Project Summary Precision identification of alcohol use disorder (AUD) is an urgent but unmet clinical need. Timely identifying AUD remains challenging, with only 1 in 6 binge drinkers is screened for AUD. AUD in individuals with lower socioeconomic status are less likely to be diagnosed and treated timely due to the fear of stigmatization. They also suffer higher risk of death and other health and social consequences. The All of Us (AoU) research dataset holds the promise for computationally identifying undiagnosed AUD cases at clinical settings. The AoU dataset covers longitudinal electronic medical records (EMR), the Alcohol Use Disorders Identification Test- Consumption (AUDIT-C) screening results which is one of the gold standards for AUD diagnosis, the genomics data that explains the heritable factors, and the rich SDoH and lifestyle surveys for the environmental exposures. However, the unique and complex AoU data domains and data patterns challenge current statistical and machine learning models. To address these challenges, we propose to develop the Graph Artificial Intelligence for Precision identification of Alcohol use disorder (GAIPA), an evidence-driven, graph- based clinical informatics approach to transfer the knowledge learned from the AoU data to identify undiagnosed AUD cases in clinical care. We propose to: Aim 1. Establish an AUD classification model using graph artificial intelligence model, the AoU multi-modality data, and the Indiana Network for Patient Care (INPC) EMR data. Specifically, we will: Aim 1.1: Build the generic model through weak supervision using AoU AUDIT-C survey and AUD diagnosis as weak ground truth; and Aim 1.2: Build hospital-specific AUD classification models through style transfer learning using INPC EMR data. Specific models will be built for Indiana University Health (IUH) and Eskenazi Health (EH). Aim 2. Validate the performance of the hospital- specific AUD classification models through clinical surveys. Current IUH and EH patients will be screened using corresponding hospital-specific AUD classification models using INPC EMR data. According to the model predictions, 200 patients who do not have AUD diagnosis (50% predicted as AUD positive and 50% as AUD negative) will be tested using AUIDT. The success of the proposed GAIPA model will provide a transferrable AUD classification model, bridge gaps between rich longitudinal EHR data and precision identification and management of AUD. This work will shift paradigms of big data and complex disease research, enabling EHR data to become part of daily AUD management.