Prediction Of Individuals’ Intelligence Quotient Based On Resting State FMRI

Intelligence quotient(IQ)is an important indicator to measure individuals’ cognitive ability,which is closely bound up with brain function.Its objective and effective measurement is not only of scientific significance but also profitable for practical applications.Resting state resonance imaging(f MRI)technology obtains spontaneous brain activity based on blood oxygen level dependence(BOLD)signals.Its non-invasive,high time and spatial resolution,and comparability between different samples make this technique extensively used in the study of brain function.With the progress of machine learning technology,the objective measurement of IQ based on f MRI has become a hot topic in the field.There are still many drawbacks in the objective prediction of IQ based on f MRI in previous studies.(1)Most of the existing researches ignore the topological features of the human brain network.The input of BrainNetCNN,which is designed for the topology of the brain network,does not conform to the property of the human brain network,which means,the human brain network is scale-free.(2)Most studies to date are based on a single template for predictions.(3)Most of the existing studies extract features based on the assumption of time stability.To this end,this paper based on resting state f MRI images from the HCP1200 dataset which has 991 samples,conducts three studies to achieve more effective predictions of individuals’ fluid intelligence,crystal intelligence,and total intelligence:(1)Prediction of individuals’ IQ based on the sparse BrainNetCNN.To solve the problem that the human brain network is not fully connected,this paper proposes to construct a sparse BrainNetCNN for predictions of individuals’ IQ.In other words,only strong functional connections are selected as the input to the BrainNetCNN.The experimental results indicate that the prediction accuracies of individuals’ fluid intelligence,crystal intelligence,and total intelligence are improved by the sparse BrainNetCNN,compared with the original model without sparse operation.(2)Prediction of individual’s IQ based on multi-template learning.To take full advantage of the complementary information of human brain function,which is provided by different templates,the present study applies multi-template learning methods based on templates among 100,200,and 300 regions of interest(ROI)for predictions of individuals’ IQ.One of the multi-template learning methods is a simple connection of features.The other is an ensemble learning strategy.The experimental results point out that multi-template learning methods are superior to single-template learning methods.Multi-template learning based on the ensemble learning strategy is superior to multitemplate learning based on the simple connection of features.The performance of deep models is better than that of shallow models.(3)Prediction of individuals’ IQ based on dynamic functional connectivity.In order to capture the dynamic changes of the functional network of the human brain,the present study extracts the mean and variance of dynamic functional connectivity in the time dimension as features and further carries out predictions of individuals’ IQ.The experimental results indicate that predictions of individuals’ IQ can be achieved based on dynamic attribute characteristics,but the prediction accuracy is lower than that based on static functional connectivity.The fusion of dynamic attribute characteristics and static functional connectivity could not effectively improve the prediction accuracies of IQ.These results suggest that static functional connectivity may be a more stable and reliable method to describe individuals’ brain function for predictions of individuals’ IQ.In summary,the contributions of this study are as follows:(1)The sparsity is introduced into BrainNetCNN to achieve a more accurate prediction of individuals’ IQ.(2)The prediction of individuals’ IQ is further improved by using multiple templates to extract complementary features of human brain function.(3)The present study tries to achieve objective prediction of individuals’ IQ based on dynamic attribute characteristics that represent the functional networks of the human brain.