Early prediction of shock in intensive care unit patients by machine learning using discrete electronic health record data

Purpose: To use machine learning to predict new-onset shock for at-risk intensive care unit (ICU) patients based on discrete vital sign data from the electronic health record. Methods and results: We included 11,305 adult cardiac, medical, neurological, and surgical ICU patients who did not have shock within 4 h of ICU admission. We used routine vital sign data collected from the first 4 h of the ICU stay to predict new-onset shock within the subsequent 4 h. We compared logistic regression with machine learning models including elastic net, random forest, boosted trees and extreme gradient boosting (XGB). Median age was 64.0 years (44.5 % females). New-onset shock after 4 h developed in 483 (4.3 %) patients, and these patients had higher ICU (8.5 % vs. 1.9 %) and in-hospital (14.3 % vs. 5.0 %) mortality. Standard logistic regression had limited discrimination for new-onset shock, with the best single predictors being the maximum shock index and the minimum blood pressure during the second 2 h of the ICU stay. Discrimination in the validation cohort (n = 2826) was better for each ML model: elastic net, 0.76; boosted tree, 0.76; random forest, 0.79; XGB, 0.82; each model had ≥ 98 % negative predictive value. Accuracy was highest (81 %) with XGB, although positive predictive value was only 14 %. The XGB model also predicted in-hospital mortality with good discrimination. Conclusions: Machine learning prediction models can achieve very good discrimination and accuracy for new-onset shock in ICU patients using vital sign data within 4 h after ICU admission.