Multi-modal, Multi-measure, and Multi-class Discrimination of ADHD with Hierarchical Feature Extraction and Extreme Learning Machine Using Structural and Functional Brain MRI

Abstract

Structural and functional MRI unveil many hidden properties of the human brain. We performed this multi-class classification study on selected subjects from the publically available attention deficit hyperactivity disorder ADHD-200 dataset of patients and healthy children. The dataset has three groups, namely, ADHD inattentive, ADHD combined, and typically developing. We calculated the global averaged functional connectivity maps across the whole cortex to extract anatomical atlas parcellation based features from the resting-state fMRI (rs-fMRI) data and cortical parcellation based features from the structural MRI (sMRI) data. In addition, the preprocessed image volumes from both of these modalities followed an ANOVA analysis separately using all the voxels. This study utilized the average measure from the most significant regions acquired from ANOVA as features for classification in addition to the multi-modal and multi-measure features of structural and functional MRI data. We extracted most discriminative features by hierarchical sparse feature elimination and selection algorithm. These features include cortical thickness, image intensity, volume, cortical thickness standard deviation, surface area, and ANOVA based features respectively. An extreme learning machine performed both the binary and multi-class classifications in comparison with support vector machines. This article reports prediction accuracy of both unimodal and multi-modal features from test data. We achieved 76.190% (p < 0.0001) classification accuracy in multi-class settings as well as 92.857% (p < 0.0001) classification accuracy in binary settings. In addition, we found ANOVA-based significant regions of the brain that also play a vital role in the classification of ADHD. Thus, from a clinical perspective, this multi-modal group analysis approach with multi-measure features may improve the accuracy of the ADHD differential diagnosis.

A systematic review of structural MRI biomarkers in autism spectrum disorder: A machine learning perspective

“…Of the studies that perform classification, SVMs were the most prevalent classification method (Almeida et al, 2017; Chaddad et al, 2017; Demirhan, 2018; Ecker et al, 2010b; Hoeft et al, 2011; Katuwal et al, 2015; Qureshi et al, 2017; Sabuncu et al, 2015; Subbaraju et al, 2015; Vidhusha and Anandhan, 2015; Vigneshwaran et al, 2013; Xiao et al, 2017), which achieved an accuracy ranging from 53 to 97%, on varying ASD datasets. Of all classification approaches, SVM obtained the highest accuracy in ASD classification (accuracy = 97.8%), albeit on a small dataset (n = 30) (Vidhusha and Anandhan, 2015).…”

“…Many of the studies outlined in Supplementary Table 8 employ no validation approach, while Leave One Out Cross Validation (LOOCV) was used frequently (Chaddad et al, 2017), however as this approach tests on only one data point it is an unreliable estimate of model generalisability. Although splitting datasets into exclusive model training and testing partitions is the most rigorous test of a models generalisability, as was done in a few of the reviewed studies all with n > 100 datasets (Almeida et al, 2017; Ghiassian et al, 2016; Qureshi et al, 2017; Vigneshwaran et al, 2013), limited data in neuroimaging studies necessitates a more efficient usage of data for the purpose of model training and testing. Therefore N‐fold cross validation, which splits the data equally into N groups, trains the model on (N‐1) of the groups and tests on the remaining one group ‘N' times, seems the optimal approach in this setting.…”

Predictive Modelling of The Dynamic Patterns of Thinking in Attention-Deficit/Hyperactivity Disorder: Diagnostic Accuracy of Spatiotemporal Fractal Measures

“…This review included other existing methods proposed to diagnose ADHD, that were not covered in the former systematic review. To make the indirect comparison possible, this in-house scoping review included only diagnostic accuracy studies that reported the chosen metrics of diagnostic for methods based on; electroencephalography and eventrelated potentials (Marcano et al, 2017;Loo et al, 2016;Snyder et al, 2015;Mohammadi et al, 2016;Gloss et al, 2016;Biederman et al, 2017;Gamma and Kara, 2016;Marcano et al, 2018;Manouilenko et al, 2017), structural and functional neuroimaging (Iannaccone et al, 2015;Rangarajan et al, 2014;de Celis Alonso et al, 2017;Qureshi et al, 2017;Serrallach et al, 2016;Hasaneen et al, 2017;Tan et al, 2017b;Uddin et al, 2017), simulated virtual reality and computer games (Negut et al, 2017(Negut et al, , 2016Berger et al, 2017;Faraone et al, 2016), and peripheral biochemical markers (Faraone et al, 2014;Scassellati and Bonvicini, 2015;Scassellati et al, 2012;Thome et al, 2012). According with this complementary "scoping review", the method presented in this research, seems to outperforms all the diagnostic accuracy metrics reported in the trials scrutinised in the aforementioned review.…”

Multimodal Neuroimaging-based Prediction of Adult Outcomes in Childhood-onset ADHD using Ensemble Learning Techniques

“…SVM has also been applied to structural MRI and DTI data collected from adults with ADHD and controls, which reported between-group differences in widespread GM and WM regions in cortices, thalamus, and cerebellum (Chaim-Avancini et al, 2017). Meanwhile, neural network-based techniques, including deep belief network, fully connected cascade artificial neural network, convolutional neural network, extreme learning machine, and hierarchical extreme learning machine, have also been utilized to structural MRI and resting-state functional MRI (fMRI) data in children with ADHD and controls (Deshpande et al, 2015; Kuang and He, 2014; Peng et al, 2013; Qureshi et al, 2016; Qureshi et al, 2017; Zou et al, 2017). The most important group discrimination predictors identified by these neural network studies included functional connectivities within cerebellum, functional connectivity, surface area, cortical thickness and/or folding indices of frontal lobe, temporal lobe, occipital lobe and insula (Deshpande et al, 2015; Peng et al, 2013; Qureshi et al, 2017).…”

“…Meanwhile, neural network-based techniques, including deep belief network, fully connected cascade artificial neural network, convolutional neural network, extreme learning machine, and hierarchical extreme learning machine, have also been utilized to structural MRI and resting-state functional MRI (fMRI) data in children with ADHD and controls (Deshpande et al, 2015; Kuang and He, 2014; Peng et al, 2013; Qureshi et al, 2016; Qureshi et al, 2017; Zou et al, 2017). The most important group discrimination predictors identified by these neural network studies included functional connectivities within cerebellum, functional connectivity, surface area, cortical thickness and/or folding indices of frontal lobe, temporal lobe, occipital lobe and insula (Deshpande et al, 2015; Peng et al, 2013; Qureshi et al, 2017). In addition, principle component-based Fisher discriminative analysis (PC-FDA) (Zhu et al, 2008), Gaussian process classifiers (GPC) (Hart et al, 2014; Lim et al, 2013), and multiple kernel learning (Dai et al, 2012; Ghiassian et al, 2016) have also been used in functional and structural MRI data to discriminate children with ADHD from controls.…”