Multi-Region Risk-Sensitive Cognitive Ensembler for Accurate Detection of Attention-Deficit/Hyperactivity Disorder

Abstract

In this paper, we present a multi-region ensemble classifier approach (MRECA) using a cognitive ensemble of classifiers for accurate identification of attention-deficit/hyperactivity disorder (ADHD) subjects. This approach is developed using the features extracted from the structural MRIs of three different developing brain regions, viz., the amygdala, caudate, and hippocampus. For this study, the structural magnetic resonance imaging (sMRI) data provided by the ADHD-200 consortium has been used to identify the following three classes of ADHD, viz., ADHD-combined, ADHD-inattentive, and the TDC (typically developing control). From the sMRIs of the amygdala, caudate, and hippocampus regions of the brain from the ADHD-200 data, multiple feature sets were obtained using a feature-selecting genetic algorithm (FSGA), in a wraparound approach using an extreme learning machine (ELM) basic classifier. An improved crossover operator for the FSGA has been developed for obtaining higher accuracies compared with other existing crossover operators. From the multiple feature sets and the corresponding ELM classifiers, a classifier-selecting genetic algorithm (CSGA) has been developed to identify the top performing feature sets and their ELM classifiers. These classifiers are then combined using a risk-sensitive hinge loss function to form a risk-sensitive cognitive ensemble classifier resulting in a simultaneous multiclass classification of ADHD with higher accuracies. Performance evaluation of the multi-region ensemble classifier is presented under the following three scenarios, viz., region-based individual (best) classifier, region-based ensemble classifier, and finally a multiple-region-based ensemble classifier. The study results clearly indicate that the proposed “multi-region ensemble classification approach” (MRECA) achieves a much higher classification accuracy of ADHD data (normally a difficult problem because of the variations in the data) compared with other existing methods.