Abstract Intradialytic hypotension (IDH) occurs in one-third of critically ill patients with acute kidney injury and treated with kidney replacement therapy in the intensive care unit (ICU). Occurrence of IDH is associated with increased resource utilization such as fluid and vasopressor administration, discontinuation of kidney replacement therapy, decreased recovery of kidney function, dependence on kidney replacement therapy and death. IDH is often unrecognized until it is well established, by which time patients are refractory to treatment or have already developed organ injury. Thus, if one could accurately predict who and when patients develop IDH, then effective preemptive treatments could be administered to reduce risk of IDH and improve outcomes. Our preliminary work showed that advanced high-frequency data modeling and waveform analysis identified patients at risk for hypotension within 2 minutes of monitoring in the ICU, and if monitored for 5 minutes, differentiated between patients who would develop hypotension or remain stable over the next 48 hours. In this proposal entitled “Artificial Intelligence Driven Acute Renal Replacement Therapy (AID-ART)”, we propose to apply predictive analytics using linked electronic health record and high-frequency monitor data to critically ill patients with acute kidney injury and undergoing intermittent and continuous kidney replacement therapies at the University of Pittsburgh Medical Center and the Mayo Clinic ICUs. We will examine the accuracy of various machine learning models to predict IDH risk-evaluating model performance, usability, alert frequency, lead time and number needed to alert, and hospital mortality and dependence on kidney replacement therapy (Aim 1a); predict response to a range of clinical interventions for IDH and subsequent clinical outcomes (Aim 1b); and perform cross validation across the two healthcare systems (Aim 1c). We will construct reinforcement learning systems to develop a rule-driven intervention for IDH alerts and measurement-driven responses to avoid and respond to IDH based on principles of functional hemodynamic monitoring (Aim 2a). We will also develop a reinforcement learning algorithm to learn an optimal intervention strategy based on the probability of events rather than in reaction to IDH events (Aim 2b). We will silently deploy and evaluate the ability of this artificial intelligence (AI) algorithm to forecast IDH risk and recommend interventions in real-time across the two healthcare systems. We will then assess the validity of recommended interventions using an expert clinician adjudication panel (Aim 3a); and will compare the AI recommended interventions with that of actual interventions performed by bedside clinicians (Aim 3b). This proposal will be the harbinger of a future multicenter randomized clinical trial to examine personalized risk prediction and AI-augmented management of IDH among critically ill patients with acute kidney injury and undergoing kidney replacement therapy in the intensive care unit.

Project Narrative Among critically ill patients with acute kidney injury and treated with dialysis in the intensive care unit, if one could accurately predict early which patients will develop low blood pressure then effective treatments could be given. In this study, we propose to develop and validate an artificial intelligence system that will predict low blood pressure even before they occur and recommend correct treatment to clinicians.