We referred to the strengthening the reporting of observational studies in epidemiology guidelines for drafting the study protocol and reporting the results of our study [14]. The present observational cohort study was designed retrospectively using the National Sample Cohort (NSC) database (version 2.0) developed by the National Health Insurance Sharing Service. The National Health Insurance Program has comprehensively covered the medical insurance and delivery system for the entire population of South Korea. The NSC database includes socio-demographic, clinical, and procedural information on 1,025,340 participants, which constitute 2.2% of the total population who claimed hospital charges, and represents the patients in South Korea from 2002 to 2015 by systematic stratified random sampling with proportional allocation to ensure representativeness [15]. The NSC database consists of the following six tables: (1) qualifications, (2) birth and death, (3) specifications, (4) medical history, (5) diagnosis, and (6) prescription history. As the NSC database includes those who were alive in 2006, we analyzed 997,173 patients who were followed up from 2006 to 2015. An event of ICU admission was identified by special codes in the claim database, which were considered highly accurate because these exclusive codes were made differently from other codes, and every medical institute must enter these codes to claim medical expenses from the National Health Insurance Review & Assessment Service (Supplementary Table 1). Procedures conducted only for critical care, such as MV, endobronchial tube management, and continuous renal replacement management, were identified using the same method. We analyzed the data using the Korean Standard Classification of Diseases and Causes of Death 7th revision (KCD-7) code, which is based on the International Statistical Classification of Diseases and Related Health Problems 10th revision, to reveal the most reliable diagnosis for ICU admission and mortality.

We identified all the patients with at least one or more events of ICU admission between January 1st, 2006 and December 31st, 2015. During this period, the subspecialty certification for critical care was started (in 2008), and a major epidemiological event, the 2009 influenza A virus subtype H1N1 pandemic, occurred, with a high ICU mortality of 42.6% [16]. There was no other considerable change in the nationwide policies for ICU admissions, management, and discharges. As our target population was critically ill patients in ICUs, we excluded the patients who were admitted to ICUs without critical illnesses, such as for close observation after procedure or surgery and only for short term intermediate care. Operational definition of critically ill patients was adults aged ≥ 18 years who either stayed at least 48 hours or expired in the ICU during admission to a tertiary care center or general hospital. Operational definition of pulmonary disorder group, the test group, was critically ill patients with a KCD-7 code for a respiratory disease as the main diagnosis at the time of admission to the ICU (Supplementary Table 1). Operational definition of non-pulmonary disorder group, the control group, was critically ill patients who had some other medical illness as the main diagnosis at the time of entering the ICU. Patients who were admitted to the ICU after surgery were not included in the control group.

We evaluated the critically ill patients’ socio-demographic characteristics such as age, sex, residence in urban or rural areas, particulate matter 10 mm or less in diameter (PM10) or ozone concentration change from the prior year, the number of total or tertiary healthcare units per area (10 km²), the decile of income, Charlson comorbidity index (CCI), and severity of disability. Data regarding air pollution was acquired from the homepage of Air Korea (https://www.airkorea.or.kr/index) and has been presented in the Supplementary Table 2. Based on the area in which the patients lived, we calculated the average number of healthcare units per 10 km² for each region. The CCI was calculated using each patient’s claim documents which recorded their comorbidities and used to categorize them into four groups. Severity of disability was determined according to the criteria for disability (grades 1 to 6) set by the Korean Ministry of Health and Welfare with grades 1 to 2 indicating severe disability, grades 3 to 6 indicating mild disability, and others indicating lack of disability [17]. To evaluate the trends of healthcare resource usage, we analyzed the course of admission to hospital, ICU patient volume, number of days spent in the hospital, number of days spent in the ICU, use of endotracheal tube and mechanical ventilator, number of days of mechanical ventilator usage, and total cost during ICU stay. The course of admission to hospital was defined as where the decision on hospitalization was made before admission. The ICU patient volume was defined as the number of patients who were admitted to an ICU per healthcare unit, and divided into tertile, “High,” Middle,” and “Low.” We assessed clinical outcomes such as the result of ICU discharge, the time to ICU re-admission and mortality, the long-term incidence rate of ICU re-admission and mortality, and most probable major cause of overall mortality. If a patient had multiple admissions to ICUs, the first episode was regarded as the major event during the study period. The cases of intra- or inter-hospital transfer between ICUs were identified using individual identifier numbers and claim bill numbers and considered one episode of ICU admission.

The primary endpoint was to elucidate current epidemiologic features, trends of healthcare resource usage, and clinical outcomes of critically ill patients who were admitted to ICU for various pulmonary disorders compared with those admitted for other disorders.

For comparisons between the pulmonary and non-pulmonary disorder groups, categorical variables were evaluated using percentage difference and 95% confidence interval (CI) between proportions. Continuous variables were analyzed using mean difference and 95% CI. Correlation analysis was conducted to evaluate Pearson’s correlation coefficient. Kaplan-Meier curve with log-rank test and hazard ratio (HR) with 95% CI were used to assess differences in cumulative rates of ICU re-admission and mortality. Poisson regression analysis was performed to compare the adjusted incidence rate ratio of ICU re-admission and mortality between the two groups. Subgroup analysis was conducted according to the various entities of pulmonary disorders to determine the number of patients, total cost during ICU stay, in-hospital mortality, and re-admission to ICU for each disorder. We analyzed the relationship of the baseline characteristics and the usage of healthcare resources with in-hospital survival. After testing the proportional hazards assumption, univariable and multivariable analyses with Cox proportional hazards model were conducted to reveal the adjusted impact of pulmonary disorder on the in-hospital mortality in the critically ill patients. Multicollinearity was measured by variance inflation factor and a cut-off value of 4 was regarded as high. All the analyses were conducted using R for Statistical Computing software, version 3.3.3 (R Core Team [2017], Vienna, Austria).

The Institutional Review Board Committee of Seoul National University Bundang Hospital examined and approved the protocol of this study and exempted it from the need for informed consent for access to electronic medical records (Institutional Review Board number: X-1706-405-903).

Among the 997,173 patients in the NSC database, 49,414 (5.0%) had at least one episode of ICU admission from 2006 to 2015, and 12,983 (1.3%) in 383 ICUs were eligible to be categorized as critically ill, as per the operational definition (Fig. 1). Among the critically ill patients, we found that 3,875 (30.0%) were eligible to be included in the pulmonary disorder group and 3,980 (30.7%) received MV. The frequency of patients receiving MV was approximately 2.6 times higher in the pulmonary disorder group than in the control group. Annual proportion of the pulmonary disorder group among the critically ill patients had increased during the 10-year follow-up (compound annual growth rate = 6.3%), with 28.7% being the lowest value, in 2006, and 55.4% being the highest value, in 2013 (Fig. 2).

The patients in the pulmonary disorder group were more likely to be old and male (Table 1). There was no significant difference according to the residence area. More patients in the pulmonary disorder group were found in the areas where the ozone concentration had increased compared with the previous year. The number of total or tertiary healthcare units per area was not significantly different between the two groups. The number of patients using ICU resources for management of critical illness linearly varied according to the incomes. Patients with higher incomes occupied a larger proportion of the critically ill group (Supplementary Fig. 1). In terms of underlying conditions, more patients in the pulmonary disorder group showed a higher CCI score and higher grade of disability than those shown by patients in the non-pulmonary disorder group.

There was no difference in the course of admission to hospital or ICU volume per healthcare unit between the pulmonary disorder group and non-pulmonary disorder group (Table 2). Patients in the pulmonary disorder group spent 6.2 days more in the hospital and 4.5 days more in the ICU than those in the non-pulmonary disorder group. Patients in the pulmonary disorder group were approximately 34.3% more likely to use an endotracheal tube and a mechanical ventilator for 2.6 days more than those in the non-pulmonary disorder group. In the pulmonary disorder group, the total cost during ICU stay and number of patients in the third tertile of ICU cost, who needed more hospital expenditure than other patients, was higher than for those in the non-pulmonary group. Among the pulmonary disorder group, patients with empyema, neoplastic disorder, and ARDS needed higher expenditure than other patients (Supplementary Table 3).

Patients in the pulmonary disorder group were more likely to die in hospital, whereas a higher proportion of patients in the non-pulmonary disorder group were more likely to be discharged to home or the previous healthcare unit (Table 3). More patients in the pulmonary disorder group were re-admitted to ICU after discharge. In the Kaplan-Meier curves used to estimate time to overall death and re-admission to ICU from the initial ICU admission, the pulmonary disorder group showed significantly lower survival rate (log-rank p < 0.001; HR, 1.78; 95% CI, 1.69 to 1.87) and higher rate of re-admission to ICU (log-rank p < 0.001; HR, 2.07; 95% CI, 1.90 to 2.25) (Fig. 3). The ratios of the unadjusted and adjusted incidence rate in terms of both overall mortality and re-admission to ICU were higher in the pulmonary disorder group (Table 4). The analysis for the most probable major cause of overall mortality revealed that, although majority of the patients in the pulmonary disorder group died of pulmonary disorder (44.9%), more than half died of other disorders, such as cardiovascular disorder (19.5%) (Supplementary Table 4).

The subgroup analysis according to entities of pulmonary disorder revealed that half of the total patients in the pulmonary disorder group were admitted to ICU for treating pneumonia, while in-hospital mortality rate was higher for interstitial lung disease (ILD) and ARDS than for other disorders (Supplementary Table 3). The two disorders with the highest rate of re-admission to ICU were pleural effusion and pneumonitis. We found three different patterns of survival curve among the various pulmonary disorders (Fig. 4). One was pattern A, in which the survival rate drops sharply at the initial stage of the ICU admission, such as that for ARDS, respiratory failure, pneumonia, pneumonitis, ILD, and neoplastic disorder. Another is pattern B, in which survival rate decreases gradually over time, such as that for pneumothorax, pleural effusion, and chronic airway disease. Pattern C is one which does not show difference in survival rate compared with the non-pulmonary disorder group, such as that for pulmonary embolism, lung abscess, and empyema.

Age, sex, air pollution, income, CCI score, disability, pulmonary disorder as the main cause of ICU admission, course of admission to hospital, and ICU patient volume were associated with in-hospital mortality (Supplementary Tables 5 and 6). The cases of critically ill patients, mainly with pulmonary disorder, were significantly related to in-hospital mortality in the unadjusted and adjusted regression analyses with Cox proportional hazards model (Table 5). In addition, higher age, male sex, decreased concentration of PM10, disability, admission to hospital via outpatient clinic, and low ICU patient volume were significantly associated with a higher risk of in-hospital mortality, after adjustment. Risk factors of in-hospital mortality were different between the pulmonary disorder group and non-pulmonary disorder group (Supplementary Table 7). Patients in the pulmonary disorder group who lived in an urban area had a higher risk of in-hospital mortality (HR, 1.12; 95% CI, 1.00 to 1.26). In contrast, disability, admission course, and ICU patient volume were not associated with in-hospital mortality in the pulmonary disorder group.