Research On The Application Of Graph Machine Learning In Predicting Behavioral Traits Through Brain Network

Functional similar,spatially separated brain regions with highly synchronized activity patterns organized as functional brain network,which not only support the neurocomputing in cognitive processing,but also determine individual differences in personality traits and emotional characteristics.Studies have a long-standing interest in understanding the relationship between network connectivity patterns and behavioral individuals’ difference under the interdisciplinary of neuroscience and machine learning.However,the application of graph machine learning(GML)in the prediction of behavioral traits based on network connectivity is still incisive.First,interpretation-oriented prediction research is mostly based on data-driven methods,theory-driven machine studies were lacked,which hinder the form of systematic theoretical.Second,behavioral prediction based on machine learning methods mainly focuses on the theoretical interpretability,which sacrificing the predictive accuracy by ensure the feature and model interpretability.For example,the topological characteristics of brain networks and the interactions of the input measures at the higher dimensional space are not fully utilized.Therefore,the current study applies graph machine learning method on behavioral prediction under both the guidance of interpretability and accuracy.Study 1 was an explanation-oriented mechanism study,which aimed to examine whether the theoretical-driven GML can be applied to the network mechanisms underlying behavioral traits.Specifically,the theory-dirven and data-driven connectome-based predictive model(td CPM and dd CPM)was constructed.In feature selection,td CPM assumes that functional integration and separation of frontoparietal network(FPN),attention network(AN)and default network(DN)might support working memory process,and functional connectivity within/between above three networks were calculated as input characteristics,while dd CPM has no specific theoretical hypothesis and takes the whole brain network measures as input characteristic.Same samples,network building and fitting method were applied to td CPM and dd CPM.A total of 130 healthy adult subjects were collected from the open source database from Consortium for Neuropsychiatric Phenomics.111 subjects were retained after thehead movement control,health screening,and data integrity checks.Spatial and verbal working memory tasks were used to measure the working memory ability.Dosenbach160 atlas was used to define nodes within DN,AN and FPN,and the average time series correlation was used to calculate the functional connectivity between nodes.The connectivity strength within and between these networks was used as the input variables,and the relationship between the input feature(functional connectivity)and the predicted variable(working memory)is fitted by the linear regression model with clear decision-making process.The results show that td CPM could predict working memory and achieve similar prediction accuracy of dd CPM,indicating that the hypothetical-driven GML was a potential method which can be applied to neural mechanism exploration.Study 2 is an accuracy-oriented predictive study which designed to verify that GML based on topological measures and artificial neural networks can achieve accurate predictions of behavioral traits.First,500 subjects were randomly selected from the Human Connectome project database,including behavioral traits and resting-state functional magnetic resonance data.Fluid intelligence,openness of the big five personality,and the sad traits were selected as representative variables of cognition,personality and emotion.Subsequently,the whole brain network was constructed using the method of study 1.In terms of feature selection,clustering coefficient,and node efficiency are calculated as input characteristics.In terms of model selection,shallow-wide and narrow-deep multilayered perceptrons(sw MLP and nd MLP)were use to fit the relationship between connection parameters and the behavioral traits.Finally,the model performance was examined by using of the 5-folder cross-validation method.The results show that sw MLP and nd MLP can predict all three behavioral characteristics,and the accuracy of sw MLP is generally higher than nd MLP.The results suggest that GML based on topological parameters and artificial neural networks can be applied to accuracy-oriented predictive studies,but the optimal model architecture needs to be considered based on characteristic types in practice.In conclusion,several findings of the study are as follows: 1)Previous explanatoryoriented mechanism studies have been based on data-driven methods,and there is a lack of theoretical-driven predictive research.By constructing td CPM and dd CPM,study 1 found that working memory ability may be related to the functional separation and integration of AN,DN and FPN.This study provided that GML with clear decision-making process can be used as mechanism studies,and can be used to guide the formation of specific theory.2)Previous studies aimed to predict behavioral traits based on GML paid efforts to the explanatory the models,which ignoring the predictive accuracy.The current study used topological parameters and neural networks to realize the prediction of multi-type behavioral traits,which shows that neural network-based GML could to be applied to behavioral trait prediction.