Lead/Presenter: Jejo Koola,
All Authors: Koola JD (VA San Diego Healthcare System)
Ho SB (VA San Diego Healthcare System)
Chen G (Vanderbilt Unversity Medical Center)
Perkins AM (Vanderbilt Unversity Medical Center)
Cao A (Vanderbilt Unversity Medical Center)
Matheny ME (VA Tennessee Valley Healthcare System)
Objectives:
We sought to develop a national mortality risk prediction model among hospitalized patients with cirrhosis. Though several mortality models exist, few use a time to event outcome with censoring, many have small sample sizes, and none for VA patients with cirrhosis. Models developed in disparate populations suffer from reduced performance, and models that are more population focused can leverage risk variables relevant only to that population.
Methods:
We analyzed a retrospective cohort of cirrhotics hospitalized from 124 medical centers in the Department of Veterans Affairs between 01/01/2005 and 12/31/2013. Data were collected from the corporate data warehouse. We excluded patients with 1) liver transplantation, 2) discharged against medical advice, 4) transfer from another facility, 5) death prior to discharge, and 6) length of stay > 30 days. Using a time dependent cox proportional hazards model, the outcome was time to all-cause death with censoring for study date ending or liver transplant. The index date was discharge date. 176 candidate variables were used from patient demographics, radiology, encounters, administrative codes, laboratory tests, and medications. Missing values were imputed using non-negative matrix factorization. Discrimination was assessed with time varying AUC and overall c-index, and optimism estimated with 200 bootstraps.
Results:
A total of 247,957 patient hospitalizations were included with an overall cohort mortality rate of 55.6%. The AUCs at 90, 183, and 365 days were 0.80-0.81 with an overall model C-index of 0.87. In the final model with 156 variables, the strongest predictors were cancer metastasis (Hazard Ratio[HR] 2.28), hepatocellular carcinoma (HR 1.76), midodrine (HR 1.56), albumin infusion (HR 1.47), unknown race (HR 1.31), lactulose (HR 1.27), hepatorenal syndrome (HR 1.25), and discharge locations to another hospital (HR 1.60), hospice (4.79), nursing home (2.40), serum albumin (HR 0.78), and lincomycin (HR 0.66).
Implications:
This study highlights a risk prediction model with high discrimination for mortality predicting among cirrhotic patients, and can be used to tailor peri-discharge management and mobilize resources. It also identified variables not reported in prior literature, including modifiable risk factors.
Impacts:
This model can be embedded within clinical decision support tools to deliver personalized medicine for patients with cirrhosis.