Multi-label Estimation Of Individuals’ Cognitive Variables Based On Magnetic Resonance Image

It is a research hotspot in cognitive neuroscience to estimate individuals’ cognitive variables based on functional magnetic resonance images(fMRI).These estimations can explore the individual differences in brain structure and function,and are helpful for diagnosis and early warning of neuropsychiatric diseases.However,a majority of these studies used single-label learning rather than multi-label learning techniques.Multi-label learning considers the correlation among labels,which can provide richer information to improve the accuracy of estimations.At present,most of the quantitative estimations of individuals’ cognitive variables is based on resting state functional magnetic resonance imaging(fMRI).Recent studies have shown that naturalistic fMRI is superior to resting-state fMRI in many key aspects,quantitative estimations of individuals’ cognitive variables based on naturalistic fMRI has not yet been performed.Therefore,in this study,we introduced the multi-label learning technologies to estimate multiple individual cognitive variables simultaneously,and we systemically explored the performance of predictive models based on naturalistic fMRI and resting state fMRI.Specifically,the details of three studies in this paper are as follows:(1)Estimations of multiple cognitive variables based on PLSR algorithm.This work systemically investigated the performance of PLSR in fMRI-based estimations of individual cognitive variables,especially its performance in simultaneous estimations of multiple cognitive variables(multi-label learning).We used resting state functional connections(RSFCs)based on partial correlation and full correlation as features,and a total of ten cognitive variables were estimated.The results showed that PLSR performed well in both single-and multi-label learning.RSFC features is helpful to inference the biological significance underlying the estimations.Besides,this work also found that the estimation accuracies based on RSFCs among 100,200 and 300 ROIs were higher than those based on RSFCs among 15,25 and 50 ROIs;the estimation accuracies based on RSFCs evaluated using partial correlation were higher than those based on RSFCs evaluated using full correlation.(2)Estimations of multiple cognitive variables based on canonical correlation analysis(CCA)and its variants.This work proposed three multi-label estimation models combining one of three CCA models with ridge regression,and used RSFCs based on partial correlation and full correlation as features.Then we used two different label sets to explore the generalization performance of the models.The results showed that the three models perform well in the multi-label learning of cognitive variables based on fMRI.Among them,the estimation model with the kernel function performs the best.Besides,it is once again proved that high-resolution ROI and the use of partial correlation RSFCs can improve the accuracy of multi-label estimation.(3)Estimations of multiple cognitive variables based on movie fMRI.In this study,we extracted three classic functional features(FC,ISC,ISFC)and two new brain functional features(natural stimulus response amplitude and residual signal functional connection features)from movie fMRI data.We then used the PLSR algorithm to evaluate individuals’ cognitive variables based on each of the 5 type of features.The results showed that the five brain function features contribute to varying degrees in the quantitative estimations of individuals’ cognitive variables.Among them,the FC and residual signal functional connection features are more suitable for multi-label estimations of cognitive variables.The innovations of this study are as follow:(1)The performance of PLSR in multi-label estimations of cognitive variables was systematically explored,and it was proved that the algorithm was universal applicability and can achieve high accuracy.(2)Multi-label learning models which were CCA models combined with ridge regression are applied to the estimations of multiple cognitive variables and obtains high prediction accuracy.(3)It is the first time to explore the multi-label learning of cognitive variables based on Naturalistic movie-watching fMRI.And proposed new brain function features for cognitive variables estimations.