A system for the prognostics of the chronic diseases after the medical examination based on the multi-label learning, comprising a data acquisition module, a data preprocessing module, a basic predicting model constructing module, and a local predicting module;
- wherein
- the data acquisition module is configured to acquire physical examination data of a physical examination user, the physical examination data comprises basic physiological indicators and routine assay indicators, all the forgoing indicators of the physical examination user are expressed as a feature vector X=[x1, x2, . . . xp]T, p is the total number of indicators, an occurrence time t0 of a current medical examination is recorded, the earliest occurrence time data T=[t1, t2, . . . tq]T of a diagnosis of various chronic diseases in diagnosis data in the chronic diseases in the user's electronic medical record before and after the medical examination is extracted, q is the number of categories of the chronic diseases preset by the system, in the forgoing data, X, t0 is necessary data, each of components in T is set as null when the diagnosis does not exist;
- a processing process of the data preprocessing module is specifically as follows:
- performing standardization processing ϕi on each component xi in X based on the component, so that a standard deviation of all data on the component is 1, and a mean value of all the data is 0, a standardized feature vector is denoted as X′=[x′1, x′2, . . . x′p]T;
- x
- i
- ′
- =
- ϕ
- i
-
- (
- x
- i
- )
- =
- (
- x
- i
- -
- λ
- i
- )
- σ
- i
- where, xi′ is a standardized data, λi is the mean value of all the data on the component xi, σi is the standard deviation of all the data on the component xi;
- performing an exponential operation on T based on a base number α(0<α<1) to generate a prognostic index vector Y=[y1, y2, . . . yq]T:
- y
- i
- =
- y
-
- (
- t
- i
- )
- =
- {
- α
- (
- t
- i
- -
- t
- 0
- )
- t
- i
- >
- t
- 0
- 1
-
- t
- i
- ≤
- t
- 0
- 0
-
- t
- i
-
-
- is
-
-
- null
- the basic predicting model constructing module is configured to construct a multi-label learning model for a physical examination scenario, and a constructing process is as follows:
- (1) constructing a multilayer neural network, and completing data processing from input to output via the multilayer neural network, wherein specific hyperparameters of the network comprise:
- a number K of layers of the network, a number n1, n2, . . . nK of nodes in each of the layers of the network, and an activation function between two adjacent layers, where n1=p, nK=q;
- transfer weight matrixes between two adjacent layers are denoted as W1, W2, . . . WK−1;
- and an output value of the nodes of the last layer is denoted as a predicted prognostic index C=[c1, c2, . . . cq]T;
- (2) designing a loss function E:
- E
- =
- 1
- N
-
- ∑
- k
- =
- 1
- N
-
- ∑
- i
- =
- 0
- 2
-
- λ
- i
- w
- ·
- E
- i
- k
- the loss function can be regarded as a weighted mean value of three loss functions set for different objectives, wherein N represents a data sample amount used in a single batch of gradient descent, and a superscript k of Eik represents that a component of the loss function is obtained by calculation of a k-th sample in the batch of gradient descent, λiw represents respective weight values of the three loss functions;
- a) E0 is a single-label loss function, which characterizes a difference between a predicted prognostic index ci and an actual prognostic index yi of disease prediction:
- E
- 0
- =
- -
- 1
- q
-
- ∑
- i
-
- y
- i
-
- log
-
- c
- i
-
-
- or
-
-
- E
- 0
- =
- 1
- q
-
- (
- y
- i
- -
- c
- i
- )
- 2
- b) E1 is an interval loss function, which characterizes a difference between a predicted prognostic index difference Δci,j and an actual prognostic index difference Δyi,j of two different chronic diseases:
- E
- 1
- =
- 1
- q
-
- (
- q
- -
- 1
- )
-
- ∑
- i
- ≠
- j
-
- [
- (
- y
- i
- -
- y
- j
- )
- -
- (
- c
- i
- -
- c
- j
- )
- ]
- 2
- =
- 1
- q
-
- (
- q
- -
- 1
- )
-
- ∑
- i
- ≠
- j
-
- (
- Δ
-
- y
- i
- ,
- j
- -
- Δ
-
- c
- i
- ,
- j
- )
- 2
- c) E2 is a ranking loss function, which characterizes a difference between a predicted occurrence order and an actual occurrence order of the two different chronic diseases:
- E
- 2
- =
- 1
- q
-
- (
- q
- -
- 1
- )
-
- ∑
- i
- ≠
- j
-
- e
- -
- (
- y
- i
- -
- y
- j
- )
-
- (
- c
- i
- -
- c
- j
- )
- =
- 1
- q
-
- (
- q
- -
- 1
- )
-
- ∑
- i
- ≠
- j
-
- e
- -
- Δ
-
- y
- i
- ,
- j
- ·
- Δ
-
-
- c
- i
- ,
- j
- (3) learning parameters:
- according to the physical examination data of a sample medical institution, several models M1, M2, . . . ML are constructed by matrix hyperparameter scanning, parameters of each of the models are learned based on a mini-batch gradient descent (MBGD), optimal parameters are determined via k-fold cross validation, an optimal model is used as a basic predicting model Mbest for migration to other medical institutions, Mbest is solidified into the basic predicting model constructing module;
- the local predicting module is arranged in a specific local medical institution and comprises a local model training unit and a predicting unit;
- the local model training unit receives the optimal basic predicting model Mbest provided by the basic predicting model constructing module, acquires the physical examination data X, t0, T of the specific local medical institution via the data acquisition module, generates X′, Y via the data preprocessing module, performs model training the same as a Mbest training method with model parameters of Mbest as initial parameters based on X′, Y, and solidifies a trained local predicting model MbestH into the local predicting module after the parameters are converged;
- the predicting unit performs the prognostics of the chronic diseases based on physical examination data of a new medical examinee according to the local predicting model MbestH, outputs a predicted prognostic index C=[c1, c2, . . . cq]T of the occurrence of the plurality of chronic diseases, and then obtains a future expected occurrence time t′i of the corresponding chronic disease via an inverse function t′i=y−1(ci) of y(ti).