ABSTRACT
Acute kidney injury (AKI) affects up to half of critically ill patients admitted to intensive care units (ICU). In
patients with AKI and hemodynamic instability, continuous renal replacement therapy (CRRT) is the
preferred dialysis modality. ICU mortality in this vulnerable population is high but kidney recovery occurs in
up to two-thirds of survivors. Universally accepted and accurate approaches for predicting survival or kidney
recovery in these patients do not exist currently. This is clinically relevant as prediction of key outcomes
could guide decision-making of CRRT delivery, goals of acute care, and personalized post-ICU care
according to kidney recovery prognosis. Since there are no proven interventions to improve outcomes in
these patients, identification of modifiable risk factors and sub-phenotypes is necessary to develop precision
medicine approaches in CRRT. Due to advances in artificial intelligence (AI) and availability of multi-modal
data, deep learning (DL) –a subset of AI– is a valuable approach that allows construction of accurate and
reliable risk prediction models. Further, the use of novel algorithms such as the Feasible Solution Algorithm
(FSA) could help identify patient sub-phenotypes and model applications. We propose to develop and
validate innovative and reproducible DL approaches to predict RRT-free survival at actionable timepoints
and use FSA to identify patient sub-phenotypes with differing RRT-free survival risk according to multi-
modal data. Our published preliminary data demonstrated superiority of DL models compared to optimized
logistic regression for RRT-free survival prediction. Prediction of 24-hour mortality was improved by
incorporating time-series data during CRRT. We hypothesize that time-series multi-modal data
(including EHR and CRRT machine data) will generate accurate and generalizable risk prediction to
guide clinical interventions and identify sub-phenotypes for model interpretation and clinical utility
testing. We will utilize datasets from 9 institutions that encompass multi-modal EHR clinical data and
programmatic and therapy data from CRRT machines for model and sub-phenotyping development, testing,
and independent validation. This innovative research will 1) assist development of clinical decision support
platforms to guide informed CRRT delivery and improve clinical outcomes and 2) identify sub-phenotypes of
patients that could benefit from more personalized and testable novel CRRT interventions.
