Analyses of hospital administrative data that use diagnosis codes overestimate the cases of acute pancreatitis

Abstract

Background & aims: Although widely used, little information exists on the validity of using hospital administrative data to code acute pancreatitis (AP). We sought to determine if discharge diagnosis codes accurately identify patients whose clinical course met the standard for AP diagnosis.

Methods: We analyzed data from 401 unique patients admitted through the emergency department who received a primary inpatient discharge diagnosis of AP at 2 University of Pittsburgh Medical Center hospitals in the years 2000, 2002, and 2005. Each patient was matched with a control patient who was admitted with abdominal pain and then discharged without a diagnosis of AP. Patients were matched based on demographics, testing for serum levels of pancreatic enzymes, year of visit to the emergency department, admission to the intensive care unit, and performance of abdominal computed tomography scan. The standard used to diagnose AP was the presence of 2 of 3 features (abdominal pain, ≥ 3-fold increase in serum levels of pancreatic enzymes, and positive results from imaging analysis).

Results: The median age of AP cases was 53 years (interquartile range, 41.5-67 years); 47.1% were male, 85% were white. The most common etiologies were biliary (33.4%), alcohol-associated (16.2%), and idiopathic (24.2%). Serum levels of pancreatic enzymes were increased by any amount, and by ≥ 3-fold, in 95.3% and 68.6% of patients diagnosed with AP and in 16.2% and 2.2% of controls, respectively. The standard for diagnosis of AP was met in 80% of cases diagnosed with this disorder; they had no history of pancreatitis. The sensitivity, specificity, and positive and negative predictive values of the AP diagnosis code were 96%, 85%, 80%, and 98%, respectively.

Conclusions: Approximately 1 of 5 patients diagnosed with AP upon discharge from the hospital do not meet the guidelines for diagnosis of this disorder. Efforts should be made to more consistently use guidelines for AP diagnosis.

Figure 1

Relationship between specificity and PPV of the criteria used to identify patients with AP using administrative datasets. PPV can improve further if natural language processing for radiology reports is incorporated. Unique patients may not be chosen when the interest is to study episodes of AP. Exclusion of patients with AP or CP diagnosis will depend on study aims.