The red cell distribution width to albumin ratio as a novel predictor of 180-day mortality in lung cancer patients.

Introduction
Lung cancer remains the deadliest malignancy worldwide, claiming around 1.8 million lives annually1–3. Despite advancements in screening and targeted therapies, most patients present with advanced disease where five-year survival plummets below 10%4–6. This grim prognosis is exacerbated during critical illness: nearly 60% of metastatic lung cancer patients requiring ICU admission die within six months7–9. Traditional early detection methods like low-dose CT effectively identify preclinical tumors but fail to stratify mortality risk during acute deterioration-a critical blind spot where timely intervention could alter trajectories10,11. Consequently, identifying accessible biomarkers that capture early mortality signals in decompensating patients represents an urgent unmet need.
Current prognostic approaches face dual limitations in this population. ICU severity scores (SOFA, APACHE II) excel at quantifying acute organ dysfunction but disregard cancer-specific drivers of long-term mortality: chronic inflammation fueling cachexia, immunosuppression, and metabolic collapse12,13. Meanwhile, isolated inflammatory markers like neutrophil-to-lymphocyte ratio (NLR) offer snapshot assessments but lack nutritional context and longitudinal stability14–16, decaying to AUC < 0.55 beyond 30 days17. This prognostic void necessitates composite biomarkers that integrate multidimensional pathophysiology through routinely available laboratory parameters.
The RDW-to-Albumin Ratio (RAR) emerges as a mechanistically grounded solution uniquely suited for lung cancer prognostication18,19. RDW elevation captures persistent inflammation-mediated erythropoietic dysfunction-driven by IL-6 overexpression in the tumor microenvironment20,21, while hypoalbuminemia quantifies PI3K/Akt-dependent proteolysis in cancer cachexia22,23. Critically, RAR’s synergy of these pathways enables unprecedented long-term predictive power: initial studies show AUC > 0.65 at 180 days versus < 0.60 for NLR or SOFA. Its stability stems from reflecting cumulative biological burden rather than transient physiological shifts, while operational simplicity (requiring only complete blood count and biochemistry) ensures clinical deployability across resource settings.
To definitively validate these advantages, this study leverages the MIMIC-IV database to investigate RAR’s performance in lung cancer ICU patients. We specifically assess its power to predict 180-day and 365-day mortality, benchmark against conventional scores across illness severity strata, establish actionable clinical thresholds, and elucidate interactions with cancer-specific comorbidities-addressing critical gaps in oncology-critical care prognostication.

Materials and methods

Database introduction
This retrospective cohort study utilized the MIMIC-IV 2.2 database, a de-identified electronic health record system capturing admissions to intensive care units (ICUs) at Beth Israel Deaconess Medical Center from 2008 to 2019. The repository comprises approximately 60,000 ICU admissions, with comprehensive documentation of demographic characteristics, medication records, laboratory results, vital signs, physiological measurements, and clinical outcomes.
Access authorization and data processing were managed by the third author, Guangdong Wang (Certification ID: 60106105), who completed the mandated training protocol and obtained database access credentials through standardized certification procedures. All data extraction and preprocessing operations adhered strictly to institutional ethical guidelines and data governance policies.

Population selection criteria
The cohort was derived from adult patients (≥ 18 years) with clinically validated lung cancer diagnoses (using International Classification of Diseases (ICD) codes) during their first recorded intensive care unit admission to ensure data independence (Supplementary Table S7). Individuals with ICU stays shorter than 24 h or lacking recorded measurements for red cell distribution width (RDW) and serum albumin were excluded. From an initial pool of 2,227 eligible lung cancer cases, 973 subjects fulfilled the selection criteria and constituted the analytical cohort (Fig. 1). Participants underwent stratification by RDW-to-albumin ratio (RAR) quartiles: Q1 (RAR < 4.21, (n = 241)), Q2 (4.21 ≤ RAR < 4.91, (n = 245)), Q3 (4.91 ≤ RAR < 6, (n = 243)), and Q4 (6 ≤ RAR < 14.29, (n = 244)).

Data extraction and RAR derivation
Patient-level variables were extracted from the MIMIC-IV database using PostgreSQL software. To standardize temporal relevance, only clinical parameters documented during the initial 24-hour window following ICU admission were retained. To establish baseline values and minimize the impact of ICU-specific interventions on laboratory parameters, the first recorded measurement for each variable within the initial 24 hours was utilized for analysis. The comorbidity “Malignant cancer” in the regression models refers to the Elixhauser/Charlson category for metastatic solid tumors and non-metastatic malignancies, and is distinct from the primary lung cancer diagnosis used to define the study cohort. We further screen for significant covariates by performing univariate Cox analyses among the meaningful covariates. Covariates exceeding a 20% missingness threshold were systematically excluded, with residual missing data addressed through multivariate imputation using the random forest imputation method via the ‘mice’ package in R. This method was selected for its ability to handle complex relationships between variables and accurately impute both categorical and continuous data. A total of 20 imputations were performed, with 5 iterations to allow for convergence, and a random seed (12345) was used to ensure reproducibility of the results (Supplementary Table S8).
Additionally, the RDW-to-albumin ratio (RAR) was mathematically derived as RDW (%)/serum albumin (g/dL), where both components underwent rigorous quality verification and necessary transformations prior to analytical utilization.

Outcomes
Predefined endpoints comprised long-term all-cause mortality outcomes, with the primary endpoint defined as death from any cause occurring within 180 days following the index hospital admission, and the secondary endpoint encompassing deaths recorded within 365 days post-admission. Mortality status was ascertained through institutional vital registry linkages, irrespective of discharge status.

Statistical analysis
Patients were categorized into quartile-based groups (Q1-Q4) according to RAR values for baseline assessments. Continuous variables were summarized as mean ± standard deviation (SD) for normally distributed parameters or median with interquartile range (IQR) for skewed distributions, while categorical variables were reported as frequencies with percentages. Appropriate statistical methods were selected based on data distribution characteristics: parametric tests for normally distributed continuous variables, nonparametric rank tests for non-normally distributed continuous variables, and categorical comparison tests for count data.
Potential predictors of 180-day mortality were initially identified through baseline characteristic analysis. Three sequential Cox proportional hazards models were developed: Model 1: Adjusted for demographic factors (Age, Gender, Race); Model 2: Model 1 covariates plus physiological parameters (Heart rate, SpO2) and severity scores SOFA; Model 3: Model 2 covariates supplemented by laboratory markers (Anion gap, BUN, AST, WBC), comorbidities (AKI, Malignant cancer, Chronic Pulmonary Disease, Diabetes), and interventions (Vasoactive Agents, Ventilator, CRRT). To assess the potential impact of BMI and infection-related diseases on the RAR indicator, Model 2 was adjusted for BMI, and Model 3 was further adjusted for sepsis and pneumonia (Supplementary Table S9). Then multicollinear variables (variance inflation factor [VIF] > 5) were eliminated. Restricted cubic spline (RCS) modeling evaluated non-linear associations.
Survival trajectories across RAR quartiles were compared via Kaplan-Meier curves with log-rank testing. The prognostic accuracy of RAR and SOFA score for the primary and secondary endpoints was assessed using standard receiver operating characteristic (ROC) curve analysis and quantified by the area under the curve (AUC) with 95% confidence intervals (95% CI). The optimal cut-off values for RAR from the ROC curves were determined by maximizing Youden’s index (J = sensitivity + specificity − 1).
Subgroup analyses were pre-specified to assess the consistency of the RAR-mortality association. Based on established clinical benchmarks defining substantial organ failure, patients were stratified by a SOFA score of 5 (SOFA < 5 vs. SOFA ≥ 5) to test for effect modification by illness severity.
All computations were executed in R (version 4.4.1) and Free Statistics (version 2.0), with statistical significance defined at two-sided P < 0.05.

Results

Patient characteristics
Significant clinical gradients emerged across RDW-to-albumin ratio (RAR) quartiles (Q1: <4.21; Q2: 4.21-4.91; Q3: 4.91-6; Q4: 6 ≤ RAR < 14.29) in our cohort of 973 ICU-admitted lung cancer patients. While demographic characteristics (age, gender, race) showed no intergroup differences (all P > 0.05), ascending RAR quartiles exhibited progressively worsening physiological derangements: Q4 patients demonstrated higher heart rates (median 97 vs. 89 bpm in Q1; P = 0.006), lower systolic (118 vs. 133 mmHg) and diastolic blood pressure (64 vs. 72 mmHg; both P < 0.001), and elevated respiratory rates (22 vs. 19/min; P < 0.001). Disease severity scores showed quartile-dependent escalation with APS III increasing from 42 (Q1) to 57 (Q4), SOFA from 3 to 5, and SAPS II from 37 to 42 (all P < 0.001). Laboratory parameters revealed critical biomarker derangements including declining albumin (3.7 g/dL in Q1 to 2.5 g/dL in Q4), elevated RDW (13.5% to 16.85%), and reduced hemoglobin (12.1 to 9.6 g/dL; all P < 0.001). For the primary endpoint of 180-day mortality, there were 483 death events and 490 censored cases. For the secondary endpoint of 365-day mortality, there were 575 death events and 398 censored cases. Clinical outcomes demonstrated striking RAR-dependent stratification with 180-day mortality rising from 31.1% (Q1) to 66.4% (Q4) and 365-day mortality from 43.2% to 73.4% (both P < 0.001), paralleled by increased intervention requirements including vasoactive agents (30.3% to 49.6%; P < 0.001) and CRRT utilization (1.2% to 7.0%; P = 0.003) (Table 1). Following Cox regression (Supplementary Table S1) and VIF (Supplementary Table S2), the multivariable Cox model was built with 17 covariates: patient demographics (age, gender, race); vital signs (heart rate, SpO2); illness-severity scores (SOFA); laboratory values (Anion gap, BUN, AST, WBC); comorbidities (AKI, malignant cancer, chronic pulmonary disease, diabetes); and therapeutic interventions (vasopressors, mechanical ventilation, CRRT).

Characteristics stratified by 180-day mortality status
Significant differences emerged between 180-day survivors (n = 490) and non-survivors (n = 483) in lung cancer ICU patients. Non-survivors were older (72.4 vs. 70.9 years, P = 0.026) with more severe physiological derangements: lower SpO₂ (95.7% vs. 96.6%, P < 0.001), higher heart rates (median 95 vs. 91 bpm, P = 0.008), and trends toward reduced diastolic BP (66 vs. 69 mmHg, P = 0.054). Critical illness scores were markedly elevated in non-survivors including APS III (55 vs. 45.5), SOFA (4 vs. 4), SAPS II (42 vs. 37), and OASIS (36 vs. 33) (all P < 0.001). Laboratory profiles revealed profound disparities: non-survivors exhibited higher RDW (15.4% vs. 14.6%, P < 0.001), lower albumin (2.9 vs. 3.2 g/dL, P < 0.001), reduced hemoglobin (10.3 vs. 11.1 g/dL, P < 0.001), elevated INR (5.3 vs. 4.6, P < 0.001), and prolonged PT (13.7 vs. 13.4s, P = 0.003). They also demonstrated higher BUN (23 vs. 19 mg/dL, P < 0.001), Anion gap (14 vs. 14 mmol/L, P = 0.016), and WBC counts (10.9 vs. 9.6 K/µL, P = 0.001). Comorbidity analysis showed non-survivors had higher malignant cancer prevalence (64.0% vs. 38.2%, P < 0.001) but lower chronic pulmonary disease (61.1% vs. 70.2%, P = 0.003). The prevalence of pneumonia and sepsis was significantly higher among the 180 deceased lung cancer patients (P < 0.05). No significant differences existed in interventions (vasoactive agents, CRRT, ventilator) (all P > 0.05) (Table 2).

Association between RAR and risk of mortality
RAR exhibits a significant association with 180-day and 365-day mortality across all models as a continuous or categorical variable (Table 3). As a continuous variable, RAR was strongly associated with 180-day mortality (HR = 1.11, 95%CI 1.05–1.18, P < 0.001) and 365-day mortality (HR = 1.10, 95%CI 1.04–1.16, P < 0.001) after comprehensive adjustment. When analyzed by quartiles, patients in the highest RAR quartile (Q4: 6 ≤ RAR < 14.29) exhibited increased 180-day mortality risk (HR = 1.2, 95%CI 1.14–1.27, P < 0.001) and 365-day mortality risk (HR = 1,19, 95%CI 1.13–1.25, P < 0.001) compared to Q1 (reference).

Kaplan–Meier survival analysis and RCS
Kaplan-Meier curves demonstrated a graded reduction in survival probability across ascending RAR quartiles. Patients in the highest RAR quartile (Q4: 6 ≤ RAR < 14.29) exhibited the most pronounced mortality risk at both 180-day and 365-day (Fig. 2A,B).

In addition, as shown in Fig. 3A,B, RCS analysis established a significant linear relationship between continuous RAR values and mortality risk at both 180-day (P < 0.001) and 365-day endpoints (P < 0.001).

Prediction of all-cause mortality by RAR
The ROC curves compare the predictive performance of RAR and clinically SOFA scores (Fig. 4A,B). It could be seen RAR exhibited a superior predictive property to SOFA for both 180-day mortality (AUC 0.65 (95%CI 0.62–0.69) vs. 0.56 (95%CI 0.53–0.60)) and 365-day mortality (AUC 0.64 (95%CI 0.60–0.67) vs. 0.55 (95%CI 0.51–0.59)) (Table 4).

Subgroup analysis of RAR-mortality association
Forest plots demonstrated consistent prognostic utility of RAR across most subgroups for 180-day and 365-day mortality in critically ill lung cancer patients, with significant effect modification observed in specific subpopulations (Fig. 5A,B, Supplementary Tables S3, S4). The association between RAR elevation and mortality risk remained robust in age-stratified groups (< 65y HR = 1.16; ≥ 65y HR = 1.12 for 180-day) and comorbidity subgroups (malignant cancer HR = 1.14; chronic pulmonary disease HR = 1.11), with no significant interactions (all P > 0.05). Disease severity substantially attenuated RAR’s predictive capacity at both endpoints (180-day P = 0.006; 365-day P = 0.011), where RAR lost prognostic significance in high-severity patients (SOFA ≥ 5: HR = 1.07, 0.99–1.15 for 180-day). This SOFA-dependent effect modification indicates RAR functions primarily as an early-risk biomarker, with its predictive utility diminishing when multi-organ failure develops.

Discussion
This study robustly validates the RAR as a critical predictor of long-term mortality in critically ill lung cancer patients, extending beyond conventional severity scores. Our findings demonstrate that elevated RAR independently stratifies mortality risk at both 180 and 365 days, with a clear dose-response gradient across quartiles. The highest RAR quartile (6 ≤ RAR < 14.29) exhibited over twice the mortality risk of the lowest quartile, underscoring its potent discriminative capacity. Notably, RAR’s predictive accuracy (AUC 0.65 for 180-day mortality) surpassed the widely adopted SOFA score (AUC 0.56), highlighting its potential as a superior prognostic tool to SOFA in this vulnerable population.
The ability of RAR to forecast long-term outcomes (180/365 days) addresses a significant gap in oncology-critical care prognostication. Traditional ICU scores like SOFA prioritize short-term organ dysfunction, often overlooking cancer-specific pathophysiology24. In lung cancer, where chronic inflammation and malnutrition drive progressive decline25,26, RAR encapsulates these processes through RDW (reflecting oxidative stress and ineffective erythropoiesis) and albumin (indicating catabolism and hepatic dysfunction). This aligns with emerging paradigms emphasizing composite biomarkers for cancer survival prediction, as seen in gastric and colorectal malignancies where similar ratios outperform single parameters27,28. Our results thus affirm RAR’s unique relevance over extended timelines in advanced cancer.
The selection of RAR as a core biomarker is mechanistically grounded in the interplay between inflammation and nutritional depletion in lung cancer. RDW elevation signifies chronic inflammation-mediated erythropoietic suppression, while hypoalbuminemia results from cytokine-driven catabolism (e.g., IL-6-induced hepatic reprioritization)29–31. Their ratio amplifies prognostic sensitivity, as demonstrated in sepsis and heart failure cohorts32,33. Notably, our study extends this rationale to lung cancer ICU patients-a population with heightened inflammatory-nutritional dysregulation due to tumor burden and critical illness synergism. This biological plausibility strengthens RAR’s validity beyond statistical association, positioning it as a pathophysiologically anchored indicator.
A pivotal finding emerged in subgroup analyses: RAR’s prognostic power diminished significantly in high-severity patients (SOFA ≥ 5; P = 0.006 for 180-day). This suggests RAR primarily identifies early-risk patients before multi-organ failure dominates the clinical trajectory. In such advanced stages, organ-specific scores may better capture immediate mortality drivers. Importantly, RAR maintained robust prediction across age and comorbidity subgroups, confirming its broad applicability. These nuances refine its clinical utility. RAR excels in early risk stratification (e.g., triggering preemptive nutrition/anti-inflammatory interventions) but may defer to SOFA in terminal organ failure.
We assessed the predictive value of the RAR for long-term mortality in critically ill lung cancer patients. Our results show that RAR significantly outperforms the SOFA score, with higher AUCs at both 180 days (0.65 vs. 0.56) and 365 days (0.64 vs. 0.55). However, while the AUC difference is statistically significant, the clinical relevance of a 0.09 AUC difference is modest and may only provide slight improvements in patient risk stratification. This modest gain in discrimination suggests that while RAR can enhance predictive accuracy, its impact on clinical decision-making should be considered in context. The strength of RAR lies in its ability to integrate inflammation (via RDW) and nutritional status (via albumin), two critical factors in lung cancer progression that are often not captured by traditional severity scores like SOFA. Future studies should also explore combining RAR with other clinical indicators or severity scores, such as lactate, APACHE II, or SAPS II, to improve the model’s predictive power. Such a combination could offer a more holistic view of a patient’s condition, enhancing the model’s clinical utility in the ICU setting. Overall, our findings support the potential of RAR as an accessible and effective tool for predicting mortality in lung cancer patients, with opportunities for further refinement and validation to maximize its clinical applicability.
Clinically, RAR’s quartile-based stratification provides actionable insights for ICU management. Patients in the highest RAR quartile (6 ≤ RAR < 14.29) face > 65% 180-day mortality risk, a threshold that may warrant more intensive monitoring and prompt a comprehensive evaluation of their nutritional and inflammatory status. The metric’s simplicity facilitates rapid risk stratification at admission. Whether this stratification can guide targeted interventions, such as personalized nutritional support or anti-inflammatory strategies, to improve outcomes constitutes a crucial area for future prospective trials.
A key strength of our study lies in the rigorous statistical approach undertaken to validate the independence of RAR as a prognostic marker. Specifically, in direct response to reviewer comments, we conducted additional analyses to account for potential confounding from nutritional status and critical illness. Even after sequential adjustment for BMI and a spectrum of infection-related comorbidities—including sepsis, pneumonia, and others—in our multivariate models, a higher RAR remained a statistically significant and independent predictor of long-term mortality. This persistence of the RAR-mortality association, despite controlling for these potent clinical variables, substantially strengthens the argument that RAR is not merely a surrogate for underlying nutritional deficit or infection severity but possesses its own unique prognostic value. By proactively testing and confirming the robustness of our findings against these potential confounders, we enhance the validity and clinical relevance of our conclusion.
Our study has limitations inherent to retrospective designs. First, its retrospective design from a single database means that unmeasured confounders (e.g., detailed cancer stage, specific treatment regimens) may persist despite multivariable adjustment. Second, the absence of serial RAR measurements precluded the assessment of dynamic trends and their relationship with outcomes. Third, and specific to the ICU setting, the interpretation of serum albumin is complex; factors such as fluid resuscitation and capillary leak can cause hemodilution and extravasation, potentially influencing albumin levels independently of the patient’s chronic nutritional-inflammation status and thereby acting as a confounding factor for RAR. Fourth, our cohort was exclusively composed of lung cancer patients. The lack of control cohorts, such as non-cancer ICU patients with sepsis, gastrointestinal tumors or autoimmune diseases, limits our ability to claim that RAR is a specific biomarker for cancer-related outcomes versus a general marker of critical illness severity. Fifth, excluding patients with ICU stays shorter than 24 h may introduce bias by potentially removing early deaths, which could result in a lower observed long-term mortality rate. This exclusion may have enriched the cohort with patients who had more complete clinical data. Future studies should consider including these patients or conducting sensitivity analyses to assess the impact of this exclusion on mortality outcomes. Sixth, despite the recognized importance of cancer cachexia as a key driver of hypoalbuminemia, our study was unable to adjust for direct measures of nutritional status or body composition, such as body mass index (BMI), due to a high proportion of missing height and weight data in the source database. This is an important limitation, as the inability to account for cachexia may confound the relationship between RAR and mortality. Future prospective studies designed to validate RAR should prioritize the systematic collection of anthropometric data to disentangle the specific contributions of cachexia and inflammation to the prognostic value of this ratio. Future prospective, multi-center studies that include such control groups, incorporate detailed cancer-specific variables, and feature serial biomarker measurements are needed to validate our findings, clarify RAR’s specificity, and explore its utility in guiding interventions.

Supplementary Information
Below is the link to the electronic supplementary material.