A Study On Neuroimaging Biomarkers Of ADHD And Its Related Individual Behaviors

With the development of neuroimaging technology and the emergence of numerous large-scale brain imaging data sets,it has become frontier and hot topics of the field of neuroscience and mental illness research to mine biomarkers related to individual specific behaviors or mental disorders based on human brain neuroimaging data.In recent years,many of the brain-individual behaviors correlations or brain-mental illnesses correlations neuroimaging studies were focused on looking for patterns of brain functional activity that were activated by stimulation of specific cognitive tasks(for example,the Go/no-go and N-back task).But because affected by medical prior knowledge and hypotheses,researchers only focus on whether regions of interest or seed regions respond to stimulation of specific cognitive task,which makes it difficult to obtain ground-breaking research results.On the one hand,to this day,little is known about these relationships between human brain function or structure and individual behavior or mental illnesses.On the other hand,a growing number of human brain magnetic resonance imaging(MRI)data are available for scientific research.Therefore,we should make the most of existing or newly developed data mining algorithms in statistics to mine the laws of data itself from high-dimensional brain MRI data by using a pure data-driven approach,so as to obtain these real brain-individual behaviors or brain-mental disorders relationship patterns.In this thesis,we use brain multi-modal MRI data(i.e.,structural MRI and functional MRI)and four individual behaviors(i.e.,inhibitory control,intelligence,inattentive,hyper/impulsive)measurements associated with attention deficit/hyperactivity disorder(ADHD),as well as diagnosis of ADHD to carry out a research on these relationships between multi-modal high-dimensional brain features and specific behaviors or ADHD disorder.The main body of this dissertation consists of two parts.In the first part,we used pure data-driven methods and statistical models to mine brain features closely related to four individual behaviors such as inhibitory control,respectively.First,after standard pre-processing of MRI data,we extracted high-dimensional brain features at the level of voxel or brain region from high-resolution brain MRI data.In this thesis,a total of six kinds of brain feature measurements were used.Among them,brain functional features included functional connectivity(FC),fractional amplitude of low frequency fluctuations(f ALFF)and regional homogeneity(Re Ho),and brain structural features included morphological connectivity(MC),gray matter volume(GMV)and white matter volume(WMV).Since the overfitting probability of a statistical model or machine learning algorithm will increase with the increase of the dimension of input feature,we will then reduce the dimension of high-dimensional brain features obtained in the previous step.Dimensionality reduction methods include feature selection and feature extraction.In this thesis,we used three feature selection or feature extraction methods,including feature selection based on Pearson's correlation coefficients,feature selection based on recursive feature elimination,and feature extraction based on principal component analysis.Next,using the reserved brain feature set,we constructed and trained the statistical prediction model on the training data set,in which individual behavior measurement was defined as dependent variable and reserved brain features measurements as independent variables while controlling for age and gender variables.We used two types of statistical prediction model,that is,multiple linear regression model and kernel support vector regression model.Later,on the test data set,we used the fitted prediction model to predict new subjects,and adopted a variety of strategies to perform model evaluation.In addition,we also used cross-validation and independent data sets to perform validation analysis.Finally,based on the assessment of the predictive model,we extracted brain features that were most explanatory of the phenotypic variable.In the second part,using a public database with seven independent data sites,we studied the possible brain functional and structural developmental defects of patients with ADHD by performing lateralization analysis.for each data site,after obtaining multiple high-dimensional brain features from brain multi-modal high-resolution MRI data,we performed an inter-group lateralization analysis for each kind of brain feature measurement.In other words,the distribution of cortical regions with delayed development or developmental defects was explored by comparing the hemispheric asymmetry of each of regional measurement between adolescents with ADHD and typically-developing control.Then,we further integrated the findings of each data site by using a random-effects meta-analysis.Generally speaking,our findings showed that an effective pure data-driven data mining algorithm could extract brain features with high explanatory ability from high-dimensional structural or functional MRI data.Compared with the prediction effect of single mode single brain features,the prediction model combining multiple brain features in multi-model has better prediction effect,which is consistent with the research on combining multi-model neuroimaging data advocated by brain scientists.Moreover,our findings highlighted that patients with ADHD had the abnormalities in brain asymmetry of many areas and the greater the degree of hemispheric asymmetry in some brain areas,the worse the subjects performed on tests of inattentive,hyper/impulsive,and intelligence.We believe that the results can provide effective evidence for the pathology of ADHD related to brain development.Data mining of brain multi-model features concerned in this thesis can become an effective way to deeply understand these real brain-individual behaviors or brain-mental disorders relationship patterns and to rationally explore the pathological mechanism of mental illness In the era of data deluge.