This study, in response to RFA-MH-25-195, aims to enhance, deploy, and rigorously validate the Duke Predictive Model of Adolescent Mental Health (Duke-PMA), which uses a novel clinical signature derived from affordable, accessible measures to identify youth at high risk for psychiatric illness in primary care settings. The Duke-PMA, a neural network-based predictive tool, has already demonstrated high accuracy in predicting psychiatric risk one year in advance in youth aged 10-15, using data from the Adolescent Brain and Cognitive Development (ABCD) study. Notably, sleep disturbances have emerged as a key modifiable predictor in the model. Unlike most predictive models that rely on current symptoms to anticipate outcomes, the Duke-PMA bases its predictions on underlying disease mechanisms and protective factors, making it better suited to inform preventive interventions. Furthermore, the model identifies an elevated p-factor, a general measure of psychopathology that spans multiple psychiatric conditions, making it broadly applicable across diverse youth populations. Our project will begin by optimizing the Duke-PMA through the incorporation of behavioral tasks from the NIH Toolbox to enhance its prediction performance. Following Duke AI Health’s Algorithm-Based Clinical Decision Support Oversight framework, we will ensure the model adheres to the highest standards of transparency, quality, and equity. Additionally, we will apply trustworthy AI techniques designed to reduce effects of distribution shifts on model performance to ensure the model remains effective and equitable across diverse clinical settings and demographic groups. After optimization, the Duke-PMA will be deployed in rural primary care and pediatric clinics, where access to mental health services and research participation is often limited. We will enroll 2,000 youth from rural clinics in the Southeast and Midwest, partnering with the Science, Technology, and Research (STAR) Clinical Research Network. We will also explore the benefit of adding a measure of home environments to the Duke-PMA through digital envirotyping, which uses an AI-driven approach to assess home environments remotely without requiring in-person visits, making it much more resource-efficient and accessible than current approaches. Model performance will be validated through psychiatric diagnostics conducted one year after the initial assessment. If successful, this project has the potential to transform mental health resource allocation particularly in underserved communities by offering an accessible, low-cost, data-driven approach to identify vulnerable youth and highlight modifiable risk factors for early intervention.

Aim 1: To prepare the Duke Predictive Model of Adolescent Mental Health (Duke-PMA) for clinical use by optimizing portability, actionability and implementation feasibility, and performance, leveraging Duke AI Health’s Algorithm-Based Clinical Decision Support Oversight to ensure the highest standards of transparency, quality, and equity. H1: Portability. Leveraging the ample site and demographic variation of the ABCD study, adaptations of Duke-PMA will prioritize predictive features and methods that are robust to variations in population demographics and data collection practices, thus ensuring health equity and high performance across clinical settings. H2: Actionability and implementation feasibility will be optimized by selecting model features that maximize performance yet are actionable and minimize implementation barriers (e.g., avoids measures poorly suited for rural settings as determined by our Community Advisory Board). H3: Performance. Continuing our work with the ABCD study, integrating NIH Toolbox (behavioral) measures will augment Duke-PMA’s predictive performance.

Aim 2: To deploy and prospectively validate Duke-PMA (refined in Aim 1) in rural, primary care clinical settings. We will develop a tablet-based platform through which youth and their parents can complete questionnaires and behavioral tasks to inform our predictive model. Youth (n=2,000 youth; ages 10-15) and a parent will be recruited from rural pediatric primary care clinics to complete our assessment and will be contacted again 1 year later to assess psychiatric outcomes. H4: Model performance for predicting which youth move from low to high risk one year in the future will not decrease upon deployment. H5: Performance will remain consistent across settings, and usability will be high as determined by adherence rates and provider and participant surveys.

Exploratory Aim 3: To explore the use of digital envirotyping to enhance Duke-PMA’s predictive capacity while maintaining resource-efficiency. Integrating AI analysis of photographs of participants’ home environments (modeled after the HOME scale1) will enhance the predictive accuracy of Duke-PMA for psychiatric outcomes, while maintaining accessibility, low cost, and ease of implementation across diverse settings.