Agreement of claims-based methods for identifying sepsis with clinical criteria in the REasons for Geographic and Racial Differences in Stroke (REGARDS) cohort

Abstract

Background: Claims-based algorithms are commonly used to identify sepsis in health services research because the laboratory features required to define clinical criteria may not be available in administrative data.

Sepsis labels defined by claims-based methods are ill-suited for training machine learning algorithms

“…International Classification of Diseases (ICD)-coding) [1]. These approaches have severe limitations [2]. For instance, labels based on claims-based methods are known to have high specificity, but low sensitivity [2].…”

“…These approaches have severe limitations [2]. For instance, labels based on claims-based methods are known to have high specificity, but low sensitivity [2]. In the context of automated sepsis diagnosis at the ED, sensitivity is crucial because underdiagnosing a patient with sepsis can have fatal consequences.…”