PROJECT SUMMARY/ABSTRACT In our parent award R01 DK133465, we leverage real-world data (RWD) from the OneFlorida+ Clinical Research Consortium to identify clinically high-benefit patient subgroups for newer glucose-lowering drugs and generate economic evidence for designing policy-level interventions to improve the quality of care and health equity in type 2 diabetes (T2D) care. OneFlorida+ contains ~20 million patient EHRs across Florida, Georgia, and Alabama, linked with data from various other sources, including Medicaid and Medicare claims. We are making progress on (1) developing research-grade computable phenotype algorithms for identifying “loyal patients,” defined as those with medical encounters and drug exposure fully documented in EHRs; (2) identifying clinically high-benefit patient subgroups for newer GLDs; and (3) refining our diabetes microsimulation model to generate economic evidence for designing policy-level interventions to improve the quality of care and health equity in T2D care. In our renewal application, we aim to construct digital twin models of T2D that consider not only clinical characteristics but also the multifaceted social determinants of health (SDoH) to support the integration of social care into health care delivery. Nevertheless, AI/ML-based digital twin models are computationally complex and data-hungry, requiring to make large amounts of real-world patient data AI/ML-ready. In this administrative supplement, in Aim 1, we will develop pipelines and associated documentation to (a) standardize RWD data into a common data model with a focus on the SDoH, and (b) make the RWD into AI/ML- ready datasets, in preparation for the development of T2D digital twin models. Built on our T2D simulation model, we will systematically identify additional factors, with a focus on SDoH, that would significantly affect individuals’ quality of care and adverse outcome, and develop pipelines to extract-transform-load (ETL) from the OneFlorida+ EHR data into the widely adopted Observational Medical Outcomes Partnership Common Data Model (OMOP CDM). In Aim 2, we will evaluate the potential bias of AI/ML models developed with different degrees of EHR data completeness. A common practice in building AI/ML models using RWD is to select only patients that have more complete data, which may introduce bias. We will systematically assess the downstream AI/ML model bias using algorithmic fairness metrics, which is critical for our future development of a fair T2D digital twin. This project will make the data generated from our NIDDK-supported project AI/ML-ready and respond directly to NOT-OD-23-082, where we will (1) prepare “SDoH information for use in AI/ML” and adopt “ontologies or other standards to improve interoperability,” (2) characterize “biases that may affect AI/ML model trained on the data,” and (3) develop “documentation for or AI/ML re-users of the data.” With the success of this administrative supplement, we will be well-posit...