This was a retrospective cohort study that used routinely collected nationwide inpatient administrative data in Japan to compare the utilization of critical care services during the COVID-19 pandemic to that before the pandemic. The Institutional Review Board of The University of Tokyo approved this study (approval number, 3501-3). No information allowing the identification of individual patients, hospitals, or physicians was obtained, and the requirement for informed consent was waived because of the anonymous nature of the data.

We used the Japanese Diagnosis Procedure Combination inpatient database (DPC database), which contains discharge abstracts and administrative claims data from more than 1500 participating public and private acute-care hospitals in Japan that voluntarily contribute to the database [12]. The database includes the following patient-level data for all hospitalizations: age, sex, route of admission, cognitive function before admission, admission type, diagnoses recorded using International Classification of Diseases, Tenth Revision (ICD-10) codes, daily procedures recorded using Japanese medical procedure codes, daily drug administrations, and discharge status. In a previous study examining the validity of the recorded procedures and diagnoses, the sensitivity and specificity of procedures exceeded 90%, whereas the sensitivity and specificity of the primary diagnoses were 78.9% and 93.2%, respectively [13].

We also used the Survey of Medical Institutions 2019 and 2020, which include the facilities’ information and statistics as of July 1, 2019 and July 1, 2020, respectively [14]. We combined this information with the data in the Japanese Diagnosis Procedure Combination inpatient database using a specific hospital identifier. The Survey of Medical Institutions included academic hospitals, tertiary emergency hospitals, types of wards [e.g., general, ICU, or high-dependency care units (HDUs)], and the number of licensed hospital beds in each ward.

In Japan, the first COVID-19-related ICU admission case was confirmed on February 9, 2020 [7]. Therefore, in this study, February 9, 2020, was used as the breakpoint separating the periods before and during COVID-19 into two 1-year periods [15]: before the COVID-19 pandemic from February 9, 2019, to February 8, 2020, and during the COVID-19 pandemic from February 9, 2020, through February 8, 2021.

We examined two separate cohorts of patients in ICUs, as well as HDUs, which are potential alternatives to ICUs [16, 17]. For the ICU patients, we included all patients admitted to an ICU between February 9, 2019, and February 8, 2021, from hospitals participating in the DPC database that also had records in the Survey of Medical Institutions and had at least one ICU bed reported in the Survey of Medical Institutions of both 2019 and 2020. In this study, an ICU was defined as a separate unit providing critical care services with at least one physician on site 24 h per day, around-the-clock nursing, the equipment necessary to care for critically ill patients, and a nurse-to-patient ratio of > 1:2 [18]. As for HDU patients, we included all patients admitted to an HDU between February 9, 2019, and February 8, 2021, from hospitals participating in the DPC database that also had records in the Survey of Medical Institutions and had at least one HDU bed in the Survey of Medical Institutions of both 2019 and 2020. The definition of an HDU in this study was almost the same as an ICU, except that the required nurse-to-patient ratio was 1:3, 1:4, or 1:5. The Japanese medical procedure codes to define ICU and HDU are listed in Additional file 1: Table S1. We did not include patients admitted to the neonatal ICU or obstetric ICU in this study.

This study presents hospital characteristics before and during the COVID-19 pandemic. Hospital characteristics were based on the Survey of Medical Institutions 2019 before the COVID-19 pandemic and the Survey of Medical Institutions 2020 during the COVID-19 pandemic. We then compared the patient characteristics and outcomes before and during the COVID-19 pandemic. Due to the large sample size in this study, the patient characteristics and outcomes were compared using standardized mean differences, with an absolute standardized mean difference of ≤ 10% denoting a negligible imbalance between the two groups [19]. Patient characteristics included age, sex, route of admission, cognitive function before admission, admission type, main etiologies for admission, and organ support at least once during ICU/HDU stay. Admission type was categorized as elective surgery, emergency surgery, or nonsurgical/acute medical problem. We defined patients who were admitted to the ICU/HDU on the day of elective or emergency surgery with general anesthesia as elective or emergency surgery patients. Main etiologies for admission were defined by the ICD-10 codes in the admission-precipitating diagnosis as follows: circulatory diseases (ICD-10 codes: I00-I99), neoplasms and diseases of the blood (C00-D89), injury, poisoning and external causes (S00-T98), abdominal diseases (K00-K93), respiratory diseases (J00-J99), COVID-19 (U071), and others. We also performed a descriptive epidemiology of COVID-19 patients. Continuous variables are presented as means and standard deviation (SD) or as medians and interquartile ranges (IQRs), as appropriate. Categorical variables are presented as frequencies and percentages. All analyses were performed using Stata/MP, Version 16.0 (StataCorp, College Station, TX, USA).

We calculated the daily occupancy of ICUs/HDUs by all patients in total, patients with invasive mechanical ventilation (IMV), and patients with extracorporeal membrane oxygenation (ECMO). We defined occupancy as the total number of relevant reimbursements in the cohort on a given day divided by the total number of licensed ICU/HDU beds in the participating hospitals. Change in occupancy before and during the COVID-19 pandemic was evaluated using patient-level segmented linear regression with interrupted time-series analysis [20]. The equation for the interrupted time-series analysis was as follows:

where \(_\) is the occupancy, \(T\) is the month since the beginning of the study period (coded from 1 to 24), and \(_\) is a dummy variable indicating before or during the COVID-19 pandemic (coded 0 or 1). In this model, \(_\) represents the level change immediately during the COVID-19 pandemic and \(_\) represents the trend change during the COVID-19 pandemic compared to the baseline trend before the COVID-19 pandemic. As the sensitivity analyses for controlling for seasonality, we performed the interrupted time-series analyses stratified by the calendar month [20].