US 12,260,350 B2
Method for constructing target prediction model in multicenter small sample scenario and prediction method
Pengjiang Qian, Wuxi (CN); Zhihuang Wang, Wuxi (CN); Shitong Wang, Wuxi (CN); Yizhang Jiang, Wuxi (CN); Wei Fang, Wuxi (CN); Chao Fan, Wuxi (CN); Jian Yao, Wuxi (CN); Xin Zhang, Wuxi (CN); Aiguo Chen, Wuxi (CN); and Yi Gu, Wuxi (CN)
Assigned to JIANGNAN UNIVERSITY, Jiangsu (CN)
Filed by JIANGNAN UNIVERSITY, Wuxi (CN)
Filed on Jan. 13, 2024, as Appl. No. 18/412,519.
Application 18/412,519 is a continuation of application No. PCT/CN2023/116931, filed on Sep. 5, 2023.
Claims priority of application No. 202310807852.0 (CN), filed on Jul. 4, 2023.
Prior Publication US 2025/0013898 A1, Jan. 9, 2025
Int. Cl. G06N 5/04 (2023.01); G06N 5/048 (2023.01)
CPC G06N 5/048 (2013.01) 7 Claims
OG exemplary drawing
 
1. A method for constructing a target prediction model in a multicenter small sample scenario, comprising steps of:
S1: inputting a feature of a training sample into a multicenter training data set with m data centers, and predicting the training sample separately by using a plurality of subclassifiers and a historical knowledge classifier, wherein the historical knowledge classifier is a classifier that integrates rules of all the subclassifiers; and there are a total of m+1 dimensions after all prediction results are integrated, each dimension represents a prediction result of one classifier, a prediction result greater than 0 is determined as a positive class, and a prediction result less than 0 is determined as a negative class;
S2: normalizing a prediction result vector into [−1, 1];
S3: training a weight of each subclassifier using a ridge regression method by using a normalized prediction vector as a prediction feature and combining a training label corresponding to the training sample;
S4: obtaining a comprehensive prediction result by combining the prediction vector and the weight of the each subclassifier obtained through training; and
S5: classifying the comprehensive prediction result to obtain a result same as the training label,
wherein training of a first subclassifier comprises a basic phase and a learning phase; training of another subclassifier is based on training of the first subclassifier, and comprises the learning phase; the basic phase comprises: acquiring initial fuzzy rule parameters c, Pg, and δ and information of error classification samples errors by using classic 0-order Takagi-Sugeno-Kang (TSK), and using a Gaussian kernel function as a membership function, wherein c and δ are antecedent parameters, c is clustering centers obtained through fuzzy c-means (FCM) clustering of input data Data, and δ is a membership function kernel width obtained according to the input data Data, a membership degree matrix U of Data obtained through the FCM clustering for the clustering centers, and c; and the learning phase comprises: first performing knowledge discarding on c and Pg, wherein the knowledge discarding comprises: first, determining a knowledge discarding ratio r1 and a knowledge chaotization feature ratio r2, randomly selecting, according to r1, a rule that requires the knowledge discarding, and randomly selecting, according to r2, a feature column lose_columns that requires knowledge chaotization; for each rule that requires the knowledge discarding, randomly determining that knowledge forgetting or the knowledge chaotization is required; directly setting c and Pg to null for a rule that requires the knowledge forgetting; and for a rule that requires the knowledge chaotization, replacing a column lose_columns of the rule with a column value corresponding to another rule.