How do experts classify sepsis cases for sepsis surveillance? Lessons learned from a Behavioural Artificial Intelligence Technology (BAIT) approach.

Objectives: To identify relevant objective variables for retrospective identification of 'suspected infection' and sepsis, using behavioural artificial intelligence technology (BAIT), and to explore the accuracy of this approach for sepsis surveillance.

Methods: BAIT uses choice behaviour analysis to make implicit expert knowledge explicit. Online choice experiments with 25-30 hypothetical retrospective patient scenarios were composed, each consisting of objective variables relevant to sepsis surveillance. Experts reviewed these scenarios and labelled them as 'sepsis' or 'no sepsis'. Two rounds were conducted: round 1 focused on a sepsis surveillance definition, round 2 only on 'suspected infection' in patients with a qSOFA≥2. Relative importance (RI) of variables was calculated using binary logistic regression. Model accuracy was assessed using an expert adjudicated sepsis database.

Results: In round 1, 22 experts participated. Temperature (RI 24 %), CRP (RI 18 %) and systolic blood pressure (RI 16 %) contributed most to sepsis identification, respectively. Model accuracy was 74 % (sensitivity 87 %, specificity 66 %). Round 2 involved 21 experts. Focusing on 'suspected infection', CRP (RI 27 %), temperature (RI 18 %) and leukocyte count (RI 11 %) were most important, respectively. Model accuracy was 75 % (sensitivity 83 %, specificity 71 %).

Conclusion: Inflammatory parameters contributed most to retrospective sepsis and 'suspected infection' identification by experts. BAIT-model accuracy for surveillance was 74-75 %.