Introduction

Sepsis is a critical medical state involving organ failure resulting from a maladaptive systemic reaction to infection that impacts millions of patients worldwide annually [1]. Despite recent advancements in sepsis recognition and treatment strategies, mortality rates remain alarmingly high and range from 16.7% to 33.3% [2,3], imposing a substantial burden on global healthcare economics. Nevertheless, early identification and intervention have led to significant improvements in clinical outcomes.

Current assessment methods predominantly rely on manual documentation or electronic medical records, which often incorporate nursing process-dependent parameters such as vital sign monitoring [4–7]. Variations in measurement and recording protocols across healthcare institutions may lead to substantial discrepancies in evaluation results. Although the levels of routinely available inflammatory markers such as procalcitonin and C-reactive protein are easily determined in clinical practice, they exhibit limited prognostic value for sepsis because of their suboptimal specificity and sensitivity [1,8]. Although conventional scoring systems such as sequential organ failure assessment (SOFA) and acute physiology and chronic health assessment scoring system II (APACHE II) demonstrate high prognostic value, their complex assessment procedures and poor operational feasibility limit their widespread clinical application [9–11]. Consequently, the pursuit of more objective and convenient prognostic indicators remains important.

Neutrophils (NEUT) and lymphocytes (LYMPH), as the principal cellular components of innate and adaptive immunity, respectively, play indispensable roles as first-line immune defenses during systemic inflammatory responses [12,13]. Bacterial or fungal infections typically induce isolated NEUT elevation with concomitant lymphopenia, whereas viral infections or malignancies may predominantly increase LYMPH counts [14]. Consequently, the neutrophil-to-lymphocyte ratio (NLR) generally increases during inflammatory states. Although clinically favorable because of its accessibility and low cost, some alterations in the NLR are not inflammation-specific; for example, acute trauma, stroke, myocardial infarction, and postoperative complications can similarly modify this ratio, which compromises its diagnostic specificity and contributes to persistent clinical skepticism regarding its utility [15]. The COVID-19 pandemic has reinvigorated research interest in the NLR [16,17], particularly as a result of a recent meta-analysis by Wu et al. [18], which reported reliable prognostic performance of the NLR in sepsis despite considerable between-study heterogeneity (I² = 87.2%, 95% CI 79.5–92; p < 0.0001). Moreover, a retrospective study conducted by Schupp et al. [19] failed to demonstrate a discriminative capacity for 30-day mortality. These contradictory prognostic interpretations in sepsis underscore the need for machine learning (ML)-based reevaluations of the ability of the NLR to predict mortality.

Machine learning algorithms have emerged in recent years as powerful tools in clinical prediction models. The 2021 sepsis guidelines highlight the superior discriminative capacity of ML algorithms for predicting in-hospital mortality in sepsis patients [1]. These algorithms offer dual advantages—the efficient management of missing data and the ability to integrate weak predictive features into robust models through data-driven learning. Among the various ML techniques, extreme gradient boosting (XGBoost) has demonstrated exceptional data mining and predictive modeling capabilities. Its dominance is evidenced by its adoption in 17 of 29 winning solutions in the 2015 Kaggle competitions and its exclusive use by all top 10 teams that participated in the 2015 KDD Cup. Several studies have validated the clinical utility of XGBoost: Dong et al. developed a ML model that predicts pediatric acute kidney injury 48 hours in advance by analyzing premorbid physiological measurements and incorporating current guidelines; Zhang et al. established an XGBoost model that effectively distinguishes between fluid-responsive and nonresponsive sepsis patients; and Yue et al. created a ML model for the early detection of sepsis-associated acute kidney injury and identified XGBoost as the optimal predictive algorithm. These findings collectively support the theoretical potential of ML algorithms to enhance predictive model development and validation in critical care. However, no studies have yet employed XGBoost modeling with large datasets to systematically evaluate the ability of the NLR to predict 28-day mortality, which could provide valuable insights into its clinical relevance under resource-constrained conditions.

Methods

Study population

Inclusion criteria for this study were as follows: (1) aged between 18 and 85 years; (2) an ICU stay duration that exceeded 24 hours; and (3) suspected or confirmed infection with a SOFA score ≥2. For patients with multiple ICU admissions, only data obtained from the first ICU stay were included in the analysis. Patients were excluded if they met any of the following conditions: (1) had missing laboratory data, including NEUT, LYMPH and platelet (PLT) data; or (2) had a documented diagnosis of malignant tumors or autoimmune diseases. The study design is schematically illustrated in Fig 1.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Overview of the study design and workflow.

BIDMC, Beth Israel Deaconess Medical Center; MIMIC-IV, Medical Information Mart for Intensive Care IV; XGBoost, extreme gradient boosting; NLR, neutrophil-to-lymphocyte ratio; RCS, restricted cubic splines; ROC, receiver operating characteristic; PR, precision-recall; SHAP, Shapley additive explanations.

https://doi.org/10.1371/journal.pone.0348676.g001

Results

Baseline characteristics

In this study, 4,376 eligible patients from the MIMIC-IV database (with 28-day, in-hospital, and ICU mortality rates of 18.40%, 17.50%, and 14.12%, respectively) were ultimately enrolled. The demographic characteristics, vital signs, laboratory tests, organ dysfunction assessments, treatments and other outcomes are summarized in Table 1. Compared with nonsurvivors, survivors were significantly younger (59.77 ± 14.97 vs. 63.37 ± 15.10; p < 0.05), were more likely to be male (63.93% vs. 59.25%), and had lower disease severity, as reflected by SOFA (3.76 ± 1.96 vs. 4.99 ± 2.71) and SAPS II scores (36.79 ± 12.90 vs. 51.74 ± 15.76). With respect to vital signs, survivors demonstrated lower HR (89.48 ± 17.68 vs. 95.55 ± 20.99 bpm; p < 0.001), lower RR (19.07 ± 5.54 vs. 22.11 ± 6.29 breaths/min; p < 0.001), and higher percutaneous oxygen saturation (SpO₂, median: 98.57% vs. 97.00%; p < 0.001) than nonsurvivors, whereas MAP did not significantly differ between the groups (81.58 ± 15.22 vs. 81.81 ± 19.81 mmHg; p = 0.755). Although survivors demonstrated marginally better vital signs (all p < 0.05 except SpO₂), the intergroup differences were clinically modest. Laboratory analyses revealed significantly elevated inflammatory markers, including the NLR (6.97 [IQR 4.11–12.59] vs. 12.70 [IQR 6.92–22.75]; p < 0.05) and the SII, alongside more pronounced acidotic patterns on blood gas analysis, in nonsurvivors. Higher creatinine and blood urea nitrogen (BUN) levels suggested increased renal dysfunction in nonsurvivors, which is consistent with the results of the organ dysfunction assessments. Erythrocyte and hemoglobin-related parameters significantly differed between the two groups. Organ dysfunction distributions revealed greater respiratory and coagulation dysfunction in survivors, whereas nonsurvivors had greater hepatic, neurological, and renal failure rates (all p < 0.05). Comorbidity analyses revealed significant differences in hypertension, heart failure, chronic obstructive pulmonary disease, chronic kidney disease, atrial fibrillation, cerebral infarction, cerebral hemorrhage, and thrombosis. Fluid resuscitation metrics indicated minimal crystalloid administration within the first 3 ICU hours in both groups, with survivors receiving less crystalloid. Although statistically significant differences were observed in weight-adjusted crystalloid ratios, these differences lacked clinical relevance. Analysis of 24-hour fluid balance revealed higher total intake and output volumes in survivors than in nonsurvivors, whereas nonsurvivors exhibited greater net positive fluid balance, suggesting potential fluid overload in this group. Nonsurvivors required more MV, CRRT, and vasopressor support (all p < 0.05), whereas survivors had higher rates of antibiotic initiation within 1 hour of ICU admission. Although survivors had shorter ICU stays, their total hospitalization duration exceeded that of nonsurvivors.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Baseline characteristics of the survivor and nonsurvivor groups.

https://doi.org/10.1371/journal.pone.0348676.t001

On the basis of the overall NLR distribution and quartile values, patients were categorized into low (<4.34), intermediate (4.34–14.70), and high (>14.70) NLR groups (as shown in Table 2). All mortality outcomes, including 28-day, hospital, and ICU mortality, significantly increased in a stepwise manner with increasing NLR (p < 0.05) until they reached approximately 30% in the high-NLR group. Similarly, both hospital and ICU length of stay increased in a concentration-dependent manner (p < 0.05). No statistically significant differences were observed in mean corpuscular hemoglobin (MCH) levels, cardiovascular failure incidence, or the incidence of comorbidities such as diabetes, atrial fibrillation, or cerebral infarction. However, all other parameters that demonstrated significant differences between the survivor and nonsurvivor groups maintained concentration-dependent statistical significance across NLR stratifications (p < 0.05).

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Characteristics and outcomes of patients categorized by NLR.

https://doi.org/10.1371/journal.pone.0348676.t002

Selection of features in the models

The results from a single Boruta run are presented in S1 Fig, whereas Fig 2A displays stability-selected features (SF ≥ 0.9) from 10 iterations; this process identified 37 confirmed and 2 tentative features, whereas the remaining features were rejected. These 39 features were subjected to XGBoost modeling, which was preceded by collinearity analysis (Pearson’s r ≥ 0.7 exclusion) and clinical relevance screening (S2 Fig and S3 Table), thus yielding 29 final parameters. As shown in S4 Table, the optimized XGBoost model achieved a test-set ROC-AUC of 0.875 (95% CI: 0.854–0.896) and a PR-AUC of 0.603 (95% CI: 0.534–0.671), thereby outperforming the individual NLR, as well as the SAPS II and SOFA scores, for 28-day mortality prediction (Fig 3A, 3B). The data in Fig 2B reveal 13 important parameters, and Model 3 ultimately incorporated SAPS II and SOFA scores, as well as 24-hour fluid output, VIS, 24-hour fluid balance, dialysis type, cerebral infarction, atrial fibrillation, red cell distribution width (RDW), partial pressure of oxygen (PO₂), hematocrit, creatinine, LYMPH, base excess (BE), and MCH. S5 Table presents the predictive contribution ranking of the NLR in the XGBoost model for 28-day mortality prediction. The NLR ranked 8th in global importance but decreased to 14th in terms of the SHAP-based marginal contribution, which indicates its relatively low individual impact and potential synergistic interactions with other parameters.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Application of machine learning algorithm for feature selection.

A Feature stability selection for the relationship between the NLR and 28-day mortality was analyzed by using the Boruta algorithm. The x-axis lists parameters (actual and shadow features), whereas the y-axis shows noise-adjusted importance scores from 10 Boruta runs. Features are color-coded by stability: light blue (confirmed, SF ≥ 90%), green (tentative, 50% ≤ SF < 90%), red (rejected, SF < 50%), and gray (shadow features). A red dashed line marks the stability threshold. B SHAP analysis of XGBoost model predictions. Features are vertically ranked by the mean absolute SHAP value (light blue bars), with individual SHAP values (colored dots) distributed horizontally and colored by normalized feature magnitude (blue: low, red: high). The gray dashed line marks the baseline.

https://doi.org/10.1371/journal.pone.0348676.g002

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Comparative analysis of predictive models and survival stratification by the NLR on 28-day mortality.

Panel A: receiver operating characteristic (ROC) curves demonstrate the discriminative performance of the NLR, SAPS II, SOFA, and XGBoost; Panel B: precision-recall (PR) curves assessing the trade-off between precision and recall; Panel C: Kaplan-Meier survival analysis stratified by NLR quartiles.

https://doi.org/10.1371/journal.pone.0348676.g003

NLR and 28-day, hospital, and ICU mortality

Multivariable logistic regression analysis demonstrated that the standardized NLR was independently associated with 28-day mortality (OR 1.16; 95% CI 1.06–1.27; p < 0.001), in-hospital mortality (OR 1.13; 95% CI 1.03–1.24; p < 0.001), and ICU mortality (OR 1.14; 95% CI 1.03–1.25; p = 0.008) after full adjustment (Model 3, Table 3). Unadjusted analysis by NLR quartiles revealed a dose‒response relationship with all of the mortality endpoints (S3A–C Fig), and stronger associations were observed with increasing NLR values. In the fully adjusted model (Model 3), patients in the highest NLR quartile exhibited significantly higher risks of 28-day mortality (OR 1.95, 95% CI 1.40–2.75), in-hospital mortality (OR 1.90, 95% CI 1.34–2.70), and ICU mortality (OR 1.79, 95% CI 1.23–2.62) compared to those in the lowest quartile, and all of the trend tests were statistically significant (p < 0.001).(Fig 3C) temporally corroborated these findings, which demonstrates early divergence among groups by day 7 (log-rank p < 0.0001), whereas the high-NLR group exhibited markedly decreased survival. Restricted cubic spline analysis (S4 Fig) demonstrated biphasic nonlinear relationships between the NLR and all of the mortality endpoints, with Wald tests confirming significant nonlinear effects across all of the models (P-nonlinear <0.05). The ROC-derived threshold (NLR = 7.3) coincided with both the median NLR value and the first inflection point, where the predicted probabilities sharply increased. This predictive probability plateaued near the second inflection point (NLR = 27.3) and increased marginally from 0.15 to 0.30, even though sample representation markedly declined in this range. Beyond the second inflection point, statistical reliability diminished because of sparse data.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. The association between NLR groups and 28-day mortality, hospital mortality and ICU mortality.

https://doi.org/10.1371/journal.pone.0348676.t003

Subgroup analysis

To further evaluate the independent association between the standardized NLR and mortality outcomes, including 28-day, in-hospital, and ICU mortality, we conducted stratified analyses and interaction tests on Model 3 covariates. As demonstrated in Fig 4 and S6 Table, the NLR maintained consistent associations with mortality, with increased predictive value being observed in patients aged >45 years, those with SOFA scores ≤4 and those with SAPS II scores > 29. Those individuals ≥65 years exhibited a 17% increased 28-day mortality risk (P = 0.019), and those with a SOFA score ≤4 exhibited a > 20% increased risk across all of the mortality endpoints (p < 0.001), whereas no significant association was observed in the subgroup of patients with SOFA scores ≥9 (p = 0.369). Significant effect modifications (interaction p < 0.05) were identified in patients with a 24-hour fluid output of 1475–3420 mL, a BE > -2, a HR < 100 bpm, or a RR of 12–20 breaths/min, and similar trends were observed in those with a 24-hour fluid balance ≤2493.27 mL, an MCH ≥ 27 pg, a MAP of 65–110 mmHg, or a creatinine concentration of 1.5–3.0 mg/dL. Although the standardized NLR demonstrated strong associations with all three mortality outcomes in patients with an HR < 60 bpm, the limited sample size within this stratified subgroup resulted in imprecise estimates, thus warranting cautious interpretation. The subgroup analyses consistently demonstrated significant associations between the NLR and all three mortality endpoints across most predefined strata, which confirms the robustness of the predictive value of this biomarker in diverse patient populations.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Forest plot of subgroup analyses for the association between the standardized NLR and mortality outcomes.

The plot displays odds ratios (ORs) with 95% confidence intervals (CIs) for 28-day, in-hospital, and ICU mortality across clinically relevant subgroups. Solid squares represent point estimates of ORs; horizontal lines indicate 95% CIs. The vertical dashed line denotes the null effect (OR = 1).

https://doi.org/10.1371/journal.pone.0348676.g004

Discussion

Machine learning algorithms have increasingly been employed to refine prognostic assessment in sepsis using large-scale ICU databases. For example, Lou et al. [29] and Zheng et al. [30] recently applied XGBoost based on the MIMIC database to explore the association between the triglyceride-glucose (TyG) index and risk of death in sepsis patients, thus further confirming the potential of ML in routine biomarker re-evaluation. Within this evolving landscape, as an accessible and widely recorded inflammatory marker, the NLR has garnered renewed interest; however, its independent prognostic value for short-term mortality remains debated, and systematic ML‑driven re-evaluations remain scarce. In the present study, we leveraged a rigorous two-stage feature selection pipeline combining the Boruta algorithm with XGBoost to specifically evaluate the ability of the early NLR to predict 28‑day mortality in a large, real‑world sepsis cohort. Our optimized model achieved robust discrimination (ROC‑AUC 0.875; PR‑AUC 0.603), thus outperforming conventional severity scores. More importantly, after full multivariable adjustment, each 1‑SD increase in the standardized NLR was independently associated with a 16%, 13%, and 14% higher risk of 28‑day, in‑hospital, and ICU mortality, respectively. Dose‑response analyses confirmed a progressive mortality increase with increasing NLR, whereas RCS revealed a nonlinear, biphasic pattern; specifically, predictive capacity became increased near the first inflection point (NLR = 7.3, approximating the optimal threshold) and plateaued beyond the second inflection point (NLR = 27.3). Subgroup analyses further elucidated that the NLR-mortality association was most pronounced in hemodynamically stable patients and those with low organ failure scores (SOFA ≤ 4), whereas it lost statistical significance in the setting of advanced multi‑organ dysfunction (SOFA ≥ 9). Collectively, these findings do not merely replicate prior prognostic observations; rather, they delineate (for the first time) a phenotype‑specific, biphasic predictive pattern of the NLR that distinguishes inflammation‑driven mortality risk from organ failure‑driven mortality. By integrating state‑of‑the‑art feature selection with granular clinical stratification, this study provides a methodologically rigorous and clinically nuanced re-evaluation of the NLR’s prognostic utility in sepsis.

Our XGBoost model achieved a test-set ROC-AUC of 0.875 (PR-AUC 0.603) for 28-day mortality prediction, thus representing a substantial improvement over conventional severity scores and individual biomarkers. To contextualize this performance within the broader landscape of emerging prognostic indices in sepsis, we compared our findings with recent studies evaluating the TyG index and the creatinine-to-albumin ratio (CAR) using the MIMIC database [29–31]. Both TyG-based studies reported significant associations with mortality, with Lou et al. [29] identifying a nonlinear relationship (inflection point: TyG = 8.9) and Zheng et al. [30] describing a linear dose-response pattern. The CAR study by Lou et al. [31] demonstrated a J-shaped association with 30-day mortality in sepsis-associated acute kidney injury, with an inflection point at CAR = 1.2 mg/dL and an AUC of 0.68–0.75 being reported. Notably, our NLR-based XGBoost model achieved higher discriminative performance (AUC 0.875) than any of these reported models, which likely reflects the synergistic integration of multiple clinical and laboratory features through a robust two-stage feature selection pipeline. Beyond performance metrics, a key methodological parallel involves the nonlinear risk patterns shared across these biomarkers. The J-shaped CAR-mortality curve and the biphasic NLR-mortality relationship observed in our cohort both suggest threshold effects that demarcate distinct pathophysiological states, including renal dysfunction with systemic inflammation for CAR and innate-adaptive immune imbalance for the NLR. However, the NLR offers distinct practical advantages; specifically, it is derived from a complete blood count without the need for additional metabolic assays, incurs no extra cost, and directly reflects the hyperinflammation–immunosuppression disequilibrium that is a central process to sepsis immunopathology [15,17,18,32]. Thus, although the TyG and CAR capture insulin resistance and renal-inflammatory crosstalk, respectively, the NLR (particularly when embedded in a ML framework) provides complementary prognostic information with superior discrimination and unique biological interpretability. These findings demonstrate the NLR not as a replacement for existing biomarkers; rather, it can function as a phenotype-specific adjunct for early risk stratification in moderate sepsis, where its predictive value is most pronounced.

Recent five-year evidence has consistently demonstrated that an elevated NLR increases sepsis mortality risk by 10–40% across general populations [33–42]. More critically, when high-NLR cohorts were compared with their low-NLR counterparts, multivariable-adjusted analyses revealed that the high-NLR group had an 80% greater mortality risk, which establishes a robust dose‒response relationship. Three studies [39,41,42] employing RCS analysis confirmed nonlinear NLR-mortality relationships and demonstrated progressive risk escalation until the NLR reached a range of 15–20, beyond which the trend plateaued. Prior evaluations of the discriminative capacity of the NLR universally relied on ROC curves, with AUROC values ranging from 0.6 to 0.7, which are consistent with our findings and indicate moderate discrimination. However, given the imbalanced outcome distribution (18.4% 28-day mortality in our cohort), ROC metrics may overestimate performance because of excessive true-negative influence, which potentially masks true positive-class identification capacity [32,43,44]. We therefore performed ROC analyses with PR curves and confirmed that the NLR, SOFA score, and SAPS II score all exhibit limited mortality discrimination under outcome imbalance. XGBoost modeling that incorporated multiple parameters significantly improved both the AUC and average precision values. While the NLR demonstrated a substantial overall model contribution, SHAP analysis revealed marginal individual predictive utility, which suggests limited direct mortality risk stratification capacity. Combined NLR stratification and subgroup analyses revealed potential interaction effects, particularly in hemodynamically stable patients for whom the correlations between the NLR and mortality remained robust.

Returning to our primary research question, the limited predictive value of the NLR for 30-day mortality observed in the study by Schupp et al. [19] may stem from two methodological factors. First, the exclusively enrolled critically ill patients demonstrated uniformly high disease severity (mean SOFA score of 11 and mean APACHE II score of 23), and no significant differences were observed in illness severity between survivors and nonsurvivors. Second, the small sample size may have compromised the stability of the results. Conversely, studies [33,35,36,38,39] that support the ability of the NLR to predict sepsis mortality but that demonstrate superior discriminatory performance for 28-day mortality (with the AUROC value being observed at 0.827 in some reports) have predominantly analyzed cohorts with extreme mortality rates (30–70%) and significant intergroup differences in illness severity scores. Notably, three additional studies reported the limited predictive value of the NLR [37,40,45] after patients with critical illness or hemodynamic instability were excluded through subgroup or sensitivity analyses. These collective observations suggest that the ability of the NLR to predict mortality may be restricted to clinically stable conditions, while its utility appears to decrease in severe cases with concomitant organ dysfunction.

The context‑dependent nature of the NLR’s prognostic value observed in our study aligns with emerging evidence on other composite biomarkers that integrate inflammatory and nutritional dimensions. Recent studies investigating the NEUT count‑to‑prognostic nutritional index ratio (NPNR) in septic patients have reported J‑ or V‑shaped associations with mortality [46,47], thus mirroring the biphasic pattern that we identified for the NLR. Both the NLR and NPNR appear to capture the delicate balance between host immune activation and physiological reserve; specifically, their predictive weight diminishes when organ failure becomes the dominant driver of the outcome. This convergence across distinct biomarker constructs reinforces the notion that inflammatory markers are not universal prognostic tools; rather, they are phenotype‑specific indicators exhibiting a utility that is most pronounced in moderate disease and wanes in the setting of terminal organ dysfunction. Therefore, future research should focus on context‑aware risk stratification that accounts for the dynamic interplay between inflammation, nutrition, and organ function.

Mechanistically, the increase in the NLR in sepsis patients precisely mirrors the dual pathological processes of hyperactivated NEUT and apoptotic LYMPH depletion; thus, the NLR serves as a dynamic biomarker of the “hyperinflammation–immunosuppression” imbalance that characterizes sepsis immunopathology [15,45]. Current evidence indicates that delayed NEUT apoptosis leads to systemic accumulation, whereas accelerated CD4+/CD8+ T-cell apoptosis and impaired NK cell function collectively contribute to this dysregulated state. Notably, the magnitude of NLR elevation may reflect the depletion of anti-inflammatory reserves, and higher values may indicate a poorer capacity to counteract inflammatory insults [45]. Clinically, the NLR demonstrates robust predictive value for moderate sepsis (SOFA scores <9), as evidenced by the finding reported by Liu et al. [37] that serial NLR measurements (AUC = 0.823) outperformed single assessments, as a day-7 NLR > 4.18 predicted significantly increased 28-day mortality. Ye et al. [40] further confirmed that an NLR > 20.25 was an independent predictor of mortality (hazard ratio = 1.22) in large cohorts. However, its ability to predict critical illness (SOFA score ≥11) is diminished, which is likely due to the predominance of multiorgan failure observed in terminal pathophysiology [40]. The primary sources of interstudy variability include discrepancies in measurement timing, divergent threshold selection criteria, and confounding effects of therapeutic interventions such as blood purification. Notably, baseline NLR values inherently differ between specific disease populations and healthy individuals. Karakonstantis et al. [48] reported that multiple factors potentially contribute to nonpathological NLR elevation, including advanced age, exogenous steroid administration, elevated endogenous hormone levels, active hematologic disorders, type 2 diabetes mellitus, and acute trauma. Future research should prioritize three objectives: standardizing measurement protocols, establishing population-specific cutoffs, and investigating combination models with emerging immune biomarkers. Fundamentally, the clinical value of the NLR lies in its noninvasive nature and ability to track immune trajectory evolution rather than as a standalone prognostic determinant.

This study implemented multiple methodological innovations in its investigation of sepsis outcomes. Given the well-documented prognostic influences of key sepsis interventions, including fluid resuscitation, antibiotic therapy, MV, and blood purification [1], we developed advanced data extraction protocols to obtain and systematically categorize relevant parameters from the MIMIC-IV database. Through stepwise feature selection, the analysis consistently identified fluid resuscitation and blood purification as clinically significant factors, while numerous other variables were excluded. A planned evaluation of crystalloid fluid administration effects within the initial three hours after ICU admission was determined to be impractical because of insufficient recorded intravenous infusion volumes. The derived weight-adjusted ratios fell below clinically meaningful thresholds, demonstrated minimal predictive value for mortality outcomes, and were consequently excluded during collinearity analysis and feature selection procedures.

Several limitations must be acknowledged. First, although the MIMIC-IV database provides a large, well-curated sample of critically ill patients, its single-center origin introduces inherent selection bias and reflects the treatment protocols and case mixture of a specific healthcare system. Thus, the absence of external validation represents a substantial limitation. Although we implemented rigorous analytical methods to ensure internal validity (particularly the proposed biphasic, phenotype‑dependent predictive pattern of the NLR), our findings must be regarded as hypothesis‑generating rather than definitive. Independent validation in prospective, multi-center cohorts is urgently required before the NLR can be considered for routine clinical risk stratification. Furthermore, external validation should specifically test the generalizability of the identified inflection points (NLR = 7.3 and 27.3) and the differential predictive utility across SOFA strata observed in our study. Second, while NLR standardization was necessary to address scale disparities in multivariable regression modeling, this processing may reduce clinical interpretability. Additionally, the exclusive calculation of the NLR from the first available measurement within six hours of ICU admission captures only a static snapshot of the inflammatory state. Given the highly dynamic nature of sepsis, this approach may underestimate the prognostic value of NLR trajectories or peak values, which could better reflect the evolving host response. Future studies incorporating serial NLR measurements or longitudinal trajectory modeling are warranted to further validate our findings. Third, the dataset lacked several potentially relevant parameters, including nutritional status indicators, metabolic markers, and established inflammatory biomarkers such as procalcitonin and C-reactive protein, which may affect the generalizability of the results. Fourth, the sepsis management guidelines were revised multiple times during the extended study period spanning 2008–2022, which may have introduced temporal confounding factors that could have influenced the outcomes. Despite these limitations, this study benefits from a substantial sample size and comprehensive statistical methodology that may partially mitigate these concerns. Future research should focus on targeted analyses of specific sepsis patient subgroups to more precisely characterize the prognostic utility of the NLR.

Conclusions

This investigation revealed a positive correlation between the magnitude of NLR elevation and sepsis-related mortality, with the strongest predictive value being observed at the highest NLR values. This correlation may primarily reflect the ability of the NLR to predict the outcomes of clinically stable patients, whereas its predictive value may be reduced for patients with severe organ failure. Thus, the utility of the NLR appears to primarily involve dynamic monitoring rather than single static measurements. However, given the absence of external validation, these conclusions remain hypothesis‑generating and require confirmation in independent, prospective, multi-center studies before clinical implementation.

Supporting information

Acknowledgments

The authors appreciate all of the investigators who organized, developed, and maintained the MIMIC database.