Multi-label Classification Of Individual Cognitive Indexes Based On Brain Image

It is a hot spot in neuroimaging research to classify individual cognitive indexes based on brain image data in combination with machine learning,especially for the classification of mental illness.However,most of such studies are based on single label classification technology,which means each sample has only one label.Multi-label classification allows a sample to have multiple related labels.In the classification of individual cognitive indexes based on brain image data,multi-label classification can not only classify multiple cognitive indexes at the same time,but also can lead to a more reliable classification results based on the full use of correlation between labels.In a wide range of brain images,resting state f MRI is considered as one of the brain imaging techniques with great potential for clinical application because of its high temporal and spatial resolution.Based on resting state f MRI data,this study carried out multi-label classification of individual cognitive indexes,and employed hamming loss,one error,ranking loss,coverage and average precision to evaluate the classification effect.This study was carried out in the following three aspects:(1)Feature extraction and fusion for classification of individual cognitive indexes.Effective feature extraction and fusion is the key to classification of individual cognitive indexes.In this study,in addition to the traditional time-domain functional connectivity and complex brain network parameters,the frequency band functional connectivity feature is introduced,and then the individual cognitive indexes were classified based on three single features.After that,two features with high classification accuracy were combined to carry out multi-label classification of individual cognitive indexes.The experimental results show that: compared with the two time-domain features,the frequency functional connection feature has a higher classification accuracy,and the classification accuracy based on a single frequency-domain feature is similar to that based on the multi feature fusion.(2)Utility analysis of classical multi-label classification algorithms in individual cognitive indexes classification.At present,it is not clear which classification method is more suitable for multi-label classification of individual cognitive indexes.In this study,the performance of five classical multi-label classification algorithms in individual cognitive indexes classification were systematically analyzed from the perspective of feature space and label space.The experimental results show that: compared with label space information,the full feature space information is more conducive to multi-label classification of individual cognitive indexes.(3)Multi-label classification of individual cognitive indexes with unbalanced data sets and missing labels.At present,the research on individual cognitive indexes classification based on brain image is still in the laboratory stage,and the problem of data sets imbalance and labels missing has not been considered.In this study,imbalanced multi-label learning and semi-supervised weak label learning were introduced into the multi-label classification of individual cognitive indexes,and improved to solve these two problems at the same time.The experimental results show that: the introduction of imbalanced multi-label learning and semi-supervised weak label learning can effectively improve the multi-label classification accuracy of individual cognitive indexes.The contributions of this study are as follows:(1)The frequency band function connection feature is introduced.The results show that: this feature is more sensitive to multi-label classification of individual cognitive indexes than the traditional f MRI time domain features;(2)By analyzing many classical multi-label classification algorithms,it is found that the information of full feature space is more conducive to multi-label classification of individual cognitive indexes than label space information;(3)Imbalanced multi-label learning and semi-supervised weak label learning are introduced for the first time to solve the problem,which also improve the multi-label classification accuracy of individual cognitive indexes.