Attention-deficit/hyperactivity disorder (ADHD) is a neurodevelopmental disorder that could persist into adulthood with known abnormalities in brain structure. Genetics also play an important role in the etiology of the disorder and could affect the disorder trajectory. In this study, we investigated the prediction power of brain image and genomic features for symptom change in 77 individuals with ADHD as part of NeuroIMAGE cohort. Gray matter components and working memory assessments at baseline, as well as gene scores of interest, were used to predict the changes in the two symptom domains: inattentive and hyperactive/impulsive, an average of 4 years. A linear regression model coupled with various feature selection approaches, including leave-one-out-cross-validation (LOOCV), stability selection with resampling, and permutation tests, was implemented to mitigate the overtraining potential caused by small sample sizes. Results showed that traditional LOOCV overestimated the prediction power. We proposed a novel stability selection with the threshold set by permutation tests, which provided more objective assessment. Using our proposed procedure, we identified a statistical promising prediction model for inattention symptom change; the consistent correlation between predicted values and measured values during model training, validating and hold out testing (r=0.64, 0.53, 0.46, respectively), but the p value is not significant in the holdout test. The selected features include age, gray matter in the insula, genes OSBPL1A, CTNNB1, PRPSAP2, ACADM, and polygenic risk score of education attainment, which have been previously reported to be associated with ADHD. We speculate that significant associations may be observed with a large sample size.