Data source
The present study was a retrospective cohort design using the administrative claims database of inpatients and laboratory test values in Japan provided by Medical Data Vision Co., Ltd. (Tokyo, Japan). This database includes approximately 190,000 episodes of admission to the ICU between April 2008 and September 2021 at Japanese acute care hospitals. Administrative data were constructed under the Diagnosis Procedure Combination (DPC) payment system [17].
The present study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of Hitachi General Hospital (2020-131). The requirement for informed consent was waived due to its retrospective design and the use of anonymized data.
Study population
We identified patients admitted to and treated in the ICU for more than 3 consecutive days. Exclusion criteria were as follows: (1) patients younger than 18 years old, (2) pregnant women, (3) patients admitted with a diagnosis of gastrointestinal bleeding, bowel obstruction, or bowel ischemia (defined as ICD-10 codes K25–29, K55–57, and K625), (4) patients who had undergone gastrointestinal surgery, (5) patients who were discharged or died between day 0 (the day of admission) and day 7, (6) patients who did not receive EN therapy between days 0 and 7. In cases where a patient had two or more episodes, only the initial episode was included in the analysis.
Patients who received EN therapy during the first three days after hospitalization (day 0, 1, or 2) were assigned to the EEN group, and those who did not receive EN during the first three days, but then received it between days 3 and 7 were assigned to the late enteral nutrition (LEN) group.
Covariates
We derived the following data as covariates from our database: age (categorized as <50, 50–59, 60–69, 70–79, or >80 years old), sex, body mass index (categorized as <18.5, 18.5–25.0, 25.0–30.0, or >30.0 kg/m2), smoking status (non-smoker or current/ex-smoker), the Charlson Comorbidity Index defined by ICD-10 codes [18] (categorized as 0–1, 2–3, or ≥4 points), ambulance use, emergency admission, surgery under general anesthesia (emergent or elective), diagnoses for ICU admission, sequential organ failure assessment (SOFA) scores (total score and scores for each component), the catecholamine index, mechanical ventilation, extracorporeal membrane oxygenation (ECMO), renal replacement therapy (continuous or intermittent), the administration of a systemic steroid, and laboratory data (the lymphocyte count, albumin, and CRP). Patients who received treatment (mechanical ventilation, ECMO, renal replacement therapy, and the administration of a systemic steroid) on day 0 or 1 were defined as having the corresponding therapy. Since information on oxygenation (PaO2 and FiO2) was not included in our database, the respiratory component of SOFA was defined as follows: patients who did not receive oxygen therapy as 0, patients who received oxygen therapy as 1, patients who received non-invasive positive pressure ventilation (including a high-flow nasal cannula) as 2, patients who received mechanical ventilation as 3, and patients who received ECMO therapy as 4. Our database also did not have records on vital signs or time-specific information on hospitalization and treatments; therefore, we calculated the dose of catecholamines based on prescription records on day 1, and defined the cardiovascular component of the SOFA score for patients who did not receive catecholamines as 0. The dose of catecholamines was converted to gamma (µg per kg per minute) units, and the catecholamine index was calculated by the sum of dopamine, dobutamine, and noradrenaline multiplied by 100 and adrenaline multiplied by 100. Diagnoses for ICU admission were categorized as sepsis, cardiovascular, pulmonary, metabolic, neurology, trauma, digestive, and others. The diagnostic codes of sepsis were based on a previous study using the DPC database [19]. Patients who had the code of sepsis were categorized as the sepsis group. Patients with the relevant codes of cardiovascular (ICD-10 code of I00–99), pulmonary (J00–99), metabolic (E00–99), neurology (I60–69 and G00–99), trauma (S00–99, T00–19, T33–88, V00–99, W00–99, X00–99, and Y00–09), and digestive (K20–93) were categorized accordingly. The category of trauma included trauma, poisoning, and external causes. Patients who did not have these codes were categorized to the others group. Laboratory examinations as covariates used the worst values obtained during the first two days, namely, the highest value for CRP and the lowest values for the lymphocyte count and albumin.
Outcomes
The primary outcome of interest was a composite of the incidence of PICS on day 14 and 14-day mortality. Although there are no widely used clinical criteria for PICS, patients who met at least two of the following conditions and were hospitalized >14 days were defined as having PICS [10]: CRP >2.0 mg/dL, albumin <3.0 g/dL, and a lymphocyte count <800/μL. The date of these laboratory examinations was referred to the nearest day to day 14 within days 11–17 (if laboratory data were available for both days 13 and 15, data on day 15 were used). Patients who were still hospitalized, but did not have these laboratory values between days 11 and 17 were not considered to have PICS. Secondary outcomes were a composite of 28-day mortality or the incidence of PICS on day 28, the Barthel index at discharge, in-hospital mortality, laboratory data on days 14 and 28, and changes in laboratory values. The Barthel score of patients who died during hospitalization was regarded as 0 [20].
Statistical analysis
We applied the multiple imputation method for missing values of covariates. Total of 20 imputed datasets were created and the estimates and standard errors were combined according to the Rubin’s rule. To adjust for confounders, we conducted a propensity score analysis. The propensity score for assignment to the EEN group was estimated using a multivariate logistic regression model. To ensure that patients with the same diagnosis for ICU admission were matched between the EEN and LEN groups, we initially divided eligible patients into subgroups based on the diagnosis for ICU admission, and then estimated the propensity score within each subgroup [21]. Matching was performed separately for each subgroup, and patients after matching were aggregated for the main analysis. We calculated the C-statistic for the aggregated population to evaluate the performance of the discrimination of propensity scores. One-to-one nearest-neighbor matching without replacement was performed with a caliper of 20% of the standard deviation for propensity scores. Balances between the EEN and LEN groups before and after matching were described using the absolute standardized mean difference (ASD).
Risk differences and the 95% confidence intervals (CIs) of binomial outcomes were estimated, followed by the null hypothesis. Continuous outcomes (the Barthel index at discharge and laboratory values) were compared using the Wilcoxon rank-sum test. The incidence of PICS and mortality on days 14 and 28 were also compared between the EEN and LEN groups for each subgroup of the diagnosis for ICU submission.
We performed an overlap weighting analysis for a sensitivity analysis[22-24]. Overlap weighting is one of the propensity score weighting methods that emphasizes the target population with the most overlap in the observed characteristics between a treatment group and control group. Under the condition that the propensity score is estimated by a logistic regression, overlap weighting achieves an exact balance on the mean of all covariates. All P-values were two-tailed; P-values <0.05 were considered to be significant. All statistical analyses were performed using R (version 4.2.1, R Foundation for Statistical Computing, Vienna, Austria). “Mice”, “MatchIt”, “PSweight”, and the “fmsb” package were used for the multiple imputation method, propensity score matching, propensity score weighting, and estimating risk differences, respectively.