Multi-center synergetic cancer prognosis prediction system based on multi-source migration learning

A multi-center synergetic cancer prognosis prediction system based on multi-source migration learning, comprising a model parameter setting module, a data screening module, and a multi-source migration learning module, wherein,

• the model parameter setting module is arranged at a management center and is responsible for setting cancer prognosis prediction model parameters, comprising a cancer category, a source center and a target center, sample characteristics, a sample data preprocessing method, and a prognosis index;
• the management center performs coordination management of resources of each clinical center and accepts access of a user;
• the source center is a clinical center with labeled samples for a specific cancer category, and is responsible for the training of a source cancer prognosis prediction model;
• the target center is a clinical center with unlabeled samples for a specific cancer category, and is responsible for training a target cancer prognosis prediction model;
• the clinical center is an institution that actually holds clinical data, and is responsible for screening sample data and training a cancer prognosis prediction model;
• the data screening module is arranged at a clinical center, the management center transmits the set model parameter to each clinical center, such that each clinical center screens data by utilizing the data screening module, inquires a sample feature and prognosis index data from the local database of the clinical center according to the model parameter, and preprocesses the sample data according to a set sample data preprocessing method, so that the source center obtains a labeled sample set and the target center obtains an unlabeled sample set;
• the multi-source migration learning module comprises a source model training unit, a migration weight calculation unit, and a target model calculation unit;
• the source model training unit is arranged at each source center, it is assumed that there are K source centers denoted by S1,S2,S3 . . . SK, and an ith source center trains a local source cancer prognosis prediction model fSi(⋅) through its source model training unit and transmits the trained source model back to the management center;
• the migration weight calculation unit is arranged at the target center, and receives the K source cancer prognosis prediction models sent by the management center, it is assumed that there are nT unlabeled samples at the target center, an ith unlabeled sample is expressed as xiT, the K source cancer prognosis prediction models are utilized to perform prognosis prediction of the sample xT, respectively, to obtain a prediction label vector HiS:
• H
• i
• S
• =
• [
• f
• S
• 1
• ⁡
• (
• x
• i
• T
• )
• ,
• f
• S
• 2
• ⁡
• (
• x
• i
• T
• )
• ,
• …
• ,
• f
• S
• K
• ⁡
• (
• x
• i
• T
• )
• ]
• weighted summation is performed on the K prediction labels in the prediction label vector HiS to obtain a pseudo label ŷiT of the sample xiT:
• y
• ^
• i
• T
• =
• H
• i
• S
• ⁢
• θ
• =
• ∑
• j
• =
• 1
• K
• ⁢
• θ
• S
• j
• ⁢
• f
• S
• j
• ⁡
• (
• x
• i
• T
• )
• where θ=[θS1, θS2, θSK]T represents the migration weight of each source model, the weight that minimizes the difference between two samples in the target center sample set can be sought based on a smoothness assumption on the sample data of the target center, which is expressed as the following optimization problem:
• min
• θ
• :
• θ
• ′
• ⁢
• e
• =
• 1
• ,
• θ
• ≥
• 0
• ⁢
• ∑
• i
• ,
• j
• =
• 1
• n
• T
• ⁢
• (
• H
• i
• S
• ⁢
• θ
• -
• H
• j
• S
• ⁢
• θ
• )
• 2
• ⁢
• W
• ij
• where θ′ is a transposition of θ, e is a unit vector, and Wij indicates the similarity among samples;
• the aforementioned optimization problem is transformed into:
• min
• θ
• :
• θ
• ′
• ⁢
• e
• =
• 1
• ,
• θ
• ≥
• 0
• ⁢
• θ
• ′
• ⁢
• H
• S
• ′
• ⁢
• L
• T
• ⁢
• H
• S
• ⁢
• θ
• where HS is a nT×K matrix, and LT represents a graph Laplacian operator related to the target center, which can be obtained by calculating according to LT=D−W, where W is a similarity matrix of the samples of the target center, and D is a diagonal matrix obtained by calculating according to Dii=Σj=1nT Wij;
• therefore, the optimization problem is transformed into a standard quadratic programming problem, which is solved to obtain a migration weight θ;
• the target model calculation unit is arranged at the target center, obtains a sample pseudo label according to the migration weight θ, trains the target cancer prognosis prediction model at the target center by utilizing the pseudo label, and transmits the trained target model back to the management center.