A large-scale genome-wide association study indicated there is substantial continuity of ADHD from childhood to adulthood (Rovira et al, 2020). It is hypothesised that as symptoms and impairments persist into adulthood for some children with ADHD, ADHD-related brain structure differences in adults may be consistent with those observed in children. To explore this hypothesis, machine-learning models were used to analyse magnetic resonance imaging (MRI) data from individuals with and without ADHD.
T1-weighted structural MRI (sMRI) neuroimaging data were provided by the ENIGMA-ADHD Working Group from multiple sites and from individuals with and without ADHD. Participant data were separated by ADHD diagnosis, age, sex and site. A total of 151 variables were used, including 34 cortical surface areas, 34 cortical thickness measurements and seven subcortical regions from each hemisphere and intracranial volume. Individuals with >50% of missing variables were removed from the analysis. Confounding factors were balanced by randomly assigning samples to training (~70%), validation (~15%) and test (15%) subsets within each diagnosis, age subgroup, sex and site to ensure that the train/validation/test subsets had the same composition of these variables. Data were presented as area under the receiver operating curve (AUC), whereby 95% confidence interval (CI) AUC values that did not overlap with 0.5 indicated significant predictive accuracy of the machine learning models.
The final machine learning dataset consisted of 4042 individuals from 35 sites; 60.7% were children (aged <18 years; n=2454) and 39.3% were adults (aged >18 years; n=1588). In total, 54.2% (n=2192; male:female ratio=2.79) of the individuals had a diagnosis of ADHD and 45.8% (n=1850; male:female ratio=1.42) did not have an ADHD diagnosis. The model that was trained and validated on only MRI of child ADHD was able to significantly separate children with ADHD from children without ADHD: AUC=0.64 (95% CI 0.58–0.69). Whereas, the model trained and validated on data of adult ADHD could not significantly separate adults with ADHD from adults without ADHD: AUC=0.56 (95% CI 0.49–0.62; p=0.057). The model-predicted brain risk score (the probability of an individual being classified as having ADHD) calculated Cohen’s d effect sizes in the test set to be 0.47 (95% CI 0.27–0.68) for child samples and 0.15 (95% CI -0.08–0.39) for adult samples.
When age and sex were added as predictors, the adult model was able to significantly predict ADHD in adults: AUC=0.62 (95% CI 0.56–0.69; p=0.002). The addition of age and sex as predictors to the child model did not affect the ability of the model to predict ADHD. The Cohen’s d effect sizes in the test set were 0.48 (95% CI 0.27–0.69) for child samples and 0.39 (95% CI 0.15–0.63) for adult samples.
The model trained on adult data that used both MRI features and age and sex could predict ADHD in the child data, AUC=0.60 (95% CI 0.58–0.62); Cohen’s d effect size was 0.17 (95% CI 0.1–0.24). On the other hand, the model trained on child data that utilised both MRI features and age and sex predictors did not predict ADHD in the adult data, AUC=0.53 (95% CI 0.49–0.56).
There were several limitations to these analyses. First, as the data from multiple studies were combined, the limitations of the original studies were inherited. Heterogeneity within the original studies could have added to the ‘noise’ of the dataset, which could have made it difficult to discriminate between the data of individuals with and without ADHD. Second, only structural imaging data were used; if other modalities had been included, clearer results may have been provided. Third, pre-defined structures from ENIGMA standard image processing were used; if other methods had been used, this could have increased the classification accuracy. Finally, the use of neural networks could make it difficult to clarify the importance of each brain region in the model’s algorithm.
The authors concluded that the neural network approach could detect case-control sMRI differences in adults with ADHD that could not be detected with standard analyses. Additionally, this study provided evidence for the continuity of sMRI findings from childhood ADHD to adulthood.
Read more about ADHD and machine learning analysis here
Rovira P, Demontis D, Sánchez-Mora, et a. Shared genetic background between children and adults with attention deficit/hyperactivity disorder. Neuropsychopharmacology 2020; 45: 1617-1626
Zhang-James Y, Helminen EC, Liu J, et al. Evidence for similar structural brain anomalies in youth and adult attention-deficit/hyperactivity disorder: a machine learning analysis. Transl Psychiatry 2021; 11: 82.