A machine learning model accurately predicted the onset of sepsis in ICU patients four to 12 hours prior to clinical recognition in a recent study.
The observational cohort study used data from more than 31,000 admissions to ICUs at two academic hospitals as a development cohort and over 52,000 ICU patients from a publicly available database as a validation cohort. Patients who already met Sepsis-3 criteria were excluded. A set of 65 variables was calculated on an hourly basis and put into the Artificial Intelligence Sepsis Expert algorithm to predict the onset of sepsis within four, six, eight, or 12 hours. The algorithm also produced a list of the most significant contributing factors. Results were published by Critical Care Medicine on Dec. 26.
For all four time periods, the algorithm achieved an area under the receiver-operating characteristic curve (AUROC) in the range of 0.83 to 0.85, leading the study authors to conclude that it “can accurately predict the onset of sepsis in an ICU patient 4–12 hours prior to clinical recognition.” This was true whether the predicted outcome was objective physiologic manifestations of sepsis (the Sepsis-related Organ Failure Assessment score) or clinical suspicion of infection. The authors noted that the algorithm's predictive performance, as measured by the AUROC, specificity, and accuracy, decreased as the prediction window lengthened from four to 12 hours.
The next step for the tool will be to have its clinical utility analyzed in a prospective study, with the goal of having patients treated with IV antibiotics, fluids, and other adjunct therapies “based on a reliable estimate of their likelihood of developing sepsis in the near future,” the authors said. They plan to test this approach using a remote monitoring team, rather than sending alerts to bedside clinicians. The tool might also be useful in ICU patients who are decompensating due to causes other than sepsis, given that study patients who had high risk scores but did not develop sepsis (false positives) had twice the mortality risk of those who had low risk scores but developed sepsis (false negatives), the authors said.
The cover story in the January ACP Hospitalist focuses on the potential and pitfalls of using machine learning to predict inpatients' outcomes, particularly the development of sepsis and septic shock.