Data sources
We conducted this study using reimbursement claims data from Taiwan’s National Health Insurance Program. This program merged former insurance systems in March 1995 and covers more than 99% of Taiwan’s 23 million residents. The National Health Research Institutes established a National Health Insurance Research Database (NHIRD) to record all beneficiaries’ inpatient and outpatient medical services. This information includes basic patient demographics, physician’s primary and secondary disease diagnoses, treatment procedures, prescribed medications and medical expenditures for all health care services. The validity of this database has been favorably evaluated, and research articles based on it have been accepted in prominent scientific journals worldwide [17-20].
Ethics
This study was conducted in accordance with the Helsinki Declaration. To protect personal privacy, the electronic database was decoded with patient identifications scrambled for further academic access for research. According to Taiwan National Health Research Institutes regulations, informed consent is not required because patient identifications were decoded and scrambled [18-20]. Ethical approval for this study (TMU-JIRB-201504008) was provided by the Institutional Review Board of Taipei Medical University.
Study design
Among 23 million beneficiaries, 1,250,300 patients were admitted to ICU between 2006 and 2012 (supplementary Figure S1). We identified 79,528 patients aged ≥ 20 years who had histories of liver cirrhosis from the National Health Insurance Research Database. Patients with liver cirrhosis were defined as having at least two visits for medical care with physician’s primary diagnosis of liver cirrhosis within the 24 months before ICU admission. To select appropriate comparison groups, we matched each ICU patient with cirrhosis with one randomly selected ICU patients without liver cirrhosis by the analysis with a propensity score-matched pair procedure (case-control ratio=1:1). These matched factors included age, sex, low income, stay in medical center or not, diabetes, hypertension, mental disorder, chronic obstructive pulmonary disease, fracture, pneumonia, stroke, asthma, traumatic brain injury, congestive heart failure, immune thrombocytopenia, renal dialysis, hyperlipidemia, epilepsy, atrial fibrillation, peripheral vascular disease and systemic lupus erythematosus, causes of admission to ICU according to physician’s primary diagnosis (digestive disease, cancer, respiratory disease, circulatory disease, infectious disease, injury and poisoning, symptom-defined conditions, genitourinary disease, endocrine disease, musculoskeletal disease, neurological disease, skin disease, mental disorder, tumors, blood diseases, congenital anomalies, disease of perinatal period, complications of pregnancy, ICU complications (such as septicemia, pneumonia, acute renal failure, urinary tract infection, stroke, acute myocardial infarction and pulmonary embolism). After matching selection, there were 37,197 patients with cirrhosis of liver in the exposure group and 37,197 people without liver cirrhosis in non-exposure group. We investigated the impact of liver cirrhosis on 30-day mortality, ICU mortality, and one-year mortality among ICU patients in this study.
Measures and definitions
Income status was identified by defining low-income patients as those who qualified for waived medical copayment, as this status is verified by the National Health Insurance Bureau. Whether patients stayed in medical center ICUs or those in other hospitals was also recorded. We used the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) to define coexisting medical conditions and ICU complications. Details codes of ICD-9-CM for these diseases were listed in supplementary Table S1. Cirrhosis of liver before ICU stay was defined as the major exposure. Coexisting medical conditions determined from medical claims within the 24-month period before ICU stay included diabetes, hypertension, mental disorders, chronic obstructive pulmonary disease, fracture, pneumonia, stroke, asthma, traumatic brain injury, congestive heart failure, immune thrombocytopenia, hyperlipidemia, epilepsy, atrial fibrillation, peripheral vascular disease, and systemic lupus erythematosus. Renal dialysis was defined by administration code (D8, D9). Seven major complications during the ICU stay were analyzed (and those having severe cases of these diseases before ICU were excluded) including septicemia, pneumonia, acute renal failure, urinary tract infection, stroke, acute myocardial infarction, and pulmonary embolism. Length of hospital stay and ICU medical expenditure were analyzed as secondary outcomes.
Causes of admission to ICU (according to physician’s primary diagnosis at admission) were also identified and described with disease codes including digestive disease, cancer, respiratory disease, circulatory disease, infectious disease, injury and poisoning, symptom-defined conditions, genitourinary disease, endocrine disease, musculoskeletal disease, neurological disease, skin disease, mental disorder, tumor, blood disease, congenital anomalies, disease of perinatal period, and pregnancy complications.
Statistical analysis
To reduce confounding bias, we used a propensity score-matched pair combined with frequency matching procedure to balance the covariates between ICU patients with and without liver cirrhosis. We developed a non-parsimonious multivariable logistic regression model to estimate a propensity score for pre-ICU cirrhosis of the liver. We matched cirrhotic patients to patients without liver cirrhosis, using a greedy matching algorithm (without replacement) with a caliper width of 0.2 SDs of the log odds of the estimated propensity score. Clinical significance guided initial choices of covariates in this multivariable logistic regression model: age, sex, low income, ICU stay in medical center or not, diabetes, hypertension, mental disorders, chronic obstructive pulmonary disease, fracture, pneumonia, stroke, asthma, traumatic brain injury, congestive heart failure, immune thrombocytopenia, renal dialysis, hyperlipidemia, epilepsy, atrial fibrillation, peripheral vascular disease, systemic lupus erythematosus, septicemia, pneumonia, acute renal failure, urinary tract infection, stroke, acute myocardial infarction, pulmonary embolism, digestive disease, cancer, respiratory disease, circulatory disease, infectious disease, injury and poisoning, symptom-defined conditions, genitourinary disease, endocrine disease, musculoskeletal disease, neurological disease, skin disease, mental disorder, tumor, blood disease, congenital anomalies, disease of perinatal period, and complications of pregnancy. A structured iterative approach was used to refine this model to achieve covariate balance within matched pairs. We used chi-square tests to measure covariate balance, and p < 0.05 was suggested to represent meaningful covariate imbalance. We matched patients with and without cirrhosis using a greedy-matching algorithm with a caliper width of 0.2 SD of the log odds of the estimated propensity score. This method could remove 98% of bias from measured covariates.
Adjusted odds ratios (aORs) with 95% confidence intervals (CIs) for 30-day mortality, ICU mortality, and one-year mortality for patients with and without cirrhosis were analyzed with multiple logistic regression models by controlling for age, sex, low income, stay in medical center or not, coexisting medical conditions, ICU complications and admission causes. To confirm associations between liver cirrhosis and ICU mortality, we also performed stratification analysis by age, sex, low income, stay in medical center or not, coexisting medical conditions, ICU complications and causes of ICU admission. The impacts of liver-related indicators and medical care on 30-day mortality in ICU patients with cirrhosis were also measured by calculating adjusted ORs and 95% CIs in the multivariate logistic regression models. SAS version 9.1 (SAS Institute Inc., Cary, NC, USA) statistical software was used for data analyses; two-sided p < 0.05 indicated significant differences.