122: Using Artificial Intelligence to Optimize RRT Machine Allocation During COVID-19-Related RRT Surge

INTRODUCTION: Amongst COVID19 patients in the intensive care unit (ICU), acute kidney injury (AKI) occurs in 40-60% with 10-30% requiring renal replacement therapy (RRT) During any ICU surge situation, an increase in the total number of patients that will require RRT should be expected as the number of ICU patients increases This surge in RRT needs can quickly exhaust RRT machine availability During the pandemic, hospitals developed RRT surge contingency plans to maximize the number of patients who can receive RRT on a given day This may involve mixing modalities of therapy to include using CRRT machines to provide prolonged intermittent RRT (PIRRT) in a shift-based model However, implementing a machine allocation system to deploy RRT machines in the most efficient manner is a vexing challenge Automated strategies for machine allocation could aid staff in quickly developing daily RRT operational plans METHODS: Efficient routing of CRRT machines is analogous to an optimization routing problem, traditionally formulated with multiple vehicles picking up varying loads at each stop We reformulate the problem with CRRT machines as vehicles and patient needs as varying loads, with a maximum load of 24 patient-hours per day per machine, including transition time between patients RESULTS: Our solution, designed to be run daily, was written in Python 3 6 In conjunction with the primary team, nephrologists determine RRT durations based on patient need (6-12 hours vs 24 hours), generating a list of needs in the upcoming 24h period on a shared server If there is more RRT demand than capacity, nephrologists will be asked to reallocate required hours Next, we process CRRT machine locations with patient needs and locations The distance matrix preferentially routes machines between patients within units, then between closeby units, with preference given towards patients and units of similar COVID status to support cohorting The system then generates a sequence of patients that each machine should serve within minutes CONCLUSIONS: We report this as the first known implementation of automated scheduling for optimizing CRRT machine utilization given scarcity constraints from COVIDrelated surges Further characterization is necessary to quantify workload benefit and machine utilization