Prognostic value and nonlinear association of the rdw/albumin ratio with mortality among COVID-19 patients: a MIMIC-IV database analysis

Abstract

Background

RDW and the RDW-to-albumin ratio (RAR) have prognostic relevance in many conditions, including COVID-19. We assessed their prognostic value, examined the potential nonlinear association of RAR with mortality, and tested whether RAR and RDW improve prediction by established ICU scores and COVID-19 prognostic markers.

Methods

In this retrospective cohort of 373 ICU COVID-19 patients from MIMIC-IV, clinical and laboratory data within 24 h of ICU admission were analyzed. RAR and RDW were divided into tertiles. Kaplan–Meier plots and Cox models were used to evaluate mortality associations after confirming model assumptions and checking collinearity. Restricted cubic splines examined RAR nonlinearity. ROC curves were used to compare the predictive accuracy of RAR and RDW. Predictive performance was then further evaluated using ROC curves, Brier scores, calibration plots, and DCA for models with and without RAR or RDW added to LDH, SOFA, SIRS, APS III, and SAPS II. Subgroup and fairness analyses were performed.

Results

A total of 31.9% of the cohort died within 28 days of ICU admission. Higher RAR was associated with increased 28-day ICU mortality, with the highest tertile showing the greatest risk. (HR = 2.12, 95% CI 1.19–3.76, P = 0.010). Trend analysis confirmed a dose-response relationship (P for trend = 0.0137). RDW did not show a significant association with 28-day ICU mortality in Model 3. RAR had better predictive performance than RDW in unadjusted and partly adjusted models (Model 1: 0.674 vs. 0.586, DBA = 0.089, P = 0.002; Model 2: 0.700 vs. 0.650, DBA = 0.051, P = 0.005). A nonlinear J-shaped relationship was identified between RAR and 28-day ICU mortality, with a threshold at 4.98. Adding RAR to LDH or SIRS improved mortality prediction across outcomes, whereas SOFA gained mainly for long-term outcomes. Subgroups confirmed more stable performance for RAR across most subgroups.

Conclusion

RAR outperforms RDW and shows a nonlinear relationship with mortality. RAR enhances risk prediction when added to simpler markers and SOFA, supporting its use in risk stratification among critically ill COVID-19 patients.

Clinical trial number

Not applicable.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12879-026-12672-5.

Introduction

As a disease resulting from the SARS-CoV-2 infection [1], coronavirus disease 2019 (COVID-19) has been a critical global health concern. Since its outbreak in 2019, COVID-19 has resulted in widespread illness and death, leading to a profound global health crisis. By February 2025, approximately 778 million confirmed cases and over 7 million deaths had been recorded globally [2]. Despite advancements in vaccination and therapeutic interventions reducing mortality rates, COVID-19 remains a significant challenge, particularly among patients in intensive care units (ICUs) where mortality rates are significantly higher than in the general population. Therefore, timely recognition of high-risk COVID-19 patients in the ICU is essential for refining treatment strategies and improving prognosis.

To enable early risk stratification, multiple clinical and laboratory indices have been proposed to predict outcomes in patients with COVID-19. Inflammatory and coagulation-related biomarkers, including C-reactive protein, interleukin-6, lactate dehydrogenase, ferritin [3, 4], and hematological measures such as lymphocyte count, neutrophil-to-lymphocyte ratio [5], and platelet-to-lymphocyte ratio [6], have all been linked to disease severity and mortality in hospitalized patients. In the intensive care setting, composite scores that combine physiological and laboratory variables, such as the Sequential Organ Failure Assessment (SOFA) score, are also widely used for prognostication. Red blood cell distribution width (RDW), a commonly used hematological parameter, reflects the degree of variation in red blood cell sizes. Numerous studies have demonstrated its strong association with disease prognosis, including cardiovascular diseases, diabetes, hepatic and renal failure, along with various acute or chronic illnesses, severe pneumonia, sepsis, and other infectious diseases [7–9].Elevated RDW has been consistently linked to worse clinical outcomes, including increased mortality and morbidity, across multiple disease states [10].

Albumin, as an important protein reflecting nutritional status and inflammation, also plays a crucial role in clinical prognosis assessment. Researches indicate that albumin levels are strongly correlated with survival and recovery in a variety of diseases [11–13]. While individual studies have examined the predictive roles of albumin and RDW in COVID-19 [14–17], suggesting that both markers may reflect the inflammatory burden and physiologic reserve in infected patients.

RAR, representing the RDW-to-albumin ratio, serves as an innovative composite indicator that may better capture the interplay between inflammatory and nutritional statuses. Recent evidence suggests that RAR has prognostic value among diverse non‑COVID critical illnesses. For example, in large ICU cohorts with sepsis [18], acute respiratory [19] and heart failure [20], elevated RAR on admission has been associated with increased mortality risk [21]. As systemic inflammation and nutritional decline often coexist in critical illness, conditions which tend to increase RDW while lowering albumin, RAR may integrate these two factors to provide a more accurate reflection of a patient’s inflammatory and metabolic reserve [22, 23]. In the context of COVID-19, several studies have extended these observations: both RDW and RAR at presentation were linked to worse outcomes including death, need for mechanical ventilation or vasopressor support [24]. However, most previous studies used simplified models that assumed linear relationships between biomarkers and outcomes, potentially overlooking clinically relevant thresholds or nonlinear risk patterns. Furthermore, no study has examined whether adding RAR or RDW to commonly used ICU scores and COVID-19 prognostic markers can improve mortality prediction in critically ill COVID-19 patients.

To bridge this gap in a well-defined ICU COVID-19 population, we leveraged the publicly available Medical Information Mart for Intensive Care IV (MIMIC-IV, version 3.0) database to assemble a cohort of critically ill COVID-19 patients. Our aims were: (1) to evaluate and directly compare the predictive value of RDW and RAR for mortality in ICU patients with COVID-19, (2) to characterize potential non-linear relationships between RAR and mortality risk using restricted cubic spline (RCS) analysis, (3) and tested whether RAR and RDW improve prediction by established ICU scores and COVID-19 prognostic markers.

Method

Research design and data collection

This cohort study retrospectively retrieved and analyzed data from MIMIC-IV database (version 3.0), which contains detailed records of over 360,000 patients and 540,000 hospital admissions from 2008 to 2022. Comprehensive patient information is available in the database, covering updates from the COVID-19 pandemic period. Access to the data was granted to one author (P. Huang) following successful completion of National Institutes of Health (NIH) human research participant protection training and the Collaborative Institutional Training Initiative (CITI) certification (certification number: 61915383). As the database contains fully anonymized patient records, informed consent was not necessary for this cohort study.

Study population

Patients diagnosed with COVID-19 were selected using International Classification of Diseases (ICD) codes. In this study, we used ICD codes U071 and J1282. The screening and enrollment procedures are outlined in Fig. 1.

Fig. 1.

Flow diagram of patient selection. Abbreviations: ICD code: International Classification of Diseases code; RDW, red cell distribution width; RDW/ALB, red cell distribution width to albumin ratio (RAR)

Patients were excluded if they met any of the following criteria: (1) age below 18 years, (2) ICU stays lasting shorter than 24 h, (3) repeated ICU admissions, where only the initial admission was considered, and (4) incomplete or abnormal RDW and albumin data during the initial 24 h of ICU stay. As a result, 373 individuals from the MIMIC-IV (version 3.0) database were qualified for inclusion and were categorized into tertiles based on RAR and RDW levels (Fig. 1).

Data collection

In this retrospective study, patient data were retrieved from MIMIC-IV (version 3.0) via Structured Query Language (SQL), focusing on patients’ average data from the initial 24 h of each patient’s initial ICU stay. Ten variable categories were analyzed: (1) Baseline variables: age, gender, race, marital status, weight and height (2) Comorbid Conditions: smoking, hypertension, type 2 diabetes mellitus (Diabetes II), acute myocardial infarction (AMI), heart failure (HF), malignant tumor (MT), acute renal failure (ARF), chronic kidney disease (CKD), chronic obstructive pulmonary disease (COPD), pneumonia and stroke. (3) Vital signs: mean arterial pressure (MAP), heart rate, respiratory rate, diastolic blood pressure (DBP), systolic blood pressure (SBP), SpO2, temperature and first-day fluid balance (input/output). (4) Laboratory Indicators: white blood cell count (WBC), neutrophil count, lymphocyte count, hemoglobin, hematocrit (HCT), red blood cell count (RBC), international normalized ratio (INR), partial thromboplastin time (PTT), prothrombin time (PT), platelets, RDW, aspartate aminotransferase (AST), alanine aminotransferase (ALT), serum sodium, serum potassium, total calcium, chloride, creatinine, blood urea nitrogen (BUN), albumin, pH, lactate, anion gap, fasting blood glucose (FBG), C-reactive protein (CRP), ferritin, lactate dehydrogenase(LDH) and RAR. (5) Score system at admission: systemic inflammatory response syndrome (SIRS), the simplified acute physiology score II (SAPSII), sepsis-related organ failure assessment score (SOFA), the acute physiology scores III (APSIII). (6) Therapeutic Interventions: continuous renal replacement therapy (CRRT), invasive mechanical ventilation (IMV), non-invasive mechanical ventilation (NIMV) glucocorticoids, antihypertensive, norepinephrine and dexamethasone. (7) Key Events: the occurrence and severity of acute kidney injury (AKI) and the comorbidity of sepsis during ICU stay. (8) Length of stay (LOS): hospital and ICU durations. (9) Primary outcome: 28-day ICU mortality. (10) Secondary outcome: 28-day in-hospital mortality, all-cause mortality at 60-day, 90-day and 180-day and 365-day.

Laboratory values were the average data taken during the initial 24 h in the ICU. Based on the missingness distribution illustrated in Supplementary Fig. 1. and detailed in Supplementary Table 1, variables with more than 20% missing data were excluded from the analysis—specifically, neutrophil count, lymphocyte count, height, input CRP, ferritin and lactate. Variables with missing rates of 20% or lower were imputed through multiple imputation using the R mice package (version 3.17.0) to minimize bias [25]. A total of 5 imputed datasets were generated, and all subsequent analyses were performed by pooling the estimates across these datasets using Rubin’s rules.

Variable definitions and outcomes

The primary exposure variables were RAR and RDW values, categorized into tertiles based on their initial measurements within 24 h of entering ICU. Mortality data in the MIMIC-IV (version 3.0) database were obtained from state and hospital death records, ensuring at least one year of follow-up for all participants. The primary outcome was 28-day ICU mortality, while all-cause mortality at 60, 90, 180, and 365 days served as secondary outcomes. The 60-day, 90-day, 180-day and 365-day all-cause mortality collectively defined as long-term mortality. Survival time was measured from ICU entry until death or withdrawal from follow-up, whichever occurred first. Individuals who were discharged before the endpoint or survived beyond the follow-up period were considered censored [26].

Statistical analysis

R software (R Core Team. R: A language and environment for statistical computing. Version 4.4.2. Vienna, Austria: R Foundation for Statistical Computing; 2024.) was utilized for all statistical analyses and data visualizations, with P < 0.05 considered statistically significant. The normality of continuous variables was evaluated using the Shapiro-Wilk test, while variance homogeneity was evaluated via Levene’s test. As all continuous variables failed to meet these assumptions, the Kruskal-Wallis test was conducted to compare continuous variables across RAR or RDW tertile groups, and the results were reported as medians and interquartile ranges (IQR). For continuous variables showing significant differences in the Kruskal-Wallis test (P < 0.05), pairwise post-hoc comparisons were conducted using Dunn’s test with Bonferroni correction to adjust for multiple comparisons [27]. Counts and percentages were used to describe categorical variables, which were compared across groups using the Chi-square test. Kaplan-Meier survival analysis was conducted to analyze all outcomes, with log-rank tests comparing mortality across RAR and RDW tertile groups.

Associations between each index and mortality outcomes were evaluated using univariate and multivariable Cox proportional hazards models. Univariate Cox regression was used for variable screening, and covariates with a p value < 0.10 were included as candidate confounders in the multivariable Cox model [28–31], in order to avoid excluding clinically important variables that might become significant only after adjustment. Proportional hazards assumptions were assessed for both individual covariates and the overall model. To assess collinearity, the variance inflation factor (VIF) was calculated, and any variable with a VIF of 10 or higher was excluded from analysis. Ultimately, we constructed three multivariate Cox proportional hazards models: unadjusted (Model 1), partially adjusted (Model 2), and fully adjusted (Model 3). RAR and RDW were analyzed as continuous and categorical variables, with P for trend calculated based on tertiles. Receiver operating characteristic (ROC) analysis was used to identify most discriminative RAR and RDW cutoffs, followed by pairwise comparisons to assess predictive accuracy [32]. Using the three Cox models, we derived estimated probabilities to plot ROC curves. We then applied DeLong’s test to compare AUC (Area Under the Curve) values for RAR versus RDW. Finally, we identified optimal cut-points via Youden’s index and calculated sensitivity, specificity, LR+, and LR−. In addition, the incremental predictive value of RAR and RDW beyond LDH and conventional severity scores (SOFA, SIRS, APSIII, and SAPSII) was quantified using DeLong’s test for paired AUC comparisons. Model calibration was evaluated using calibration plots [33] and Brier scores, while clinical utility was assessed using decision curve analysis (DCA).

RCS regression model [34, 35]were employed to evaluate nonlinear associations between RAR and outcomes. RCS regression with three knots was applied to model potential nonlinear associations. The spline function was defined as: f(X)=β₀ + β₁X + β₂S₁(X)+β₃S₂(X) where S₁(X) and S₂(X) denote the truncated cubic spline basis functions derived from the selected knots. Full-adjustment Cox models (Model 3) were utilized for all outcome analyses to control confounders. Lastly, Subgroup analyses evaluated variations in the primary outcome by gender, age (≤ 64 or > 64 years), smoking history, and comorbidities including hypertension, CKD, ARF and Diabetes II. The analyses also accounted for AKI occurrence and severity, sepsis occurrence, and therapeutic interventions like IMV and glucocorticoid use during ICU stay. Subgroup analyses were performed by adding an interaction term between RAR and each stratification variable to assess potential effect modification. Subgroup specific HRs and 95% confidence intervals (CI) were reported. This analytic strategy aligns with contemporary recommendations for subgroup assessment in observational studies [36]. Additionally, we performed a fairness analysis by comparing the predictive performance of RAR and RDW across prespecified clinical subgroups. Within each subgroup, we derived the AUC for RAR and RDW separately. Differences in AUC between RAR and RDW within each subgroup were evaluated using DeLong’s test for comparison of ROC curves. To evaluate whether exclusion of patients with missing RAR values might have biased the results, we compared mortality outcomes, hospital length of stay, and ICU length of stay between patients included in the main analysis (with available RAR) and those with missing RAR who were excluded.

Results

Patient characteristics at baseline

The clinical characteristics of enrolled 373 COVID-19 patients were summarized in Table 1 and Supplementary Table 2. The median age of the enrolled patients was 64 years (IQR: 54–74), with 62.47% (233) of the patients being male. The 28-day ICU mortality rate was 31.9%. Patients were stratified into RAR tertiles: T1 ≤ 4.482, 4.482 < T2 < 5.56, 5.56 ≤ T3 ≤ 11.75, and RDW tertiles: G1 ≤ 13.5, 13.5 < G2 < 15.07, 15.07 ≤ G3 ≤ 19.77.

Table 1.

Baseline characteristics grouped by RAR levels

Characteristics	RAR			Total	P-value
	T1 (≤ 4.482)	T2 (4.482–5.560)	T3 (5.560–11.750)	(2.698–11.750)
Number of patients, n	124	124	125	373
Baseline variables
Age (year)	60 (52, 68)	65 (53, 73)	67 (59, 76)	64 (54, 74)	0.001
Weight (kg)	86.0 (77.8, 99.5)	89.0 (71.7, 102.2)	82.5 (70.3, 98.1)	86.0 (73.0, 99.4)	0.212
Gender, n (%)					0.446
Male	83 (66.935)	74 (59.677)	76 (60.800)	233 (62.466)
Female	41 (33.065)	50 (40.323)	49 (39.200)	140 (37.534)
Race, n (%)					0.785
White	52 (41.935)	42 (33.871)	48 (38.400)	142 (38.070)
Black	19 (15.323)	22 (17.742)	21 (16.800)	62 (16.622)
Others	53 (42.742)	60 (48.387)	56 (44.800)	169 (45.308)
Marital Status, n (%)					0.027
Married	46 (37.097)	35 (28.226)	41 (32.800)	122 (32.708)
Single	42 (33.871)	34 (27.419)	29 (23.200)	105 (28.150)
Widowed/Divorced	15 (12.097)	23 (18.548)	13 (10.400)	51 (13.673)
Others	21 (16.935)	32 (25.806)	42 (33.600)	95 (25.469)
Comorbid Conditions, n (%)
Smoking, n (%)	27 (21.774)	31 (25.000)	27 (21.600)	85 (22.788)	0.772
Hypertension, n (%)	46 (37.097)	35 (28.226)	38 (30.400)	119 (31.903)	0.295
Diabetes II, n (%)	46 (37.097)	47 (37.903)	41 (32.800)	134 (35.925)	0.665
HF, n (%)	13 (10.484)	32 (25.806)	30 (24.000)	75 (20.107)	0.004
AMI, n (%)	10 (8.065)	16 (12.903)	20 (16.000)	46 (12.332)	0.159
MT, n (%)	7 (5.645)	10 (8.065)	10 (8.000)	27 (7.239)	0.704
CKD, n (%)	20 (16.129)	31 (25.000)	24 (19.200)	75 (20.107)	0.209
ARF, n (%)	58 (46.774)	84 (67.742)	102 (81.600)	244 (65.416)	< 0.001
Pneumonia, n (%)	39 (31.452)	54 (43.548)	52 (41.600)	145 (38.874)	0.110
COPD, n (%)	7 (5.645)	18 (14.516)	16 (12.800)	41 (10.992)	0.060
Stroke, n (%)	5 (4.032)	6 (4.839)	5 (4.000)	16 (4.290)	0.934
Vital signs
Heart rate (beats/min)	79 (70, 92)	81 (70, 94)	87 (76, 98)	82 (72, 95)	0.007
MAP (mmHg)	83 (80, 89)	81 (76, 85)	80 (74, 85)	82 (76, 86)	< 0.001
SBP (mmHg)	119 (111, 128)	114 (107,125)	114 (105, 123)	115 (108, 126)	0.002
DBP (mmHg)	70 (66, 76)	66 (62, 72)	67 (60, 71)	68 (63, 73)	< 0.001
Respiratory rate (times/min)	22 (20, 26)	23 (19, 27)	24 (20, 28)	23 (19, 27)	0.278
SpO2 (%)	95.4 (94.2, 96.5)	95.4 (94.2, 96.9)	96.1 (94.4, 97.3)	95.6 (94.2, 97.0)	0.312
Temperature (℃)	36.9 (36.8, 37.2)	36.9 (36.7, 37.2)	36.9 (36.6, 37.2)	36.9 (36.7, 37.2)	0.562
Output (ml)	1360.0 (955.0, 2100.0)	1365.0 (650.0, 2125.0)	1350.0 (700.0, 2020.0)	1360.0 (775.0, 2055.0)	0.726
Laboratory Indicators:
WBC (K/µL)	8.9 (6.7, 11.4)	9.9 (7.6, 12.5)	11.5 (8.8, 15.0)	10.0 (7.4,13.2)	< 0.001
RBC (m/uL)	4.4 (3.9, 4.9)	3.8 (3.1, 4.3)	3.3 (2.9, 4.0)	3.9 (3.2, 4.5)	< 0.001
Hb (g/dl)	12.9 (11.8, 14.2)	11.2 (9.2, 12.5)	9.5 (8.4, 11.0)	11.4 (9.3, 13.0)	< 0.001
HCT (%)	38.9 (35.4, 43.2)	35.0 (28.8, 39.1)	29.7 (26.1, 35.7)	35.3 (29.1, 40.3)	< 0.001
PT(s)	13.2 (12.2, 13.9)	13.6 (12.4, 15.2)	13.7 (13.0, 15.0)	13.5 (12.5, 14.8)	< 0.001
PTT (s)	29.7 (27.8, 35.4)	32.2 (28.3, 40.9)	35.6 (31.4, 43.7)	32.5 (28.6, 39.3)	< 0.001
INR	1.20 (1.10, 1.30)	1.25 (1.13, 1.40)	1.25 (1.20, 1.40)	1.23 (1.15, 1.35)	0.002
Platelets (K/µL)	234.0 (192.5, 289.5)	218.3 (149.5, 281.0)	200.8 (118.0, 286.5)	218.3 (156.5, 287.0)	0.018
RDW (%)	13.2 (12.6, 13.7)	14.5 (13.7, 15.4)	15.9 (14.3, 17.2)	14.3 (13.2, 15.7)	< 0.001
ALT (IU/L)	38.0 (23.5, 53.0)	30.0 (18.0, 46.0)	29.0 (18.0, 41.0)	31.7 (20.5, 47.0)	0.004
AST (IU/L)	48.5 (33.0, 72.5)	42.8 (29.0, 58.3)	38.3 (26.0, 59.5)	42.7 (30.0, 64.0)	0.039
BUN (mg/dL)	20.0 (14.0, 26.5)	29.0 (16.5, 40.8)	29.7 (20.8, 42.7)	25.3 (15.6, 36.5)	< 0.001
Creatinine (mg/dL)	0.90 (0.75, 1.18)	1.10 (0.80, 1.65)	1.18 (0.88, 1.88)	1.03 (0.80, 1.60)	< 0.001
LDH (IU/L)	457.00 (309.50,573.50)	470.00 (369.00,607.50)	393.00 (301.00,529.00)	440.00 (318.50, 575.00)	0.046
Albumin (g/dL)	3.40 (3.20, 3.60)	2.90 (2.80, 3.10)	2.35 (2.05, 2.60)	2.95 (2.55, 3.30)	< 0.001
Sodium (mEq/L)	137.0 (135.0, 139.5)	138.7 (136.0, 141.8)	138.0 (134.5, 141.0)	138.0 (135.0, 140.7)	0.023
Potassium (mEq/L)	4.17 (3.95, 4.40)	4.27 (3.97, 4.60)	4.30 (3.95, 4.63)	4.20 (3.95, 4.55)	0.074
Total Calcium (mg/dL)	8.55 (8.30, 8.90)	8.30 (7.95, 8.85)	8.00 (7.60, 8.50)	8.35 (7.93, 8.80)	< 0.001
Chloride (mEq/L)	101.5 (98.0, 104.0)	103.0 (99.7, 106.0)	102.7 (99.2, 105.8)	102.0 (99.0, 105.5)	0.011
Anion gap(mEq/L)	12.50 (11.00, 14.67)	13.25 (11.00, 15.00)	12.20 (10.00, 15.67)	12.67 (10.75, 15.00)	0.290
FBG (mg/dL)	145.0 (120.0, 198.5)	149.0 (124.8, 190.5)	143.0 (117.0, 178.2)	144.5 (119.5, 187.0)	0.192
RAR	3.968 (3.649, 4.219)	4.983 (4.700, 5.227)	6.527 (6.011, 7.389)	4.983 (4.226, 6.011)	< 0.001
Score system
SOFA	4 (2, 7)	6 (4, 10)	8 (5, 10)	6 (3, 9)	< 0.001
SIRS	2 (2, 3)	2 (2, 3)	3 (2, 4)	3 (2, 3)	< 0.001
APSIII	39 (30, 50)	54 (36, 72)	61 (46, 76)	51 (36, 69)	< 0.001
SAPSII	30 (24, 39)	43 (32, 53)	47 (39, 58)	40 (29, 51)	< 0.001
Therapeutic Interventions, n (%)
IMV	59 (47.581)	81 (65.323)	86 (68.800)	226 (60.590)	0.001
NIMV	93 (75.000)	78 (62.903)	58 (46.400)	229 (61.394)	< 0.001
CRRT	14 (11.290)	28 (22.581)	35 (28.000)	77 (20.643)	0.004
Glucocorticoids	96 (77.419)	82 (66.129)	83 (66.400)	261 (69.973)	0.086
Antihypertensive	100 (80.645)	103 (83.065)	110 (88.000)	313 (83.914)	0.273
Norepinephrine	38 (30.645)	49 (39.516)	57 (45.600)	144 (38.606)	0.051
Dexamethasone	78 (62.903)	53 (42.742)	52 (41.600)	183 (49.062)	< 0.001
Key Events, n (%)
AKI	106 (85.484)	112 (90.323)	118 (94.400)	336 (90.080)	0.062
AKI stage					< 0.001
0	18 (14.52)	12 (9.68)	7 (5.60)	37 (9.92)
1	13 (10.48)	8 (6.45)	15 (12.00)	36 (9.65)
2	63 (50.81)	47 (37.90)	35 (28.00)	145 (38.87)
3	30 (24.19)	57 (45.97)	68 (54.40)	155 (41.55)
Sepsis	73 (58.871)	94(75.806)	105(84.000)	272 (72.922)	< 0.001
Length of stay (LOS)
LOS in hospital (days)	15.51 (8.82, 29.27)	18.70 (9.84, 28.70)	23.21 (11.78, 40.55)	17.97 (9.64, 32.89)	0.069
LOS in ICU (days)	6.80 (2.82, 14.95)	8.22 (3.04, 16.92)	7.60 (3.17, 19.60)	7.33 (3.01, 16.92)	0.514
Primary outcome
28-day ICU mortality, n (%)	18 (14.516)	44 (35.484)	57 (45.600)	119 (31.903)	< 0.001
Secondary outcome
28-day in-hospital mortality, n (%)	18 (14.516)	42 (33.871)	53 (42.400)	113 (30.295)	< 0.001
60-day all-cause mortality, n (%)	23 (18.548)	52 (41.935)	69 (55.200)	144 (38.606)	< 0.001
90-day all-cause mortality, n (%)	24 (19.355)	53 (42.742)	70 (56.000)	147 (39.410)	< 0.001
180-day all-cause mortality, n (%)	26 (20.968)	54 (43.548)	74 (59.200)	154 (41.287)	< 0.001
365-day all-cause mortality, n (%)	29 (23.387)	56 (45.161)	77 (61.600)	162 (43.432)	< 0.001

Abbreviations: See Abbreviation Table. AKI stage, stage of acute kidney injury during ICU stays; AKI stage 0–3, no AKI after ICU (0) and stages 1–3 defined by KDIGO. P values in bold meant significantly different (P < 0.05). P values represent comparisons among RAR tertiles and were calculated using the Kruskal-Wallis test for continuous variables and the chi-square test for categorical variables. RAR tertiles: T1 ≤ 4.482, 4.482 < T2 < 5.560, 5.560 ≤ T3 ≤ 11.750

Higher RAR and RDW levels correlated with a greater proportion of older age, and higher prevalence of HF and ARF. These patients had lower MAP, SBP, RBC, HCT, platelets, ALT, albumin, LDH and pH levels, alongside higher heart rates, PT, PTT, INR, BUN, creatinine levels, and SOFA, APSIII, and SAPSII scores. Post-hoc analyses showed that many of these differences were statistically significant, particularly between the highest and lowest tertiles of both RAR and RDW (Adjusted P < 0.05) (Supplementary Tables 3 and 4). Meanwhile, they were more likely to undergo more severe AKI stages and CRRT treatment and experienced higher mortality rates (all P < 0.05). Compared to RDW, increased RAR levels were further associated with elevated WBC, hemoglobin, neutrophil counts, and SIRS, scores. Patients with higher RAR levels were more frequently treated with IMV and dexamethasone, yet a significantly lower proportion underwent NIMV. This may result from higher RAR levels leading to a greater requirement for invasive mechanical ventilation. These patients also had higher sepsis comorbidity rates, AST, and total calcium levels. These findings indicate that RAR may serve as a more sensitive marker of inflammation compared to RDW alone.

Outcomes of survival analyses

The Kaplan–Meier curves revealed that elevated RAR correlated with higher 28-day ICU mortality,28-day in-hospital and other all-cause mortality in COVID-19 patients (log-rank P < 0.05, Fig. 2, Supplementary Fig. 2). For RAR and 28-day ICU mortality, a significant correlation was observed between elevated RAR levels and increased mortality (log-rank P < 0.001), with significant differences between T1 and T2 (P < 0.001) and T1 and T3 (P < 0.001). No significant difference was found between T2 and T3 (P = 0.075) (Fig. 2). A similar pattern was observed for RAR in relation to 28-day in-hospital, 60-day, 90-day, 180-day, and 365-day all-cause mortality, with significant differences across all tertiles (Supplementary Fig. 2).For RDW and 28-day ICU mortality, the G3 group exhibited significantly higher mortality than G1 (log-rank P = 0.0162), while no significant differences were observed between G1 vs. G2 (P = 0.211) or G2 vs. G3 (P = 0.185) (Fig. 2). This trend persisted over 60, 90, 180, and 365 days for all-cause mortality (Supplementary Fig. 3). For 365-day all-cause mortality, G2 patients showed significantly higher mortality than G3 patients (log-rank P = 0.04135).

Fig. 2.

Kaplan-Meier curves for primary outcomes. This figure illustrates Kaplan–Meier (K–M) survival curves for 28-day ICU mortality stratified by RAR (Panel A) and RDW (Panel B) levels. In Panel A, patients were categorized into RAR tertiles: T1 ≤ 4.482, 4.482 < T2 < 5.560, 5.560 ≤ T3 ≤ 11.750., with an overall log-rank P < 0.001. Similarly, in Panel B, RDW was stratified into three groups: G1 ≤ 13.5, 13.5 < G2 < 15.07, 15.07 ≤ G3 ≤ 19.77, with an overall log-rank P = 0.0162

The results of Cox regression models

A complete list of covariates evaluated in the univariable Cox regression for 28-day ICU mortality is provided in Supplementary Table 5. The VIF [37] were calculated for all covariates to assess multicollinearity (Fig. 3A). For several variables including HCT, hemoglobin, MAP, DBP, PT, INR and RBC, the VIF values exceeded 10. Then, to examine the correlation between RAR, RDW, and mortality in COVID-19 patients, three Cox regression models were applied (Table 2). Model 1 unadjusted; Model 2 adjusted for age, gender; and Model 3 adjusted for age, gender, ARF, first day output, WBC, platelets, BUN, SOFA, and the use of IMV. For Model 3, we plotted the VIF values for each variable (Fig. 3B) and visualized the correlation structure among covariates with a heatmap (Fig. 3C). All covariates included in Model 3 had VIFs below 10, indicating no evidence of problematic multicollinearity.

Fig. 3.

Correlation and collinearity assessment: and VIF (A and B) and Volcano Plot (C). Panel A shows a bar plot of VIF for all variables. In this panel, HCT, hemoglobin, MAP, DBP, PT, INR, and RBC all exhibited VIF values greater than 10. Panel B displays a bar plot of the VIFs for all predictors included in Model 3. In this adjusted model, all predictors had VIF values below 10, indicating no concerning multicollinearity. Panel C illustrates a heatmap of pairwise correlations among the predictors in Model 3. The color gradient spans from blue (negative correlation) to red (positive correlation), with deeper shades indicating stronger correlations. Based on this heatmap, all pairwise correlations among the predictors in Model 3 were relatively low (≤ 0.44), suggesting acceptable independence among covariates. Abbreviations: See Abbreviation Table

Table 2.

Cox proportional hazard ratios (HR) for outcomes

	Model1	P-value	P for trend	Model2	P-value	P for trend	Model3	P-value	P for trend
	HR (95% CI)			HR (95% CI)			HR (95% CI)
28-day ICU mortality
RAR
Continuous	1.32(1.19,1.47)	< 0.001		1.29(1.15,1.43)	< 0.001		1.13(1.00,1.29)	0.049
Tertiles			< 0.001			<0.001			0.0137
T1 (N = 124)	Ref			Ref			Ref
T2 (N = 124)	2.80(1.62,4.85)	< 0.001		2.53(1.45,4.39)	0.001		1.78(1.00,3.16)	0.050
T3 (N = 125)	4.02(2.36,6.83)	< 0.001		3.55(2.08,6.05)	< 0.001		2.12(1.19,3.76)	0.010
RDW
Continuous	1.13(1.03,1.23)	0.009		1.12(1.02,1.22)	0.020		1.04(0.94,1.16)	0.456
Tertiles			0.0047			0.0241			0.757
G1 (N = 127)	Ref			Ref			Ref
G2 (N = 122)	1.35(0.84,2.18)	0.210		1.31(0.81,2.10)	0.271		0.97(0.59,1.59)	0.902
G3 (N = 124)	1.90(1.21,2.97)	0.005		1.68(1.07,2.64)	0.025		1.07(0.66,1.74)	0.789
28-day in-hospital mortality
RAR
Continuous	1.27(1.14,4.62)	< 0.001		1.24(1.11,1.38)	< 0.001		1.08(0.95,1.23)	0.242
Tertiles			< 0.001			< 0.001			0.0995
T1 (N = 124)	Ref			Ref			Ref
T2 (N = 124)	2.66(1.53,4.62)	< 0.001		2.40(1.37,4.18)	0.002		1.59(0.89,2.83)	0.119
T3 (N = 125)	3.61(2.11,6.16)	< 0.001		3.16(1.84,5.42)	< 0.001		1.71(0.95,3.08)	0.072
RDW
Continuous	1.09(1.00,1.20)	0.06		1.08(0.98,1.18)	0.123		0.99(0.89,1.11)	0.895
Tertiles			0.0268			0.1014			0.677
G1 (N = 127)	Ref			Ref			Ref
G2 (N = 122)	1.27(0.79,2.05)	0.329		1.22(0.75,1.97)	0.426		0.86(0.52,1.43)	0.567
G3 (N = 124)	1.67(1.06,2.63)	0.028		1.47(0.93,2.33)	0.103		0.89(0.54,1.46)	0.642
60-day all-cause mortality
RAR
Continuous	1.34(1.22,1.48)	< 0.001		1.31(1.19,1.44)	< 0.001		1.17(1.04,1.30)	0.006
Tertiles			< 0.001			< 0.001			0.0026
T1 (N = 124)	Ref			Ref			Ref
T2 (N = 124)	2.69(1.65,4.39)	< 0.001		2.42(1.48,3.97)	< 0.001		1.76(1.05,2.95)	0.031
T3 (N = 125)	3.99(2.49,6.40)	< 0.001		3.50(2.18,5.64)	< 0.001		2.21(1.33,3.68)	0.002
RDW
Continuous	1.09(1.00,1.20)	0.060		1.13(1.04,1.23)	0.003		1.07(0.97,1.17)	0.181
Tertiles			0.00128			0.0089			0.556
G1 (N = 127)	Ref			Ref			Ref
G2 (N = 122)	1.42(0.92,2.18)	0.112		1.38(0.90,2.13)	0.144		1.05(0.67,1.65)	0.816
G3 (N = 124)	1.95(1.29,2.94)	0.001		1.73(1.15,2.62)	0.009		1.14(0.73,1.78)	0.564
90-day all-cause mortality
RAR
Continuous	1.34(1.22,1.47)	< 0.001		1.30(1.18,1.43)	< 0.001		1.16(1.04,1.29)	0.010
Tertiles			< 0.001			< 0.001			0.0032
T1 (N = 124)	Ref			Ref			Ref
T2 (N = 124)	2.64(1.63,4.27)	< 0.001		2.37(1.46,3.86)	< 0.001		1.72(1.03,2.85)	0.037
T3 (N = 125)	3.90(2.45,6.21)	< 0.001		3.43(2.15,5.47)	< 0.001		2.14(1.30,3.54)	0.003
RDW
Continuous	1.15(1.06,1.24)	< 0.001		1.14(1.05,1.24)	0.002		1.07(0.97,1.18)	0.152
Tertiles			< 0.001			0.0066			0.526
G1 (N = 127)	Ref			Ref			Ref
G2 (N = 122)	1.38(0.90,2.12)	0.138		1.35(0.88,2.07)	0.173		1.02(0.66,1.60)	0.919
G3 (N = 124)	1.97(1.31,2.95)	0.001		1.75(1.17,2.64)	0.007		1.14(0.74,1.77)	0.547
180-day all-cause mortality
RAR
Continuous	1.35(1.23,1.47)	< 0.001		1.31(1.19,1.44)	< 0.001		1.16(1.05,1.30)	0.005
Tertiles			< 0.001			< 0.001			0.0022
T1 (N = 124)	Ref			Ref			Ref
T2 (N = 124)	2.50(1.56,3.99)	< 0.001		2.26(1.41,3.62)	< 0.001		1.61(0.98,2.64)	0.058
T3 (N = 125)	3.87(2.47,6.05)	< 0.001		3.40(2.16,5.33)	< 0.001		2.12(1.30,3.44)	0.002
RDW
Continuous	1.15(1.06,1.24)	< 0.001		1.14(1.05,1.23)	0.002		1.07(0.98,1.17)	0.151
Tertiles			< 0.001			0.0047			0.475
G1 (N = 127)	Ref			Ref			Ref
G2 (N = 122)	1.37(0.90,2.08)	0.138		1.34(0.88,2.04)	0.167		1.02(0.66,1.57)	0.938
G3 (N = 124)	1.98(1.33,2.93)	< 0.001		1.77(1.19,2.63)	0.005		1.16(0.76,1.77)	0.500
365-day all-cause mortality
RAR
Continuous	1.35(1.23,1.47)	< 0.001		1.31(1.19,1.44)	< 0.001		1.18(1.06,1.31)	0.002
Tertiles			< 0.001			< 0.001			0.0015
T1 (N = 124)	Ref			Ref			Ref
T2 (N = 124)	2.34(1.50,3.67)	< 0.001		2.11(1.34,3.31)	0.001		1.54(0.96,2.47)	0.073
T3 (N = 125)	3.67(2.39,5.64)	< 0.001		3.21(2.09,4.94)	< 0.001		2.08(1.31,3.31)	0.002
RDW
Continuous	1.16(1.07,1.25)	< 0.001		1.15(1.06,1.24)	< 0.001		1.09(0.99,1.19)	0.071
Tertiles			< 0.001			0.00137			0.257
G1 (N = 127)	Ref			Ref			Ref
G2 (N = 122)	1.40(0.93,2.11)	0.109		1.37(0.91,2.08)	0.130		1.07(0.69,1.64)	0.772
G3 (N = 124)	2.10(1.43,3.09)	< 0.001		1.88(1.27,2.77)	0.002		1.26(0.83,1.91)	0.276

Model 1: unadjusted

Model 2: adjusted for age, gender

Model 3: adjusted for age, gender, ARF, output, WBC, platelets, BUN, SOFA, and the use of IMV

RAR tertiles: T1 ≤ 4.482, 4.482 < T2 < 5.560, 5.560 ≤ T3 ≤ 11.750

RDW tertiles: G1 ≤ 13.5, 13.5 < G2 < 15.07, 15.07 ≤ G3 ≤ 19.77

The bold part indicates that P < 0.05, which is statistically significant

Abbreviations: See Abbreviation Table

When modeled as continuous variables, elevated RAR was consistently and significantly associated with increased 28-day ICU mortality across all three Cox models: Model 1, HR = 1.32 (95% CI 1.19–1.47, P < 0.001); Model 2, HR = 1.29 (95% CI 1.15–1.43, P < 0.001), and Model 3, HR = 1.13 (95% CI 1.00–1.29, P = 0.049). In contrast, RDW lost its significance in Model 3 (HR = 1.04, 95% CI 0.94–1.16, P = 0.456), even though it had associations in Model 1 and 2. When categorized into tertiles, RAR exhibited a clear dose–response trend: in Model 3, compared with the lowest tertile (T1), the highest tertile (T3) had HR = 2.12 (95% CI 1.19–3.76, P = 0.010). By contrast, for RDW the highest tertile (G3) did not show a significant association in Model 3 (P = 0.789). Trend tests remained significant for RAR in all models (P for trend < 0.05), but for RDW the p for trend became non-significant in Model 3 (P for trend = 0.757).

Subsequent analyses of secondary outcomes showed that after full adjustment (Model 3), the association between both RAR and RDW with 28-day in-hospital mortality was no longer significant. However, for long-term all-cause mortality at 60, 90, 180 and 365 days, RAR demonstrated a robust and increased association with mortality: in Model 3, T3 was significantly associated with increased mortality at 60 days (HR = 2.21, P = 0.002), 90 days (HR = 2.14, P = 0.003), 180 days (HR = 2.12, P = 0.002) and 365 days (HR = 2.08, P = 0.002). Trend tests across these outcomes were all highly significant (P for trend = 0.0026, 0.0032, 0.0022, 0.0015, respectively). By contrast, RDW exhibited no consistent or significant association with long-term mortality after full adjustment, with all P for trend > 0.05. Together, these findings suggest that RAR retains significant prognostic value for both 28-day ICU mortality and long-term all-cause mortality, whereas RDW does not.

Proportional hazards (PH) assumptions were assessed using scaled Schoenfeld residuals. RAR and RDW met PH assumptions for 28-day ICU mortality, 28-day in-hospital mortality and for 60-, 90-, 180- and 365-day all-cause mortality (Supplementary Fig. 4; Supplementary Tables 6–7). For all outcomes, RAR and RDW satisfied the PH assumption (all P > 0.05). Although, gender and SOFA showed minor deviations from proportionality, global Schoenfeld tests indicated no PH violation for RAR or RDW across any outcome (all global P > 0.05).

Comparative analysis of ROC curves

As outlined in Table 3 and Supplementary Table 8, ROC analysis compared RAR and RDW for predicting 28-day ICU, in-hospital, and 60-, 90-, 180-, 365-day all-cause mortality.

Table 3.

Comparative analysis of ROC curves of RAR and RDW for 28-day ICU Mortality

	AUC	95% CI	95%CI
		Lower Limit	Upper Limit	Sensitivity	Specificity	LR+	LR-	DBA	P-value
RAR (unadjusted)	0.674	0.618	0.731	0.790	0.531	1.69	0.40	0.089	0.002
RDW (unadjusted)	0.586	0.524	0.648	0.639	0.563	1.46	0.64	Ref	Ref
RAR + Model 2	0.700	0.64	0.75	0.62	0.71	2.13	0.53	0.051	0.005
RDW + Model 2	0.650	0.59	0.71	0.79	0.47	1.49	0.45	Ref	Ref
RAR + Model 3	0.776	0.73	0.82	0.820	0.660	2.41	0.28	0.003	0.569
RDW + Model 3	0.773	0.72	0.82	0.810	0.680	2.50	0.29	Ref	Ref

Abbreviations: AUC, area under the curve; CI, confidence interval; DBA, difference between areas; LR+, positive likelihood ratio; LR-, negative likelihood ratio; RAR, red cell distribution width to albumin ratio; RDW, red cell distribution width

P values in bold meant significantly different (P < 0.05)

For 28-day ICU mortality, RAR demonstrated better predictive performance than RDW in Models 1 and 2. For Model 3, although the AUC for RAR was slightly higher, the difference was not statistically significant: [Model 1:area under ROC curve (AUC): 0.67, 95% CI: 0.62–0.73 vs. AUC: 0.59, 95% CI: 0.52–0.65,P = 0.002.;Model 2:AUC: 0.70, 95% CI: 0.64–0.75 vs. AUC: 0.65, 95% CI: 0.59–0.71, P = 0.005; and Model 3:AUC: 0.776, 95% CI: 0.73–0.82 vs. AUC: 0.773, 95% CI: 0.72–0.82, P = 0.569]. Initially, RAR had greater sensitivity (0.79 vs. 0.64) and a comparable specificity (0.53 vs. 0.56), resulting in a higher LR+ (1.69 vs. 1.46). Once partially adjusted, RAR’s specificity improved further, while RDW enhanced its sensitivity at the expense of specificity. Under full adjustment, RDW posted a higher LR+ (2.50) but also a higher LR− (0.29), which could reduce overall sensitivity. The ROC curves for 28-day ICU mortality (Fig. 4) illustrate these trends, and the final optimal cutoffs were 4.73%/(g/L) for RAR and 14.29% for RDW.

Fig. 4.

ROC curves of RAR and RDW for predicting the 28-day ICU mortality. Unadjusted model (Panel A): The area under the curve (AUC) for RAR was 0.67 (95% CI: 0.62–0.73), outperforming RDW (AUC = 0.59, 95% CI: 0.52–0.65, DBA = 0.089, p = 0.002). Fully adjusted model (Panel B): Both RAR and RDW achieved high predictive accuracy, with RAR at 0.776 (95% CI: 0.73–0.82) and RDW at 0.773 (95% CI: 0.72–0.82, DBA = 0.003, p = 0.569). The improvement in AUC was significant in the unadjusted analysis, while the difference between RAR and RDW was not significant after full adjustment. Abbreviations: ROC, receiver operating characteristic; RAR, red cell distribution width to albumin ratio; RDW, red cell distribution width; AUC, area under the curve; DBA, difference in the area under the curve

For 28-day in-hospital and 60-,90-,180- and 365-day all-cause mortality, the ROC results (Supplementary Table 8, Supplementary Fig. 5) further confirmed RAR’s consistent advantage. Overall, RAR displayed more stable performance and sensitivity across time points and adjustment levels. Thus, RAR may offer a slight advantage in identifying high-risk COVID-19 patients.

Improvement in prediction performance for mortality

In Supplementary Table 9, we present whether adding RAR or RDW improved the predictive performance of the various established severity scores (SOFA, APS III, SAPS II, SIRS) and LDH. In the primary outcome of 28-day ICU mortality (Fig. 5), combining LDH with RAR, or SIRS with RAR, improved the prognostic performance of LDH and SIRS alone. For LDH, the AUC rose from 0.56 to 0.687 (P < 0.001), and for SIRS, from 0.557 to 0.672 (P < 0.001). Adding RAR to SOFA, APS III, or SAPS II improved their performance for 28-day ICU mortality, but the changes were not statistically significant (all P > 0.05). RDW, when combined with these scores and LDH, did not improve their predictive power for 28-day ICU mortality (all P > 0.05).

Fig. 5.

ROC Curves Showing Improved 28-Day ICU Mortality Prediction by Adding RAR or RDW. Panel A–E show ROC curves for SOFA, LDH, SIRS, APS III, and SAPS II with RAR or RDW added for predicting 28-day ICU mortality. Adding RAR consistently achieved higher AUC values than adding RDW. Abbreviations: See Abbreviation Table

As shown in Supplementary Tables 9 and Supplementary Figs. 6 and 7, we evaluated whether adding RAR or RDW improved the predictive performance of established severity scores (SOFA, APS III, SAPS II, SIRS) and LDH for the secondary outcomes. For the long-term secondary outcomes (60-day, 90-day, 180-day, and 365-day all-cause mortality), RAR added to SOFA significantly improved the predictive power of SOFA. The AUC increased from 0.709 to 0.743 for 60-day mortality (P = 0.0157), from 0.707 to 0.739 for 90-day mortality (P = 0.02), from 0.708 to 0.741 for 180-day mortality (P = 0.0219), and from 0.713 to 0.749 for 365-day mortality (P = 0.0134). RAR did not improve the prediction of 28-day in-hospital mortality when added to SOFA. RDW did not improve the predictive performance of SOFA for any of the secondary outcomes (all P > 0.05). RAR improved the prognostic value of LDH for 28-day in-hospital mortality (P = 0.0012), and for 60-day, 90-day, 180-day, and 365-day all-cause mortality (P < 0.001). Combining RDW with LDH improved the long-term mortality prediction (60-day, 90-day, 180-day, and 365-day) (all P < 0.05), but the improvement was more significant when RAR was combined with LDH. The combination of RAR and SIRS significantly increased the predictive performance of SIRS for all secondary outcomes (all P < 0.05). Although RDW improved the prognostic performance for long-term secondary outcomes, the effect was not as significant as the combination of RAR and SIRS. Adding RAR or RDW to APS III and SAPS II did not improve their performance for all secondary outcomes (all P > 0.05).

As shown in Supplementary Table 10, adding RAR or RDW to LDH, SOFA, SIRS, APS III, or SAPS II consistently reduced the Brier scores for all mortality outcomes compared with using LDH or these scores alone, with RAR generally achieving lower Brier scores than RDW. According to the DCA (Supplementary Figs. 8–9), models incorporating RAR or RDW also provided higher standardized net benefit than the corresponding baseline models using SOFA, LDH, SIRS, APS III, or SAPS II alone across most threshold probabilities for all outcomes, indicating superior clinical utility. The gain in net benefit was particularly pronounced when RAR, rather than RDW, was added to LDH or SIRS based models. As shown in the calibration plots (Supplementary Figs. 10 and 11), predicted and observed mortality probabilities were generally well aligned for SOFA-, APS III-, and SAPS II–based models, with only minimal changes after adding RAR or RDW. In contrast, for LDH- and SIRS-based models, incorporating RAR shifted the calibration curves closer to the 45° reference line across the clinically relevant risk range, indicating improved calibration compared with LDH and SIRS alone.

The detection of nonlinear relationships

To assess potential nonlinear associations between RAR and mortality in COVID-19 patients, we conducted RCS analyses using both unadjusted and fully adjusted models (Model 3) (Fig. 6). In unadjusted analyses, RAR exhibited significant nonlinearity with 28-day ICU mortality, 28-day in-hospital mortality, and 60-, 90-, 180- and 365-day all-cause mortality (P for nonlinearity = 0.0012, < 0.001, 0.0045, 0.0062, 0.0175, and 0.0154, respectively) (Fig. 6A, C, E, G, I, K). However, after full adjustment, only 28-day ICU mortality retained significant nonlinearity (P for non-linearity = 0.0392) (Fig. 6.B), while 28-day in-hospital mortality and other long-term all-cause mortality (60-day, 90-day, 180-day and 365-day mortality) no longer exhibited significant nonlinearity (P for non-linearity = 0.0734,0.1019,0.1403, 0.2966 and 0.2511, respectively) (Fig. 6.D, F, H, J, L).In summary, our RCS analysis demonstrated a robust J-shaped association between RAR and 28-day ICU mortality among COVID-19 patients, with a critical threshold at 4.98. While a similar arcuate trend was noted for long-term mortality (60-day to 365-day), the nonlinearity lost significance after adjusting for major confounders, transitioning toward a linear increase. This suggests that the nonlinear effect of RAR is predominantly evident in short-term outcomes, with both extremely low and high RAR values linked to an elevated mortality risk.

Fig. 6.

Nonlinear relationships between RAR and outcomes. Panels A, C, E, G, I, and K represent unadjusted models, displaying the association between RAR and 28-day ICU mortality (A), 28-day in-hospital mortality (C), 60-day (E), 90-day (G), 180-day (I), and 365-day all-cause mortality (K). Panels B, D, F, H, J, and L show fully adjusted models (Model 3) for the corresponding outcomes. A J-shaped nonlinear association was observed across models, with mortality risk sharply rising beyond the RAR threshold. After full adjustment, the association remained significant only for 28-day ICU mortality, while nonlinear trends in other outcomes showed no statistical significance (P for non-linearity > 0.05). Abbreviations: See Abbreviation Table. The solid line represents the estimated hazard ratio, and the shaded red area indicates the corresponding 95% confidence intervals

Subgroup analyses

Subgroup analyses were carried out to investigate potential differences in the associations of RAR and RDW with 28-day ICU mortality across clinical characteristics (Fig. 7). For RAR (Fig. 7A), the overall hazard ratio (HR) remained significant (HR = 1.32, 95% CI: 1.19–1.47, p < 0.001) across all COVID-19 patients. No significant interactions were identified across age, gender, smoking status, hypertension, or glucocorticoid use (p for interaction > 0.05). However, RAR tended to show stronger associations in patients without sepsis (HR = 1.84, 95% CI: 1.22–2.78, p = 0.004) compared to those with sepsis (HR = 1.21, 95% CI: 1.08–1.36, p < 0.001; p for interaction = 0.047), indicating a potential differential effect by infection status. Similarly, RAR tended to show stronger associations in patients without ARF (HR = 1.61, p < 0.001) compared to those with ARF (HR = 1.18, p = 0.01; p for interaction = 0.016) and in those without Diabetes II (HR = 1.42, p < 0.001) versus those with Diabetes II (HR = 1.12, p = 0.276; p for interaction = 0.042). Additionally, IMV modified the effect of RAR, with a higher HR in patients not receiving IMV (non-IMV group) (HR = 1.51, p < 0.001) compared to the IMV group (HR = 1.19, p = 0.011; p for interaction = 0.034). For RDW (Fig. 7B), while a significant association with 28-day ICU mortality was observed in the overall cohort (HR = 1.13, 95% CI: 1.03–1.23, p = 0.009), its effect size was smaller compared to RAR. A significant relationship between RDW and mortality was observed in younger patients (HR = 1.20, P = 0.007), whereas no such association was found in older individuals (HR = 1.06, P = 0.376; P for interaction = 0.176).Similarly, RDW was significant in the male subgroup (HR = 1.23, p = 0.004) but not in females (HR = 1.06, p = 0.366; p for interaction = 0.085). Notably, the association between RDW and mortality was stronger in patients without sepsis (HR = 1.36, P = 0.023) than in those with sepsis (HR = 1.09, P = 0.105; P for interaction = 0.103), exhibiting a pattern consistent with RAR. No significant interactions were found for CKD, ARF, Diabetes II, or AKI stage (p for interaction > 0.05), indicating that RDW’s effect showed a similar pattern across subgroups.

Fig. 7.

Subgroup analysis of RAR (A) and RDW (B) for 28-day ICU mortality. Subgroup variables include age (≤ 64 vs. >64), gender, smoking status, glucocorticoid (GC) use, hypertension (HT), sepsis, chronic kidney disease (CKD), acute renal failure (ARF), type II diabetes mellitus (Diabetes II), acute kidney injury (AKI), AKI stage (AKI_stage: I, II, III), and the use of invasive mechanical ventilation (IMV). Abbreviations: RAR, red cell distribution width to albumin ratio; RDW, red cell distribution width; HR, hazard ratio; CI, confidence interval; ICU, intensive care unit

To assess the fairness of predictive performance across clinical subgroups, we evaluated the discrimination of RAR and RDW alone using AUCs within each predefined subgroup (Supplementary Table 11). For 28-day ICU mortality, RAR showed significantly better discrimination in patients without sepsis than in those with sepsis (AUC: 0.829 vs. 0.620, P < 0.001). And its performance was lower in patients with AKI stage 3 than in those without AKI (AUC: 0.552 vs. 0.833, P = 0.0193). RAR also discriminated mortality better in patients not receiving IMV than in those receiving IMV (AUC: 0.765 vs. 0.611, P = 0.0076). Similarly, RDW demonstrated higher prognostic performance in patients without sepsis (AUC: 0.765 vs. 0.553, P = 0.0014) and in those without Diabetes II (P = 0.0365). The difference in prognostic performance between RAR and RDW for 28-day ICU mortality observed in the overall cohort was also evident across multiple subgroups (Supplementary Table 12), including patients aged ≤ 64 years, females, non-smokers, those treated with glucocorticoid, those with or without hypertension, those with sepsis, those without CKD, those with ARF, those with or without Diabetes II, those with or without AKI, and those without IMV. Taken together, these findings suggest that RAR is a more reliable and consistent predictor than RDW across most subgroups, underscoring its stronger prognostic role for 28-day ICU mortality in critically ill COVID-19 patients.

Comparison of outcomes in patients with missing RAR data

In total, 385 patients had RAR values missing during the first 24 h after ICU admission. Compared with patients with missing RAR, the study cohort had longer ICU stays (17.97 vs. 15.61 days, P = 0.020), longer hospital stays (7.33 vs. 6.28 days, P = 0.033), higher 28-day ICU mortality (31.9% vs. 23.64%, P = 0.011), and higher 28-day in-hospital, 60-, 90-, 180-, and 365-day mortality (P = 0.013, 0.002, 0.002, 0.002, and 0.003, respectively; Supplementary Table 13).

Discussion

This retrospective cohort study demonstrates that higher RAR and RDW at admission are linked to higher 28-day ICU mortality among ICU COVID-19 patients. Moreover, our findings consistently suggest that RAR demonstrates stronger predictive power than RDW across unadjusted and partly adjusted models and over different outcomes, including 28-day ICU, 28-day in-hospital mortality, 60-day, 90-day, 180-day and 365-day all-cause mortality. The RCS analysis identified a nonlinear J-shaped relationship between RAR and 28-day ICU mortality, suggesting a threshold effect around 4.98, beyond which mortality risk sharply increases. Adding RAR to LDH or to severity scores such as SOFA or SIRS consistently improved mortality prediction compared with adding RDW. LDH- or SIRS-based models gained across all outcomes, while SOFA-based models showed improvement primarily for long-term outcomes. In contrast, incorporating RAR or RDW into APS III or SAPS II did not enhance their prognostic performance for 28‑day ICU mortality or other outcomes in our cohort. Subgroup analyses revealed that while RAR remained significantly associated with 28-day ICU mortality in the overall cohort, its discriminatory performance varied across clinical subgroups. RAR appeared to perform better in non-sepsis, non-IMV, non-ARF, and non-Diabetes II patients, whereas its predictive accuracy was lower in sepsis or IMV subgroups. These findings suggest potential heterogeneity in prognostic utility depending on clinical context and indicate that RAR’s risk-stratification value may be limited or less robust in more severely ill or mechanically ventilated patients. RDW also retained significance in the full cohort but showed weaker and more inconsistent effects across most subgroups, underscoring the added prognostic value contributed by the albumin component in RAR.

Our data resonate with established evidence that elevated RDW indicates poor outcomes in severely ill patients [8, 38] including those with COVID-19 [14, 15]. Previous research has demonstrated that higher RDW levels at ICU admission are linked to increased 30-day and 90-day mortality rates in COVID-19 patients [39]. Similarly, studies have reported that non-survivors of COVID-19 tend to have higher RDW levels at ICU admission, and RDW values greater than 14.5% are linked to a 31% mortality risk [40]. Albumin, on the other hand, is recognized as a nutritional and inflammatory marker. A decline in albumin levels is associated with severe metabolic failure, alongside increased vascular permeability, hepatic dysfunction, and excessive protein loss, all of which contribute to poor prognosis in critically ill patients. Studies have consistently linked progressive hypoalbuminemia to increased mortality in sepsis [41] and has also been implicated in pulmonary edema development among patients at risk for acute respiratory distress syndrome(ARDS) [42].Additionally, RAR has recently emerged as a more reliable biomarker, with elevated levels linked to higher mortality and a greater likelihood of requiring IMV or vasopressor support in critically ill pediatric COVID-19 patients [24]. This study validated RAR as a superior and more reliable predictor of 28-day ICU mortality in adult Covid-19 patients and underscores its potential as an affordable and widely accessible indicator for early risk stratification in critically ill COVID-19 individuals, consistent with the findings of Shan X et al. [43]. Furthermore, this aligns with recent findings suggesting that RAR predicts mortality and the risk of septic shock among critically ill individuals with community-acquired bacteremia [44]. Another notable contribution of our research is the comparative analysis of RAR and RDW. What we found in the subgroup analysis suggests that RAR may be a more sensitive indicator of inflammation and disease severity compared to RDW alone. This observation aligns with studies highlighting RAR’s superior predictive value over RDW and albumin individually [45].

Multiple factors may explain why RAR outperforms RDW alone in predicting critically ill COVID-19 patients’ mortality. RDW reflects erythropoiesis dysfunction [46], systemic inflammation, and oxidative stress [47], all of which heighten mortality risk [7]. Its elevation correlates with cytokines (e.g., IL-6, TNF-α) that suppress erythropoietin production and disrupt iron metabolism [48]. However, RDW alone omits nutritional and metabolic status—two essential considerations in COVID-19 prognosis. On the other hand, albumin serves as a biomarker of both inflammation and nutritional depletion. Hypoalbuminemia can signify protein-energy malnutrition, capillary leakage, or liver dysfunction, each contributing to worse outcomes in COVID-19 [7]. Thus, RAR integrates RDW’s reflection of inflammatory RBC changes with albumin-based metabolic and nutritional imbalances, offering a more comprehensive assessment of disease severity.

Findings from our subgroup analysis suggest that the variations in RAR’s prognostic value across different patient groups may be driven by distinct pathophysiological mechanisms. In patients without sepsis, RAR appears to have a stronger predictive role than in those with sepsis, which may reflect the fact that sepsis substantially affects both RDW and albumin levels via inflammatory cascades, fluid imbalances, and metabolic dysregulation. These variations driven by disease progression may reduce RAR’s predictive ability in affected patients. Conversely, in those without such conditions, RAR likely serves as a more accurate indicator of baseline inflammatory and nutritional status, preserving its prognostic value. Similarly, IMV impacts RAR’s predictive ability due to its association with increased disease severity, prolonged ICU stays, and multi-organ dysfunction. The resultant metabolic instability may obscure RAR’s prognostic role. In contrast, among non-IMV group, RAR serves as a more stable and reliable predictor of mortality, as fewer confounding metabolic disruptions interfere with its prognostic performance.

Our RCS analysis demonstrated a distinct J-shaped association between RAR and 28-day ICU, with mortality risk elevated at both low and high extremes. The lowest mortality risk was observed at a threshold of 4.98. However, for long-term mortality (60-day, 90-day, 180-day and 365-day all-cause mortality), the nonlinear trend observed in unadjusted models became attenuated after adjusting for confounders, indicating that RAR’s impact on mortality is more pronounced in the short term. This aligns with the linear association between RAR and 1-year mortality in ICU patients with CHD and DM, as reported by Sheng Chen et al. [49]. One possible explanation is that patients with extreme RAR values either experience early mortality or undergo treatment interventions that alter their condition, leading to a reduced proportion in long-term follow-up. The observed J-shaped nonlinear association between RAR and short-term prognosis in COVID-19 patients may be attributed to several underlying mechanisms. An extremely low RAR typically reflects a normal or even reduced RDW, with albumin levels remaining relatively stable or slightly elevated. However, in the setting of acute critical illness, pathological factors such as severe vascular extravasation or fluid resuscitation interventions may result in falsely elevated albumin levels [50]. Furthermore, specific RBC compensatory mechanisms [51, 52] may prevent RDW from increasing, despite substantial disease severity. Consequently, an extremely low RAR does not inherently indicate a lower mortality risk. Extremely high RAR results from a substantial elevation in RDW, indicative of severe inflammation or abnormal hemoglobin distribution [53–55], alongside a significant reduction in albumin, reflecting profound malnutrition or organ failure [56, 57]. This combination frequently signals severe pathological processes, including cytokine storms [58] or multiple organ dysfunction syndrome [59], contributing to a sharp rise in mortality risk. When RAR exceeds 4.98, disturbances in protein metabolism, oxidative stress, and inflammation become exacerbated, precipitating rapid physiological decline. As albumin depletion surpasses a critical threshold, metabolic stress worsens, further diminishing survival probability. In contrast, within the intermediate range, where neither severe inflammatory response nor extreme nutritional deficiencies are present, RAR maintains a relatively balanced level, corresponding to the lowest mortality and defining the curve’s optimal range.

When added to LDH or SIRS, which mainly reflect acute inflammation or tissue injury, RAR produced the largest gains in discrimination, calibration, and net clinical benefit for both 28-day ICU and long-term mortality. RDW yielded smaller but directionally similar improvements, consistent with RAR capturing additional inflammatory and nutritional information beyond these indices. By contrast, SOFA and especially APS III and SAPS II are already highly informative composite scores with strong baseline discrimination for ICU mortality, so adding RAR only modestly improved SOFA for long-term outcomes and did not materially enhance APS III or SAPS II, suggesting a ceiling effect and indicating that RAR is particularly useful as an adjunct to simpler indices such as LDH or SIRS rather than to already complex scoring systems.

This study provides a fresh perspective on the RDW-to-albumin ratio (RAR) as a predictor of mortality in critically ill COVID-19 patients. We demonstrated that RAR has superior prognostic value compared with RDW in unadjusted and partially adjusted models, characterized its nonlinear association with mortality, and examined whether adding RAR to established ICU scores and the COVID-19–related biomarker LDH improves mortality risk prediction. By evaluating RAR in both short-term (28-day ICU and 28-day in-hospital) and long-term (60-, 90-, 180-, and 365-day) mortality, this study enhances the understanding of red cell–based indices and offers a more comprehensive approach to risk stratification in critically ill COVID-19 individuals. The stronger correlation of RAR with early outcomes highlights its value as an effective screening marker for identifying high-risk patients requiring prompt intervention. Furthermore, to enhance the reliability of the Cox regression results, multiple confounder-adjusted models were established using a comprehensive statistical approach to control for confounding factors.

However, this study has certain limitations. Primarily, the retrospective, single-center design may limit generalizability, as variations in demographic profiles and clinical practices across different regions or institutions may lead to varying outcome patterns. Additionally, the lack of external or multicenter validation further restricts the applicability of our findings to broader ICU populations, especially given the heterogeneity in clinical settings. Secondly, RDW, albumin, all covariates used in the multivariable Cox models and other laboratory values were all measured at baseline and treated as time-invariant. We did not incorporate dynamic covariates or conduct longitudinal analyses, which may have offered deeper insight into temporal trends and causality. In addition, the limited sample size of 373 patients may have reduced the statistical power to detect subtle associations, particularly in subgroup analysis. This limitation, combined with nine variables included in the multivariate Cox regression model, may have contributed to an increased risk of overfitting and should be recognized as a methodological concern. Moreover, we cannot exclude selection bias related to missing RAR values. Thus, our study predominantly reflects a sicker population in whom RDW and albumin were more frequently ordered, and the observed associations between RAR and mortality may not be fully generalizable to all ICU patients or to settings with different testing patterns.

Future studies should focus on multicenter validation of the RAR threshold, its predictive value and its J-shaped association in diverse ICU populations. Longitudinal analyses are needed to track RAR fluctuations over time. Additionally, RAR guided intervention trials should evaluate whether albumin supplementation, RBC transfusions, or erythropoiesis support improve outcomes. Future research can also consider integrating RAR into machine-learning models could enhance ICU triage and risk stratification, strengthening its role as a prognostic and therapeutic tool.

Conclusion

This study demonstrates that RAR provides stronger and more consistent predictive value for critically ill COVID-19 patients than RDW across unadjusted and partially adjusted models, and RCS analysis revealed a nonlinear J-shaped association with 28-day ICU mortality, with a risk inflection around an RAR of 4.98. Adding RAR to LDH or to simpler severity measures such as SIRS substantially improved mortality prediction, while SOFA-based models benefited mainly for long-term outcomes. Subgroup analyses further indicated that RAR performed better in patients without sepsis or those not receiving IMV than in their counterparts, and in most subgroups RAR’s 28-day ICU mortality predictive performance exceeded that of RDW. In conclusion, RAR appears to be a valuable prognostic tool for guiding clinical decision-making in critically ill COVID-19 patients.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AKI

Acute kidney injury

ALT

Alanine aminotransferase

AMI

Acute myocardial infarction

APSIII

The acute physiology scores III

ARF

Acute renal failure

AST

Aspartate Aminotransferase

AUC

Area Under the Curve

BUN

Blood urea nitrogen

CI

Confidence interval

CIE

Change-in-estimate

CITI

The Collaborative Institutional Training Initiative

CKD

Chronic kidney disease

DCA

Decision Curve Analysis

COPD

Chronic obstructive pulmonary disease

COVID-19

Coronavirus disease 2019

CRP

C-reactive protein

CRRT

Continuous renal replacement therapy

DBA

Difference between areas

DBP

Diastolic blood pressure

Diabetes II

Type 2 diabetes mellitus

FBG

Fasting blood glucose

HCT

Hematocrit

HF

Heart failure

HR

Hazard ratio

ICD

International Classification of Diseases

ICU

Intensive care units

IMV

Invasive mechanical ventilation

INR

International normalized Ratio

IQR

Interquartile ranges

LDH

lactate dehydrogenase

LOS

Length of stay

MAP

Mean arterial pressure

MIMIC-IV

The Medical Information Mart for Intensive Care IV

MT

Malignant tumor

NIH

National Institutes of Health

NIMV

Non-invasive mechanical ventilation

non-IMV group

patients not receiving IMV

PH

Proportional hazards

PT

Prothrombin time

PTT

Partial thromboplastin time

RAR

Red cell distribution width to albumin ratio

RBC

Red blood cell count

RCS

Restricted cubic spline

RDW

Red cell distribution width

ROC

Receiver operating characteristic

SAPSII

Simplified acute physiology score II

SBP

Systolic blood pressure

SIRS

Systemic inflammatory response syndrome

SOFA

Sepsis-related organ failure assessment score

SQL

Structured Query Language

VIF

Variance inflation factor

WBC

White blood cell

Author contributions

MYZ, KK, JNZ, PH designed this study. PH extracted data. JNZ, PH analyzed the data, prepared the tables and figures. PH, QYY drafted and wrote the manuscript. XTW and QQL participated in writing the manuscript. MYZ, KK and JNZ revised the manuscript. All authors have made an intellectual contribution to the manuscript. The final version was approved by all authors.

Funding

This study is supported by Heilongjiang Province Key R&D Program (GA23C007, GA21C011) and the National Natural Scientific Foundation of China (82172164).

Data availability

Data supporting the results of this study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

Access to the MIMIC-IV database was granted to one author (P. Huang) after successful completion of the National Institutes of Health (NIH) human research participant protection training and the Collaborative Institutional Training Initiative (CITI) certification (ID: 61915383). The database was approved by the Institutional Review Boards of the Massachusetts Institute of Technology (Cambridge, MA) and Beth Israel Deaconess Medical Center (Boston, MA). As the database contains fully de-identified patient data, informed consent was not required for this retrospective cohort study. This study was conducted in accordance with the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.