Model Teens !!BETTER!!
Little is known about processes through which behavior therapy (BT) for adolescent ADHD improves outcomes. The purpose of this study was to build a theoretical model for the processes through which a BT for adolescent ADHD (Supporting Teens' Autonomy Daily; STAND) impacts functioning. Seventy-eight audio recordings from a standard therapeutic task in the final STAND session were analyzed as parents and adolescents (ages 11-16) reflected upon what changed during STAND and why. Qualitative coding sorted parent and teen statements into orthogonal categories of perceived changes. Network analysis examined inter-relations between categories. Results indicated twenty-one categories of perceived change areas. Parent use of behavioral strategies, adolescent motivation, and adolescent organization skills were central nodes in the network of perceived changes, with strong relations to academic and parent-teen relationship outcomes. A model is proposed in which skills training in STAND increases parent behavioral strategy use and teen organization skills, while Motivational Interviewing (MI) in STAND increase parent behavioral strategy use and initial adolescent motivation. In turn, parent behavioral strategy use is proposed to further reinforce teen motivation through contingency management, thereby increasing teen application of organization skills to daily life. As a result of improved teen motivation and organization skills, the model proposes that ADHD symptoms, academic problems, and parent-teen conflict abate. We discuss secondary mechanisms and outcomes in this model, the possibility of person-specific processes, implications for community-based adaptation of STAND, and plans to validate this conceptual model using sophisticated mediational models.
Aggressive and suicidal behaviours cause significant disruptions on a personal and societal level [1,2,3]. Early identification of those at risk for these behaviours in the general population would likely improve prognosis by assisting with risk stratification as well as targeting clinical interventions, more efficiently than with clinical interviews for each child. However, many risk prediction models focus on clinical samples, i.e. those who have already sought treatment. By expanding this approach to the general population, prevention effectors could be improved. To this end, it is of interest to create a model that can determine who is at risk to exhibit suicidal behaviour, aggressive behaviour or both, versus who is not.
Many studies have looked into predicting suicide or aggressive behaviour. Although many models exist for prediction of suicidal behaviours in the clinical population, there are relatively fewer studies which use population-based samples [25, 26]. Whereas research in forensic psychology has worked to predict recidivism and violent criminal behaviour in general [27,28,29], an approach to prevent aggressive behaviours in adolescence, a developmentally sensitive period, is less common. Moreover, only a handful of studies using a clinical population have examined suicidal and aggressive behaviours together as an outcome [30, 31]. While aggressive behaviour tends to be childhood- and adolescent-limited, there is a subset of individuals for whom aggressive behaviour persists into adulthood. This trajectory is associated with poorer outcomes in adulthood . Thus identifying those who remain aggressive at late adolescence or older are of clinical importance . Additionally, given the significant overlap between the risk factors and co-occurrence of suicidal behaviours and aggression, creating a combined model that could be used in practice would reduce the need for separate questionnaires or assessments.
Our aim is to create a multi-class model that can predict who will report suicidal behaviours, aggressive behaviours, both, or neither in young adulthood within large scale, epidemiological samples. Using a combination of genetic, environmental, and psychosocial factors obtained from epidemiological cohorts would theoretically allow for a highly generalisable, comprehensive model that could improve understanding of risk factors for these behaviours, as well as further inform future models for clinical prediction and decision-making.
PGS were derived using LDpred . LDpred accounts for the linkage disequilibrium between single nucleotide polymorphisms (SNPs) to avoid inflation of effect sizes. For NTR the LD structure was determined on a subset of unrelated individuals and using a set of well imputed variants, while in CATSS, data from 1000 genomes phase 3 version 5 was used as an external reference sample [46, 47]. The weighted effect sizes were used as a basis for the polygenic scores. LDpred requires the specification of prior probabilities corresponding to the fraction of SNPs from the discovery samples considered causal with the trait, and we created scores at a range of priors (0.01, 0.05, 0.1, 0.2 0.3, 0.5, 1). In order to reduce the complexity inherent in having multiple PGS predictors and outcome variables, we performed principal component analysis (PCA) on all priors for each trait PGS, and included the first principal component (PCA-PGS) for each trait in our model according to Coombes and colleagues . PCA analysis is an unsupervised machine learning technique which reduces the dimensionality of datasets while maintaining as much variability as possible; the resulting principal components (PCs) represent a certain amount of variation within the dataset. The first PC can be interpreted to represent the most variation within the data . This method has been shown to prevent overfitting each PGS to each outcome and removes the need to select a single prior across all PGS.
We created a stacked ensemble model, i.e. a model that combines input predictions from separate models, which included a gradient boosted machine, random forest, elastic net, and a neural network . These models were selected based on their availability in H2O. This package was chosen based on its compatibility with multiclass ensemble models. The stacked ensemble model did not contain any parameters other than the number of folds for cross validation.
Variable importance scores can be created from tree-based models and thus were obtained from the gradient boosted machines and random forest models that were included within the ensemble model. The overall variable importance rankings were determined using the average of scaled importance scores across the random forest and gradient boosted machine model. The variable importance in models built with H2O can be interpreted as the improvement in the squared error when the variable is split on a node, i.e. the decision points of the tree .
In order to determine the extent to which all of the PGS variables contributed to the model we completed the analysis with all genetic variables removed. The same tuning procedure in the main analysis was used to create the final ensemble model. We then performed a Venkatraman test using the R package pROC to determine whether the performance difference between both models (with and without genetic variables) was significant . Additionally, in order to assess the discrepancy between the proportion of suicidal behaviours in the CATSS and NTR we performed a logistic regression using suicidal behaviours as an outcome with cohort and measurement year as predictors.
We created a model to identify adolescents at a high risk of suicidal and/or aggressive behaviour using a wide range of predictors: questionnaires relating to home environment, behaviours, psychiatric symptoms, and genetic data for various traits. By training the model in the CATSS sample and validating it in the NTR sample, we tested the cross-cultural/external prediction of the model. Moreover, we examined the extent to which PGS variables contributed to the model through variable importance scores and by creating a model without genetic variables.
Out of the 25 highest variable importance scores 18 were PGS variables, with four PGS variables among the top ten. The highest performing PGS variables were for psychiatric related variables, but other traits such as birth weight, childhood body mass index, IQ, and educational attainment were also ranked highly. Additionally, removing all PGS variables from the model showed attenuated performance compared to the main model, however the difference in model performance was not statistically significant for most classes. Thus, while PGS of traits generally do not have clinical utility within psychiatry on their own, the results suggest that they can contribute to model performance when combined with other variables/risk factors . Notably, AUC of the NTR outperformed the test set in the sensitivity analysis. This is likely a result of the model over relying on the variable with the highest importance, aggression at age 15/16, as the NTR had a higher proportion of individuals in the aggressive class and the one-vs-all aggression AUC was the top performing class in the NTR.
The primary strength of this study was the use of longitudinal data from both questionnaires and genetic data. Moreover, we were able to validate our model through an externally collected data source , indicating that our results are generalisable to Northern Europe and did not suffer from overfitting. However, our study comes with caveats. First, our model would likely be improved by a larger sample size and additional variables related to psychiatric symptoms, such as impulsivity and emotion-dysregulation. Second, our measure of suicidal behaviours were somewhat inconsistent across both cohorts, which likely affected the performance between the data sets.
This study adds to the growing literature around genetically informed prediction models for mental health in adolescents. The results from the variable importance scores suggest that self-reported psychiatric symptoms in mid-adolescence, sex, and psychiatric PGS are key indicators for predicting later aggression and self-harm. Through the sensitivity analysis we found that removing genetic variables led to attenuated but not statistically significant differences in model performance for most of the classes. As of now prediction models are not ready for clinical use in psychiatric clinics and our model is no different . Moreover, the current cost and processing time of genotyping means that clinical utility of genetically informed models may be limited, especially given our models comparable performance to models using non-genetic biomarkers . 041b061a72