By comparing four methods for predicting septic shock in children hospitalized with sepsis, Johns Hopkins researchers have found a newer machine learning approach that is superior to an older one and to two conventional methods.
The top performer, the open-source XGBoost (for eXtreme Gradient Boosting), delivered accurate early predictions that, in clinical practice, would have given intensive care teams nearly nine hours to intervene preventively.
The researchers used data from more than 6,100 past patients from the Johns Hopkins pediatric ICU to retrospectively train and test the model.
XGBoost also had early prediction performance of 0.90 area under the receiver efficacy curve, 43% overall positive predictive value, and patient-specific positive predictive value as high as 62%, the authors report.
The other AI contender was a generalized linear model, while the field was complemented by Cox proportional hazard modeling and sequential assessment of organ failure.
Senior author Raimond Winslow, PhD, and colleagues describe the work in the June issue of Critical Care Explorations, a journal of the Society of Critical Care Medicine.