US 12,517,057 B2 · Interaction of spectroscopy and artificial intelligence for serological analysis and its applications

Abstract

A spectroscopy and artificial intelligence-interaction serum analysis method includes: collecting bulk SERS spectral data of clinical serum samples, performing dimension reduction on the spectral data by using a covariance matrix to obtain spectral different peak positions of cancer patients and normal individuals, and performing spectral data processing and algorithm identification by using an svm model of an artificial intelligence algorithm to obtain a cancer identification rate. Compared with the conventional serum analysis method, the spectroscopy and artificial intelligence-interaction serum analysis method requires no antibody-antigen or other biological specificity modification processes, and the serum of cancer patients and normal individuals can be identified more cheaply, rapidly and accurately. Also the different peak positions in SERS spectra of a large amount of serum samples can be located, which provides an entirely novel detection and analysis method at a molecular bond energy level for the field of liquid biopsy of clinical cancers.

A spectroscopy and artificial intelligence-interaction serum analysis method comprising:

(1) preparing a purified silver nanowire solution, wherein the purified silver nanowire solution includes silver nanowires without an intrinsic Raman signal as surface enhanced Raman scattering (SERS) probes;
and in addition, centrifuging peripheral blood plasma samples of patients with different types of diseases and normal people to obtain corresponding serum samples;
(2) performing liquid-phase mixing and incubation on the silver nanowire solution and the serum samples according to a same proportion to ensure that the silver nanowires are fully contacted with the serum samples, and after incubation, performing bulk SERS spectral data collection on the serum samples by using a Raman spectrometer, wherein to implement the bulk SERS spectral data, a laser having a wavelength of 532 nm is employed such that a spectrum collection range is 600 cm−1-1800 cm−1, and each sample is subjected to the bulk SERS spectral data collection for 5 times;
(3) after the performing the bulk SERS spectral data collection, performing dimension reduction on serum SERS spectral data from different sources to remove irrelevant items in sample data points, and screening effective dimensions capable of reflecting data difference by calculating an original data dimension relevancy among different samples by using a covariance matrix, and then taking data points with a lowest relevancy as effective dimensions after the dimension reduction, the effective dimensions corresponding to different peak positions among different sample types of the serum samples;
(4) performing algorithm training:
performing binary classification processing to obtain identification accuracy rates of different diseased samples of the serum samples compared with normal samples of the serum samples by taking the data points subjected to the dimension reduction as characteristic values during algorithm training and identification, obtaining a training set of samples and utilizing at least a portion of the serum samples as a test set, and scaling data of each sample, a scaling range being [0, 1], a normalization formula used in a scaling process being:
y
′
=
lower
+
(
upper
-
lower
)
*
y
-
min
max
-
min
;
wherein, y is data before scaling, y′ is data after scaling, lower and upper are minimum and maximum values of the data after scaling, and min/max are minimum/maximum values of the data before scaling;
a corresponding support vector expansion being:
f
⁢
(
x
)
=
w
T
⁢
ϕ
⁢
(
x
)
+
b
=
∑
i
=
1
N
α
i
⁢
y
i
⁢
ϕ
⁡
(
x
i
)
T
⁢
ϕ
⁢
(
x
)
+
b
=
∑
i
=
1
N
α
i
⁢
y
i
⁢
k
⁢
(
x
,
x
i
)
+
b
;
wherein, k(x, xi) is a kernel function, N is a number of support vectors, x denotes an input sample data set, f(x) denotes a decision function, wT denotes a normal vector, ϕ(x) maps the input sample data set to a feature space, b is a displacement term, αi is a Lagrange multiplier, and yi denotes a support vector label and the above formula shows that an optimal solution of the support vector machine model is expanded through the kernel function of training samples;
the kernel function used in the algorithm processing being a radial basis kernel function (that is, an RBF kernel function), that is:
K
⁡
(
x
j
,
x
j
)
=
exp
⁡
(
-
γ
⁢

x
ι
-
x
j

2
)
,
γ
>
0
;
γ being a hyperparameter of a Gaussian kernel function and xi and xj denoting two sample vectors in a feature space;
specifically:
first, converting an original problem into a convex optimization problem:
the original problem:
min
w
,
b
,
ξ
1
2
⁢

w

2
+
C
⁢
∑
i
=
1
N
⁢
ξ
i
s
.
t
.
y
i
⁢
(
w
·
x
i
+
b
)
≥
1
-
ξ
i
,
i
=
1
,
2
,
⋯
,
N
ζ
i
≥
0
,
i
=
1
,
2
,
⋯
,
N
;
then solving the convex optimization problem;
{circle around (1)} For a dual problem of the original problem, constructing a Lagrangian function:
L
⁡
(
w
,
b
,
ξ
,
α
,
μ
)
≡
1
2
⁢

w

2
+
C
⁢
∑
i
=
1
N
⁢
ξ
i
⁢
1
2
-
∑
i
=
1
N
⁢
α
i
(
y
i
(
w
·
x
i
+
b
)
-
1
+
ξ
i
)
-
∑
i
=
1
N
⁢
μ
i
⁢
ξ
i
;
wherein, α is a Lagrangian multiplier;
w is a normal vector on plane and determines a direction of a hyperplane;
b is a displacement term and represents a distance from the hyperplane to an origin;
ξ represents a relaxation variable;
u is a dual variable, minimum values of w, b, and ξ are firstly solved, partial derivatives are solved respectively and the derivatives are let to be 0, then results are substituted into the original function, the maximum value of α is solved for the minimum value, and then maximum value solving is converted into minimum value solving to get the dual problem:
min
a
1
2
⁢
∑
i
=
1
N
⁢
∑
j
=
1
N
⁢
α
i
⁢
α
j
⁢
y
i
⁢
y
j
⁢
(
x
i
·
x
j
)
-
∑
i
=
1
N
⁢
α
i
s
.
t
.
∑
i
=
1
N
⁢
α
i
⁢
y
i
=
0
0
≤
α
i
≤
C
,
i
=
1
,
2
,
⋯
,
N
;
selecting
⁢
K
⁡
(
x
j
,
x
j
)
=
(
-
γ
⁢

x
i
-
x
j

2
)
,
γ
>
0
as the kernel function;
{circle around (2)} from Karush-Kuhn-Tucker (KKT) condition establishment, obtaining:
w
*
=
∑
i
=
1
N
⁢
α
i
*
⁢
y
i
⁢
x
i
b
*
=
y
j
-
∑
i
=
1
N
⁢
α
i
*
⁢
y
i
(
x
i
·
x
j
)
;
wherein, C is a penalty coefficient, that is a tolerance to an error, g is a parameter of the Radial Basis Function (RBF) function after the RBF function is selected as the kernel function, * denotes the optimal solution, and best parameters for C and g are selected by grid optimization through a parameter optimization tool grid. py in libsvm;
a relation between γ and g being deduced from the following formula:
k
⁡
(
x
,
z
)
=
exp
⁡
(
-
d
⁡
(
x
,
z
)
2
2
*
σ
2
)
=
exp
⁢
(
-
gamma
·
d
⁡
(
x
,
z
)
2
)
⇒
gamma
=
1
2
·
σ
2
;
wherein, d(x, z) is the distance, gamma=γ, that is, a value of g is equal to the hyperparameter value of the Gaussian kernel function, and σ is a width parameter of the function;
after the kernel function and the parameters C and g are selected, performing training by using the training set to obtain an svm model for the serum SERS spectral data, a classification decision function used in this process being:
f
⁡
(
x
)
=
sign
⁡
(
∑
i
=
1
N
α
i
*
⁢
y
i
⁢
K
⁡
(
x
i
,
x
)
+
b
*
)
;
wherein a* is obtained by an smo algorithm, K (xi, x) corresponds to the Gaussian kernel function, and b* is a threshold;
selecting a hinge loss function as the loss function,
λ
⁢

w

2
being a regularization term, wherein λ is a regularization parameter, that is:
min
w
,
b
∑
i
=
1
N
max
⁡
(
0
,
1
-
y
i
(
w
·
x
i
+
b
)
)
+
λ
⁢

w

2
;
and (5) testing the obtained model by using the test set, comparing an actual situation with a model prediction result, and finally obtaining identification accuracy rate and outputting a result.

562 more elements

Interaction of spectroscopy and artificial intelligence for serological analysis and its applications

Life of this patent

Interaction of spectroscopy and artificial intelligence for serological analysis and its applications

Life of this patent

Abstract

Claims7