How do experts classify sepsis cases for sepsis surveillance? Lessons learned from a Behavioural Artificial...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Decision Support Systems

Background:

• Sepsis surveillance requires accurate identification of suspected infections.
• Behavioral Artificial Intelligence Technology (BAIT) offers a novel method to operationalize expert knowledge for clinical surveillance.

Purpose of the Study:

• To identify objective variables for retrospective identification of 'suspected infection' and sepsis using BAIT.
• To evaluate the accuracy of BAIT for sepsis surveillance.

Main Methods:

• Online choice experiments with hypothetical patient scenarios were used to elicit expert knowledge.
• Experts labeled scenarios as 'sepsis' or 'no sepsis', with two rounds focusing on different definitions.
• Relative importance of variables was calculated, and model accuracy was assessed against an expert-adjudicated database.

Main Results:

• In round 1 (sepsis identification), temperature, CRP, and systolic blood pressure were most important (RI 24%, 18%, 16%). Model accuracy was 74% (sensitivity 87%, specificity 66%).
• In round 2 ('suspected infection'), CRP, temperature, and leukocyte count were most important (RI 27%, 18%, 11%). Model accuracy was 75% (sensitivity 83%, specificity 71%).

Conclusions:

• Inflammatory parameters are crucial for retrospective identification of sepsis and suspected infections.
• BAIT demonstrated a surveillance accuracy of 74-75% in this study.