Research On Classification Of ADHD Based On Feature Selection In Resting-state FMRI

Attention deficit hyperactivity disorder(ADHD)is a common mental disorder in children,which characterized by lack of attention,hyperactivity,and difficulty in controlling one's behavior.The pathogenesis of ADHD is very complicated and it is not clear at present,and ADHD will have a great impact on children if it cannot be found and treated in time.In addition,the current diagnosis of ADHD mainly depends on the assessment scale,which leads to great subjectivity.Hence,finding the pathology of and correctly diagnosing ADHD bears importance in improving the lives of affected children.Resting state functional magnetic resonance imaging(fMRI)is widely used in the research of ADHD because it can reflect the activity of the brain in the absence of tasks.As an important biomarker of fMRI,fractional amplitude of low-frequency fluctuation(fALFF)can reflect the intensity of spontaneous neuronal activity.The diagnosis of ADHD is essentially a two classification problem,thus fALFF combined with feature selection methods and classifiers can be used in finding the pathology of and correctly diagnosing ADHD.The main research contents of this paper are as follows:First of all,fALFF of each voxel is used to construct features.After data preprocessing and calculate fALFF in 0.01-0.08 Hz,respectively;multiple linear regression method is introduced for the reduction of confounding effects.Our experimental results indicate that age and sex will bring extra noise when using a small number of samples,and thus have a negative impact on the classification of ADHD children and neurotypicals.However,multiple linear regression is a good method which can control these confounding effects.By analyzing the regression coefficients of the two categories,we found that age and sex have different effects on fALFF.fALFF of neurotypicals decreased significantly in posterior cingulate gyrus with age,and healthy girls usually show higher fALFF in posterior cingulate gyrus,caudate nucleus and superior temporal gyrus compared with healthy boys.Secondly,considering that huge number of voxels are consisted in one sample,classification and feature selection using whole voxels are time-consuming.In this work,principal component analysis(PCA)is used for feature extraction.Results of study show that PCA significantly decreased the dimension and computation cost.Thirdly,considering that the number of samples is far less than the number of voxels and the imbalance problem of samples,R-RELIEF(Reliable RELIEF)algorithm is proposed for finding the pathology of and correctly diagnosing ADHD.After feature selection using R-RELIEF algorithm,we analyzed the association between these abnormal brain areas and ADHD symptoms.Our experimental results indicate that abnormal fALFF of ADHD children are located in cerebellum,brainstem,prefrontal gyrus and superior temporal gyrus compared with neurotypicals.Fourthly,a novel method is proposed to classify ADHD children and neurotypicals.After the calculation of fALFF,principal components,volatility index,fractal dimension,kurtosis and interquartile range are used for building feature sets.A classification accuracy of 86.53%(SEN,85.61%,SPE,87.5%,F-score,87.49%)is obtained by using our method.Our experimental results indicate that our R-RELIEF algorithm has a better performance compared with RELIEF algorithm and minimum redundancy maximum relevance algorithm.In the classification of ADHD children and neurotypicals,most of studies were based on structural features and functional connectivity features.In this paper,fALFF is used as a biomarker.Multiple linear regression method and binarization are introduced to process data.PCA is used to calculate principal components.After that,principal components,volatility index,fractal dimension,kurtosis and interquartile range are used to achieve the classification of ADHD subjects and neurotypicals.Consequently,a good classification performance is achieved by using R-RELIEF algorithm and the processed data.Our experimental results indicate that fALFF is a reliable fMRI marker for investigation of ADHD.