Pattern classification of response inhibition in ADHD: toward the development of neurobiological markers for ADHD

Abstract

The diagnosis of Attention Deficit Hyperactivity Disorder (ADHD) is based on subjective measures despite evidence for multisystemic structural and functional deficits. ADHD patients have consistent neurofunctional deficits in motor response inhibition. The aim of this study was to apply pattern classification to task-based functional magnetic resonance imaging (fMRI) of inhibition, to accurately predict the diagnostic status of ADHD. Thirty adolescent ADHD and thirty age-matched healthy boys underwent fMRI while performing a Stop task. fMRI data were analyzed with Gaussian process classifiers (GPC), a machine learning approach, to predict individual ADHD diagnosis based on task-based activation patterns. Traditional univariate case-control analyses were also performed to replicate previous findings in a relatively large dataset. The pattern of brain activation correctly classified up to 90% of patients and 63% of controls, achieving an overall classification accuracy of 77%. The regions of the discriminative network most predictive of controls included later developing lateral prefrontal, striatal, and temporo-parietal areas that mediate inhibition, while regions most predictive of ADHD were in earlier developing ventromedial fronto-limbic regions, which furthermore correlated with symptom severity. Univariate analysis showed reduced activation in ADHD in bilateral ventrolateral prefrontal, striatal, and temporo-parietal regions that overlapped with areas predictive of controls, suggesting the latter are dysfunctional areas in ADHD. We show that significant individual classification of ADHD patients of 77% can be achieved using whole brain pattern analysis of task-based fMRI inhibition data, suggesting that multivariate pattern recognition analyses of inhibition networks can provide objective diagnostic neuroimaging biomarkers of ADHD.

Keywords: ADHD; Gaussian process classifier; biomarker; diagnosis; functional magnetic resonance imaging; inhibition.

Figure 1

A: Receiver operating characteristic (ROC) curve. Each point on the ROC curve describes the performance of the classifier at a particular decision threshold. Points on the lower left‐hand side of the ROC graph are “conservative,” requiring strong evidence for a positive classification and therefore have low true and false positive rates (TPR/FPR). Points on the upper right‐hand side of the graph are “liberal,” making positive classifications with weak evidence and therefore have a high TPR but also a high FPR. A classifier is more accurate than another if it is closer to the point (1,0), which reflects perfect classification. Chance level is indicated by the dotted diagonal line. B: Classification accuracies for GPC predictions for classifying individuals into either the ADHD or the control group. The x‐axis describes the probability with which each subject is predicted to be an ADHD patient (equal to 1‐ the probability of being a control). The dotted line indicates the decision threshold.

Figure 2

A: Unthresholded GPC weights overlaid on an anatomical template. The color code shows the relative weight of each voxel for the decision boundary (red/yellow scales: higher weights for ADHD boys and blue scales: higher weights for healthy control boys). B: The univariate ANOVA group comparison map at P < 0.05 for voxel and P < 0.01 for clusters, showing brain areas that are decreased in activation in ADHD relative to healthy control boys in blue and brain areas that are increased in activation in ADHD relative to healthy control boys in red/yellow. [Color figure can be viewed in the online issue, which is available at http://wileyonlinelibrary.com.]