System and method for privacy preserving predictive models for lung cancer survival analysis
View Patent ↗A computer-implemented method for privacy-preserving data mining to determine cancer survival rates includes providing a random matrix B agreed to by a plurality of entities, wherein each entity i possesses a data matrix A i of cancer survival data that is not publicly available, providing a class matrix D i for each of the data matrices A i , providing a kernel K(A i , B) by each of said plurality of entities to allow public computation of a full kernel, and computing a binary classifier that incorporates said public full kernel, wherein said classifier is adapted to classify a new data vector according to a sign of said classifier.
1. A computer-implemented method for privacy-preserving data mining to determine cancer survival rates, said method comprising the steps of:
providing a random matrix B agreed to by a plurality of entities, wherein each entity i possesses a data matrix A i of cancer survival data that is not publicly available;
providing a class matrix D i for each of the data matrices A i ;
providing a kernel K(A i , B) by each of said plurality of entities to allow public computation of a full kernel; and
computing, with a computer system, a binary classifier that incorporates said public full kernel, wherein said classifier is adapted to classify a new data vector according to a sign of said classifier, wherein computing comprises:
extracting covariates x from the data matrix A i ; and
maximizing a quantity w′K(x, B′) to solve for vector w, wherein said quantity w′K(x, B′) is an effect parameter in a survival model that characterizes an effect on said cancer survival rates.
2. The method of claim 1 , wherein said class matrix D is of size m×m wherein in is a number of data points and has a value +1 on a main diagonal for each survival datum that exceeds a predetermined number of years, and has a value of −1 on said main diagonal for each survival datum that does not exceed the predetermined number of years.
3. The method of claim 1 , wherein said random matrix B is real valued of size {tilde over (m)}×n, wherein n is a dimensionality of each data point, and {tilde over (m)}<n, and each data matrix A i is of size m×n, wherein m is a number of data points.
4. The method of claim 1 , wherein said full kernel is
K
(
A
,
B
′
)
=
K
(
[
A
1
A
2
⋮
A
q
]
,
B
)
=
[
K
(
A
1
,
B
)
K
(
A
2
,
B
)
⋮
K
(
A
q
,
B
)
]
,
where q is a number of entities.
5. The method of claim 1 , wherein said classifier is defined as K(x′,B′)w−γ=0 wherein w is a vector normal to a hyperplane separating the two classes of the binary classifier, γ determines the location of the separating hyperplane relative to the origin, and K(x′, B′) is a row vector of K(A, B′).
6. The method of claim 5 , wherein the classifier is solved for using a Newton-Lagrangian method wherein a square of a 2-norm of a slack variable is minimized with weight ν/2 wherein ν>0 and a distance between bounding planes is measured in an (n+1)-dimensional space of (w,γ)εR n+1 .
7. A computer-implemented method for privacy-preserving data mining to determine cancer survival rates, said method comprising the steps of:
providing a random matrix B agreed to by a plurality of entities, wherein each entity i possesses a data matrix A i of cancer survival rates that is not publicly available;
providing a kernel K(A i , B) by each of said plurality of entities to allow public computation of a full kernel; and
maximizing a quantity w′K(x,B′) wherein x is a row in one of said data matrices A i to solve for vector w, wherein said quantity w′K(x,B′) is an effect parameter in a survival model that characterizes an effect on said cancer survival rates.
8. The method of claim 7 , where said effect parameter w′K(x,B′) is an effect parameter of a cox regression model log h(t)=α(t)+w′×B′, wherein h(t) is a hazard function and α(t) represents an unspecified baseline hazard function.
9. The method of claim 7 , wherein said random matrix B is real valued of size {tilde over (m)}×n, wherein n is a dimensionality of each data point, and {tilde over (m)}<n, and each data matrix A i is of size m×n, wherein m is a number of data points.
10. The method of claim 7 , wherein said full kernel is
K
(
A
,
B
′
)
=
K
(
[
A
1
A
2
⋮
A
q
]
,
B
)
=
[
K
(
A
1
,
B
)
K
(
A
2
,
B
)
⋮
K
(
A
q
,
B
)
]
,
where q is a number of entities.
11. A program storage device readable by a computer, tangibly embodying a program of instructions executable by the computer to perform the method steps for privacy-preserving data mining to determine cancer survival rates, said method comprising the steps of:
providing a random matrix B agreed to by a plurality of entities, wherein each entity i possesses a data matrix A i of cancer survival data that is not publicly available;
providing a class matrix D i for each of the data matrices A i ;
providing a kernel K(A i , B) by each of said plurality of entities to allow public computation of a full kernel; and
computing a binary classifier that incorporates said public full kernel, wherein said classifier is adapted to classify a new data vector according to a sign of said classifier, wherein computing comprises:
extracting covariates x from the data matrix A i , and
maximizing a quantity w′K(x,B′) to solve for vector w, wherein said quantity w′K(x, B′) is an effect parameter in a survival model that characterizes an effect on said cancer survival rates.
12. The computer readable program storage device of claim 11 , wherein said class matrix D is of size m×m wherein in is a number of data points and has a value +1 on a main diagonal for each survival datum that exceeds a predetermined number of years, and has a value of −1 on said main diagonal for each survival datum that does not exceed the predetermined number of years.
13. The computer readable program storage device of claim 11 , wherein said random matrix B is real valued of size {tilde over (m)}×n, wherein n is a dimensionality of each data point, and {tilde over (m)}<n, and each data matrix A i is of size m×n, wherein in is a number of data points.
14. The computer readable program storage device of claim 11 , wherein said full kernel is
K
(
A
,
B
′
)
=
K
(
[
A
1
A
2
⋮
A
q
]
,
B
)
=
[
K
(
A
1
,
B
)
K
(
A
2
,
B
)
⋮
K
(
A
q
,
B
)
]
,
where q is a number of entities.
15. The computer readable program storage device of claim 11 , wherein said classifier is defined as K(x′,B′)w−γ=0 wherein w is a vector normal to a hyperplane separating the two classes of the binary classifier, γ determines the location of the separating hyperplane relative to the origin, and K(x′, B′) is a row vector of K(A, B′).
16. The computer readable program storage device of claim 15 , wherein the classifier is solved for using a Newton-Lagrangian method wherein a square of a 2-norm of a slack variable is minimized with weight ν/2 wherein ν>0 and a distance between bounding planes is measured in an (n+1)-dimensional space of (w, γ)εR n+1 .