Identifying Ascites in Patients with Cirrhosis Using Administrative Codes and Diuretic Use: A Multicenter Study.

ABSTRACT

BACKGROUND: Ascites is associated with significantly increased morbidity, mortality, and health care costs. Large population studies are necessary to determine the burden and impact of ascites; however, ascites ICD-10 codes perform poorly in the identification of patients. METHODS: We utilized three independent retrospective cohorts at the University of Michigan (cohorts 1 and 2) and Duke University (cohort 3). Cohort 1 Child A5-6 patients followed up to 10 years (n = 150); cohort 2 Child A5-B7 patients with portal hypertension followed for up to 1 year (n = 65); cohort 3 cross-sectional cohort of patients evaluated for liver transplant (n = 100). We computed performance characteristics for ascites-related ICD-10 codes (K70.31, K70.11, K71.51, R18.8), as well as loop and/or potassium-sparing diuretics. RESULTS: A total of 315 patients were included across three cohorts. Algorithms including any ascites code provided better sensitivity and equivalent specificity to R18.8 alone for all cohorts. In cohort 2, we found that loop diuretics, potassium-sparing diuretics, and a combination of both with a cirrhosis code were highly sensitive (82.3% for each) and specific (89.1-93.5%). In contrast, ascites codes were insensitive. In patients with moderate-severe ascites, a combination of recorded diuretics showed high sensitivity and specificity (95.2% and 86.8%). In Cohort 3's transplant evaluation patients, we found that loop diuretics, potassium-sparing diuretics, and a combination of both with a cirrhosis code were highly sensitive (90.4%, 78.8% and 75.0%, respectively) and specific (85.0%, 90.0% and 95.0%, respectively). For moderate-severe cirrhosis, loop diuretics and R18.8 showed higher sensitivity (77.8%) and specificity (88.9%), respectively. CONCLUSION: Diuretic records with a cirrhosis code improve the identification of ascites. This method for identifying ascites should be used in future large dataset studies.