Diagnosis-dependent misclassification of infections using administrative data variably affected incidence and mortality estimates in ICU patients R. Gedeborg a, * , M. Furebring b , K. Michae ¨lsson c a Department of Surgical Sciences, Section of Anaesthesiology and Intensive Care, Uppsala University Hospital, SE-751 85, Uppsala, Sweden b Department of Medical Sciences, Section of Infectious Diseases, Uppsala University Hospital, Uppsala, Sweden c Department of Surgical Sciences, Section of Orthopaedics, Uppsala University Hospital, Uppsala, Sweden Accepted 15 May 2006 Abstract Objective: To determine the accuracy of hospital discharge diagnoses in identifying severe infections among intensive care unit (ICU) patients, and estimate the impact of misclassification on incidence and 1-year mortality. Study Design and Setting: Sepsis, pneumonia, and central nervous system (CNS) infections among 7,615 ICU admissions were iden- tified using ICD-9 and ICD-10 diagnoses from the Swedish hospital discharge register (HDR). Sensitivity, specificity, and likelihood ratios were calculated using ICU database diagnoses as reference standard, with inclusion in sepsis trials (IST) as secondary reference for sepsis. Results: CNS infections were accurately captured (sensitivity 95.4% [confidence interval (CI) 5 86.8e100] and specificity 99.6% [CI 5 99.4e99.8]). Community-acquired sepsis (sensitivity 51.1% [CI 5 41.0e61.2] and specificity 99.4% [CI 5 99.2e99.6]) and primary pneumonia (sensitivity 38.2% [CI 5 31.2e45.2] and specificity 98.6% [CI 5 98.2e99.0]) were more accurately detected than sepsis and pneumonia in general. One-year mortality was accurately estimated for primary pneumonia but underestimated for community-acquired sepsis. However, there were only small differences in sensitivity and specificity between HDR and ICU data in the ability to identify IST. ICD-9 appeared more accurate for sepsis, whereas ICD-10 was more accurate for pneumonia. Conclusion: Accuracy of hospital discharge diagnoses varied depending on diagnosis and case definition. The pattern of misclassifi- cation makes estimates of relative risk more accurate than estimates of absolute risk. Ó 2007 Elsevier Inc. All rights reserved. Keywords: Sepsis; Pneumonia; Central nervous system (CNS) infections; Diagnosis; Classification; Intensive care 1. Introduction Evaluations of health care delivery and effective plan- ning of clinical trials in intensive care require nationwide estimates of incidence rates and diagnosis-specific mortal- ity. Specialized national databases of intensive care unit (ICU) patients exist, but this is a costly and labor-intensive method for these purposes. An alternative, low-cost approach would be to use readily available administrative databases to extract information required to calculate incidence rates and diagnosis-specific mortality. However, detailed knowledge of the nature and size of diagnosis misclassification is imperative for correct design and inter- pretation of such studies. The accuracy of diagnoses in administrative databases has been found to be highly variable, and depend on the diagnosis and clinical setting, as well as the method used for evaluation [1e6]. Severe infections are major issues in intensive care research but may be difficult to study us- ing available sources of administrative data. Sepsis and pneumonia may be primary diagnoses, or they may be sec- ondary complications of the patient’s main condition. A previous study of the accuracy of hospital discharge diag- nosis of sepsis for the identification of bacteremia in a gen- eral hospital population showed an excessively high degree of misclassification [7], whereas another study found high accuracy in identifying patients with meningococcal dis- ease from two nationwide registers [8]. Different strategies can be applied to increase the accuracy of case definitions. Adjusting restrictions in the sample definition [9] or adding secondary data sources [5,10] can improve the accuracy. It is also essential that patient transfer between departments and hospitals can be traced to avoid overestimation of disease frequency [11]. One way of increasing the accuracy of data from a national hospital discharge register (HDR) would be to * Corresponding author. Tel.: þ46-18-6114893; fax: þ46-18-559357. E-mail address: rolf.gedeborg@surgsci.uu.se (R. Gedeborg). 0895-4356/07/$ e see front matter Ó 2007 Elsevier Inc. All rights reserved. doi: 10.1016/j.jclinepi.2006.05.013 Journal of Clinical Epidemiology 60 (2007) 155e162