Table 8.
Final predictors of cases of ODD at ages 11–12 years.
| (A) Age of case determination | Ranked final predictors | Importance |
|---|---|---|
| New onset at age 11–12 years | Youth has ever received MH/SU services | 0.1581 |
| Disorder of excessive somnolence | 0.1290 | |
| Disorder of sleep-wake transition | 0.0968 | |
| Either parent has nerves or nervous breakdown problem | 0.0871 | |
| Prosocial behavior scale score | 0.0516 | |
| Parent externalizing problems | 0.0290 | |
| Parent aggressive behavior problems | 0.0032 | |
| Mean | 0.0694 | |
| All cases at age 11–12 years | Child has received MH/SU services in last 6 months | 0.1856 |
| Parent reports getting along very well with child | 0.0951 | |
| Total sleep disturbance score | 0.0827 | |
| Mother has depression problem | 0.0819 | |
| Either parent has been to doctor or counselor due to emotional/mental problem | 0.0420 | |
| Parent externalizing problems | 0.0367 | |
| Prosocial behavior mean score | 0.0358 | |
| Disorder of sleep-wake transition | 0.0347 | |
| Mean | 0.0743 | |
| (B) Neural data type | Ranked final predictors | Importance |
| New onset at age 11–12 years | Correlation between salience network and left caudate ROI | 0.0750 |
| Correlation between retrosplenial temporal network and right caudate ROI | 0.0743 | |
| Correlation between default network and left hippocampal ROI | 0.0509 | |
| Correlation between cinguloopercular network and left amygdala ROI | 0.0262 | |
| Correlation between cinguloopercular network and right putamen ROI | 0.0256 | |
| Correlation between fronto-parietal network and left caudate ROI | 0.0207 | |
| Correlation between sensorimotor hand network and left amygdala ROI | 0.0158 | |
| Mean | 0.0410 | |
| All cases at age 11–12 years | Average FA in GM in right lateral orbitofrontal ROI | 0.0044 |
| Cortical area in mm2 of left inferior parietal ROI | 0.0007 | |
| Mean | 0.0026 |
Final predictors of all prevailing cases of ODD at 11–12 years, as well as new onset cases only at 11–12 years of age, are shown for the most accurate models obtained using deep learning optimized with IEL obtained with (A) multimodal features and (B) only neural features. Final predictors are ranked in the order of importance, where the relative importance of each predictor is computed with the Shapley additive explanation technique and presented here averaged across all participants in the sample. Features in red indicate an inverse relationship with ODD verified with the Shapley method. MH, mental health; SU, substance use; ROI, region of interest; GM, gray matter.