The proportion of pediatric emergency department or inpatient visits due to suicide-related behaviors (SRB) has doubled in recent years. This public health emergency appears to be affecting youth of underrepresented racial, ethnic, and linguistic backgrounds (i.e., REL minority youth) in particular. Trends in suicide attempts over the past decade have remained higher or increased among Black and Latinx youth compared to white peers. Machine learning (ML) and natural language processing (NLP) with electronic health records (EHR) have advanced suicide risk identification, with potential to fill the need for sustainable and scalable practice- based resources with near-term impact. Despite the near-term promise of ML and NLP with EHR data for advancing suicide risk prediction, several important issues need to be addressed before ML-derived tools can be broadly and optimally implemented in practice-based settings. Of particular concern is that such ML algorithms may worsen health disparities in that they may lead to improvements in care for majority populations without corresponding advancements for REL minority populations insofar as minority populations are underrepresented in the development of these ML algorithms. Within this context, algorithms trained on EHR data linked to geocoded data of social determinants of health (SDOH) and computerized adaptive testing (CAT) has potential to improve suicide risk prediction. The portability (and thus scalability) of ML algorithms across settings needs also to be demonstrated. Our current objective is therefore to develop EHR-derived ML suicide risk algorithms that (i) minimize bias against underserved groups; (ii) are portable across settings; and (iii) incorporate low-burden indices of general and population-relevant suicide risk. We will develop ML algorithms with EHR data for prospective prediction of suicide risk for REL minority youth within 3 months of inpatient discharge across 2 psychiatric inpatient services; determine the portability of these algorithms across the 2 sites; and evaluate the added predictive value of SDOH and CAT data collected as part of standard care. Our first aim is to develop prediction algorithms using EHR data to classify risk for STBs within 3 months of discharge. We will evaluate the relative performance of co-trained versus locally trained ML algorithms from two busy pediatric psychiatric inpatient sites to assess their portability. Our second aim is to evaluate the incremental value of geocode-derived SDOH and CAT data in classifying suicide risk in REL minority youth. As an exploratory aim, we will evaluate the performance of the algorithms developed in Aims 1 and 2 for specific REL minority subsamples. The current project directly addresses NIMH/National Action Alliance for Suicide Prevention’s Aspirational Goal to improve evaluation of suicide risk among diverse populations and in diverse settings through feasible and scalable assessment strategies. It does so by focu...