Support vector regression for censored data

• US 7,505,948 B2
• Filed: 11/17/2004
• Issued: 03/17/2009
• Est. Priority Date: 11/18/2003
• Status: Active Grant

• Current Assignee: David Sans
• Original Assignee: Aureon Biosciences, Inc.
• Inventors: Verbel, David A., Saidi, Olivier
• Primary Examiner(s): Vincent; David R
• Assistant Examiner(s): Rifkin; Ben M
• Application Number:
  US10/991,240
• Publication Number:
  US 20050108753A1
• Time in Patent Office: 1,581 Days
• Field of Search: 706/14
• US Class Current: 706/14
• CPC Class Codes: G06F 17/18 for evaluating statistical data , e.g. average values, frequency distributions, probability functions, regression analysis forecasting specially adapted for a specific administrative, business or logistic context G06Q10/04
• CPC Class Codes: G06F 18/2178 based on feedback of a supervisor
• CPC Class Codes: G06F 18/2411 based on the proximity to a decision surface, e.g. support vector machines
• CPC Class Codes: G06V 10/764 using classification, e.g. of video objects
• CPC Class Codes: G06V 10/7784 based on feedback from supervisors
• CPC Class Codes: G16H 10/20 for electronic clinical trials or questionnaires

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.

Claims

1. A computer-implemented method of producing a model for use in predicting time to occurrence of a health-related condition, the method comprising:

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 health-related condition 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 occurrence of the health-related condition based upon at least some of the information contained in the vectors of information;

wherein 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;

wherein performing the regression includes using penalty functions that include linear functions of a difference between a predicted value of the model and a target value for the predicted value, and wherein a first slope of the linear function for positive differences between the predicted and target values for the censored-data penalty function is lower than a second slope of the linear function for positive differences between the predicted and target values for the non-censored-data penalty function.

2. The method of claim 1 wherein the regression comprises support vector machine regression.

3. The method of claim 1 wherein the censored-data penalty function has a larger positive epsilon value than the non-censored data penalty function does.

4. The method of claim 1 wherein the first slope is substantially equal to a third slope of the linear function for negative differences between the predicted and target values for the censored-data penalty function and a fourth slope of the linear function for negative differences between the predicted and target values for the non-censored-data penalty function, and wherein positive and negative epsilon values of the non-censored-data penalty function and a negative epsilon value of the censored-data penalty function are substantially equal.

5. The method of claim 1 wherein the data of the vectors are associated with categories based on at least one characteristic of the data that relate to the data'"'"'s ability to help the model provide the output value such that the output value helps predict time to occurrence of a health-related condition, the method further comprising performing the regression using the data from the vectors in sequence from the category with data most likely, to the category with data least likely, to help the model provide the output value such that the output value helps predict time to occurrence of a health-related condition.

6. The method of claim 5 wherein the at least one characteristic is at least one of reliability and predictive power.

7. The method of claim 5 wherein the regression is performed in a greedy-forward manner in accordance with the features of the data to select features to be used in the model.

8. The method of claim 7 further comprising performing a greedy backward procedure to the features of the vectors, after performing the regression, to further select features to be used in the model.

9. The method of claim 7 wherein the regression is performed in the greedy-forward manner with respect to only a portion of the features of the vectors.

10. The method of claim 9 wherein the vectors include categories of data of clinical/histopathological data, biomarker data, and bio-image data, and wherein the regression is performed in the greedy-forward manner with respect to only the biomarker data and the bio-image data of the vectors.

11. The method of claim 1 wherein the vectors of information are indicative of status of test subjects that are at least one of living, previously-living, and inanimate.

21. The method of claim 1 wherein the first slope is lower than a slope of the linear function for negative differences between the predicted and target values for the censored-data penalty function.

22. The method of claim 1 wherein the second slope is greater than a slope of the linear function for negative differences between the predicted and target values for the non-censored-data penalty function.

23. The method of claim 1 wherein the first slope is lower than a third slope of the linear function for negative differences between the predicted and target values for the censored-data penalty function and a fourth slope of the linear function for negative differences between the predicted and target values for the non-censored-data penalty function.

24. The method of claim 1 wherein performing the regression comprises using penalty functions that include epsilon values which control how much deviation between predicted and target values is tolerated before a penalty is assessed, wherein at least two of the epsilon values are different.

25. The method of claim 24 wherein a epsilon value of the censored-data penalty function for positive differences between the predicted and target values is greater than a epsilon value of the censored-data penalty function for negative differences between the predicted and target values.

26. The method of claim 24 wherein a epsilon value of the non-censored-data penalty function for negative differences between the predicted and target values is greater than a epsilon value of the non-censored-data penalty function for positive differences between the predicted and target values.

27. The method of claim 24 wherein a epsilon value of the censored-data penalty function for positive differences between the predicted and target values is greater than a epsilon value of the censored-data penalty function for negative differences between the predicted and target values, a epsilon value of the non-censored-data penalty function for negative differences between the predicted and target values, and a epsilon value of the non-censored-data penalty function for positive differences between the predicted and target values.

28. The method of claim 1 wherein the model provides an output value indicative of at least one of a time to occurrence of a health-related condition and a probability of occurrence of the health-related condition.

29. The method of claim 1 wherein the vectors include categories of data of clinical/histopathological data, biomarker data, and bio-image data from a computer image of tissue.

12. A computer program product producing a model for use in predicting time to occurrence of a health-related condition, the computer program product residing on a computer readable medium, the computer program product comprising computer-readable, computer-executable instructions for causing a computer to:

obtain 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 health-related condition with respect to the corresponding test subject; and

perform 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 occurrence of the health-related condition based upon at least some of the information contained in the vectors of information;

wherein 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;

wherein the instructions for causing the computer to perform the regression include instructions for causing the computer to use penalty functions that include linear functions of a difference between a predicted value of the model and a target value for the predicted value, and wherein a first slope of the linear function for positive differences between the predicted and target values for the censored-data penalty function is lower than a second slope of the linear function for positive differences between the predicted and target values for the non-censored-data penalty function.

13. The computer program product of claim 12 wherein the regression comprises support vector machine regression.

14. The computer program product of claim 12 wherein the censored-data penalty function has a larger positive epsilon value than the non-censored data penalty function does.

15. The computer program product of claim 12 wherein the first slope is substantially equal to a third slope of the linear function for negative differences between the predicted and target values for the censored-data penalty function and a fourth slope of the linear function for negative differences between the predicted and target values for the non-censored-data penalty function, and wherein positive and negative epsilon values of the non-censored-data penalty function and a negative epsilon value of the censored-data penalty function are substantially equal.

16. The computer program product of claim 12 wherein the instructions for causing the computer to perform regression cause the regression to be performed using the data from the vectors in sequence from a category with data most likely, to a category with data least likely, to help the model provide the output value such that the output value helps predict time to occurrence of a health-related condition.

17. The computer program product of claim 16 wherein the instructions for causing the computer to perform regression cause the regression to be performed in a greedy-forward manner in accordance with features of the data to select features to be used in the model.

18. The computer program product of claim 17 further comprising instructions for causing the computer to perform a greedy backward procedure to the features of the model, after performing the regression, to further select features to be used in the model.

19. The computer program product of claim 17 wherein the instructions for causing the computer to perform regression in the greedy-forward manner cause the computer to perform the greedy-forward feature selection with respect to only a portion of the features of the vectors.

20. The computer program product of claim 19 wherein the vectors include categories of data of clinical/histopathological data, biomarker data, and bio-image data, and wherein the instructions for causing the computer to perform regression in the greedy-forward manner cause the computer to perform the greedy-forward feature selection with respect to only the biomarker data and the bio-image data of the vectors.

30. The computer program product of claim 12 wherein the first slope is lower than a slope of the linear function for negative differences between the predicted and target values for the censored-data penalty function.

31. The computer program product of claim 12 wherein the second slope is greater than a slope of the linear function for negative differences between the predicted and target values for the non-censored-data penalty function.

32. The computer program product of claim 12 wherein the first slope is lower than a third slope of the linear function for negative differences between the predicted and target values for the censored-data penalty function and a fourth slope of the linear function for negative differences between the predicted and target values for the non-censored-data penalty function.

33. The computer program product of claim 12 wherein the instructions for causing the computer to perform the regression comprise instructions for causing the computer to perform the regression using penalty functions that include epsilon values which control how much deviation between predicted and target values is tolerated before a penalty is assessed, wherein at least two of the epsilon values are different.

34. The computer program product of claim 33 wherein a epsilon value of the censored-data penalty function for positive differences between the predicted and target values is greater than a epsilon value of the censored-data penalty function for negative differences between the predicted and target values.

35. The computer program product of claim 33 wherein a epsilon value of the non-censored-data penalty function for negative differences between the predicted and target values is greater than a epsilon value of the non-censored-data penalty function for positive differences between the predicted and target values.

36. The computer program product of claim 33 wherein a epsilon value of the censored-data penalty function for positive differences between the predicted and target values is greater than a epsilon value of the censored-data penalty function for negative differences between the predicted and target values, a epsilon value of the non-censored-data penalty function for negative differences between the predicted and target values, and a epsilon value of the non-censored-data penalty function for positive differences between the predicted and target values.

37. The computer program product of claim 12 wherein the model provides an output value indicative of at least one of a time to occurrence of a health-related condition and a probability of occurrence of the health-related condition.

38. The computer program product of claim 12 wherein the vectors include categories of data of clinical/histopathological data, biomarker data, and bio-image data from a computer image of tissue.