Anxiety disorders in youth are highly prevalent [1] and impairing [2-4]. Left untreated, these disorders confer substantial additional risk for the development of a wide range of negative sequelae, including substance use [5], suicidal ideation and attempts [6], and additional mental health comorbidities [7]. Although several treatments have demonstrated efficacy for anxiety in youth, including individual cognitive behavioral therapy (ICBT), family CBT (FCBT), medication (MED), and combination of CBT and medication (COMB) [8], a meaningful portion of youth are classified as non-responders after a full course of treatment [9]. The identification of baseline predictors and moderators of response is critical to improve treatment efficacy and reduce burden on families. Increased anxiety severity, comorbidity (behavioral problems, depression), and family psychopathology, along with older age, female gender and anxiety diagnosis, have been highlighted as potential predictors and moderators of outcome. However, studies have been underpowered and findings are inconsistent [10]. To date, all studies have taken a traditional analytic approach, which typically provides conservative estimates as a result of imposed explanatory constraints [11]. Machine Learning (ML) represents a promising complementary statistical technique to traditional analyses, given its focus on predictive fit rather than explanatory inference [12] and will facilitate identification of non-linear, complex patterns of predictors and moderators at the individual level [13, 14]. These methods have shown promise in identification of treatment outcome predictors in other medical [e.g., 15-18] and psychiatric samples [e.g., 19-22], but to date have not been implemented in a sample of anxious youth. The proposed project will aggregate datasets from at minimum ten peer-reviewed and published randomized controlled trials (N=1444) and train and validate two models along overlapping features, including (1) demographics, (2) diagnosis, (3) anxiety severity (4) behavioral problems, and (5) family psychopathology. Models will also be used to examine differential response to ICBT, FCBT, MED and COMB. Aggregated data will be uploaded into a centralized dataset, in line with the NIMH RDoC db and NDAR [23] datasets, and then used to predict outcome for individual anxious youth (N=80) completing ICBT and COMB at the Child and Adolescent Anxiety Disorders Clinic at Temple.
The aims of this study are consistent with calls issued in the NIMH strategic plan (Objective 3) and will help facilitate the development of person-centered interventions for anxious youth [24]. An individualized approach to treatment is important to further increase treatment efficacy and reduce the financial and emotional burden associated with non-response [25, 26]. A training plan has been designed that consists of mentorship, formal classwork and experiential learning to develop the applicant's expertise in machine learning and dataset aggregation. The proposed study will take place within Temple University's clinical psychology program, which has a successful track record of conducting impactful NIMH-funded research and training research scientists.
Although several efficacious treatments have been identified for youth anxiety, there is significant heterogeneity in treatment response and this heterogeneity is poorly understood. To help clarify differences in treatment response and facilitate the development of person-centered interventions, novel and advanced statistical approaches should be implemented using adequately powered datasets. The proposed study will (1) use machine-learning methods to identify predictors and moderators of outcome in an aggregated dataset of at minimum ten Randomized Controlled Trials (RCTs) examining youth anxiety treatments (individual and family cognitive behavioral therapy, medication, and combination treatments), (2) create a centralized dataset of aggregated RCTs to foster continued cross-site collaboration and (3) test the model's predictive accuracy at the individual level in a real-world sample.
