(PLEASE KEEP IN WORD, DO NOT PDF)
Acute Kidney Injury (AKI) is a life-threatening clinical syndrome prevalent in hospitalized patients (10-15% affected), especially among critically ill patients (>50% affected). AKI patients are 6.5-fold more likely to die in the hospital and at much higher risk for developing poor long-term outcomes including incident and progressive chronic kidney disease, cardiovascular disease, and death. Early and reliable risk assessment is the key to proactive intervention and prevention.
With the ever-growing availability of electronic health records (EHR), machine learning has made substantial progress in modeling the complex data for disease risk prediction including AKI. However, the majority of existing prediction models are built on data from a predefined patient cohort, also known as a global prediction model, optimized for the supposedly “average” patient. This one-size-fits-all prediction model may not work for all patients, and is especially inadequate for heterogeneous diseases such as AKI that have multiple etiologies, variable pathogenesis, and diverse outcomes. Combining patients with different etiologies in training a prediction model may hide subgroups that are more tightly associated with the clinical outcome of interest and may conceal unique pathophysiological processes specific to certain subgroups.
In our previous work, we found that a global model can make completely wrong AKI risk predictions for patients in high-risk and heterogeneous subgroups. Personalized modeling is a promising alternative in which a prediction model is dynamically trained for each incoming patient by using retrospective data of an individualized cohort of similar patients. Our previous work demonstrated that personalized modeling can capture patient heterogeneity with an improved AKI risk prediction for various subpopulations, but we also identified critical challenges to address to ensure model reliability and robustness.
The overall goal of this project is to develop personalized transfer learning methods to achieve equitable AKI prediction across subpopulations in the hospital setting. Specifically, we propose to develop new machine learning techniques to address two challenges in personalized transfer learning: (1) what is the best way to construct individualized patient cohort? and (2) how to avoid negative knowledge transfer during learning? Methods developed in this project are broadly applicable to other diseases and study findings can advance personalized clinical decision support for improving patient outcome.