Differentiating Boys With ADHD From Those With Typical Development Based On Whole-brain Functional Connectivity Using A Machine Learning Approach

Introduction: A growing body of studies suggest that neuropsychiatric diseases stems not only from circumscribed anomalies in discrete brain regions,but also from impairments in distributed neural networks.Previous resting-state functional magnetic resonance imaging(rs-f MRI)studies have demonstrated that there are brain network abnormalities in attention-deficit/hyperactivity disorder(ADHD),such as the frontostriatal-cerebellum pathway,cognitive control network,default mode network and other brain networks.However,these studies mianly used traditional group-level statistical methods,which fails to identify ADHD children at the individual level.In addition,these rs-f MRI studies assumed that resting-state functional connectivity(FC)is stationary over the entire scanning and ignored the fact that FC might change during this time.As a data-drivern technique,machine learning can complement the traditional statistical analyses.Therefore,current study aimed to use machine learning methods to explore the abnormal FC in children with ADHD.We conducted the following works: firstly,we explored the resting-state FC pattern of ADHD,which can be used to differate children with ADHD from typical development controls.Secondly,we explored the changes in the dynamic functional connections of ADHD.Material and Methods: All rs-f MRI data were from ADHD-200 database(http://fcon?1000.projects.nitrc.org/indi/adhd200/).Study 1,we measured the restingstate FC interregional functional connections from 40 ADHD patients(mean age 11.83 ± 2.88 years old)and 28 matched typical developed controls(TDC,mean age 11.99 ± 3.05 years old).Support vector machine based on whole brain resting-state FC was used to classify patients from typical developed controls.Classifier performance was assessed by permutation tests.To explore the correlation between the brain imaging index and the scales(inattention and hyperactivity/impulsivity scales),pearson correlation analysis was performed between behavioural scales and functional connectivity.Study 2,we performed time-varying connectivity analysis followed by kmeans clustering on rs-f MRI of 32 ADHD patients(mean age 12.66 ± 1.75 years old)and 52 matched healthy controls(mean age 12.09 ± 1.61 years old).Moreover the relationship between behavioural scales(inattention and hyperactivity/impulsivity scales)with amplitude of FC in regional level was explored.Results: The results of study 1 showed that 85.3% of participants were correctly identified using leave-one-out cross-validation of support vector machine.The majority of the most discriminative functional connections were located within or across the cerebellum,default mode network(DMN)and frontoparietal network(FPN).About half of the most discriminative connections were involved in the cerebellum.The cerebellum,right superior orbitofrontal cortex,left olfactory,left gyrus rectus,right superior temporal pole,right calcarine gyrus and bilateral inferior occipital cortex showed the highest discriminative power in classification.In the brain-behavior relationships,some functional connections between the cerebellum with DMN were significantly correlated with behavioral scales(inattention and hyperactivity/impulsivity scales)in ADHD(p < 0.05).Study 2,two distinct dynamic functional connectivity states(strong and week interaction state)were identified.There were significant group difference between ADHD with TDC in occurrence in each state and transitions of states.In week interaction state,ADHD presented decreased FC between frontal with insula,parietal and cerebellum as well as between thalamus with anterior cingulate cortex and parietal compared to TDC.The amplitude of FC in vm PFC and am PFC was significantly correlated with hyperactivity/impulsivity scores and in the precuneus was significantly correlated with inattention scores.By comparing with the results of static and dynamic FC,the changed static and dynamic FC showed a similar anatomical trend,and the spatial distribution of this trend was characterized by the change of large-scale brain network centered on the frontal-striatal-cerebellar pathway.Conclusions: Our results suggested that ADHD-related brain network is mainly distributed in brain regions centered on frontal-striatum-cerebellar pathway.The study 1 results indicated that whole-brain resting-state FC may provide potential effective biomarkers for clinical diagnosis of ADHD.Our findings in study 2 suggested that the ADHD-caused dynamic FC alterations mostly appeared in the cool interaction state.This paper provides evidence for the f MRI study of ADHD from the perspective of machine learning based on resting-state FC.