Machine learning classification of ADHD and HC by multimodal serotonergic data

Serotonin neurotransmission may impact the etiology and pathology of attention-deficit and hyperactivity disorder (ADHD), partly mediated through single nucleotide polymorphisms (SNPs). We propose a multivariate, genetic and positron emission tomography (PET) imaging classification model for ADHD and healthy controls (HC). Sixteen patients with ADHD and 22 HC were scanned by PET to measure serotonin transporter (SERT‘) binding potential with [(11)C]DASB. All subjects were genotyped for thirty SNPs within the HTR1A, HTR1B, HTR2A and TPH2 genes. Cortical and subcortical regions of interest (ROI) were defined and random forest (RF) machine learning was used for feature selection and classification in a five-fold cross-validation model with ten repeats. Variable selection highlighted the ROI posterior cingulate gyrus, cuneus, precuneus, pre-, para- and postcentral gyri as well as the SNPs HTR2A rs1328684 and rs6311 and HTR1B rs130058 as most discriminative between ADHD and HC status. The mean accuracy for the validation sets across repeats was 0.82 (±0.09) with balanced sensitivity and specificity of 0.75 and 0.86, respectively. With a prediction accuracy above 0.8, the findings underlying the proposed model advocate the relevance of the SERT as well as the HTR1B and HTR2A genes in ADHD and hint towards disease-specific effects. Regarding the high rates of comorbidities and difficult differential diagnosis especially for ADHD, a reliable computer-aided diagnostic tool for disorders anchored in the serotonergic system will support clinical decisions.