TW201303618A - Method and computer program product of producing a model for use in predicting to an event, apparatus for determining time-to-event predictive information, method and computer program product for determining a predictive diagnosis for a patient - Google Patents

Abstract

A method of producing a model for use in predicting time to an event includes obtaining multi-dimensional, non-linear vectors of information indicative of status of multiple test subjects, at least one of the vectors being right-censored, lacking an indication of a time of occurrence of the event with respect to the corresponding test subject, and performing regression using the vectors of information to produce a kernel-based model to provide an output value related to a prediction of time to the event based upon at least some of the information contained in the vectors of information, where for each vector comprising right-censored data, a censored-data penalty function is used to affect the regression, the censored-data penalty function being different than a non-censored-data penalty function used for each vector comprising non-censored data.

Description

A method for generating a model for estimating the time of occurrence of an event, a computer program product, a device for determining an estimated time of occurrence of an event, a method for determining a patient's estimated diagnosis, and a computer program product

The present invention relates to time-to-event analysis, and more specifically to time-event analysis for right-bound data.

In many cases, we want to estimate the likelihood of an event occurring (initiating and/or recurring) and/or the amount of time it seems to have occurred during a certain period of time. For example, in the medical field, this would be used to predict the recurrence, if any, of a patient who has been treated for a particular disease. Mathematical models have been developed to accomplish this time-event estimate based on data obtained from actual cases. In the above example, this predictive model can be switched by studying a group of patients who have been treated for a particular disease and indicating that there are no recurrences and recurrences of the patient's common characteristics or "characteristics". By considering the actual time of recurrence in this group of patients, characteristics and characteristic values can also be specified to correlate patients who occur at a particular time. These characteristics can also be used to estimate the recurrence time of a future patient based on the individual characteristics of the patient. This time-event estimate can assist a treating physician in assessing and planning for the event to occur.

The unique feature of time-event data is that the desired event (in this case, the recurrence of the condition) has not been observed. For example, this will happen when one of the patients in the group visits the doctor, but the condition does not recur. The information related to this patient visit is called "right limit" because the part of the data that is wanted is missing (ie, the event is expected, for example, the recurrence of the disease has not occurred). Although the limit data lacks certain information by definition, it would be useful to limit the data to be taken into account when estimating the model, as it provides more information on the parameters used to adopt the model. In fact, time-event data, especially right-limited time-event data, is one of the most common types of data in diagnosis, pharmacy, and biomedical research.

In forming or training mathematical models, we generally want to add as much information as possible from many sources. So, for example, for a health time-event estimate, we generally want to get as much information as possible from as many patients as possible and as much as possible from each patient. However, with regard to this large amount of information, it is difficult to handle all the information as available. Although there are various models, none of them can be fully satisfied to process high-dimensional heterogeneous data sets containing right-limit data. For example, the Cox proportional hazard model is a known model used to analyze the limits of the indicated data due to differences in patient characteristics, which is in the entire structure, assuming that the failure rate of any two patients is proportional and the patient The independent nature affects this hazard in a variety of ways. Although the Cox model can properly handle right-limit data, the Cox model is not ideal for analyzing high-dimensional data sets because it is limited by the total regression degree of freedom in the model and if a complex model is processed, A sufficient number of patients. In the other direction, support vector machines (SVMs) perform well on high-dimensional data sets, but they are not suitable for setting data.

In general, in one aspect, the present invention provides a method of generating a model for estimating an event occurrence time, the method comprising: obtaining a representation a multi-dimensional nonlinear vector of information about the state of the test object, at least one vector being a right-limit, which is related to the test object, lacking a representation of the time at which the event occurred; and using a vector of information to perform regression to produce a kernel-based The model provides an output value for an estimated time of the event based on at least part of the information contained in the information-based vector, wherein for each vector containing the right-limit data, a limit data penalty function is used to influence In the regression, the limit data penalty function is different from the non-limit data penalty function for each vector containing non-restricted data.

Implementations of the invention may include one or more of the following features. The regression includes support vector machine regression. The limit data penalty function has a larger positive slack variable than the non-limit data penalty function. Performing regression involves using a majority penalty function that includes a linear function of the difference between the estimated value of the model and the target value of the estimate, and the linearity used to estimate the positive difference between the penalty function and the target value. The first slope of the function is lower than the second slope of the linear function for the positive difference between the predicted and target values of the non-limited data penalty function. The first slope is substantially equal to the linearity of the third slope of the linear function used to limit the negative difference between the estimate and the target value of the data penalty function and the linear difference between the estimate and the target value for the non-limited data penalty function The fourth slope of the function, and the positive and negative relaxation variables of the unrestricted data penalty function are substantially equal to the negative relaxation variable of the limit data penalty function.

Implementations of the invention may also include one or more of the following features. The vector data is based on at least one feature of the data, related to the classification, and has the ability to assist the model to provide an output value, so that the output value assists in estimating the time-event, and the method further includes: using the most Subdivision of information Classify the vector data that is least like the classification of the data to perform regressions to assist the model, providing output values so that the output values assist in estimating the time to the event. The at least one feature is at least one of reliability and predictive power. Regression is performed in a positively greedy manner based on the characteristics of the data to select the characteristics to be used in the model. The method further includes performing a negative greedy program on the characteristics of the vector after performing the regression to further select characteristics to be used in the model. This regression is performed in a positive greedy direction only with respect to a part of the characteristics of the vector. These vectors contain clinical/histopagoographic data, biomarkers, and data classification of biometric data, and the regression system is only performed in a positively greedy manner for biomarker data of biomarkers and vectors. The vector of information represents the state of the test object of at least one of the living body, the pre-living body, and the inanimate body.

In general, in another aspect, the present invention provides a computer program product that generates a model for estimating the time of an event, the computer program product being internal to a computer readable medium, the computer program product Computer-readable, computer-executable instructions for causing a computer to: obtain a multi-dimensional, non-linear vector of information indicative of the state of a majority of the test object, at least one of which is a right-limit, lacking an indication of the time at which the event of the test object occurred And using the vector of information to perform regression to generate a kernel-based model to provide an output value for the estimated time of occurrence of the event based on at least some of the information contained in the vector of information, where for each containment The vector of limited data, a limit data penalty function is used to influence the regression, and the limit data penalty function is different from the non-limit data penalty function used for each vector containing the non-limited data.

Implementations of the invention include one or more of the following features. The regression contains support vector machine regression. The limit data penalty function has a larger positive slack variable than the owner of the non-limited data penalty function. The instruction to cause the computer to perform the regression includes instructions to cause the computer to use a penalty function that includes a linear function of the difference between the predicted value of the model and the target value for the predicted value, and a pre-limitation of the penalty function The first slope of the linear function that estimates the positive difference between the target values is lower than the second slope of the linear function used for the positive difference between the predicted and target values of the unset data compensation function. The first slope is substantially equal to a third slope of a linear function used to limit the negative difference between the estimate and the target value of the data penalty function and a negative difference between the estimate and the target value for the non-limited data penalty function. The fourth slope of the linear function, and the positive and negative relaxation variables of the non-limited data penalty function are substantially equal to the negative relaxation variable of the limit data penalty function.

Implementations of the invention may also include one or more of the following features. The instructions used to cause the computer to perform regression can cause regression: sequentially use vector data from the most categorical classification to the least categorical data to perform regression to assist the model, provide output values, and make the output values assist Estimate the time of the event. The instruction to cause the computer to perform the regression causes the regression to be performed in a positively greedy manner based on the characteristics of the data to select the characteristics to be used in the model. The computer program product further includes instructions to cause the computer to perform a negative greedy program on the characteristics of the vector after performing the regression to further select characteristics to be used in the model. The instruction that causes the computer to perform a regression toward greedy mode causes the computer regression system to perform only in a positive direction with respect to a part of the characteristics of the vector. These vectors contain clinical/tissue pathology data, Biomarkers, and data classification of biometric data, and the instructions that enable the computer to perform regressions perform regression in a positively greedy manner, so that the computer system only performs positive greedy characteristics selection for biomarkers of biomarkers and vectors.

In general, in one aspect, the present invention provides a method of generating a model for predicting the time to an event, the method comprising: obtaining a multi-dimensional nonlinear vector representing information of a state of a majority of the test object; and using The vector of information is used to perform regression to generate a kernel-based model, and at least part of the information contained in the information-based vector provides an estimated time output value for the event, wherein the vector data is based on at least the data A classification of features associated with the ability of the data assistance model to provide an output value, such that the output value assists in estimating the time to the event, and wherein the regression system is sequentially classified from the most image to the least The data of the vector from the data classification is executed to assist the model in providing the output value so that the output value assists in estimating the time of the event.

Implementations of the invention may also include one or more of the following features. Regression is performed in a positively greedy manner based on the characteristics of the data to select the characteristics to be used in the model. The method further includes performing a negative greedy program on the characteristics of the vector after performing the regression to further select characteristics to be used in the model. This regression is performed in a positive greedy direction only with respect to a part of the characteristics of the vector. The vectors include clinical/histopathological data, biomarkers, and data classification of biometric data, and wherein the regression is performed in a non-positive greedy manner for clinical/biological histopathological data, and vector-only biomarkers Biophotographic data Execution. At least one vector is right-defined and lacks a time indication of the occurrence of the test object.

In general, in another aspect, the present invention provides a computer program product for generating a model for estimating an event time, the computer program product being readable on a computer readable medium and including a computer readable Computer executable instructions to cause the computer to: obtain a multidimensional non-linear vector of information indicative of the state of the majority of the test object, at least one vector being a right-limit, related to the test object, lacking a representation of the time at which the event occurred; and using information The vector performs regression to generate a kernel-based model, and provides at least part of the information about the estimated time of the event based on at least part of the information contained in the information-based vector, wherein the vector data is based on at least one characteristic of the data The classification is related, the at least one feature has the ability to provide an output value for the data assistance model, such that the output value assists in estimating the time to the event, and wherein the regression system sequentially uses the most categorized data to the least similar data. The vector data from the classification is executed to assist the model in providing output values such that the output values assist in estimating the time of the event.

Implementations of the invention may include one or more of the following features. Regression can be carried out in a positive greedy manner based on the characteristics of the data to select the characteristics to be used in the model. The computer program product further includes instructions to perform a negative greedy procedure on the computer to vector characteristics after performing the regression to further select characteristics to be used in the model. This regression is performed in a positive greedy manner only with respect to a part of the characteristics of the vector. The vector contains clinical/histopathological data, biomarker data, and data classification of biological imaging data, and in this order, the clinical/tissue pathological data is non- It is carried out in a greedy manner, and biomarker data and biometric data only for vectors are performed in a positive greedy manner.

In general, in another aspect, the invention provides a method of determining a predicted diagnosis of a patient, the method comprising: receiving at least one of clinical and histopathological data about the patient; receiving the patient Biomarker; receiving biometric data about the patient; and applying at least a portion of the clinical and histopathological data, at least a portion of the biomarker data, and at least a portion of the bioimage to a nuclear mathematical model To calculate the value indicating the diagnosis of the patient.

Implementations of the invention may include one or more of the following features. At least a portion of the biomarker data contains less than all of the biomarker characteristics of the patient. At least a portion of the biomarker data contains less than ten percent of all biomarker characteristics of the patient. At least a portion of the biomarker data contains less than five percent of all biomarker characteristics of the patient. At least a portion of the biomarker data contains less than all of the patient's biometric characteristics. At least a portion of the biomarker data contains less than one percent of all bioimage characteristics of the patient. At least a portion of the biomarker data contains less than 0.2% of the biometric characteristics of the patient. The value is at least one of a recurrence time of the health-related condition and a recurrence probability of the health-related condition.

In general, in another aspect, the present invention provides an apparatus for determining time-event estimation information, the apparatus including an input, an architecture to obtain multi-dimensional, non-linear first data related to a possible future event, and a processing device, the architecture to use the first data in the kernel-based mathematical model, the number The learning model is derived, at least in part, by a regression analysis of the multidimensional nonlinear right-limit second data, which determines the model parameters affecting the model calculation to calculate an estimate information indicating the estimated time of the possible future event. And at least one of the probabilities of possible future events.

Implementations of the invention may include one or more of the following features. The input and processing device comprises a portion of a computer program product embodied in a computer readable medium, the computer program product comprising computer readable, computer executable instructions for causing the computer to obtain the first data and in the mathematical model Use the first data to calculate the estimated information. The first data includes at least one of clinical and histopathological data, biomarker data, and biological imaging data relating to a patient, wherein the processing device is configured to use at least a portion of the clinical and histopathological data, the biological At least a portion of the marker data, and at least a portion of the biometric image to a core is the primary mathematical model to calculate the patient's estimated information. At least a portion of the biomarker data contains less than five percent of the biomarker characteristics of the patient. At least a portion of the biomarker data contains less than all of the biometric characteristics of the patient. At least a portion of the biomarker data contains less than 0.2% of all bioimage characteristics of the patient.

In general, in another aspect, the present invention provides a computer program product for determining a predicted diagnosis of a patient, the computer program product being readable on a computer readable medium and including a computer readable Computer executable instructions for causing the computer to: receive at least one of clinical and histopathological information about the patient; receive biomarker data about the patient; receive biometric information about the patient; Applying clinical and histopathological resources At least a portion of at least one of the materials, at least a portion of the biomarker data, and at least a portion of the biometric image to a nuclear-based mathematical model to calculate a representation of the diagnosis of the patient.

Implementations of the invention may include one or more of the following equivalences. At least a portion of the biomarker data contains less than all of the biomarker characteristics of the patient. The computer program product of claim 50, wherein at least a portion of the biomarker data contains less than ten percent of all biomarkers of the patient. At least a portion of the biomarker data contains less than five percent of all biomarker characteristics of the patient.

Implementations of the invention include one or more of the following features. At least a portion of the biomarker data contains less than all of the biological images of the patient. At least a portion of the biomarker data contains less than one percent of all biometric characteristics of the patient. At least a portion of the biomarker data contains less than 0.2% of the biometric characteristics of the patient. This value is at least one of the recurrence time of the health-related condition and the recurrence probability of the health-related condition.

The present invention provides novel techniques, such as utilizing the high dimensional capabilities of the SVR, while applying it to the limit data, particularly the right limit data. Support Vector Regression (SVRc) for setting data provides a variety of advantages and capabilities. Because much of the information available to form or train a predictive model may be limited, SVRc can increase the accuracy of the model estimate by using the limit data and unbound data in the SVR. With SVRc, high-dimensional data containing a small number of result data points observed by the right limit can be used to generate a temporary Inter-event prediction model. The characteristics of the high-dimensional data can be reduced to leave a set of characteristics for the decrement in the time-event prediction model, so that the time-event prediction accuracy can be improved.

These and other abilities of the present invention, as well as the scope of the invention and the detailed description of the invention, are fully understood.

Embodiments of the present invention provide techniques for improving the accuracy of estimated time-event probabilities. In order to develop an improved model for estimating time-event probability, a novel modified loss/penalty function is used in a support vector machine (SVM) for right-bound heterogeneous data. Using this newly modified loss/penalty function, the SVM can meaningfully process the right limit data to perform support vector regression (hereinafter referred to as SVRc) on the limit data. The data used to develop the model can come from a variety of measurements, depending on the event you want to estimate. For example, the test article can be a living or pre-living body, such as a human or other animal for medical applications. The test substance can also be an inanimate body for medical or non-medical applications. For example, inanimate test objects can be automotive parts for wear analysis, accounting reports, such as stock market performance for financial services, and the like.

In an exemplary embodiment, SVRc can be used to generate a model that predicts cancer recurrence. This model can be characterized by three different characteristic domains taken by a patient group: (i) clinical/histopathological properties; (ii) biomarker properties; and (iii) biometric properties, where the properties are Each stage is added to the model, with characteristics selected from different domains as subsequent stages. Anchor.

The clinical characteristics indicate the specific patient data collected by the doctor during a routine office visit. Such information may include information such as age, race, sex, etc., as well as information about some of the conditions, such as clinical stage or experimental parameters, such as prostate specific antigen (PSA).

Histopathological properties are information that is part of the pathology that describes the essential nature of the condition, particularly the structural and functional changes in the tissues and organs of the body caused by the condition. Examples of histopathological features include Gleason scores, surgical margin status, and ploidy information.

Biomarker properties are indicative of biochemicals in the body of a particular molecular property that can be used to measure the progress of a condition or the effect of a treatment. The type of biomarker property is exemplified by the use of antibodies to indicate a particular cell type, cell organ, or cell unit. Biomarker properties include, for example, the percentage of cells that are positive for several biomarkers in sample staining and the staining intensity of these biomarkers.

Bioimaging features represent information derived from mathematics and computational science to produce a bitmap image from tissue or cells. Examples of this information are lumen average, maximum, minimum, and standard deviation. Clinical/histopathological properties, biomarker properties, and bioinformatic properties are presented in the appendix. Various characteristics can be obtained and analyzed by using, for example, Cellenger, which is commercially available from Definiens AG (www.definiens.com), and MATLAB software, which is commercially available from MathWorks, Inc. (www.mathworks.com).

In this example, the characteristics from these three domains are added to the three phases. In the model (eg first stage: clinical/histopathological data; second stage: selection of clinical/histopathological characteristics used as anchoring and biomarker characteristics added; stage 3: selected clinical/histopathology and selection Biomarker properties are used as anchoring and bioimage (IMG) properties to be added. The resulting model contains selected features and model parameters that are iteratively adjusted/tuned for these characteristics. Other embodiments are still within the scope of the invention.

Embodiments of the invention may be used in a variety of applications. In the medical field, for example, embodiments can be used to predict time-events such as recurrence of prostate specific antigen (PSA). Embodiments can also be used to predict various chronic disease diagnoses or other health related events, including responses to a drug or hormone, or radiation or chemotherapy. Other applications generally involve the use of tissue-based clinical trials and clinical trials. Other applications that want to estimate the occurrence of an event are also possible. Examples from the health field include estimates of infections in dialysis patients, infections in burn patients, and weaning in newborns. In other areas, for example, engineers can estimate when the wagon will fail. In the medical field embodiment shown in FIG. 1, an SVRc system 10 includes clinical/histopathological measurement/data collection 12, biomarker data collection 14, and biometric data measurement/collection data sources, and data regression. And an analysis device 18 that provides an estimated diagnostic output 26. The data sources 12, 14, 16 may contain appropriate individuals (eg, doctors), data records (eg, medical databases), and/or machines (eg, camera devices, staining devices, etc.). The regression and analysis device 18 includes a computer 20 including a memory 22 and a processor 24 configured to perform computer readable, computer executable software code Instruction to execute SVRc. The computer 20 is represented by a personal computer, but other forms of computing devices are also available. Device 18 is further structured to provide as output 26 material that contains or can be processed to represent a predetermined time-event. For example, output 26 can be an estimated diagnosis of the time (including recurrence) of a cancer in a patient. Output 26 can be provided on display screen 28 of regression and analysis device 18.

The computer 20 of the regression and analysis device 18 is architected to perform SVRc by providing an SVM that is modified to analyze the bound and non-restricted data. The computer 20 can process data in accordance with the following structure of the SVRc.

SVRc structure

The data set T has an N sample, Where z _i ={x _i ,y _i ,s _i }, where x _i R ⁿ (R is a real array) is a sample vector, and y _i R is the target value (that is, the time when you want to estimate), and s _i {0, 1} is the set state of the relevant sample. The sample vector is the vector property for the ith sample/patient in (N). The target value y is the last known time of the actual time of the detection event (eg, reoccurrence) and the limit data for the non-limited data. If the limit state s _i is 1, the i-th sample z _i is a set limit sample, and if s _i is 0, the i-th sample z _i becomes a non-limit sample. When i = 1, ... N, s _i =0, the data set T becomes normal, and no data set is completely set. In addition, the limit state s _i =1 indicates that the unrestricted sample and s _i =0 indicate that the data set of the set limit sample is valid; at this time, the SVRc is controlled to consider the limit in the opposite manner.

The SVRc formula constructs a linear regression function f ( x )= W ^T Φ( x )+ b (1)

In a property space F, f(x) is the estimated time-event for sample x. Where W is the vector in F, and Φ(x) maps the input (x) to the vector in F. W and b in (1) are obtained by solving an optimization problem, and the general form is:

However, this equation assumes that the convex set optimization problem is always available, but this may not be the case in this case. Furthermore, we want to allow small errors in the regression estimates. For these reasons, a loss function is used for the SVR. Losses in the regression estimate allow for some margin. Ideally, the model built will accurately and correctly calculate all results, which is not feasible. The loss function allows for the departure of the ideal wide range error, which is controlled by the relaxation variables ξ and ξ ^* with a penalty C. Errors that deviate from the ideal error but are still within the range defined by ξ and ξ ^* are calculated, but their contribution is moderated by C. The more the example error, the greater the penalty. The smaller the example error (close to ideal), the smaller the penalty. Penalty results in a slope with the concept of error, and C controls this slope. Although various loss functions can be used, for an ε insensitive loss function, the general equation is converted to:

For the ε insensitive loss function according to the invention (with different loss functions applied to the set and unbound data), the equation becomes:

The optimization criterion penalty data indicates that its y value leaves f(x) beyond ε. The slack variables ξ and ξ ^* correspond to the size of the excess offset of the positive and negative offsets, respectively. This penalty mechanism has two components, one for non-restricted data (ie, non-right limit) and one for limit data. The two elements are represented by the form of the loss function, which is referred to as the ε insensitive loss function. An exemplary loss function 30 for setting data is defined in (3) and illustrated in FIG.

Where e=f(x)-y.

Therefore, e=f(x)-y represents the difference between the predetermined time-event and the actual time-event (detection/hypothetical event). The C and ε values are adjusted to the amount of penalty caused by the various deviations between the estimated and actual time-event. The C value controls the slope of the relevant portion of the loss function 30. The negative and positive ε offset values (ε _s ^* and - ε _s ) control how much offset is controlled before the penalty is paid. A limited data sample is processed in a different way than a traditional SVR because it only provides "one-sided information." For example, in the survival time estimation, where y _i in z _i represents the survival time, the limit data sample Z _i only indicates that the event does not occur until y _i does not indicate when it will be after y _i occur. The loss function of equation (3) reflects this reality. For the limit data, it is estimated that the time before the current time - the event (when the event is about to happen) is estimated to be worse than the current time (because this estimate may come true). Therefore, the projected limit data depends on whether the estimate is positive or negative for the actual/current time and is treated differently. The ε and C values are used to distinguish between penalties for f(x)>0 versus f(x)<0 (and to distinguish between bound and unbound data estimates). For a time-event estimate earlier than the current time, e < 0, the penalty is applied to the current time, ie, e>0 and then predicted to be a small offset (ε _s < ε _s ^* ). Furthermore, the larger difference between the predicted time-increment of the time before the current time (and greater than ε _s ) results in a later time between the predicted time and the event that is later than the current time (ie greater than ε _s ^* ). The big penalty is C _s >C _s ^* . As a result, the estimate before the current time has resulted in a larger penalty than the estimate after the current time.

Figure 2 shows that

(1) If e [-ε _x ,0], no penalty is imposed; if e (-∞, -ε _s ), then a linear increase penalty with a slope C _s is applied.

(2) If e [0, -ε _s ^* ], no penalty is imposed, if e (ε _s ^* , ∞), then a linear increase penalty with a slope C _s ^* is applied.

Since ε _s ^* > ε _s , and C _s ^* < C _s , when the estimated value f(x) < y, there are more penalties than f(x) > y. This mechanism assists in the full utilization of the SVRc regression function performed by the computer 20 to provide one-sided information in the set of data samples.

Furthermore, the modified loss function for non-restricted data can be represented in ε insensitive form. This loss function is a good consideration for recording time - the event may not be the reality of the actual time - event. Although the target value y _i roughly represents a time-event, when the event is detected, y _{i is} actually time, and the precise time at which the event occurs is often a period of time before y _i . The computer 20 can take this into account in the loss function of the non-limited data samples. An exemplary non-restricted data loss function 32 is provided in equation (4) and is illustrated in FIG.

Where e=f(x)-y.

Note that ε _n ^* ≦ ε _n and C _n ^* ≧ C _n , otherwise the explanation of Fig. 3 is substantially the same as Fig. 2.

Several simplifications and/or approximations can be used to simplify the calculation. For example, because the difference between the detection event time and the precise event time is small and usually negligible, ε _n ^* = ε _n and C _n ^* = C _n can be set, which simplifies the loss function of the unbound data sample. . In order to further reduce the number of free parameters formulated by SVRc and make it easy to use, in most cases, ε _s ^(*) , ε _n ^(*) , C _s ^(*) and C _n ^(*) can be set. ε _s ^* > ε _s = ε _n ^* = ε _n

C _s ^* <C _s =C _n ^* =C _n

As is known in the art and as noted above, the standard SVR uses a loss function. The loss function provided above is an ε-insensitive loss function and is used as an example only. Other ε-insensitive functions (eg, having different ε and/or C values), and other forms of loss functions can also be used. The exemplary loss function is discussed in S. Gunn's Support Vector Machine for Classification and Regression on page 29 of the Electronic and Computer Science Technical Report of the Department of Engineering and Applied Sciences, May 1998. Except for the ε-insensitive function, the exemplary loss function includes a quadratic, Laplacian or Huber loss function. Regarding the loss function 30, 32, the penalty used to estimate the estimate that is later than the actual/current time may be different (eg, for f(x) values below zero and zero, there are different slopes/shapes). The shape can be used to prepare for use as a no penalty next to f(x) = 0, and to prepare for different increments depending on whether f(x) is greater or less than zero.

Implementation of SVRc structure

In operation, referring to FIG. 4, with further reference to Figures 1 through 3, using system 18, a program 40 using SVRc to develop an estimate model includes The stage shown. However, this program 40 is for illustrative purposes only and is not limiting. Program 40 can be varied, for example, by adding, removing, or rearranging the stages.

At stage 42, the training of the starting model of model 1 is performed. Clinical/histopathological data 12 of relevant clinical/histopathological properties are supplied to system 18 to determine a set of calculus parameters and model parameters for the corresponding set of Model 1. The calculus parameters are parameters that govern the regression performed by the computer 20 to determine model parameters and select characteristics. Examples of the calculation parameters are the kernel used for regression, and the margins - ε _s , ε _s ^* , - ε _n , ε _n ^* , and the loss function slopes C _n , C _n ^* , C _s , C _s ^* . The model parameters affect the output value of the model f(x) for a given input x. The calculation parameters are set in stage 42 and fixed at the set values in other stages of the program 40.

Referring to Figure 5, with reference to Figures 1 through 4, a routine 60 for implementing the implementation phase 42 of Figure 4 is used to determine the model 1 using the system 18 using SVRc. However, this program 60 is for illustrative purposes only and is not limiting. Program 60 can be varied, for example, by adding, removing, or rearranging the stages.

At stage 62, the calculus parameters are initiated. The first time phase 62 is executed, the calculation parameters are the initial settings, and the subsequent performance at stage 62 is reset. Each time phase 62 is executed and a set of unused calculus parameters are selected for use in the model to train the model parameters.

In stage 64, the model parameters are initiated. The model parameters may be a higher group of model parameter values, but are preferably based on SVR knowledge to reduce the time spent on computer 20 to train model parameters. Although this stage It is displayed separately from the other phases, but the actions can be performed in conjunction with other phases, such as calculus parameter selection and/or phase 66 at stage 42 of FIG.

At stage 66, the model parameters are trained using the current selection set calculus parameters. To train the model parameters, portions of the data vector (and possibly all of the data) in a set of data vectors are fed into the computer 20. The data vector contains information about various characteristics. For example, the patient data vector preferably includes clinical/tissue pathology, biomarkers, and bio-image characteristics as well as corresponding values for these characteristics for each patient. For the selection of the calculation parameters in the procedure 60, it is preferred to use only clinical/tissue pathological characteristics and related data. These values are used as input x of model f to determine the value of f(x). The vectors also contain a target value y corresponding to the target value of f(x). Computer 20 determines the f(x) value for each patient and the difference between the model output and the target value, i.e., f(x)-y. The computer 20 determines whether the input vector x is set or not, and uses the loss function 30, 32. The computer 20 uses the information from the loss functions 30, 32, in accordance with equation (2), to perform the SVR to determine a set of model parameters corresponding to the current set of calculus parameters. Using the determined model parameters, the computer 20 uses 5 fold cross validation to calculate and store the key table index (CI) for the set of calculus parameters and model parameters.

At stage 68, a query is made as to whether there is still a set of calculus parameters to try. Computer 20 determines if each of the available set of calculus parameters has been used to determine a corresponding set of model parameters. If not, program 60 returns to stage 62 where a new set of calculation parameters is selected. If the calculation parameters of all groups have been used The model 60 is determined to correspond to the model parameters of the group.

At stage 70, computer 20 selects a desired set of calculus parameters to further train the model. The computer 20 analyzes the stored vocabulary index for the model corresponding to the various sets of calculus parameters and the associated model parameters determined by the computer 20. Computer 20 finds the maximum stored CI and fixes the relevant calculation parameters to the calculation parameters of the model that will be used in the other stages of program 40 as shown in FIG. This version of the model and the selected calculation parameters and the associated model parameters form the model 1. Model 1 is output by stage 42 and forms an anchor for stage 44.

Referring to Figure 4 and Figures 1 through 3, in stage 44, a supplemental model, Model 2, is trained. Model 1 is used to determine the anchor of Model 2, where the calculus parameters are set at stage 42, which will remain unchanged in other model training. Model 1 is an anchor in which the characteristics used in Model 1 (here, the pro/bed histopathological properties) will be used to form other further models, in particular, to provide a basis for Model 2.

In order to form model 2 based on model 1, feature selection (FS) is performed using a forward greedy (GF) algorithm, and only features that are considered to improve the predictive correctness of the model are retained in the model. In the exemplary case of cancer prediction, biomarker data is fed to device 18 at stage 44 to determine which biomarker is added to model 1 to form model 2. A data vector x containing values for clinical/tissue pathological properties and selected biomarker properties was used for the SVRc setup described above. A five-fold cross-validation system is used to determine model parameters that contain new features. The correctness of the revised model and the previous model is specified by individual CIs. If the revised model is pre- The correctness is better than the previous model (for the biomarker feature, the former model is model 1), the characteristics of the revised model are maintained, and new features are added for evaluation. If the correctness of the estimate is not improved, the features of the recent additions are abandoned and other new features are added for evaluation. This will continue until all biomarker features have been tested and abandoned or added to the model. The model with the relevant model parameters is output by the device 18 as a model 2 at stage 44.

At stage 46, a supplemental model, Model 3, is trained. Model 2 is used as an anchor for determining model 3. Model 2 is an anchor, and the characteristics contained in Model 2 (here, clinical/tissue pathological properties, if any, plus biomarker properties) will be used to form Model 3.

In order to form model 3 based on model 2, the feature selection (FS) is performed using a forward greedy (GF) algorithm, and only features that are considered to improve the correctness of the model prediction are retained in the model. Preferably, in terms of individual and/or group, the characteristics evaluated by Model 1 to form Model 2 are considered to be more reliable and/or pre-optimal than the characteristics of Model 2 being evaluated to form Model 3. Estimate functionality (data value versus time and/or event similarity). In the illustrated case of cancer prediction, the biometric image is fed into device 18 at stage 46 to determine which biometric feature is added to model 2 to form model 3. The data vector x containing the values of the clinical/histopathological properties, the biomarker characteristics selected at stage 44, and the selected bioimage characteristics are used in the SVRc structure described above. Five-fold cross-validation is used to determine model parameters that contain new features. The revised model and the correctness of the prediction of the previous model are individual CI tables. Show. If the estimated correctness of the revised model is better than that of the previous model (for the bio-image feature, the former model is model 2), the characteristics just added to the model are maintained, and new features are added for evaluation. If the correctness of the estimate is not improved, the latest additions are discarded and another new feature is added for evaluation. This will continue until all biometric features have been tested and discarded or added to the model. The model with the relevant model parameters is output by the device 18 as a model 3 at stage 46.

At stage 48, a negative greedy (GB) program is executed to refine the model from model 3 into the final model. When Model 3 executes a GB algorithm to perform feature selection, one feature is removed from the model and the model is retested for correctness. If the predictive correctness of the model increases when a feature is removed, then the feature is removed from the model and the GB program is applied to the revised model. This continues until the GB program is unable to obtain an increase in the estimated correctness when any of the features in the current feature are removed. The final model parameters are then used with the test data to determine the correctness of the final model estimate. The resulting final model with the potential to reduce the set of characteristics and determine the model parameters is the output of stage 48 and can be used by device 18 to provide a time-event probability when providing information for the characteristics in the final model.

Other embodiments are still within the scope and spirit of the appended claims. For example, due to the nature of the software, the above functions can be implemented using software, hardware, firmware, or any combination. The feature enforcement function can also be physically located at various locations, including being distributed such that portions of the functionality are implemented at different physical locations. Furthermore, although in the program 60, the model parameters are adjusted, but the mode Type parameters can be set based on the knowledge of the SVR and will not change subsequently. This can reduce the processing capacity and/or time of developing the SVRc model. Furthermore, one or more criteria can be placed on most of the features that are considered to be added to the model. For example, only a key table index with a threshold (eg, 0.6) or more can be added to the model and tested to affect the correctness of the model. Therefore, the set of characteristics to be tested can be reduced, which can reduce the processing capacity and/or time at which a model is generated. Furthermore, the model can be developed without using the property domain as an anchor. The characteristics can be added to the model and after each characteristic domain, the impact of not establishing the model as an anchor on the accuracy of the estimate has been considered.

Experimental and experimental results Lab 1: Internal Verification

Modern machine learning algorithms were applied to post-operative prostate cancer patients for 540 patients treated at the Bayer University Medical Center. These patients underwent prostate eradication surgery at the Bayer University Medical Center. Clinical and histopathological variables were provided for 539 patients, and the number of patients lost data for patients and variables. Similarly, tissue microarray sheets (including ternary normal and ternary cancer nucleus) were provided for these patients; these were used for H&E staining for photography, and the remaining sheets were used for biomarker studies.

Regarding the image analysis components of the study, only nuclei containing at least 80% of the cancer were used to preserve the integrity of the signals that were intended to be measured in these tissue samples (and to improve the signal-to-noise ratio). The signal to be measured consists of an abnormality in the cancer microdissection. (Conversely, the "noise" in image analysis is Normal tissue microanatomy measurements. ) 80% deletion was chosen to maximize the size of the group while preserving the integrity of the results. Therefore, the effective sample size of the study was based on patients who had access to information from clinical data, biomarker data, and biometric data. Therefore, the total number of patients that can be used to integrate the predictive system is 130.

SVRc is applied to this group of patients and their related data. The SVRc system was applied separately to clinical/histopathological data (17 characteristics), biomarker data (43 characteristics from 12 markers), and from bioimaged software phantom company (manufactured by Aureon Biotech, Inc., New York). Bio-image data obtained by Script4 (496 characteristics). The SVRc algorithm is applied to each of the three data types to find individual predictive power for each data type. In each case, two models were created: one model used all of the original features; the other used a set of selected features obtained for the negative greedy feature selection (SVRc-GB). The SVRc algorithm is also applied to all three data types in accordance with the above procedure 40.

Experiment 1: Results, Synthesis and Conclusions

The results are summarized in Tables 1 and 6.

It shows an increasing tendency to predict the ability to add molecular and bioinformatic information to individual clinical/tissue pathology information. This result supports a systematic pathology analysis that integrates patient information at different stages (ie, clinical/histopathology, microdissection, and molecular) to improve the concept of predictive power across the system. When compared to traditional multivariate modeling techniques, the analysis also demonstrates that advanced monitored multivariate modeling techniques can be used to build improved predictive systems. with In addition to clinical/histopathological properties, some molecular and bioimage characteristics of PSA recurrence can be selected.

In contrast to the survival analysis of the Cox model, which was traditionally applied to clinical data alone, SVRc has the advantage of exhibiting high-dimensional data sets that can be processed in a small group of diseased strings. In this research data set, SVRc provides results that are more realistic and superior than those produced by the standard Cox model.

Lab 2: External Validation of Expert Domain Knowledge

To evaluate overall system performance, a fairly conservative two-level verification procedure was used to simulate external verification. The 140 pairs of training and test groups were generated by randomly picking up 100 records as training groups and using the remaining 30 unselected records as test groups.

(1) For each pair, the training group uses program 40 to build an estimate model.

(2) The model is then applied to the test group to assess the correctness of the final model.

(3) Steps (1) and (2) were repeated 40 times to obtain 40 estimated correctness and final estimated performance was reported as the average estimated correctness of the 40 specific final models.

The most frequently selected features in the 40 different final models were then used to train three other models for each pair of training and test groups using SVRc: one model based only on clinical/tissue pathology; one model based on clinical/organization Pathological characteristics and biomarker characteristics; and a model based on clinical/histopago/biomarker characteristics and bio-image characteristics.

Experiment 2: Results, Summary and Conclusions

The experimental results are shown in Table 2. The results can be summarized as follows: For 40 execution, the average generalization correctness (ie, the predictive correctness of the model when applied to a test group) is: (1) for clinical/tissue pathological data of 0.74; (2) for clinical/tissue pathology And the biomarker information was 0.76; and (3) the clinical/hiortopathology/biomarker plus bioimage data was 0.77.

A complete list of the characteristics and frequencies of the final model is provided in the appendix.

As before, it shows an increasing trend of predictive power from molecular and biological imaging information to clinical/tissue pathology information. This result further supports the notion that systematic pathology analysis of patients at different levels (ie, clinical/histopathology, microdissection, and molecular) can improve the predictive power of the entire system. Compared with the traditional multivariate model technology of the traditional group of small group patients who only use clinical data, the analysis also shows that the advanced monitoring multivariate model technology can improve the estimation system.

It can be concluded that adding a layer of domain specific advice can assist in the selection of features to improve the predictive power of the system.

For the organizational segmentation of the phantom system of Aureon Biotech of New York, the image object is classified as using spectral characteristics, shape Histopathological grades of specific characteristics and special relationships between histopathological objects. For a given histopathological object, its properties are calculated and output as biometric features. Features include spectrum (color channel values, standard deviation and brightness) and general shape (area, length, width, compactness, density, etc.) characteristics. Statistics (minimum, maximum, mean, and standard deviation) are calculated for each characteristic of a histopathological object. The above is reflected in the name of the appendix. For example, the characteristics "Lumen.StdDevAreaPxl", "Lumen" are represented as histopathological objects, "StdDev" is expressed as a standard deviation, and "AreaPxl" is expressed as a guest characteristic.

Statistics and characteristics were calculated for the following histopathological objects. “Background” is the portion of the digital image that is not occupied by the organization. "Cytoplasm" is an amorphous "pink" region that surrounds an epithelial nucleus. "Epithelial nuclei" is the "circle" object surrounded by this cytoplasm. "Lumen" is the sealed white area surrounded by the epithelial cells. Alternatively, the cavity can be filled with prostate fluid (pink) or other "residues" (eg, macrophages, dead cells, etc.). The cavity and epithelial nucleus form a glandular unit. "Stroma" is a form of connected tissue having different densities that maintain the architecture of the prostate tissue. The matrix system appears between the glandular units. A "matrix nucleus" is an elongate cell with minimal or no cytoplasmic (fibroblasts). This classification may also include endothelial cells and inflammatory cells, and if a cancer occurs, the epithelial nucleus is also found to be interspersed between the stroma. "Red blood cells" are usually small red round objects located in blood vessels (arteries or veins), but can be found throughout the tissue, AK.1, AK.2, AK.3, AK.4 and AK.5 User-specific labels are not meaningful. "C2EN" is the relative ratio of nuclear area to cytoplasm example. The more regressive/malignant epithelial cells, the larger the area occupied by the nucleus. "EN2SN" is the percentage or relative amount of epithelial cells to stromal cells that appear in digital tissue images. "L2Core" is the number or area of cavities that appear in the tissue. The higher the Gleason score, the less the number of cavities. "C2L" is a cytoplasmic versus cavity. "CEN2L" is the cytoplasmic endothelial cell to the lumen.

The name part after the object is for the challenger developed by the Definiens company in Munich, Germany, to develop the voice 4.0 software.

End appendix

10‧‧‧SVRc system

12‧‧‧Clinical/hitopathological data

14‧‧‧Biomarker data collection

16‧‧‧Biometric data measurement/collection

18‧‧‧ Data regression and analysis device

20‧‧‧ computer

22‧‧‧ memory

24‧‧‧Processing machine

26‧‧‧ Output

28‧‧‧ Display screen

Figure 1 is a simplified block diagram of an estimated diagnostic system for right-limit data.

Figure 2 is an illustration of an exemplary loss function for setting data.

Figure 3 is an illustration of an exemplary loss function for non-restricted data.

Figure 4 is a block diagram of a process for developing a program for estimating time-event information.

Figure 5 is a block flow diagram of the production of the start model program as shown in Figure 4.

Figure 6 is a three-dimensional map of the model's performance as summarized by the vocabulary index determined by the embodiment of the present invention and the traditional Cox proportional hazard model using experimental data.

Claims (34)

A method of generating a model for estimating the time at which an event occurs, the method comprising: obtaining a multidimensional non-linear vector of information representative of a state of a majority of the test object; and using a vector of information to perform regression to generate a kernel-based model to At least part of the information contained in the information-based vector is provided with an estimated value of the estimated time of the event; wherein the data of the vector is correlated based on the classification of at least one feature of the data, the at least one feature being related to the data assisting model The ability to provide an output value such that the output value assists in estimating the time to the event, and wherein the regression system is sequentially executed using data from the most material-like to the least-classified data classification to assist in providing the model. The output value is such that the output value assists the estimated time of the event. The method of claim 1, wherein the regression is performed in a positive greedy manner based on the characteristics of the data to select characteristics to be used in the model. For example, the method described in claim 2 includes, after performing the regression, performing a negative greedy procedure on the characteristics of the vector to further select characteristics to be used in the model. The method of claim 2, wherein the regression is performed in a positive greedy direction only with respect to a part of the characteristics of the vector. The method of claim 4, wherein the vectors comprise clinical/tissue pathological data, biomarkers, and biological imaging materials. Data classification, and wherein the regression system is performed in a non-positive greedy manner for clinical/biological histopathological data, and biomarker data and bio-image data for only vectors are performed in a positive greedy manner. The method of claim 1, wherein the at least one vector is right-defined and lacks a time indication of an occurrence of the test object. A computer program product for generating a model for estimating the time of occurrence of an event, the computer program product being in a computer readable medium and including computer readable computer executable instructions for causing the computer to: obtain a majority of the test object a multidimensional nonlinear vector of state information, at least one vector being right-defined, with respect to the relevant test object, lacking a representation of the occurrence time of the event; and using a vector of information to perform regression to generate a kernel-based model based on At least part of the information contained in the vector of information provides an output value for an estimated time to the event; wherein the data of the vector is related to the classification based on at least one feature of the data, the at least one feature having a data assistance model The ability to provide an output value such that the output value assists in estimating the time to the event; and wherein the regression system is sequentially executed using data from the most material-like to the least-classified data classification to assist in providing the model. The output value is such that the output value assists the estimated time of the event. For example, the computer program product described in claim 7 of the patent scope, wherein the regression is performed in a positive greedy manner according to the characteristics of the data to select characteristics to be used in the model. The computer program product described in claim 8 of the patent application further includes instructions to perform a negative greedy procedure on the computer-to-vector characteristics after performing the regression to further select characteristics to be used in the model. For example, the computer program product described in claim 8 wherein the regression is performed in a positively greedy manner only with respect to a part of the characteristics of the vector. The computer program product according to claim 10, wherein the vectors comprise clinical/histopathological data, biomarker data, and data classification of biological imaging materials, and wherein the regression system is for clinical/tissue pathological data, In a non-positive greedy manner, and only biomarker data and bio-image data for vectors are performed in a positive greedy manner. A method of determining a predicted diagnosis of a patient, the method comprising: receiving at least one of clinical and histopathological data about the patient; receiving a biomarker related to the patient; receiving a biological activity about the patient Image data; and applying at least a portion of at least one of the clinical and histopathological data, at least a portion of the biomarker data, and at least a portion of the biological image to a core as a mathematical model for computational representation The value of the diagnosis of the patient. The method of claim 12, wherein at least a portion of the biomarker data comprises less than all biomarker characteristics of the patient. The method of claim 13, wherein the creature At least a portion of the marker data contains less than about 10% of all biomarker characteristics of the patient. The method of claim 14, wherein at least a portion of the biomarker data comprises less than about five percent of all biomarker characteristics of the patient. The method of claim 12, wherein at least a portion of the biomarker data comprises less than all of the biometric characteristics of the patient. The method of claim 16, wherein at least a portion of the biomarker data comprises less than one percent of all biometric characteristics of the patient. The method of claim 17, wherein at least a portion of the biomarker data comprises less than about 0.2 percent of all bioimage characteristics of the patient. The method of claim 12, wherein the value is at least one of a recurrence time of a health-related condition and a recurrence probability of the health-related condition. An apparatus for determining estimation information of an event occurrence time, the apparatus comprising: an input, an architecture to obtain multi-dimensional, non-linear first data related to a possible future event; and a processing device configured to use the core-based mathematics The first data in the model is used to calculate the estimation information, and the mathematical model is at least partially derived from the regression analysis of the second data of the multidimensional nonlinear right limit, the data determines the shadow The model parameters calculated by the model, the estimated information indicating at least one of an estimated time of the possible future event and a probability of a possible future event. The device of claim 20, wherein the input and processing device comprises a portion of a computer program product embodied in a computer readable medium, the computer program product comprising a computer readable, computer executable instruction In order to enable the computer to obtain the first data and use the first data in the mathematical model to calculate the estimated information. The device of claim 20, wherein the first data comprises at least one of clinical and histopathological data, biomarker data, and biological imaging data related to a patient, wherein the processing device is At least a portion of the clinical and histopathological data, at least a portion of the biomarker data, and at least a portion of the biometric image to the core are used as the primary mathematical model to calculate the patient's estimated information. The device of claim 22, wherein at least a portion of the biomarker data comprises less than all biomarker characteristics of the patient. The device of claim 23, wherein at least a portion of the biomarker data comprises less than about five percent of all biomarker characteristics of the patient. The device of claim 22, wherein at least a portion of the biomarker data comprises less than all of the biometric characteristics of the patient. The device of claim 25, wherein at least a portion of the biomarker data contains less than all biometric characteristics of the patient About 0.2% of the data. A computer program product for determining a patient's predictive diagnosis, the computer program product being readable on a computer readable medium and containing computer readable, computer executable instructions for causing the computer to receive information about the patient At least one of clinical and histopathological data; receiving biomarker data about the patient; receiving biometric data about the patient; and applying at least a portion of at least one of clinical and histopathological data, the organism At least a portion of the marker data and at least a portion of the biometric image are subjected to a nuclear-based mathematical model to calculate a representation of the diagnosis of the patient. The computer program product of claim 27, wherein at least a portion of the biomarker data contains less than all biomarker characteristics of the patient. The computer program product of claim 28, wherein at least a portion of the biomarker data comprises less than about 10% of all biomarkers of the patient. The computer program product of claim 29, wherein at least a portion of the biomarker data comprises less than about five percent of all biomarker characteristics of the patient. The computer program product of claim 27, wherein at least a portion of the biomarker data comprises less than all biometric characteristics of the patient. A computer program product as described in claim 31, At least a portion of the biomarker data contains less than about one percent of all bioinformatic properties of the patient. The computer program product of claim 32, wherein at least a portion of the biomarker data comprises less than about 0.2 percent of all biometric characteristics of the patient. The computer program product of claim 27, wherein the value is at least one of a recurrence time of a health-related condition and a recurrence probability of a health-related condition.