Relationship between the Geriatric Nutrition Risk Index and the Prognosis of Severe Coronavirus Disease 2019 in Korea

1. Study design and population

This retrospective observational study collected data from patients aged 19 years or older who were diagnosed with severe COVID-19 and respiratory failure. Severe COVID-19 was classified according to the guidelines of the Korea Centers for Disease Control and Prevention (K-CDC), which stipulate that severity requires at least one of the following forms of respiratory support: high-flow nasal oxygen therapy, non-invasive MV, invasive MV, or ECMO. These criteria align with K-CDC standards for assigning dedicated intensive care unit (ICU) beds to severe COVID-19 cases [14]. The classification criteria were applied upon admission to the dedicated ICU beds at participating hospitals. Patients with missing data on height, weight, or albumin levels were excluded from the analysis (Figure 1). Data collection spanned 22 tertiary and university-affiliated hospitals from January 2020 to August 2021. The Institutional Review Board (IRB) of each participating hospital approved this study, including the IRB of Pusan National University Yangsan Hospital (Yangsan, South Korea; approval number: 04-2021-042; approval date: October 7, 2021). Given the study’s observational nature and minimal risk, the requirement for informed consent was waived.

Fig. 1.

Flowchart of the patient selection process. Of the 1,114 patients with severe coronavirus disease 2019 (COVID-19), 26 with missing data on height (n=21), weight (n=2), and albumin levels (n=3) were excluded. GNRI: Geriatric Nutritional Risk Index.

2. COVID-19 diagnosis and data collection

COVID-19 was confirmed via a real-time reverse transcription polymerase chain reaction assay, detecting the presence of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) in upper respiratory tract specimens. The collected data encompassed demographic details, comorbidities, disease severity (assessed using the Sequential Organ Failure Assessment [SOFA] score and other baseline hemodynamic and laboratory variables), the source and type of infection (both community and hospital-acquired), treatment regimens, resource utilization, and outcome data, including ICU and hospital mortality rates. Patients were identified as having hospital-acquired COVID-19 infections if they tested positive for COVID-19 between 5 and 7 days post-hospitalization, aligning with the typical incubation period. Ventilator-free days (VFDs) during the initial 30 days post-MV initiation were calculated. VFDs represent the number of days a patient remained alive without the need for MV within this period. A VFD value of zero was assigned to patients who died within the initial 28 days. The Clinical Frailty Scale (CFS) was utilized to evaluate functional status prior to hospitalization and upon discharge.

3. Nutritional status assessment

The ideal body weight was calculated using the Lorentz formulas as follows: ideal body weight=height (cm)−100−[(height−150)/4] for men and ideal body weight=height (cm)−100−[(height−150)/2] for women [15]. The GNRI was determined as follows: GNRI=[1.489×serum albumin (g/L)+41.7×(present weight/ideal body weight)] [11]. If the patient’s current weight exceeded the ideal weight, the present weight-to-ideal body weight ratio was set to 1.

4. Group division and outcome measures

Patients were stratified into no-risk (GNRI >98) and risk (GNRI ≤98) groups based upon their GNRI assessed at admission to a dedicated ICU bed. This stratification was consistently applied to all patients, including those diagnosed with hospital-acquired COVID-19. Hospital mortality was identified as the primary outcome, while secondary outcomes included changes in ICU mortality, MV duration, and CFS scores at discharge.

5. Statistical analysis

Continuous variables are expressed as either the mean±standard deviation or the median and interquartile range (IQR), based on their distribution. Analyses were performed using the Student’s t-test or the Mann–Whitney U test as appropriate. Categorical variables were reported as frequencies and percentages and analyzed using the chi-square or Fisher’s exact test as appropriate. Propensity score matching (PSM) was utilized to emulate random assignment in both the total cohort and the elderly subset by estimating the likelihood of classification into the risk (GNRI ≤98) group. To ensure balanced comparison between groups, PSM was conducted using the nearest neighbor matching method with a 1:1 ratio. The optimal caliper width was set at 0.2 times the standard deviation of the log-transformed propensity scores, an empirically common threshold to minimize differences between matched pairs while maintaining a sufficiently large number of matched units. Following matching, covariate balance was assessed using the standardized mean difference (SMD), which measures the extent of mean differences between groups relative to the combined standard deviation. Covariates were considered well-balanced if the SMD was below 0.1, indicating negligible imbalance. The matching process, including SMD values for covariates before and after matching, is detailed in Table 1.

Table 1.

Clinical characteristics of the total cohort

The propensity score model incorporated variables that reflected pre-admission characteristics (e.g., age), admission data (e.g., CFS score, site of infection acquisition, and comorbidities such as cardiovascular disease, chronic lung disease, and chronic neurological disease), and post-admission severity markers (e.g., the use of HFNC therapy or MV). This comprehensive approach was adopted to balance baseline characteristics and adjust for confounding variables associated with nutritional risk and mortality [16]. A Cox proportional hazards regression model analyzed factors influencing hospital mortality. To identify independent predictors of hospital mortality, univariate analyses were conducted on all variables presented in Table 1. Variables demonstrating a p<0.05 in the univariate analysis were selected for inclusion in the multivariate regression model. We implemented a stepwise backward selection method for the multivariate analysis to systematically identify significant variables, minimizing overfitting risk. This strategy was employed to balance the exploratory nature of the study with the need to identify robust predictors of mortality. Variables retained in the final model were assessed for their biological plausibility and relevance to severe COVID-19 outcomes to confirm their suitability. Risk-adjusted survival curves were generated using the proportional hazards model combined with the mean of the covariate method. Statistical analyses were carried out using SPSS version 27.0 (IBM Corporation, Armonk, NY, USA) and R version 3.6.3 (R Foundation for Statistical Computing, Vienna, Austria). A p<0.05 was deemed significant.

1. Patient characteristics

Overall, 1,114 patients with severe COVID-19 were enrolled in the study. After the exclusion of 26 patients due to incomplete height, weight, and albumin data, 1,088 patients were included in the final analysis (Figure 1). Baseline characteristics at ICU admission are displayed in Table 1. Among the included patients, 463 (42.6%) received HFNC therapy alone, while 498 (45.8%) required MV support throughout the duration following the COVID-19 diagnosis. Additionally, 127 patients (11.7%) received both MV and ECMO support. Based on GNRI scores at admission, the patients were categorized into the no-risk group (GNRI >98, n=446) and the risk group (GNRI ≤98, n=642, 59%). The latter group was significantly older (mean age 70.5 years vs. 63.2 years, p<0.001), had a higher percentage of elderly individuals (70.1% vs. 49.3%, p<0.001), lower body mass index (BMI) (23.1 vs. 27.3, p<0.001), and higher baseline CFS scores (median 3 [IQR, 2 to 4] vs. 3 [IQR, 2 to 3], p<0.001) (Table 1). This group also exhibited higher rates of infections in nursing homes, nursing hospitals, and hospitals (p<0.001), as well as higher prevalence of comorbidities such as cardiovascular disease, chronic lung disease, chronic neurological disease, and solid malignancy. In terms of clinical management, the malnutrition risk group showed lower utilization of HFNC therapy and greater reliance on MV, indicating more severe respiratory compromise. After PSM, no significant differences remained between the groups, except for BMI and GNRI. The SMD for most covariates was reduced to below 0.1 (10%), indicating that a covariate balance was largely achieved.

1) Primary outcomes

The primary outcome was hospital mortality. The risk group exhibited significantly higher hospital mortality rates compared to the no-risk group (30.7% vs. 19.5%, p<0.001) (Table 2). Despite PSM, hospital mortality remained significantly higher in the risk group (34.3% vs. 19.4%, p<0.001).

Table 2.

Clinical outcomes

2. Cox regression for hospital mortality

Before propensity matching, GNRI did not exhibit a significant relationship with hospital mortality in both univariate and multivariate Cox regression analyses (Supplementary Table S1). Following propensity adjustment, belonging to the risk group (GNRI ≤98) was significantly associated with increased hospital mortality in both the univariate (hazard ratio [HR], 1.55; 95% confidence interval [CI], 1.17 to 2.05; p=0.002) and multivariate Cox regression analyses (HR, 1.64; 95% CI, 1.24 to 2.18; p=0.001) (Table 3). Both before and after propensity adjusting, a lower GNRI (≤98) correlated with higher mortality rates (Figure 2A: total cohort, χ²=5.17, p=0.023; Figure 2B: propensity-matched cohort, χ²=9.71, p=0.002).

Table 3.

Hazard ratio for mortality in the propensity-matched cohort determined using the Cox proportional hazards model

Fig. 2.

Time-to-event curve of hospital mortality by Geriatric Nutritional Risk Index (GNRI) scores in the propensity-matched group. Those at risk for malnutrition had a higher risk of hospital mortality (A: total cohort; generalized Wilcoxon, χ²=5.17, p=0.023; B: propensity-matched cohort; log rank, χ²=9.71, p=0.002).

3. Subgroup analysis of elderly patients

In this study, 670 elderly patients were examined, and 450 (67.2%) were identified as at risk of malnutrition based on the GNRI score. The risk group was older (mean 77 vs. 74.7, p<0.001) and had a higher proportion of men (60.2% vs. 50.9%, p=0.022) compared to the no-risk group (Supplementary Table S2). Furthermore, the risk group displayed higher CFS scores before admission (median 3 [IQR, 3 to 5] vs. 3 [IQR, 2 to 4], p<0.001) and lower BMI (22.9 vs. 26.7, p<0.001). Moreover, they exhibited a higher rate of infections acquired in care facilities and a higher prevalence of certain comorbidities, such as chronic neurological diseases and solid malignancies. Subsequent to PSM, no significant differences were observed between the groups, except for BMI and GNRI (Supplementary Table S2).

4. Clinical outcomes of elderly patients

Prior to PSM, the risk group exhibited elevated median CFS scores at discharge (5 [IQR, 3 to 7] vs. 4 [IQR, 3 to 6], p<0.001) and elevated tracheostomy rates (24.9% vs. 17.4%, p=0.029) (Table 4). After PSM, the risk group displayed significantly higher rates of hospital (52.1% vs. 29.5%, p<0.001) and ICU mortality (47.2% vs. 29.1%, p<0.001).

Table 4.

Clinical outcomes of the elderly cohort

5. Cox regression analysis for hospital mortality in elderly patients

In the analysis of the elderly subgroup, GNRI did not demonstrate a significant impact on hospital mortality in either univariate or multivariate Cox analyses prior to propensity matching (Supplementary Table S3). Nonetheless, following propensity matching, being in the risk group (GNRI ≤98) was significantly correlated with increased hospital mortality rates in both univariate (HR, 1.76; 95% CI, 1.30 to 2.39; p<0.001) and multivariate analyses (HR, 1.66; 95% CI, 1.22 to 2.27; p=0.001) (Table 5). Elderly patients in the risk group exhibited a significantly elevated risk of hospital mortality compared to those in the no-risk group (after propensity matching χ²=13.65, p<0.001) (Supplementary Figure S1).

Table 5.

Hazard ratio for hospital mortality in the elderly cohort determined using the Cox proportional hazards model after propensity matching