Baseline clinical characteristics of the study cohort

In total, 476 patients with sepsis were evaluated; 398 of these patients presented with explicit sepsis and 78 with implicit sepsis. According to the exclusion criteria, 74 patients were excluded. The remaining 402 sepsis patients were included in the final study (Fig. 1). The median age of patients in this study cohort was 68 years, and 249 (61.9%) were men. Hypertension (38.8%) and diabetes mellitus (34.3%) were the two main co-morbidities. The median RDW was 13.80%, the median albumin level was 2.71 g/dL, and the median RAR was 5.28%/g/dL. In this study population, 158 (39.3%) patients died during hospitalization, 163 (40.6%) patients died within 28 days, and 214 (53.2%) patients died within 90 days. Non-survivors of in-hospital deaths had higher levels of RDW, RAR, CRP, AST, PT, and D dimer but lower levels of albumin, calcium, fibrinogen, and hemoglobin than did survivors. Table 1 presents the baseline characteristics of the study population, including age, gender, co-morbidities, SOFA score, laboratory test data, and outcome events.

Fig. 1

Flowchart of study patient enrollment

Table 1 Clinical characteristics and outcomes in sepsis patients in ICUsIndependent risk factors for in-hospital mortality

After performing univariate logistic regression analysis, a total of 16 variables were found to have a significance level of P < 0.1, including RDW, serum albumin level, and RAR. All continuous independent variables were associated with the logit of the dependent variable (assessed using the Box–Tidwell procedure). A correlation was observed among RAR, serum albumin, and RDW. RDW and serum albumin levels were excluded from the model. The remaining 14 variables were included in multivariate models, which showed that the independent risk factors included two co-morbidities (active malignancy and chronic cerebrovascular diseases), SOFA score, and laboratory results (CRP and RAR). Notably, RAR was one of five independent risk factors after accounting for any potential confounding variables (adjusted OR: 1.383; 95% CI: 1.164–1.645; P < 0.001). Table 2 lists the detailed ORs and 95% CIs of the univariate and multivariate analyses. The results of the multivariate Cox proportional hazards regression models showed that, after accounting for other risk factors, RAR remained an independent risk factor for the 28- and 90-day mortalities with significant prognostic value in the univariate Cox regression analysis. Table S1 provides a summary of the precise HRs, 95% CIs, and P values from the Cox proportional hazards models.

Table 2 Results of univariate and multivariate logistic regression analysis of in-hospital mortalityClinical characteristics across quartile of RAR

Tables 3 and 4 list the characteristics of all the participants in this study. In the entire cohort, the median RAR was 5.28%/g/dl. The ranges of RAR levels were 0–4.339%/g/dl, 4.340–5.279%/g/dl, 5.280–6.779%/g/dl, and > 6.779%/g/dl in quartile 1 (Q1), quartile 2 (Q2), quartile 3 (Q3), and quartile 4 (Q4), respectively. Some clinical characteristics showed significant linear trends across the RAR quartiles (P for trend < 0.05). Patients in Q4 had higher levels of AST, PT, and APTT but lower levels of calcium and hemoglobin than those in Q1. Chronic renal disease and chronic hepatitis were substantially more prevalent in people with a high RAR than those with a low RAR. Consistent with the expected trends, the in-hospital, 28-day, and 90-day mortality rates were considerably greater in individuals with a high RAR than those with a low RAR (all P < 0.001; all P for trend < 0.001).

Table 3 Laboratory examinations across quartiles of RAR in sepsis patients in ICUsTable 4 Clinical characteristics, disease score and outcomes across quartiles of RAR in sepsis patients in ICUsRelationship between RAR and mortality

As can be seen from Table 4, the higher the quartile of RAR, the higher the mortality. The Q1 group was used as the reference group; the unadjusted and multivariable-adjusted ORs for in-hospital mortality for individuals from Q2–Q4 in Models I–IV are shown in Fig. 2. After adjusting for confounding factors, the association was no longer significant in Q2, but it was still significant in Q3 and Q4. In Model IV, compared with Q1, ORs (95% CIs) of in-hospital mortality for Q2, Q3, and Q4 were 1.027 (0.413–2.553), 3.632 (1.579–8.354), and 4.175 (1.025–10.729), respectively. In Models I–IV, we found statistically significant linear trends in in-hospital mortality across the RAR quartiles (all P < 0.001). Similar outcomes were observed in the multivariate Cox regression models assessing the risk of 28- and 90-day mortalities. After adjusting for confounding factors, multivariate models showed a significant association between the RAR in Q3 and Q4 and the 28- and 90-day mortality rates. Detailed HRs and 95% CIs of the multivariate models are shown in Table 5.

Fig. 2

A Logistic regression was used to evaluate the association between quartiles of RAR and in-hospital mortality. Model I adjusted for nothing. Model II adjusted for sex, age. Model III adjusted for model II plus diabetes, chronic renal diseases, chronic hepatic disease, and the SOFA score (Sequential Organ Failure Assessment score). Model IV adjusted for Model III plus hemoglobin, calcium, CRP (C-reaction protein), procalcitonin, fibrinogen, AST (aspartate aminotransferase), PT (prothrombin time), APTT (activated partial thromboplastin time). Abbreviations: RAR, red blood cell distribution width to albumin ratio; OR: odds ratio; 95% CI: 95% confidence interval

Table 5 Association of quartiles of RAR with 28- and 90-day mortality

Patients with sepsis were categorized according to RAR quartiles, and 28- and 90-day survival curves were constructed to analyze cumulative survival at varying RAR levels. As illustrated in Fig. 3a and b, the 28- and 90-day survival rates for patients in the high RAR group were significantly higher than patients in the low RAR, according to a Kaplan–Meier analysis (P < 0.001). The relationship between RAR and outcomes was further analyzed using restricted cubic splines. After adjusting for confounding factors (age, active malignancy, chronic cerebrovascular diseases, SOFA score, CRP, and AST), spline analyses suggested a linear association between RAR and in-hospital mortality (Fig. 4, P-non-linear = 0.170).

Fig. 3

Survival curves of sepsis patients in ICUs at 28-day and 90-day follow-up. (a) 28-day mortality; (b) 90-day mortality. Abbreviations: RAR, red blood cell distribution width to albumin ratio

Fig. 4

Receiver operating characteristic (ROC) curves for in-hospital mortality in patients with sepsis in intensive care units. Abbreviations: RDW, red blood cell distribution width; RAR, RDW to albumin ratio

Predictive capabilities of RAR for mortality

Using ROC curves, the predictive capabilities of RAR, RDW, and albumin levels for in-hospital mortality were examined (Fig. 5). The results demonstrated that the area under the receiver operating characteristic curve (AUC) for RAR was substantially larger than that for RDW or albumin alone. Table 6 summarizes the AUC value, cut-off value, specificity, and sensitivity for predicting mortality from RAR, RDW, and albumin. The AUC value of RAR for predicting in-hospital mortality was 0.761 (P < 0.001), which was higher than that of albumin alone, which was 0.697 (P < 0.001), and RDW alone, which was 0.708 (P = 0.025). Similar conclusions were obtained from the ROC analysis examining the accuracy of RAR, RDW, and albumin in predicting 28- and 90-day mortalities (Figure S1). RAR’s AUC value for predicting 28- and 90-day mortalities was higher than that of albumin or RDW alone. The cut-off RAR values were 5.209%/g/dl (77.2% sensitivity and 66.0% specificity) to discriminate the in-hospital mortality, 5.223%/g/dl (80.4% sensitivity and 69.7% specificity) to discriminate the 28-day mortality, and 5.076%/g/dl (81.3% sensitivity and 76.9% specificity) to discriminate the 90-day mortality.

Fig. 5

Receiver operating characteristic (ROC) curves for in-hospital mortality in patients with sepsis in intensive care units

Table 6 Receiver operating characteristic curve analysis for mortality of RAR, RDW, albuminSubgroup and sensitivity analyses

We performed subgroup analyses by selecting subsets of patients based on age, sex, and the two most common sites of infection to assess the robustness of our findings. There was a significant interaction between age and RAR in regard to in-hospital mortality risk (p for interaction<0.001). The association of RAR with in-hospital mortality was stronger in the patient group younger than 65 years old (OR, 1.673; 95% CI, 1.259–2.224). The association between RAR and in-hospital mortality was not modified by sex (P for interaction = 0.168), pulmonary sepsis (P for interaction = 0.433), or hepatobiliary sepsis (P for interaction = 0.399). In each subgroup, RAR remained an independent risk factor associated with in-hospital mortality. ROC analysis performed on these subgroups showed that RAR had moderate diagnostic value. The relevant data from these analyses are summarized in Table 7.

Table 7 Receiver operating characteristic analysis and multivariate logistic regression analysis for the associations between in-hospital mortality and the RAR level among different patient subgroups

Human serum albumin levels may be affected by chronic renal or hepatic diseases. After removing individuals with chronic renal disease or chronic hepatic disease, another sensitivity analysis was performed (Table S2). These results were in line with the main finding, showing that RAR had a strong relationship with the in-hospital mortality (adjusted OR: 1.362; 95% CI: 1.107–1.676; P = 0.004), 28-day mortality (adjusted HR: 1.101; 95% CI: 1.017–1.192; P = 0.018), and 90-day mortality (adjusted HR: 1.144; 95% CI: 1.074–1.218; P<0.001).

Finally, case-wise deletion was used for missing values. Sensitivity analyses were conducted to determine whether our findings depended on how missing data were handled. We conducted regression analysis after excluding cases that had at least one missing variable. RAR and clinical outcomes remained strongly associated (Table S3).