Machine Learning for Sepsis Identification

Early sepsis identification improves outcomes, but currently there is no early definition of sepsis with acceptable sensitivity and specificity. These authors used machine learning to derive and validate a new “Risk of Sepsis” score in a large emergency department (ED) population.

The authors randomly divided a retrospective cohort of ED encounters from a large healthcare system into a training cohort (67%) and a testing cohort (33%). The definition of sepsis presumed infection plus organ dysfunction, and notably did not include lactate results. Risk of Sepsis scores were reported at 1, 3, 6, 12, and 24 hours and were compared against several other sepsis scores.

Among nearly 3 million patients, 54,661 (2.0%) had sepsis according to the study…