Artificial Intelligence to Predict Outcomes in Patients with Acute Kidney Injury on Continuous Renal Replacement Therapy

ABSTRACT Acute kidney injury (AKI) affects up to half of all 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 for solute and volume control. ICU mortality in this vulnerable population is high (~75%) but kidney recovery occurs in up to two-thirds of survivors. Fluid overload is a potentially modifiable risk factor associated with these outcomes. There are currently no universally accepted approaches for predicting patient response to fluid removal, survival or kidney recovery. Due to recent advances in computer science and widespread big data usage, deep learning -a subset of the artificial intelligence universe- has risen as a valuable approach. Deep learning allows construction of risk prediction models using time-series data, incorporating thousands of variables and dynamic (time-varying) changes in these variables derived from multi-dimensional sources and not only static values of these variables. We propose to develop and validate innovative deep learning approaches to dynamically predict these outcomes. We demonstrated the superiority of deep learning models without a-priori variable selection compared to optimized logistic regression (C-Statistic of 0.72 for DL vs. 0.62 for logistic regression) for the prediction of RRT liberation using clinical data prior to CRRT initiation. We showed that model performance improved by incorporating dynamic changes in clinical data within 6-hour intervals after CRRT initiation. We also identified distinctive mortality risk according to quintiles of achieved net ultrafiltration rates, after adjustment by patient’s weight, duration of CRRT, and other critical parameters: OR 8.0 (95% CI: 2.7-25.1) when the highest quintile (>36 ml/kg/day) was compared to the lowest quintile (<13 ml/kg/day). We hypothesize that our innovative deep learning approach using time-series big data will generate more accurate and generalizable models and performance will be superior to traditional approaches. We will utilize a multi-institutional dataset that encompasses comprehensive clinical data and programmatic and therapy data from the CRRT machine (CRRTnet registry, n=1500 patients) for model development and an independent multi-institutional dataset for validation (n=1500 patients) to: 1) continuously predict short-term (7-day) and medium-term (28-day) liberation from RRT due to kidney recovery; 2) continuously predict 24-hour mortality; and 3) identify and validate sub-phenotypes of patients with AKI on CRRT with differing achieved net ultrafiltration rates. This innovative research will assist the development of novel clinical decision support platforms for guiding informed CRRT delivery and promoting survival and kidney recovery in this debilitated population and identify sub-phenotypes of CRRT patients that benefit from precision-medicine approaches to fluid removal and inform the design of interventional studies focusing on fluid removal to impact patient-centered outcomes.