Study design, setting, and ethical considerations

This population-based cohort study followed the Strengthening the Reporting of Observational Studies in Epidemiology guidelines [12]. The Institutional Review Board (IRB) of Seoul National University Bundang Hospital waived the requirement for IRB approval of this study because of its use of public data available to all researchers (IRB number: X-2205-758-901). The requirement for informed consent was waived by the IRB because the study was based on a retrospective analysis of anonymized data.

KDCA-COVID19-NHIS cohort (data source)

We used data from the KDCA and National Health Insurance Service (NHIS). The KDCA-COVID19-NHIS cohort was generated for academic purposes through the collaboration of the KCDA and NHIS in South Korea. The KCDA initially extracted data regarding patients confirmed with COVID-19 by polymerase chain reaction (PCR) test from October 8, 2020, to December 31, 2021. The data from the KCDA contained information regarding age, sex, date of COVID-19 diagnosis using the PCR test, date of death, date of vaccination (1st, 2nd, and 3rd), and type of infection route. The type of infection route was classified into six groups: (1) inflow from foreign countries, (2) contact with person-related inflow from foreign countries, (3) outbreak in hospitals or nursing care centers, (4) outbreak in local communities, (5) contact with a confirmed patient, and (6) unknown. In addition to data from the KDCA, the NHIS extracted information regarding demographic and socioeconomic status, all disease diagnoses using the International Classification of Diseases (ICD)-10 codes, and prescription information of any procedures or drugs until March 31, 2022. Both KDCA and NHIS approved the data sharing for this study (research grant number: KDCA-NHIS-2022-1-489).

Study population

This study included adult patients (≥ 18 years old) confirmed with COVID-19 by PCR test who were admitted to the ICU with a main diagnosis of COVID-19. To focus on the last episode of ICU admission in a patient, in multiple (≥ 2) admission cases the earlier ICU admissions were excluded. These multiple ICU admissions included transfer from hospital to hospital. For example, if a patient was transferred from the ICU of a general hospital to the ICU of a tertiary general hospital, only the last ICU admission at the tertiary general hospital would be included in this study.

Trained intensivist system in South Korea

In this study, a trained intensivist was defined as an individual who was certified by the Korean Society of Critical Care Medicine after completing a fellowship training program in intensive care medicine. The fellowship program includes a compulsory one-year immersion in the ICU at the training hospital of the intensivist, an examination, an interview, the presentation of an abstract at the annual congress of the Korean Society of Critical Care Medicine, and the publication of an original article in the official publication of the society. Doctors specializing in internal medicine, anesthesiology and pain medicine, pediatrics, neurology, neurosurgery, emergency medicine, general surgery, and thoracic surgery are eligible to apply for the fellowship training course. As reported in previous studies [10, 13], South Korea implemented a special payment system for trained intensivist coverage from August 2015. This special payment system applies to medical centers that hire trained intensivists for the management of ICU patients. Trained intensivists must work only in the ICU (not in the ward) for at least ≥ 8 h/day and ≥ 5 days/week. Moreover, there should be at least one intensivist per ICU in a hospital to receive special payments.

In the present study, critically ill COVID-19 patients admitted to ICUs that hired trained intensivists were assigned to the intensivist group, whereas all other critically ill COVID-19 patients were assigned to the non-intensivist group. Critically ill COVID-19 patients who were admitted to an ICU that did not have a trained intensivist were managed by doctors other than trained intensivists.

ICU management during the COVID-19 pandemic in South Korea

As all patients admitted to the ICU must be registered in the NHIS database using ICU prescription codes, there were no missing cases. Since the start of the COVID-19 pandemic, from 2020 to the present, the Central Disease Control Headquarters in South Korea have overseen the national-level countermeasures against infectious diseases [14]. First, general ICUs (i.e., those not dedicated to patients with COVID-19) were required to be designated as ICUs for COVID-19 patients; this decision was based on the number of critically ill patients and the increasing social isolation level in the community. Second, a system for transferring and treating critically ill patients to other regions, according to the number of COVID-19 patients in certain areas with available ICU beds, was established nationwide. Through this strategy, the Central Disease Control Headquarters attempted to ensure that there would not be a lack of ICU beds for critically ill COVID-19 patients during the pandemic. Additionally, to prevent a shortage of medical staff to care for such patients, doctors in the military, public health doctors, and resting nurses were assigned to ICUs as required.

Although ICUs were newly and rapidly constructed to take care of critically ill COVID-19 patients, they still had to meet certain legal requirements. First, emergency resuscitation devices, intubation devices, mechanical ventilators, defibrillators, electrocardiograms, and respiratory function measuring devices had to be available for use at all times. Second, they were required to include a dedicated space that was larger than the general ward, with at least 10 square meters per patient. Third, they were required to always have a dedicated doctor and at least one nurse for every two patients. Fourth, as patients may die during power outages, they needed to be equipped with uninterruptible power supply. The critically ill COVID-19 patients who were admitted to the ICU in this study included patients who were admitted to both existing and newly constructed ICUs in South Korea.

Study outcomes

The primary outcome of this study was in-hospital mortality, defined as death during hospitalization-associated COVID-19.

Included covariates

Age and sex were collected as demographic information. To determine the socioeconomic status of critically ill COVID-19 patients, employment status, the household income level, and residence at the time of admission due to COVID-19 infection were collected. The household income level was evaluated differently for employee-insured and self-employed insured individuals. For employee-insured individuals, the insurance premium was determined solely based on their income. On the other hand, for self-employed insured individuals, the insurance premium was determined based on their income, property, living standards, and rate of participation in economic activities. Of note, those who could not afford insurance premiums or had difficulty financially supporting themselves were included in the medical aid program instead. In addition to the medical aid program group, the household income level was classified into four quartile ratios (from Q1, the lowest, to Q4, the highest). The residences of critically ill COVID-19 patients were classified as urban (Seoul and other metropolitan cities) or rural (all other areas). Data regarding hospital level (A, B, C, and D; detailed information is presented in the statistical methodology section) and total case volume of ICU admissions due to COVID-19 during the study period were collected. All patients were classified into four groups according to the hospital where they were admitted to the ICU using quartile ratios (Q1: 0–150, Q2: 151–257, Q3: 258–408, and Q4: ≥ 409). To reflect the severity of critically ill patients with COVID-19, we utilized the WHO clinical progression scale without the P/F ratio [15]. Treatment information from the day of ICU admission or the day after ICU admission was used to indicate the initial severity of such patients. The clinical progression scale used in this study classified patients as follows: 1 point (no oxygen therapy), 2 points (oxygen by mask or nasal prongs), 3 points (oxygen by non-invasive ventilation or high flow nasal cannula [HFNC]), 4 points (intubation and mechanical ventilation), 5 points (mechanical ventilation with vasopressor use), and 6 points (mechanical ventilation with vasopressor use, dialysis, or extracorporeal membrane oxygenation [ECMO]). The vasopressors used included norepinephrine, epinephrine, dopamine, dobutamine, and vasopressin. In addition, ICD-10 coding-based acute respiratory distress syndrome (ARDS, J80) diagnosis during hospitalization due to COVID-19 was collected as a covariate. With regard to the comorbid status of patients, the Charlson comorbidity index (CCI) and disability at ICU admission were collected. CCI scores were calculated using ICD-10 codes—2020–2021 which were registered in the NHIS database (Additional file 1: Table S1). All individuals with any disabilities should be registered in the NHIS database in order to receive various social welfare benefits. Disabilities were categorized into the following 15 types: physical and brain lesion disabilities; visual disturbances; hearing and speech disabilities; autism; intellectual, mental, renal, heart, and respiratory disorders; hepatopathies; facial disfigurements; intestinal and urinary fistulae; and epilepsy. The degree of each disability was divided into the following two groups according to the severity criteria: severe disability, and mild-to-moderate disability. Disability was diagnosed and determined according to the laws of a specialty physician in each field. The most important criterion for determining disability was whether it interfered with maintaining daily life.

Statistical methodology

The Mann–Whitney U test for continuous variables and the Chi-square test for categorical variables were used to compare the clinicopathological characteristics of patients between the intensivist and non-intensivist groups. We first selected covariates based on patient factors that might affect the prognosis of patients with COVID-19, such as age, sex, CCI, disability, and history of vaccination. We also selected covariates related to socioeconomic status factors, as socioeconomic status can have an impact on COVID-19 patients’ outcomes [16]. The type of infection route and WHO clinical progression scale were also collected because the characteristics and prognosis of COVID-19 infection may differ based on these factors [17]. Lastly, total case volume and hospital levels were collected to reflect the capacity of each hospital where critically ill patients with COVID-19 were admitted to the ICU.

Among the covariates, hospital levels for the critically ill COVID-19 patients were identified using a hierarchical approach. For hierarchical cluster analysis, agglomerative clustering was performed using hospital-related variables, such as the type of hospital (general hospital or long-term care facility), total number of doctors, specialist doctors, nurses, pharmacists, hospital beds, operating room beds, adult ICU beds, and emergency room beds. Four groups were created based on the results of the hierarchical clustering analysis; the characteristics of the four hospital groups are presented in Additional file 2: Table S2.

We constructed a multivariate logistic regression model for in-hospital mortality among critically ill COVID-19 patients. The selected covariates were included in the model for multivariable adjustment, and the results are presented as odds ratios (OR) with 95% confidence intervals (CI). We also performed subgroup analyses according to the WHO clinical progression scale, ARDS diagnosis, and hospital level. Hosmer–Lemeshow statistics were used to confirm that the goodness of fit in the multivariable model was appropriate. There was no issue regarding multicollinearity within the variables with the criterion of variance inflation factors < 2.0. All statistical analyses were performed using IBM SPSS Statistics for Windows (version 25.0; IBM Corp., Armonk, NY, USA). A P-value of < 0.05 was considered statistically significant.