Development and validation of a diagnostic model for early differentiation of sepsis and non-infectious SIRS in critically ill children - a data-driven approach using machine-learning algorithms

BMC Pediatr. 2018 Mar 15;18(1):112. doi: 10.1186/s12887-018-1082-2.

Abstract

Background: Since early antimicrobial therapy is mandatory in septic patients, immediate diagnosis and distinction from non-infectious SIRS is essential but hampered by the similarity of symptoms between both entities. We aimed to develop a diagnostic model for differentiation of sepsis and non-infectious SIRS in critically ill children based on routinely available parameters (baseline characteristics, clinical/laboratory parameters, technical/medical support).

Methods: This is a secondary analysis of a randomized controlled trial conducted at a German tertiary-care pediatric intensive care unit (PICU). Two hundred thirty-eight cases of non-infectious SIRS and 58 cases of sepsis (as defined by IPSCC criteria) were included. We applied a Random Forest approach to identify the best set of predictors out of 44 variables measured at the day of onset of the disease. The developed diagnostic model was validated in a temporal split-sample approach.

Results: A model including four clinical (length of PICU stay until onset of non-infectious SIRS/sepsis, central line, core temperature, number of non-infectious SIRS/sepsis episodes prior to diagnosis) and four laboratory parameters (interleukin-6, platelet count, procalcitonin, CRP) was identified in the training dataset. Validation in the test dataset revealed an AUC of 0.78 (95% CI: 0.70-0.87). Our model was superior to previously proposed biomarkers such as CRP, interleukin-6, procalcitonin or a combination of CRP and procalcitonin (maximum AUC = 0.63; 95% CI: 0.52-0.74). When aiming at a complete identification of sepsis cases (100%; 95% CI: 87-100%), 28% (95% CI: 20-38%) of non-infectious SIRS cases were assorted correctly.

Conclusions: Our approach allows early recognition of sepsis with an accuracy superior to previously described biomarkers, and could potentially reduce antibiotic use by 30% in non-infectious SIRS cases. External validation studies are necessary to confirm the generalizability of our approach across populations and treatment practices.

Trial registration: ClinicalTrials.gov number: NCT00209768; registration date: September 21, 2005.

Keywords: Diagnosis; Intensive care unit; Pediatric; Random Forest; SIRS; Sepsis.

Publication types

  • Randomized Controlled Trial
  • Research Support, Non-U.S. Gov't

MeSH terms

  • Adolescent
  • Algorithms*
  • Child
  • Child, Preschool
  • Critical Illness
  • Decision Support Techniques*
  • Diagnosis, Differential
  • Early Diagnosis
  • Female
  • Follow-Up Studies
  • Humans
  • Infant
  • Infant, Newborn
  • Machine Learning*
  • Male
  • Prospective Studies
  • Reproducibility of Results
  • Sepsis / diagnosis
  • Systemic Inflammatory Response Syndrome / diagnosis*

Associated data

  • ClinicalTrials.gov/NCT00209768