Congenital Heart Disease and Pediatric Cardiology
Session Title: Devices, Artificial Intelligence, Digital and Machine Learning
Abstract 16730: Computerized Prediction of Avoidable Serum Potassium Testing in Critically Ill Cardiac Patients
Introduction: Electrolytes are frequently monitored among critically ill patients, especially among cardiac patients due to higher risk of arrhythmias. However, a percentage of these blood draws may potentially be saved. To date, no studies have investigated strategies to safely reduce the density of electrolyte monitoring in critically ill patients.
Hypothesis: We hypothesized that machine learning models can identify potentially avoidable blood draws for serum potassium (K) among pediatric patients following cardiac surgery.
Methods: We retrospectively reviewed data of all patients admitted to the CICU at Boston Children’s Hospital during 2010-2018, having a length of stay ≥4 days and ≥2 recorded serum K measurements. We collected variables related to K homeostasis, including serum chemistry, hourly K intake, diuretics, and urine output. Using established machine learning techniques (Random Forest classifiers and hyperparameters) we created models predicting whether a patient’s K would be normal or abnormal based on the most recent K level, medications administered, urine output, and markers of renal function. We developed multiple models based on different age categories and temporal proximity of the most recent K measurement. We assessed the predictive performance of the models using an independent test set.
Results: Of the 7,269 admissions (6,196 patients) included, 95,674 serum K was measured on average of 1 (IQR 0-1) time per day. 96% of patients received at least one dose of IV diuretic and 83% received a form of K supplementation. Our models predicted a normal K value with a median positive predictive value of 0.90. A median percentage of 2.1% measurements (mean 2.5%, IQR 1.3%-3.7%) were incorrectly predicted as normal when they were abnormal. A median percentage of 0.0% (IQR 0.0%-0.4%) were incorrectly predicted as normal while being critically low or high. A median of 27.2% (IQR 7.8%-32.4%) of samples were correctly predicted to be normal and could have been potentially avoided.
Conclusions: Machine-learning methods can be used to accurately predict avoidable blood tests for serum K in critically ill pediatric patients. A median of 27.2% of samples could have been saved, with decreased costs and risk of infection or anemia.