Prediction of sleep side effects following methylphenidate treatment in ADHD youth

Sleep problems is the most common side effect of methylphenidate (MPH) treatment in ADHD youth and carry potential to negatively impact long-term self-regulatory functioning. This study aimed to examine whether applying machine learning approaches to pre-treatment demographic, clinical questionnaire, environmental, neuropsychological, genetic, and neuroimaging features can predict sleep side effects following MPH administration.

The present study included 83 ADHD subjects as a training dataset. The participants were enrolled in an 8-week, open-label trial of MPH. The Barkley Stimulant Side Effects Rating Scale was used to determine the presence/absence of sleep problems at the 2nd week of treatment. Prediction of sleep side effects were performed with step-wise addition of variables measured at baseline: demographics (age, gender, IQ, height/weight) and clinical variables (ADHD Rating Scale-IV (ADHD-RS) and Disruptive Behavior Disorder rating scale) at stage 1, neuropsychological test (continuous performance test (CPT), Stroop color word test) and genetic/environmental variables (dopamine and norepinephrine receptor gene (DAT1, DRD4, ADRA2A, and SLC6A2) polymorphisms, blood lead, and urine cotinine level) at stage 2, and structural connectivities of frontostriatal circuits at stage 3. Three different machine learning algorithms ((Logistic Ridge Regression (LR), support vector machine (SVM), J48) were used for data analysis. Robustness of classifier model was validated in the independent dataset of 36 ADHD subjects.

Classification accuracy of LR was 95.5% (area under receiver operating characteristic curve (AUC) 0.99), followed by SVM (91.0%, AUC 0.85) and J48 (90.0%, AUC 0.87) at stage 3 for predicting sleep problems. The inattention symptoms of ADHD-RS, CPT response time variability, the DAT1, ADRA2A DraI, and SLC6A2 A-3081T polymorphisms, and the structural connectivities between frontal and striatal brain regions were identified as the most differentiating subset of features. Validation analysis achieved accuracy of 86.1% (AUC 0.92) at stage 3 with J48.

Our results provide preliminary support to the combination of multimodal classifier, in particular, neuroimaging features, as an informative method that can assist in predicting MPH side effects in ADHD.