Dysregulation of behavior, cognition, and affect are features of most common psychopathology in youth. By overlaying well-developed transdiagnostic behavioral dimensions onto clinical diagnostics, we aim to both improve efficiency of diagnostic discovery and clinical prediction. From a syndromic perspective, we target children with attention-deficit/hyperactivity disorder (ADHD), disruptive behavior problems (oppositional, conduct), anxiety and mood disorders, and substance use. We refer to these various disruptions of attention, emotion, and impulse (behavior) dysregulatory psychopathology for purposes of this application.

The need for effective, efficient, reliable, accessible, and person-centric diagnostics and predictors of dysregulatory psychopathology and associated outcome is urgent. Using advanced computational models, we will integrate analyses of the electronic medical record (EMR) with laboratory-derived novel transdiagnostic phenotype measures and environmental adversity phenotypes to address two challenges: First, can we rapidly and accurately improve clinical diagnostics by combining these new phenotypes in computational models? Second, can we predict which children with dysregulatory psychopathology will face clinical deterioration? Each aim includes tests of generalizability and reproducibility and consideration of model bias. This study should dramatically improve clinical diagnostic and prognostic capacity, in a way translatable to real-world clinics.

Specific Aim 1. Refine novel diagnostic and prognostic phenotypes in the lab. Using existing prospective research cohorts (Oregon-ADHD-1000, ABCD), we will computationally refine the prediction of diagnoses and clinical outcomes for dysregulatory psychopathology by augmenting standard clinical measures with novel behavioral phenotype measures of: (a) emotional dysregulation along both positive and negative valence dimensions (via trait profiles and indicators), (b) computational cognitive phenotypes, and (c) environmental adversity measures. Hypothesis 1a: Both (a) new behavioral measures and (b) environmental adversity indices will improve the accuracy of identifying high confidence diagnoses relative to standard clinical measures alone when measured against a gold standard diagnosis (KSADS + multiple informant ratings). Hypothesis 1b: These same measures will improve the prediction of impairment, worsening clinical symptoms, suicidality and other key outcomes over a 1-year period. Refined phenotypes will be extended in Aims 3-4 to prospective data in clinics.

Specific Aim 2. Develop advanced computational models using complex EMR data to refine diagnostic and prognostic utility. Using the TriNetX Research Network, which includes 2.6 million youth with a prior mental health diagnosis and 10.4 million without a prior diagnosis (www.TriNetX.com) (see methods) we will create computational models for case identification and prediction relying only on EMR data. We will identify predictors associated with current diagnosis (by diagnostic cluster: ADHD, disruptive behavior disorder, anxiety disorder, or mood disorder) and suicidality. Hypothesis 2a: We will create cross sectional and longitudinal prediction models with sufficient levels of sensitivity and positive predictive power to test in new clinics in Aims 3-4. Hypothesis 2b: Transfer learning will bring effective benefit to prediction within archival local clinic EMR records.

Specific Aim 3. Test diagnostics in the clinic. Combine Aim 1 and Aim 2 findings with new prospective data collected in psychiatric and pediatric ambulatory clinic populations (N>7500) not previously examined, and use transfer learning to test models with EMR data and novel phenotypes for confident, accurate, and low-cost diagnostic identification. Hypothesis 3a: Combining novel phenotypes and environmental adversity measures with EMR data and models from Aims 1 and 2 will improve predictive diagnostic accuracy beyond what is possible from either data source alone. Hypothesis 3b: Against a gold standard (clinician KSADS + informant ratings pooled), computational algorithms will distinguish (without the KSADS) between (a) patients who are readily diagnosable or classified as high risk/intervene with minimal data and (b) patients where diagnoses are uncertain and require further evaluation.

Specific Aim 4: Test 1-year follow-up prediction using prospective clinic data. We will conduct a 1-year follow up of patients studied in Aim 3 and address the same issues using the same logic for outcome one year later. Hypothesis 4a: The pooled algorithm from Aims 1 and 2 (modified from what is learned in Aim 3) will improve prediction of outcomes beyond what is possible from either data source alone. We will predict worsening of clinical symptoms, impairment, suicidality, ER visits, substance use, and service utilization.

Overall, if successful, the proposed work should sharply enhance the readiness of novel behavioral phenotypes for clinic use, and markedly improve prediction of dysregulatory diagnoses (which remain central to the workflow and treatment decisions of real-world clinics) and transdiagnostic outcomes that drive real world impairment, irrespective of diagnostic boundaries. We do so by moving from laboratory discovery and refinement in Aims 1 and 2 toward clinical testing and generalizability in Aims 3 and 4, bringing both novel behavioral phenotypes and advanced analysis of EMR records to a position of being deployable in the clinic setting.