Self-paced learning based multi-kernel KRR for brain structure analysis in patients with different blood pressure levels

Various features can provide rich information for analysis of brain structural changes in hypertension. However, the extraction of multiple features involves complex data processing and results in long time-consuming. It is worthy of in-depth study that how to make a single feature achieves the same diagnostic effect as multiple features. Kernel ridge regression (KRR) has shown faster learning speed and generalization ability for classification tasks, which integrates multiple features as additional privileged information (PI) to help train an efficient classifier. This allows using only a single feature during test stage. In order to make the classifier effect better, the influence of sample attributes needs to be considered in the process of feature fusion. In this work, we construct a novel self-paced learning based multi-kernel KRR framework for analysis of brain structural changes in patients with different blood pressure levels. Specifically, one type of feature is taken as the main feature, and the rest of the features are first fed into the multi-kernel KRR and the output is serviced as PI. These two inputs are fed into the final KRR classifier together. Furthermore, self-paced learning is introduced to reduce the noise in multi-source data, which can prevent the model from falling into poor local solution and improve the generalization performance of classifier. Experimental results show that the proposed method can maximize the use the information of various types of features and achieve better classification performance. It suggests that the proposed self-paced based KRR can help to analyze the brain structure in patients with different blood pressure levels.