Use of Systematic Methods to Improve Disease Identification in Administrative Data: The Case of Severe Sepsis.

Background: Selection of International Classification of Diseases (ICD)-based coded information for complex conditions such as severe sepsis is a subjective process and the results are sensitive to the codes selected. We use an innovative data exploration method to guide ICD-based case selection for severe sepsis.

Methods: Using the Nationwide Inpatient Sample, we applied Latent Class Analysis (LCA) to determine if medical coders follow any uniform and sensible coding for observations with severe sepsis. We examined whether ICD-9 codes specific to sepsis (038.xx for septicemia, a subset of 995.9 codes representing Systemic Inflammatory Response syndrome, and 785.52 for septic shock) could all be members of the same latent class.

Results: Hospitalizations coded with sepsis-specific codes could be assigned to a latent class of their own. This class constituted 22.8% of all potential sepsis observations. The probability of an observation with any sepsis-specific codes being assigned to the residual class was near 0. The chance of an observation in the residual class having a sepsis-specific code as the principal diagnosis was close to 0. Validity of sepsis class assignment is supported by empirical results, which indicated that in-hospital deaths in the sepsis-specific class were around 4 times as likely as that in the residual class.

Conclusions: The conventional methods of defining severe sepsis cases in observational data substantially misclassify sepsis cases. We suggest a methodology that helps reliable selection of ICD codes for conditions that require complex coding.