PROJECT SUMMARY Acute Kidney Injury (AKI) is a heterogeneous syndrome that has multiple etiologies, variable pathogenesis, and diverse outcomes. For example, congestive heart failure and dehydration can produce identical changes in serum creatinine level and urine output (i.e. parameters used to define AKI); however, they differ vastly in their physiological contexts and demand completely opposite treatments. AKI is common in hospitalized patients, affecting 10% to 15% inpatients in the general units and >50% in the intensive care units. Regardless of the underlying cause, even mild forms of AKI are associated with 6.5-fold increase in mortality. The current clinical management guideline for AKI is based on minimizing the risk of developing AKI and providing supportive care. However, the sheer number of known and potentially unknown risk factors of AKI and the complex interactions among them make it impossible for physicians to analyze and forecast AKI risk for a single patient in real time. Machine learning has demonstrated its success in modeling complex electronic health record (EHR) data for disease risk predictions, including AKI. Overwhelming majority of the clinical risk prediction models are trained on data from a predefined patient cohort, also known as a global prediction model, optimized for the supposedly “average” patient. However, the one-size-fits-all prediction model may not work for all patients. Findings from our work in current funding cycle revealed that a global model can make completely wrong predictions for patients in high-risk and heterogeneous (variable pathogenesis) subgroups because it only captures knowledge generalizable to a study population but miss subtle risk drivers specific to an individual patient. Personalized modeling is a promising approach in which a model is trained on-demand for each patient by identifying an individualized retrospective cohort of similar patients. In the current funding cycle, we developed and validated a novel personalized AKI prediction framework and demonstrated its ability to capture patient heterogeneity with an improved AKI risk prediction for individuals. Building on the success of our current project, in this renewal application, we propose to focus the development of new machine learning methods within the personalized modeling framework to gain deeper understanding of the personalized AKI risk and its predictors from individualized cohorts. We designed three specific aims within the personalized modeling framework to answer three important clinical questions: (1) what is an individual patient’s risk for developing AKI during hospital stay? (2) why is an individual patient at risk for developing AKI during hospital stay? (3) when will an individual patient’s kidney function recover after AKI onset? The proposed research has the potential to advance personalized decision support, inform personalized intervention, and facilitate shared decision making for providers, AKI patient...