Prediction of methylphenidate treatment response for ADHD using conventional and radiomics T1 and DTI features: Secondary analysis of a randomized clinical trial

•We used conventional and radiomics MRI models to predict MPH treatment response.•Both conventional and radiomics models outperformed chance in predicting response during treatment.•Prediction accuracy dropped significantly post-treatment after one-week washout.•Baseline basal ganglia radiomics were key predictive features for MPH response. Attention-Deficit/Hyperactivity Disorder (ADHD) is commonly treated with methylphenidate (MPH). Although highly effective, MPH treatment still has a relatively high non-response rate of around 30%, highlighting the need for a better understanding of treatment response. Radiomics of T1-weighted images and Diffusion Tensor Imaging (DTI) combined with machine learning approaches could offer a novel method for assessing MPH treatment response. To evaluate the accuracy of both conventional and radiomics approaches in predicting treatment response based on baseline T1 and DTI data in stimulant-naive ADHD participants. We performed a secondary analysis of a randomized clinical trial (ePOD-MPH), involving 47 stimulant-naive ADHD participants (23 boys aged 11.4 ± 0.4 years, 24 men aged 28.1 ± 4.3 years) who underwent 16 weeks of treatment with MPH. Baseline T1-weighted and DTI MRI scans were acquired. Treatment response was assessed at 8 weeks (during treatment) and one week after cessation of 16-week treatment (post-treatment) using the Clinical Global Impressions − Improvement scale as our primary outcome. We compared prediction accuracy using a conventional model and a radiomics model. The conventional approach included the volume of bilateral caudate, putamen, pallidum, accumbens, and hippocampus, and for DTI the mean fractional anisotropy (FA) of the entire brain white matter, bilateral Anterior Thalamic Radiation (ATR), and the splenium of the corpus callosum, totaling 14 regional features. For the radiomics approach, 380 features (shape-based and first-order statistics) were extracted from these 14 regions. XGBoost models with nested cross-validation were used and constructed for the total cohort (n = 47), as well as children (n = 23) and adults (n = 24) separately. Exact binomial tests were employed to compare model performance. For the conventional model, balanced accuracy (bAcc) in predicting treatment response during treatment was 63 % for the total cohort, 32 % for children, and 36 % for adults (Area Under the Receiver Operating Characteristic Curve (AUC-ROC): 0.69, 0.33, 0.41 respectively). Radiomics models demonstr