appropriate within the scope of the present invention and is not limited to the following exemplary embodiments.

  The survival time analysis apparatus according to the present embodiment will be specifically described below with reference to FIGS.

  FIG. 1 is a diagram showing a schematic configuration of functional blocks of a survival time analysis apparatus 100 according to the present embodiment. As shown in the figure, the survival time analysis apparatus 100 includes an input unit 10, a probability calculation unit 20, a curve creation unit 30, an expected value calculation unit 40, and an output unit 50.

  The input unit 10 functions as input means for inputting gene expression profile data as feature amount data of the examinee. In other words, the input unit 10 inputs the gene expression profile of the examinee whose survival time is to be analyzed to the probability calculation unit 20, and the captured data is in the form of electronic data. Specifically, for example, it is taken in via a network or medium from a medical examinee who manages the gene expression profile data with electronic data, a medical institution, a research institution, a data management contract service organization, or printed on paper After importing the gene expression profile data with an image scanner, import the digitized data that recognizes the written characters, numerical values, figures, etc., or the terminal operator describes it while looking at the document on which the paper gene expression profile data is written Realized by capturing the result of keyboard input of the contents.

  The probability calculation unit 20 functions as a probability calculation unit that calculates the survival rate of a single example for the survival information of the examinee based on the gene expression profile data input by the input unit 10. The probability calculation unit 20 includes an estimator 21 that can perform binary classification processing with probability output by supervised machine learning. As will be described later, a plurality of estimators 21 are provided at predetermined time points.

  Here, the word “supervised machine learning” means that machine learning is performed using a plurality of learning data sets, that is, pairs of learning inputs and outputs (answers), and the learning results are used. It is a technique to make an appropriate output for an unknown input. The wording “binary classification process with probability output” classifies a state where a specific phenomenon has occurred and a state where it does not occur into binary values, and outputs the classification result as a probability. For example, the classification result may be a process in which “a specific phenomenon occurs with a probability of 70%” or “a specific phenomenon does not occur with a probability of 30%”.

As a specific method of supervised machine learning of the binary classification with probability output, a conventionally known method can be suitably used and is not particularly limited. For example, as shown in the examples described later, a combination of gene selection based on Fisher scores, one-dimensional linear discriminant analysis, weighted voting, and logistic regression can be used. In addition to this, for example, a conventionally known method such as a K nearest neighbor method, a multilayer perceptron, a Gaussian process regression method, a support vector machine with a probability output, or the like can be used (for example, Pattern Classification (2nd Edition), Richard O Duda, Peter E. Hart, David G. Stork, Wiley-Interscience (October, 2000), ISBN: 0471056693)
The basic concept of the supervised machine learning technique of binary classification with probability output in the present embodiment will be described. Assume that a feature vector xi is given for each example i. If this is a cancer case, the tissue specimen and clinical information acquired at the time of surgery are examined from various viewpoints, and the features are summarized in the form of vector data. In the present embodiment, an expression level (gene expression profile) of an important gene is used as a typical example of a feature quantity.

  First, let us consider the prediction of life and death at the time of 5 years from this feature vector. All cases are classified into 3 types of “survival”, “death”, and “unknown” at 5 years from the survival data. Of these, the binary label of survival / death excluding “unknown” is used as a learning case. According to the supervised machine learning method, the discriminant function f (x) of the label can be configured using the combination of the feature vector xi of each example i and the binary label li as learning data. The discriminant function is a function that takes a feature vector as an input and outputs a real value, and corresponds to predicting survival if f (x)> 0, and predicting death if f (x) <0.

  The real value of the discriminant function not only indicates the label, but also the absolute value indicates the probability of prediction. For example, when f (x) takes a value close to 0, the probability that the determination by the sign is correct is low, and the larger the absolute value, the higher the correct answer rate of the determination by the sign is generally. Using this property, the real value f (x) of the discriminant function can be converted into a probability value P of 0 or more and 1 or less. For a case in which the feature vector x is observed, the probability value P means the 5-year predicted survival rate of that case.

  Specifically, the estimator 21 according to the present embodiment uses, as learning data, gene expression profile data and survival information (whether or not it is alive about the individual who acquired the gene expression profile data at a predetermined time point). And the correlation between the gene expression profile data and the survival information is obtained by supervised machine learning.

A process of machine learning in the estimator 21 is schematically described as shown in FIG. First, plural (N samples) gene expression profile data x ⁽¹⁾ , x ⁽²⁾ ,... X ^(N) are prepared as learning data. Numbers in superscript parentheses correspond to samples. In addition, survival information at a plurality of predetermined time points is prepared. Here, the “predetermined time point” can be set as appropriate, for example, at the time when half a year has passed or one year has passed after the cancer treatment (surgery). In FIG. 2, the predetermined time points are set such as 0.5 year point, 1 year point, 1.5 year point,...

The survival information at _0.5 years is expressed as t _0.5 ⁽¹⁾ , t _0.5 ⁽²⁾ ,... T _0.5 ^(N), and the survival information at 5.0 years is t _{5. 0} ⁽¹⁾ , t _5.0 ⁽²⁾ , ... t _5.0 ^(N) . ^Further, 0. Survival information 5 years time points ^{x (1)} corresponds with _t ^{0.5 (1),} the survival information 5. 0 years time of ^{x (1)} corresponding with _t ^{5.0 (1)} To do. The “survival information” is information about life and death, and for example, “1” is set for the “live (alive) state” and “0” is set for the “dead (not alive) state”.

  Then, estimators 21 corresponding to the plurality of predetermined time points are prepared. Specifically, as shown in FIG. 2, a 0.5-year time estimator, a 1-year time estimator,... 5 year time estimator, etc. are prepared, and the gene expression profile data described above in each estimator 21. And supervised machine learning with a correlation between the survival information at a predetermined time.

Specifically, as shown in FIG. 2, 0. For the 5-year time estimator, the gene expression profile data ^{x (1), x (2} ), ... and ^{x (N),} 0. 5 years time As survival information, t _0.5 ⁽¹⁾ , t _0.5 ⁽²⁾ ,..., T _0.5 ^(N) are input. Similarly, the gene expression profile data x ⁽¹⁾ , x ⁽²⁾ ,... X ^(N) and the survival information at one year are t _1.0 ⁽¹⁾ , t _1. Enter ₀ ⁽²⁾ ,... T _1.0 ^(N) . This is performed for each estimator corresponding to each predetermined point in time, and the correlation between the gene expression profile and the survival information is trained by machine learning.

  As the supervised machine learning method, a conventionally known method can be suitably used as described above, and is not particularly limited. For example, a univariate linear discriminant function weighted voting method can be used. The processing of the weighted voting method of the univariate linear discriminant function will be described in detail in an embodiment described later.

  According to the estimator 21 that has been machine-learned as described above, when a gene expression profile of an arbitrary individual other than the individual used as the learning data is input, the learning has a correlation with the arbitrary gene expression profile. At the same time as the predetermined time point in the data, the probability information about the survival information of the individual who acquired the gene expression profile can be output.

This process will be schematically shown in FIG. That is, the gene expression profile data x to be analyzed (single example) is input to each of the estimators 21 that have undergone supervised machine learning. Since each of the estimators 21 is machine-learned so that binary classification processing with probability output is possible, the survival information to be analyzed can be output as a probability value (a survival rate is output) at a predetermined time point. Specifically, as shown in FIG. 3, when the gene expression profile data x is input to each of the above-described supervised machine learning estimators 21. The survival rate P _{0.5 at} the year, the survival rate P _1.0 at the year 1 from the 1-year estimator, and the 5-year survival rate P _5.0 are output from the 5-year estimator. The

  FIG. 4 schematically shows a processing procedure in which the estimator 21 calculates the survival rate P using the gene expression profile data x to be analyzed. That is, first, the gene expression profile data x to be analyzed is input to the discriminant function processing unit 21a. The discriminant function processing unit 21a receives the gene expression profile data as a feature vector and outputs a real value using the discriminant function whose estimation accuracy is increased by the learning data. If the output result is f (x)> 0, it is determined to be alive, and if f (x) <0, it is determined to be dead. The specific method of the discriminant function is not particularly limited. For example, in an example described later, a discrimination function based on “univariate linear discrimination” and “weighted voting method using Fisher score” is used.

A univariate linear discriminant function is a discriminant function using only one component of a feature vector. The simplest form of f _j (x) = Lx _j −b _j for adding a sign to the value of a certain component and performing bias correction. Here, x _j is a j-th component of x, L is a code of +1 or −1, and b _j is a bias component of a scalar value. The code L is determined so that if f _j (x)> 0, it is predicted to be live, and if f _j (x) <0, it is predicted to be dead. In the weighted voting, a discriminant function of the following formula (1) is formed in a form in which each is weighted and summed.

Here, w _j is a real number representing the weight of the jth univariate linear discriminant function. As the specific value of the weight, “the difference between the average values of the j-th component of the feature vector (gene expression profile data) in the living cases and the dead cases normalized by the variance” is used.

  Next, the output result f from the discriminant function processing unit 21a is input to the fP conversion processing unit 21b. The fP conversion processing unit 21b performs fP conversion processing on the output result f, converts it to a probability value P of 0 or more and 1 or less, and outputs the result. For a case in which gene expression profile data x is observed, the probability value P means the predicted survival rate of the case at a predetermined time point. Although the specific method of the fP conversion process is not particularly limited, in this embodiment described later, a logistic regression process (logit conversion) process is performed.

  In the logit conversion process for converting the value f of the discriminant function into the probability value P, the following formula (2) is used.

However, b ₀ and b ₁ are parameters for logit conversion, and optimum values are set based on the data.

  If the gene expression profile data x to be analyzed is input to the estimators 21 obtained by performing supervised machine learning of binary classification with probability output in this way, a plurality of predetermined examples for the analysis target single example are input. Probability output of survival information at the time can be performed.

  As the estimator 21 described above, estimators 21 that have been machine-learned in advance (learned) can be used. In addition, for example, the learning data is input to the estimators 21 through the input unit 10, and after the estimator 21 is machine-learned in the lifetime analysis apparatus 100, the data to be analyzed is input. Survival analysis can also be performed. In particular, the estimators 21... Improve the reliability of processing by supplementing, adding, and exchanging learning data as needed and repeating machine learning as appropriate. For this reason, it is more preferable that the estimators 21... Input the learning data via the input unit 10 and repeatedly perform machine learning appropriately.

  Note that the “time when feature data is acquired” in the present embodiment refers to a time when a biological sample is collected from a patient, and is not a time when gene expression profile data is acquired from the biological sample. That is, in the present embodiment, the time when the feature data is acquired is important when the biological sample is collected, not when the gene expression profile data is acquired. This is true for both learning data and feature quantity data to be analyzed.

  In the survival time analysis apparatus according to the present embodiment, from the time when the biological sample is collected (when the feature amount data is acquired) until the death (predetermined phenomenon occurs) and / or without the death (the phenomenon is This is a configuration in which the time until the observation is finished is regarded as the “lifetime”. Therefore, the time point when the biological sample is collected from the patient as the learning data and the time point when the biological sample is collected from the patient to be analyzed may be arbitrary. That is, the survival rate at each predetermined time point measured from an arbitrary time point when a biological sample is collected from the patient to be analyzed can be calculated.

  The curve creation unit 30 functions as a curve creation unit that creates a survival rate curve for a single patient to be analyzed using the survival rate values at a plurality of time points calculated by the probability calculation unit 20. . “Single case survival curve (survival curve)” is the survival curve for one sample (single case). The vertical axis shows the survival rate and posterior (posterior probability) of the sample. Is a graph representing the survival time. Note that the term “curve” in the present specification includes a line graph expressed by simply connecting a plurality of points.

  Specifically, for example, the curve creation unit 30 graphs the survival rate at a plurality of time points calculated by the probability calculation unit 20 with the vertical axis representing the survival rate and the horizontal axis representing time. As a method for creating such a curve, a conventionally known method can be suitably used, and the specific configuration is not particularly limited. For example, it can be graphed by connecting the survival rate values at each predetermined time point and expressing it as a curve.

  The expected value calculation unit 40 uses the survival rate curve created by the curve creation unit 30 to calculate the expected value of the survival time of each individual patient to be analyzed.

  The “expected value” in this specification is a basic term in statistics and probability theory, and is not particularly limited. For example, it is expected as an average value of infinite trials under the same conditions. Refers to the value to be set. For example, the expected value of the dice roll is 3.5.

  For example, the expected value calculation unit 40 can calculate the expected value of the survival time as follows. First, a survival probability is predicted for a patient whose observation value x is obtained, and a single case survival curve is drawn.

  Next, based on this survival curve, a specific survival period is predicted, but since it is originally a probabilistic prediction, various survival periods can be assumed probabilistically. For example, when there are a plurality of patients who have obtained the same observation value x, for example, the survival time up to the time of death is 3, 4, 5, 5, 5, 6, 10, 3, 5, 4, 4,. I can imagine something like a year.

  Subsequently, when a survival period (a normal survival curve for a plurality of cases instead of a single case survival curve) is drawn based on the assumed survival period data, it coincides with the first case survival curve created. In other words, the lifetime data is assumed so as to coincide with the single case survival curve.

  Finally, when the survival time data is taken so that the number of patients becomes infinite, the average value of the survival time is calculated as the expected value of the survival time.

Specifically, the expected value calculation unit 40 can calculate the expected value of the lifetime as follows. For example, for a case x, the estimated survival probabilities at the elapsed time points _0.5 , _1.0 ,..., _5.0 years are P _0.5 , P _1.0 ,. . At this time, by approximating the survival probability when the survival period y exceeds _5.0 years from the elapsed time by P _5.0 exp (−λ (y−5.0)), the survival curve as shown in FIG. Is obtained. This is expressed as g (y) as a function of the period y.

  Since the probability that a death event (phenomenon) occurs from time y to y + dy can be expressed as (g (y) −g (y + dy)) / dy, the expected value E [y] of the lifetime is expressed by the following formula (3) ).

Although there is a portion where g ′ (y) <0 in the estimated survival curve, there is no problem because the influence disappears after integration. Further, the above integration can be easily calculated by the sum of numerical integration at 0 <y <5.0 and exponential integration at 5.0 <y <∞. Also, λ, which is a parameter for determining the exponential distribution when y> 5.0, is λ = P _5.0 /5.0. This assumes that mortality similar to the average mortality in the period of 0 <y <5.0 will continue with y> 5.0.

  The output unit 50 functions as a display unit that displays the survival rate curve created by the curve creating unit 30. As the display means, a conventionally known CRT display, liquid crystal display, or the like can be used as appropriate. The output unit 50 may be configured to display a plurality of survival rates, survival rate curves, and expected values calculated by the probability calculation unit 20 singly or in combination as appropriate. When displaying data, for example, a publicly known user authentication process may be added so that browsing is not possible unless the viewer inputs a user name and a password.

  Furthermore, it is possible to print the calculated survival rate, survival rate curve, and the like of the examinee at each time point on paper using a conventionally known printing unit such as a printer. At that time, on the paper on which the survival rate and survival rate curves of the examinee at each time point are printed, for example, on a postcard, etc., the prescribed text used for administrative communication and the address and name of the examinee are printed at the same time. So that it can be shipped. In this case, in order to prevent the printed information from being viewed by a third party during the delivery process to the patient, it may be possible to add paper to conceal the printed surface or send it in a sealed letter. Good. Also in this printing process, user authentication and access right management may be introduced in the same way as the display procedure so that only a specific operator such as a person in charge of printing work can print.

  Examples of the output of the survival rate curve created by the curve creation unit 30 by the output unit 50 include a chart as shown in FIG.

  FIG. 8 shows the results of prognosis prediction of each case by microarray measurement of 136 cases of neuroblastoma using the survival time analysis apparatus 100 according to this embodiment, as shown in the examples described later. Is.

  The left panel of FIG. 8 shows the case information of each case. In particular, a solid bar graph (red bar graph in the case of color) indicates a death example, and a broken line bar graph (blue bar graph in the case of color) indicates a survival period. The two lines on the left show information on survival (black circle (blue for color)) and death (white circle (red for color)) at 2 years and 5 years. The 6 lines on the right side are cancer prognostic markers (see Examples below for details), and white (red in the case of color) circles indicate that the prognostic marker is poor (not good), A black dot (blue in the case of color) indicates that “the prognostic marker is good”.

  Further, the right panel in FIG. 8 is a line graph showing a single case survival curve of each case. Specifically, the survival rate is calculated in 0.5-year increments and shown as a line graph. Information on survival (black circle (blue for color)) and death (white circle (red for color)) Is also shown.

  As shown in the figure, it can be seen that the results of prognosis prediction by the survival time analysis apparatus 100 and the actual survival information are in good agreement.

  For example, it can be seen from the left panel that samples S001 and S022 have a good prognosis and survive for a long time. On the other hand, in the right panel showing the prognosis prediction by the survival time analysis device, the result that the risk of death is almost zero for 5 years is obtained for both S001 and S022, and the actual situation and the prediction are almost the same. I understand.

  In addition, sample S014 has a good prognosis at present, but in the prognosis prediction by the survival time analyzer, it is predicted that the mortality rate slightly increases from around the 4th year.

  In addition, samples S057 and S078 have a good prognosis at present, but in the prognosis prediction by the survival time analyzer, it is predicted that the risk of death will increase at a constant pace.

  Samples S108 and S109 have died in 2 to 3 years, respectively, and the prognosis is not good. On the other hand, in the prognosis prediction by this survival time analyzer, the risk of death is certainly increasing at a constant pace. From this result, it can be seen that the actual situation and the prediction almost coincide.

  Samples S114 and S133 have a very poor prognosis and die at a very early stage after the operation. On the other hand, in the prognosis prediction by this survival time analyzer, it is predicted that there is indeed a high risk of death at the very beginning. From this result, it can be seen that the actual situation and the prediction almost coincide.

  In addition, both samples S199 and S122 have a poor prognosis and die within 6 months to 1 year after the operation. On the other hand, in the prognosis prediction by the survival time analyzer, it is certainly predicted that there is a high risk of death after 6 months. From this result, it can be seen that the actual situation and the prediction almost coincide.

  In addition, the sample S118 is alive at the present time with a poor prognosis. However, according to the prognosis prediction by the survival time analyzer, there is a high risk of death in the first 6 months, but the risk remains low after that point. This also shows that the actual situation and the prediction are almost the same.

  Next, an example of a specific processing flow of the survival time analysis apparatus 100 according to the present embodiment will be described with reference to FIGS. First, supervised machine learning in the estimator 21 will be described, and then survival time analysis processing in the survival time analysis apparatus 100 will be described.

  FIG. 6 shows an example of a processing flow of supervised machine learning. First, as shown in the figure, a plurality of gene expression sample data (gene expression profile data) and case survival data (survival information) are input to the estimator 21.

  Subsequently, in the estimator 21, supervised machine learning (binary classification with probability output) processing is performed on the correlation between the gene expression profile data and the lifetime. The specific content of supervised machine learning is as described above.

  In this way, the estimator 21 that has undergone supervised machine learning of binary classification with probability output can be acquired.

  FIG. 7 shows an example of a processing flow of survival time analysis in the survival time analysis apparatus 100. First, as shown in the figure, the input unit 10 inputs the gene expression profile data of the diagnosis subject (analysis target) to the probability calculation unit 20 (S1). As described above, this process can be performed by the user via the input unit 10.

  Next, the probability calculation unit 20 inputs and processes the gene expression profile data of the diagnosis target (analysis target) to the estimators 21 provided at predetermined time points (S2). That is, the estimator 21 for each predetermined time point calculates the survival rate of the person to be diagnosed at a predetermined time point corresponding to each learned time point. The specific content of the survival rate calculation is as described above.

  Subsequently, the probability calculation unit 20 outputs the survival rate calculated by the plurality of estimators 21 to the curve creation unit 30 (S3).

  Next, the curve creation unit 30 creates a survival rate curve for a plurality of survival rates (S4). The specific content for creating the survival rate curve is as described above.

  Then, the expected value calculation unit 40 calculates the expected value of the survival time (survival rate) of the patient to be analyzed at an arbitrary time point using the survival rate curve (S5). The specific content for calculating the expected value is as described above.

  Finally, the curve creation unit 30 outputs the created survival rate curve and / or the expected value to the output unit (S6), and ends the process.

  As described above, according to the survival time analysis apparatus 100 according to the present embodiment, it is possible to calculate a unique survival rate for each case and further create a survival rate curve. The conventional survival time analysis apparatus / method employs a method of selecting and outputting a survival rate curve calculated in advance for each layer from a database. For this reason, for each patient (each case), the survival time is evaluated based only on the high / low risk, and a thorough risk analysis, especially on the time axis, should be performed. I could not.

  On the other hand, this survival time analyzer creates a unique (single case) survival rate curve for each case. The single case survival curve created for each case reflects the risk change in each period after diagnosis of the disease. For this reason, more detailed time series risk prediction can be obtained, which can be a new standard for clinical judgment. In other words, it is possible to analyze when the risk of each examinee fluctuates greatly, which is very useful.

  Such a single case survival rate curve can be used for, for example, customized medicine. In other words, when custom-made medical treatment is applied clinically, more detailed risk analysis is required for each case. However, according to this survival time analysis device, a detailed process for generating such risks can be accurately performed for each case. Can be analyzed. In particular, this survival time analysis apparatus is characterized by high time resolution that can analyze a risk on a time axis. Thus, when the time resolution is high, the difference between the risk types can be analyzed more accurately, and the difference between the cases can be accurately grasped.

  Further, in the present embodiment, as the gene expression profile data, data obtained by quantifying the expression level of a predetermined gene is used as feature data. At this time, the type and number of genes to be used, the method for analyzing the gene expression level, the numerical processing of the analysis results, etc. can be appropriately set within a reasonable range based on conventionally known methods and technical common sense. It is not limited. For example, in the examples described later, for 5340 genes, the gene expression profile was analyzed using a microarray, and the data obtained as a result of measuring the logarithmic ratio of the gene expression level between the control cells and the cells of the analysis target was obtained. Used as gene expression profile data.

  Further, in this embodiment, the prognostic survival rate of cancer patients is analyzed, but the specific types of cancer are not particularly limited, and the same applies to cell proliferative diseases such as conventionally known cancers and tumors. Can be done. “Cell proliferative diseases” are various diseases resulting from the proliferation of cells in an uncontrollable state due to abnormal cell cycle, such as squamous cell carcinoma, lung cancer (including small cell lung cancer), gastric cancer, etc. List various cancers and tumors such as liver cancer, breast cancer, esophageal cancer, bladder cancer, prostate cancer, colon cancer, kidney cancer, brain tumor, retinoblastoma, osteosarcoma, neurofibromatosis, malignant melanoma, leukemia Can do. In particular, it is preferable to analyze malignant tumors, but the present invention is not limited to this, and benign tumors can also be analyzed.

  In the above description, human (human) is described as an object. However, the present invention is not limited to this, and various types such as rat, mouse, rabbit, monkey, goat, sheep, pig, horse, and cow are used. The same analysis can be performed for mammals. In particular, since rats, mice, rabbits, monkeys and the like are used as experimental animals, survival time analysis of these is very useful in the field of drug development and the like.

  In addition, the present survival time analysis apparatus can use not only the above-described gene expression profile data but also other biological data and pathological diagnosis data as feature data. The term “biological data” here is not particularly limited as long as it is biological / physiological data that can be acquired from the examinee, such as genetic data and pathological diagnosis data. “Genetic information” includes gene expression profiles obtained as a result of transcriptome analysis, such as chromosome information (information on genetic predisposition to disease), gene polymorphisms, SNPs (single nucleotide polymorphisms) information, etc. Various information such as information on protein structure obtained by proteome analysis, such as inter-gene interactions (related to transcription promotion and transcription repression), information on protein interactions, information related to post-translational modifications such as sugar chain modifications to proteins Can be mentioned.

  Further, the term “pathological diagnosis data” refers to data obtained from the results of an inquiry by a doctor or a test result at a hospital. Interview results are information on the health status of the patient, mainly obtained by filling out the patient's questionnaire and questions from the doctor to the patient. The patient's medical condition, illness being treated, and subjective symptoms (palpitations, swelling, and fatigue) Etc.), diet (type, intake, regularity, etc.), exercise (type, intensity, frequency, etc.), smoking (presence / absence, amount of smoking, years of smoking, past smoking history, smoking cessation period, etc.), alcohol consumption (frequency) , Type, amount of alcohol, etc.) and daily life habits such as work (work contents, working hours, return time, etc.). Test results are judgments obtained mainly by examination equipment and doctors, such as body measurements (height, weight, obesity, etc.), visual acuity, blood pressure, pulse, urinalysis, blood tests (white blood cell count, red blood cell count) , Liver function test, lipid metabolism, gout test, sugar metabolism, etc.), chest X-ray examination, digestive organ X-ray examination, electrocardiogram, abdominal echo examination, dental examination, and other pathological samples collected during surgery And various pathological data (gene expression analysis, immunostaining results, etc.) obtained by using a biological method, a cell biological method, and an immunological method.

  The biological data and pathological diagnosis data described above can be acquired by a conventionally known method, and the specific acquisition means is not particularly limited. For example, a gene expression profile can be easily obtained by using a conventionally known analysis method such as a microarray holding oligo DNA such as GeneChip (registered trademark) manufactured by Affimetrix, cDNA (ORF) microarray, oligo macroarray, macroarray, etc. Can be obtained. Furthermore, necessary information can be acquired using a conventionally known genome database, SNPs database, expression profile database, or the like. Such a technique can be easily executed by those skilled in the art.

  Further, the temporal phenomenon occurrence analysis by the temporal phenomenon occurrence analysis apparatus according to the present invention is not only about the survival and death of living organisms, but also about the occurrence of cell proliferative disease metastasis, disease morbidity, industrial product failure, etc. The same can be done.

  According to the temporal phenomenon occurrence analysis of “metastasis of cell proliferative disease”, for example, it is possible to analyze the risk of when cancer or tumor metastasis occurs. In addition, according to the analysis of occurrence of a phenomenon with respect to “morbidity” over time, it is possible to analyze at what time the disease develops and the like, which is very useful particularly in terms of preventive medicine. The type of “disease” to be analyzed is not particularly limited. For example, diabetes, cerebral infarction, myocardial infarction, arteriosclerosis, hyperlipidemia, osteoporosis (and fracture associated therewith). The risk of onset can be predicted for conventionally known diseases such as indirect rheumatism.

  In addition, in the analysis of the occurrence of phenomena over time regarding “metastasis of cell proliferative disease” and “morbidity of disease”, data such as the above-described biological data, pathological diagnosis data, gene expression profile, etc. are suitable as feature data. Can be used.

  Further, according to the analysis of the occurrence of a phenomenon with respect to “failure of an industrial product”, it is possible to analyze at which point the failure of the industrial product occurs and at which point the product life is exhausted. The specific types of industrial products are not particularly limited. For example, in addition to simple daily miscellaneous goods, electric appliances such as TVs, refrigerators, and microwave ovens, and various devices such as automobiles, watches, and computers. Thus, it is possible to analyze the occurrence of phenomena over time for the lifetime of conventionally known industrial products.

  As the feature amount data used in the analysis of occurrence of the phenomenon of “industrial product failure” over time, various data relating to the lifetime of the industrial product can be suitably used, and is not specifically limited. For example, in addition to information such as the number of parts, shape, size, material, usage status, and years of use, information such as date of manufacture, country of manufacture, manufacturer name, and factory can be used.

  Further, by using the temporal phenomenon occurrence analysis device according to the present invention, for example, a temporal phenomenon occurrence analysis system (for example, a survival time analysis system) via a communication network can be developed. A temporal phenomenon occurrence analysis service (for example, a survival time analysis service) using a dynamic phenomenon occurrence analysis system can also be implemented.

  Of the temporal phenomenon occurrence analysis system and the temporal phenomenon occurrence analysis service, an example of the survival time analysis service using the survival time analysis system will be described. First, in a medical institution, feature amount data (for example, gene expression profile) is acquired from a biological sample collected at the time of cancer surgery. Thereafter, the acquired feature amount data is transmitted to a survival time analysis center installed at a location different from the medical institution via a communication network. The survival time analysis center is provided with a plurality of survival time analysis apparatuses according to the present invention, and analyzes the survival time of the person to be analyzed using the feature data transmitted from the medical institution. The result is transmitted to a medical institution or a direct analysis target person via a communication network. In the case of such a service, it is preferable to take various known security measures in order to avoid, for example, information leakage.

  Such a survival time analysis system or survival time analysis service includes a temporal phenomenon occurrence analysis device (survival time analysis device) according to the present invention and an arithmetic device or terminal (for example, a personal computer, Servers, routers, etc.) can be easily constructed. Therefore, the present invention includes such a survival time analysis system and survival time analysis service, that is, a temporal phenomenon occurrence analysis system and a temporal phenomenon occurrence analysis service.

  Finally, each block of the above-described survival time analysis apparatus 100, in particular, the probability calculation unit 20, the curve creation unit 30, and the expected value calculation unit 40 may be configured by hardware logic, and uses a CPU as follows. It may be realized by software.

  That is, the survival time analysis apparatus 100 includes a CPU (central processing unit) that executes instructions of a control program that realizes each function, a ROM (read only memory) that stores the program, and a RAM (random access memory) that expands the program. ), A storage device (recording medium) such as a memory for storing the program and various data. An object of the present invention is a recording medium in which the program code (execution format program, intermediate code program, source program) of the control program of the lifetime analysis apparatus 100, which is software that realizes the functions described above, is recorded so as to be readable by a computer. Can also be achieved by reading the program code recorded on the recording medium and executing it by the computer (or CPU or MPU).

  Examples of the recording medium include tapes such as magnetic tapes and cassette tapes, magnetic disks such as floppy (registered trademark) disks / hard disks, and disks including optical disks such as CD-ROM / MO / MD / DVD / CD-R. Card system such as IC card, IC card (including memory card) / optical card, or semiconductor memory system such as mask ROM / EPROM / EEPROM / flash ROM.

  Further, the lifetime analysis apparatus 100 may be configured to be connectable to a communication network, and the program code may be supplied via the communication network. The communication network is not particularly limited. For example, the Internet, intranet, extranet, LAN, ISDN, VAN, CATV communication network, virtual private network, telephone line network, mobile communication network, satellite communication. A net or the like is available. Also, the transmission medium constituting the communication network is not particularly limited. For example, even in the case of wired such as IEEE 1394, USB, power line carrier, cable TV line, telephone line, ADSL line, etc., infrared rays such as IrDA and remote control, Bluetooth ( (Registered trademark), 802.11 wireless, HDR, mobile phone network, satellite line, terrestrial digital network, and the like can also be used. The present invention can also be realized in the form of a computer data signal embedded in a carrier wave in which the program code is embodied by electronic transmission.

  Hereinafter, examples will be shown, and the embodiment of the present invention will be described in more detail. Of course, the present invention is not limited to the following examples, and it goes without saying that various aspects are possible in detail. Furthermore, the present invention is not limited to the above-described embodiments, and various modifications can be made within the scope shown in the claims, and the embodiments obtained by appropriately combining the respective technical means disclosed are also included in the present invention. It is included in the technical scope of the invention.

  In this example, survival time data and microarray gene expression level data were prepared for 136 cases of neuroblastoma.

  Survival data is a collection of information on the survival period (in units of one month) for each case and whether or not the death of the case has been confirmed.

  The gene expression data by the microarray is obtained by measuring the log ratio of the expression level between the control cell and the target cell for 5340 genes in each case.

  Hereinafter, the expression level of the i-th gene in the j-th case is assumed to be Xij.

  The n-year survival probability estimator outputs the probability that a specific case will survive for n years. Specifically, an n-year survival probability estimator is configured for each of a total of 10 time points of n = 0.5, 1.0,..., 5.0 based on the database.

  That is, when gene expression data by a microarray is obtained for a new case, the output of the n-year survival probability estimator is set to each of a total of 10 time points of n = 0.5, 1.0,. To obtain a single case survival curve.

  To construct the n-year survival probability estimation, a binary classification method with a probability output by machine learning is used. As a method for this, in this embodiment, a combination of gene selection based on Fisher scores, one-dimensional linear discriminant analysis, weighted voting, and logistic regression is used.

  For cases where survival or death is confirmed at n years, label survival = (1), death = (0). Let D0 be a set of death cases and D1 be a set of survival cases.

[A] Gene selection by Fisher score The Fisher score ci of the i-th gene is defined by the following equation (4).

  However, the following mathematical formula (5) is the average expression level of each of the death cases and survival cases.

  The following formula (6) represents the magnitude (standard deviation) of noise. Note that # D0 and # D1 represent the number of elements of D0 and D1, respectively. Moreover, ci is also called an S / N ratio (signal to noise ratio) with another name.

  A gene with a higher Fisher score has a stronger meaning in separating death and survival cases, and a gene with a lower Fisher score is considered irrelevant in determining survival and death. . Therefore, Ntop genes from the top of the absolute value of the Fisher score are used for the following analysis. This is called gene selection. A set of genes selected here is written as G.

[B] One-dimensional linear discriminant analysis Let x be the vector of gene expression level data obtained from this new case. The i-th gene component is xi. At this time, the one-dimensional linear discriminant function fi (x) relating to the expression level of the gene i is defined by the following formula (7).

  Case x is estimated to be alive when fi (x)> 0, and dead when fi (x) <0. This estimation is based on only gene i. Accuracy can be improved by performing estimation based on a plurality of genes. The method for that is the weighted voting method described below.

[C] Weighted voting method A discriminant function F (x) by weighted voting is defined as the following formula (8).

  This is a vote that intentionally has one vote of disparity (weight) with respect to judgment by one-dimensional linear discriminant analysis.

[D] Logistic regression (logit conversion)
The probability p (x) that the case x will survive for n years is estimated using the value of F (x) as in the following formula (9).

  This is called a logistic model.

Β and α are determined as follows.
(I) Prepare database Xij and n-year survival death label (ii) Configure discriminant function F (x) by weighted voting based on the database (iii) n-year survival death label corresponding to different database Xik Prepare. Then, the survival case is D1 ′ and the death case is D0 ′. (Iv) The output Fk as a result of inputting Xik to F (x) is calculated. (V) The gradient method using the following formula (10) is used. Is obtained by maximizing by β and α.

[E] Demonstration by Leave One Out In order to evaluate the performance of the n-year survival probability estimator based on the above procedure, Leave One Out (LOO) analysis was performed on 136 cases. Specifically, the procedure was as follows.
(I) One case is extracted from 136 cases for testing, and the remaining 135 cases are used as learning cases. (Ii) An n-year survival probability estimator is configured based on the 135 learning cases. (Iii) One test case. Is input to the n-year survival probability estimator, and the probability p (x) is calculated and output. (Iv) The above (i) to (iii) are repeated for each of 136 cases extracted for testing (v). (Vi) Perform (i) to (v) above for n = 0.5, 1.0,..., 5.0 (vii) Draw a single case survival curve for each case The reason why the LOO analysis is required is to prevent an information leakage problem that an unpredictably good prediction performance is obtained by putting a case to be estimated into a learning target.

  Although the details of the experiment conducted in this example are not shown, it should be noted that the paper is currently being submitted and that its accuracy is extremely reliable.

  As described above, the survival time analysis apparatus according to the present invention can be used for quality control by estimating the survival time and the time of failure of industrial products in disease treatment prognosis for living organisms such as humans. For this reason, this invention has wide industrial applicability, such as quality control of the whole medical field and industrial products.

It is a figure which shows the functional block of the survival time analyzer which concerns on this Embodiment. It is a figure which represents typically the basic concept of the process of machine learning in the estimator of this Embodiment. It is a figure which represents typically an example of the process which calculates a survival rate in the estimator of this Embodiment. It is a figure which illustrates typically the process sequence which the estimator of this Embodiment calculates a survival rate using the gene expression profile data of analysis object. It is a figure which shows an example of the survival curve in this Embodiment. It is a figure which shows typically about an example of the processing flow of the machine learning of the estimator used for the lifetime analysis apparatus which concerns on this Embodiment. It is a figure which shows typically about an example of the processing flow of the lifetime analysis apparatus which concerns on this Embodiment. In a present Example, it is a figure which shows the single example survival rate curve created with the survival time analyzer.

Explanation of symbols

10 Input section (input means)
20 Probability calculator (probability calculator)
21 Estimator 21a Discriminant Function Processing Unit 21b f-P Conversion Processing Unit 30 Curve Creation Unit (Curve Creation Unit)
40 Expected value calculation unit (expected value calculation means)
100 Survival time analysis device

Claims (12)

An input means for inputting feature data obtained from the analysis target;
Probability calculation means for calculating the occurrence probability of a predetermined phenomenon for the analysis target based on the feature amount data input by the input means,
The probability calculation means is:
As learning data, a plurality of sets of feature amount data and phenomenon information on whether or not the predetermined phenomenon has occurred at a predetermined elapsed time from the time when the individual who acquired the feature amount data is earned are used. , An estimator obtained by supervised machine learning of the correlation between the feature data and the phenomenon information,
When feature amount data of an arbitrary individual at any time other than the individual used for the learning data is input, a predetermined elapsed time in the learning data from the time point correlated with the arbitrary feature amount data And an estimator that predicts whether or not the predetermined phenomenon occurs in the individual that acquired the feature amount data, and outputs the probability.
The probability calculation means includes a plurality of estimators corresponding to a plurality of predetermined elapsed time points, and calculates a probability that a predetermined phenomenon occurs for the analysis target at a plurality of time points corresponding to the estimator. An apparatus for analyzing the occurrence of a phenomenon over time.   Further, a curve creation means for creating a phenomenon occurrence probability curve over time from the time of acquisition of feature amount data in an arbitrary analysis target, using the occurrence probability values at a plurality of elapsed times calculated by the probability calculation means. The temporal phenomenon occurrence analysis apparatus according to claim 1, comprising:   The temporal phenomenon according to claim 1 or 2, further comprising an expected value calculation means for calculating an expected value of a time that elapses until the predetermined phenomenon occurs using the phenomenon occurrence probability curve. Generation analysis device. The measurement object is a living organism, the feature amount data is biological data,
The occurrence of the predetermined phenomenon is the death of the organism to be analyzed, the morbidity of the disease, or the metastasis of the cell proliferative disease, or the occurrence of the temporal phenomenon according to any one of claims 1 to 3. Analysis device.   5. The temporal phenomenon occurrence analysis apparatus according to claim 4, wherein the biological data is pathological diagnosis data of an organism.   6. The temporal phenomenon occurrence analysis apparatus according to claim 4, wherein the biological data is a gene expression profile of an organism.   7. The temporal phenomenon occurrence analysis apparatus according to any one of claims 4 to 6, wherein the biological data relates to a prognosis of a cell proliferative disease. The estimator is a discriminant function whose estimation accuracy is increased by the learning data, and whether or not a predetermined phenomenon occurs using a discriminant function that receives the feature quantity data to be analyzed and outputs a real value. A discriminant function processing unit for calculating
2. An f-P conversion processing unit that performs an f-P conversion process on an output value f from the discriminant function processing unit to calculate a probability of occurrence of a predetermined phenomenon. 8. The temporal phenomenon occurrence analysis device according to any one of items 7 to 9. The discriminant function in the discriminant function processor uses one-dimensional linear discriminant analysis and a weighted voting method,
9. The temporal phenomenon occurrence analysis apparatus according to claim 8, wherein the fP conversion processing in the fP conversion processing section uses logistic regression. An input process for inputting feature data obtained from the analysis target;
A probability calculating step of calculating a probability of occurrence of a predetermined phenomenon for the analysis target based on the feature amount data input by the input means,
The probability calculation step is
As learning data, a plurality of sets of feature amount data and phenomenon information on whether or not the predetermined phenomenon has occurred at a predetermined elapsed time from the time when the individual who acquired the feature amount data is earned are used. , An estimator obtained by supervised machine learning of the correlation between the feature data and the phenomenon information,
When feature amount data of an arbitrary individual at any time other than the individual used for the learning data is input, a predetermined elapsed time in the learning data from the time point correlated with the arbitrary feature amount data A step of using an estimator that predicts and outputs a probability as to whether or not the predetermined phenomenon occurs in the individual that acquired the feature data,
The probability calculating step calculates a probability of occurrence of a predetermined phenomenon for the analysis target at a plurality of elapsed time points corresponding to the estimator using a plurality of the estimators corresponding to a plurality of predetermined elapsed time points. A method for analyzing the occurrence of a phenomenon over time, characterized in that:   Furthermore, a curve creation step of creating a phenomenon occurrence probability curve over time from the time of acquiring feature amount data in an arbitrary analysis target using the occurrence probability values at a plurality of elapsed times calculated by the probability calculation step The method for analyzing the occurrence of a phenomenon over time according to claim 10.   12. The temporal phenomenon according to claim 10 or 11, further comprising an expected value calculation step of calculating an expected value of a time elapsed until the predetermined phenomenon occurs using the phenomenon occurrence probability curve. Occurrence analysis method.

Images