Association between the monocyte-to-lymphocyte ratio and 28-day all-cause mortality in sepsis-associated delirium patients: a retrospective study and machine learning

Abstract

Background

The monocyte-to-lymphocyte ratio (MLR) has been associated with the prognosis of various diseases. However, evidence delineating its relationship with adverse outcomes in sepsis-associated delirium (SAD) remains sparse. The present study seeks to elucidate the association between MLR and 28-day all-cause mortality in patients with SAD.

Methods

Patients diagnosed with SAD were identified from the MIMIC-IV and eICU-CRD databases according to the Sepsis-3 criteria and Confusion Assessment Method for the Intensive Care Unit (CAM-ICU). Patients were grouped by MLR quartiles. The primary outcome was 28-day all-cause mortality. Nonlinearity was assessed with restricted cubic splines (RCS). Survival was compared with Kaplan-Meier (K-M) curves and the log rank test. Hazard ratios (HR) and 95% confidence interval (CI) were estimated with stratified and adjusted Cox models. We also built survival machine learning models that included MLR and clinical features selected with the Random Forest to predict 28-day all-cause mortality. Model performance was quantified using the concordance index (C-index), time-dependent area under the curve (td-AUC), and median survival time ROC curves. External validation in the eICU-CRD database was used to assess generalizability. SHapley Additive exPlanations (SHAP) analyses were performed using a reduced background dataset derived through K-means clustering.

Results

A total of 3463 SAD patients were included in the study, of whom 748 (21.60%) died within 28 days. Cox regression analysis demonstrated that higher MLR levels were significantly associated with an increased risk of 28-day all-cause mortality (HR: 1.08; 95% CI: 1.02–1.15; p = 0.013). RCS suggested a nonlinear association (P overall = 0.0038, P nonlinearity = 0.0026). K-M curves showed that the Q4 group exhibited a significantly higher risk of 28-day all-cause mortality (HR: 1.44; 95% CI: 1.15–1.80; p = 0.001). Subgroup analyses indicated the robustness of the observed associations, with no interaction effects between MLR and any subgroup (p > 0.05). The best machine learning model, the Random Survival Forest model, achieved an AUC of 0.76 in internal validation and 0.67 in external validation. SHAP analysis confirmed that elevated MLR was linked to a high predicted risk of death.

Conclusions

Elevated MLR is independently associated with 28-day all-cause mortality in patients with SAD. Survival machine learning models established based on MLR can effectively predict 28-day all-cause mortality in SAD patients.

Clinical trial

Not applicable.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12879-026-13080-5.

Background

Sepsis is a life-threatening syndrome caused by dysregulated host responses to infection. It is the leading cause of death among patients in the intensive care units (ICUs) and also a major risk factor for delirium [1]. Sepsis-associated delirium (SAD), an early manifestation of sepsis-associated encephalopathy, is characterized by inattention, disorganized thinking, altered levels of consciousness, and a fluctuating clinical course [2]. The pathogenesis of SAD is complex, and its exact pathophysiological mechanisms remain unclear; however, neuroinflammation, abnormal cerebral perfusion, and neurotransmitter imbalances are recognized as key factors contributing to the development and progression of SAD [3]. Approximately 50% of septic patients are affected by SAD, which not only increases the risk of short-term mortality but also impairs patients’ cognitive function, elevates the long-term risk of dementia, and imposes a heavier burden on healthcare systems and society [4]. Therefore, investigating the prognosis of SAD carries significant clinical implications.

Previous studies have confirmed that inflammation-related biomarkers, such as C-reactive protein, interleukin (IL)-6, and tumor necrosis factor-alpha (TNF-α), are associated with the development and prognosis of SAD [5]. However, the detection of these biomarkers often requires invasive, costly, and time-consuming procedures, which limits their application in risk assessment for SAD patients. Thus, identifying valuable, clinically widely available biomarkers associated with prognosis, developing simple and easy-to-use survival prediction models, identifying SAD patients at high risk of all-cause mortality, and implementing timely interventions are crucial for improving the prognosis of SAD patients.

To address these challenges, whole blood cell-derived inflammatory markers have been developed in recent years and have garnered increasing attention for their roles in disease diagnosis, prognostic evaluation, and risk stratification. The monocyte-to-lymphocyte ratio (MLR), an integrated whole blood cell-derived inflammatory marker, has been validated as a reliable predictor of both cardiovascular mortality and all cause mortality in patients with chronic kidney disease and also independently predicts the development of cardiovascular diseases [6, 7]. Recent studies have further confirmed that MLR is associated with the risk of cardiovascular diseases in septic patients and all cause mortality in patients with sepsis-associated acute kidney injury [8, 9]. Nonetheless, the relationship between MLR and 28-day all-cause mortality in patients with SAD has not yet been elucidated.

Therefore, this study aims to comprehensively analyze the association between MLR and 28-day all-cause mortality in SAD patients, construct a 28-day all-cause mortality prediction model by integrating MLR with survival machine learning (ML) approaches, and explore the potential pathophysiological mechanisms underlying SAD-related adverse prognosis. This work is expected to provide a more solid foundation for clinical decision-making and management strategies for SAD.

Method

Study population

We used patient data from two open critical care databases (MIMIC IV version 2.2 and the eICU Collaborative Research Database (eICU-CRD version 2.0) [10, 11]. Both resources provide deidentified ICU data and are freely available for research through PhysioNet. The MIMIC-IV database contains information on all patients admitted to Beth Israel Deaconess Medical Center between 2008 and 2019, while the eICU-CRD is a multicenter telemedicine database containing data from more than 200,000 patients admitted to 335 ICUs in 208 hospitals across the United States between 2014 and 2015. Prior to accessing the data, Z.Y.H. (the first author) completed the Collaborative Institutional Training Initiative Program (Record ID: 67,867,250). This study was conducted in adherence to the Declaration of Helsinki. Given the anonymized nature and standardized structure of the database, additional ethical approval was not required for the present study.

This study enrolled patients diagnosed with SAD, defined as new-onset delirium arising after a confirmed diagnosis of sepsis within 7 days. Sepsis was identified according to the Sepsis-3 criteria [12], requiring documentation of a suspected or confirmed infectious source together with a Sequential Organ Failure Assessment (SOFA) score ≥ 2. Delirium was assessed using the Confusion Assessment Method for the Intensive Care Unit (CAM-ICU) [13]. Prior studies have demonstrated that the CAM-ICU exhibits a pooled sensitivity of 84% (95% CI: 57–93%) and a pooled specificity of 95% (95% CI: 86–100%) for delirium detection, underscoring its robustness and reliability as a diagnostic tool for accurately identifying delirium in critically ill populations [14, 15].

Patients were excluded from the study if they met any of the following criteria: age < 18 years, ICU length of stay < 24 hours, lack of monocyte count or lymphocyte count data, diagnosis of delirium prior to sepsis, presence of malignant tumor, intracranial infection, dementia, or blood system diseases. Relevant diseases were diagnosed based on the International Classification of Diseases, Ninth Revision (ICD-9) and Tenth Revision (ICD-10) codes.

Variables extraction

Only variables available in the MIMIC-IV and eICU-CRD databases were selected as potential predictors. The collected variables included demographic characteristics: age, gender, race, and weight; vital signs: heart rate, respiratory rate, temperature, mean arterial pressure (MAP), and oxygen saturation (SpO₂); laboratory tests: hemoglobin (Hb), white blood cell (WBC) count, platelet count (PLT), monocyte count, lymphocyte count, partial pressure of oxygen (PO₂), partial pressure of carbon dioxide (PCO₂), lactate, blood urea nitrogen (BUN), potassium, sodium, calcium, bicarbonate (HCO₃^-), anion gap, glucose, creatinine, aspartate transaminase (AST), partial thromboplastin time (PTT), and international normalized ratio (INR); comorbidities: myocardial infarction, heart failure (HF), peripheral vascular disease, cerebrovascular disease, chronic pulmonary disease, diabetes, severe liver disease, and renal disease; severity of illness scores: Charlson Comorbidity Index (CCI), SOFA score, Glasgow Coma Scale (GCS) score, and Simplified Acute Physiology Score II (SAPS II); interventions: mechanical ventilation, vasopressor use, midazolam use, continuous renal replacement therapy (CRRT), and fluid input and output volume.

For the severity of illness scoring systems, the lowest GCS score within 24 hours of ICU admission was selected; the highest SOFA score, SAPS II score, and CCI within 24 hours of ICU admission were chosen. For other data, the first measured value within 24 hours of ICU admission was used. All data were extracted using structured query language (SQL) queries executed in PostgreSQL (version 17), combined with Navicat Premium (version 16.3.2) for database management and interface navigation.

The MLR was calculated from routinely available hematological parameters [16], with the formula: MLR = monocyte count/lymphocyte count.

Outcome

The primary outcome measure of this study was 28-day all-cause mortality. Defined as mortality occurring within 28 days following ICU admission, 28-day all-cause mortality was the core endpoint of interest. Study participants were followed up from the time of ICU admission until the 28th day post-admission or until death if it occurred within this timeframe. Notably, for patients with multiple hospitalizations, only data from their first ICU admission were included in the analysis.

Management of missing data and outlier

To minimize potential bias arising from incomplete data, variables with a missing rate exceeding 20% were excluded from the analysis, and the missing rates of all variables are presented in Table S1. For variables with a missing rate of less than 20%, we applied multiple imputation by chained equations (MICE) using the “mice” package in R, assuming data were missing at random (MAR). We generated five complete datasets (m = 5) using predictive mean matching (PMM) for continuous variables and logistic regression for categorical ones. Results were pooled using Rubin’s rules to obtain final estimates and confidence intervals [17]. Outliers were defined as data points less than the 1st percentile or greater than the 99th percentile of the median, and Winsorization was applied for outlier handling [18].

Statistical analysis

Given the retrospective nature of this study, no sample size calculation was performed. Patients were stratified into four groups according to quartiles of MLR. Normality was assessed using the Kolmogorov-Smirnov test. All continuous variables in this study were non-normally distributed and thus are presented as median (interquartile range, IQR). Comparisons were made using the Mann-Whitney U test or Kruskal-Wallis H test, as appropriate. Categorical variables are presented as counts and percentages (%), with group comparisons conducted using Pearson’s chi-square test or Fisher’s exact test, depending on the data characteristics.

Kaplan-Meier (K-M) survival curves were constructed to compare 28 day survival rates across the four groups, with significance evaluated by the log-rank test. Cox regression models were employed to estimate the hazard ratio (HR) and 95% confidence interval (95% CI) for the endpoint. The stability of the Cox regression models was evaluated using the proportional hazards test. A two-step approach was used to identify prognostic factors. First, univariate Cox regression was performed to select candidate variables. Variables with a p-value < 0.05 in the univariate analysis or known clinical relevance were included in the multivariate models (Table S2). Subsequently, multivariate Cox regression models were constructed to determine independent risk factors for 28-day all-cause mortality. Model 1 was the unadjusted model. Model 2 adjusted for gender, age, weight, and race based on Model 1. Model 3 further incorporated variables identified from univariate analysis and clinically relevant variables, including SOFA score, MAP, respiratory rate, temperature, SPO₂, INR, PTT, anion gap, BUN, potassium, sodium, AST, WBC count, Hb, lactate, PO₂, myocardial infarction, HF, peripheral vascular disease, cerebrovascular disease, chronic pulmonary disease, renal disease, severe liver disease, CCI, vasopressor use, CRRT, fluid input, and fluid output volume.

We used a four-knot RCS to explore a possible non-linear relation between baseline MLR and 28-day all-cause mortality. The same adjustments as in Model 3 were applied. Multicollinearity was assessed using the variance inflation factor (VIF), with a threshold of VIF < 5 to ensure variable independence [19]. Receiver operating characteristic (ROC) curves and the area under the curve (AUC) were were employed to evaluate the predictive accuracy of the MLR for 28-day all-cause mortality.

Subgroup analyses were performed to investigate the prognostic value of MLR across different subgroups, such as age (≤60 years, >60 years), gender, HF, cerebrovascular disease, chronic pulmonary disease, severe liver disease, renal disease. Cox regression analysis was conducted for each subgroup. The multivariate analysis was adjusted as in Model 3, and the results were visually presented using forest plots, illustrating HR and 95% CI.

All statistical analyses were performed using R version 4.4.2 (R Foundation for Statistical Computing) and Python version 3.10.18 (Python Software Foundation). Statistical tests were two-tailed, with p-value < 0.05 considered statistically significant.

Establishment and validation of the prediction models

In this study, the target variables for survival analysis, namely “28-day survival time” (as the time variable) and “survival status” (as the event indicator), were extracted from the dataset. “survival status” was encoded as a boolean variable, with True indicating an event occurrence (death within 28 days) and False indicating censoring (alive at 28 days). These two variables were concatenated into a structured array, formatted as required by the Python scikit-survival library.

We randomly split the MIMIC dataset into a training set and an internal validation set in a 7:3 ratio and used the eICU-CRD database for external validation. To optimize stratification and ensure that the training and test sets maintained similar distributions of both event outcomes and time-to-event variables, a modified stratified sampling strategy was employed.

Features for the ML models were selected from demographic characteristics, vital signs, laboratory tests, comorbidities, illness severity scores, and interventions. A tree-based random forest approach with default parameters was employed for feature selection. Initially, all feature variables were included in the random forest model to construct a baseline model. Within this initial model, permutation importance was utilized to quantify the contribution of each feature to model performance. This method assesses feature importance by randomly permuting the values of a single feature and observing the resulting decrease in model performance. A significant performance decline following permutation indicates a highly influential feature, whereas a minimal change suggests low relevance.

Subsequently, the eli5 library was used to visualize feature importance weights, and features were ranked accordingly. To identify the optimal feature subset and enhance model performance, a stepwise feature selection strategy was implemented [20]. Starting with individual features, the algorithm incrementally added features, constructing models with different feature combinations. During each iteration, the Concordance Index (C-index) was recorded for both the training and test sets, along with the corresponding number of selected features. This process culminated in the identification of the feature combination that maximized model performance.

Based on the selected optimal feature subset, five ML survival models, Cox Proportional Hazards (PH) Survival, Fast Kernel Survival SVM, Gradient Boosting Survival, Random Survival Forest (RSF), and Extra Survival Trees were constructed to predict the 28-day all-cause mortality risk. Except for the Cox PH Survival model, hyperparameter tuning was performed during the establishment of the other ML models. Given the time-to-event nature of the outcomes analyzed in this study, traditional classification metrics such as accuracy, sensitivity, specificity, and precision were not calculated. Instead, model performance was assessed using the C-index, time-dependent AUC curves (td-AUC), and median survival time ROC curves [21, 22]. The C-index measures the discriminative ability of each model, where higher values indicate better differentiation between patients with varying survival times. The td-AUC was utilized to graphically demonstrate the models’ predictive accuracy over time, with higher mean AUC values indicating better performance.

Model explainability

To identify and interpret the most influential features in predicting 28-day all-cause mortality, the SHapley Additive exPlanations (SHAP) framework was employed to explain the optimal predictive model. Considering the high computational cost inherent to SHAP analysis, a dimensionality reduction step was first implemented for the background dataset: K-means clustering was applied to the training data to generate 150 cluster centers, which were then used as the background data for subsequent SHAP computations [23]. A Kernel Explainer was constructed based on this reduced background dataset and the pre-established survival models, followed by the calculation of SHAP values for the test dataset to quantify the marginal contribution of each feature to the model’s predictive outcomes. We then ranked features by their SHAP importance and visualized how different feature levels influenced the predicted risk.

Results

Study population and baseline characteristics

We included 3463 patients with SAD from the MIMIC IV database, of whom 748 (21.60%) died within 28 days. The patient selection process is illustrated in Fig. 1. The median age of the study population was 66.54 (IQR: 54.61–77.06) years, with males accounting for 1983 (57.3%) of the cohort. The majority of patients were White (2,063 [59.6%]). Baseline characteristics of the study participants are presented in Table 1.

Fig. 1.

Study flowchart for cohort selection and model development

Table 1.

Baseline characteristics of the study populations

Variables	Overall (n = 3463)	Q1 (n = 865)	Q2 (n = 865)	Q3 (n = 865)	Q4 (n = 868)	P
Demographic characteristics
Male, n(%)	1983 (57.3)	461 (53.3)	483 (55.8)	506 (58.5)	533 (61.4)	0.005
Race (White), n(%)	2063 (59.6)	526 (60.8)	516 (59.7)	496 (57.3)	525 (60.5)	0.451
Age, years (M[IQR])	66.54 [54.61, 77.06]	65.77 [54.11, 77.22]	65.43 [53.57, 76.08]	66.35 [53.49, 76.23]	68.31 [56.82, 78.52]	0.001
Weight, kg (M [IQR])	85.00 [69.85, 102.20]	86.10 [70.20, 102.20]	86.30 [70.60, 103.50]	84.90 [70.10, 100.60]	83.70 [68.20, 102.30]	0.451
Vital signs (M [IQR])
MAP, mmHg	81.00 [70.00, 94.00]	79.00 [69.00, 92.00]	83.00 [71.00, 96.00]	81.00 [69.00, 94.00]	82.00 [70.00, 94.00]	<0.001
Respiratory rate	20.00 [16.00, 24.00]	18.00 [16.00, 23.00]	20.00 [16.00, 24.00]	20.00 [16.00, 24.00]	20.00 [17.00, 25.00]	<0.001
Temperature, °C	36.83 [36.50, 37.22]	36.78 [36.39, 37.17]	36.89 [36.50, 37.33]	36.80 [36.50, 37.17]	36.83 [36.56, 37.22]	<0.001
SPO2, %	98.00 [95.00, 100.00]	99.00 [96.00, 100.00]	98.00 [95.00, 100.00]	98.00 [95.00, 100.00]	97.00 [95.00, 100.00]	<0.001
Heart rate, beats/min	90.00 [78.00, 106.50]	87.00 [78.00, 104.00]	90.00 [77.00, 106.00]	91.00 [77.00, 107.00]	94.00 [79.00, 108.00]	0.007
Laboratory tests (M [IQR])
INR	1.30 [1.10, 1.70]	1.40 [1.20, 1.70]	1.30 [1.10, 1.50]	1.30 [1.10, 1.70]	1.40 [1.20, 1.80]	<0.001
PTT, seconds	31.30 [27.50, 39.30]	31.30 [27.60, 38.40]	30.80 [27.60, 37.40]	31.20 [27.20, 40.20]	32.10 [27.50, 41.60]	0.047
Anion gap, mmol/L	15.00 [13.00, 19.00]	14.00 [12.00, 18.00]	15.00 [13.00, 18.00]	16.00 [13.00, 19.00]	16.00 [14.00, 19.00]	<0.001
Bicarbonate, mmol/L	22.00 [19.00, 25.00]	22.00 [19.00, 24.00]	22.00 [19.00, 25.00]	21.00 [18.00, 24.00]	21.00 [18.00, 24.00]	0.005
Calcium, mg/dL	8.20 [7.60, 8.70]	8.20 [7.60, 8.60]	8.20 [7.70, 8.70]	8.20 [7.70, 8.80]	8.20 [7.60, 8.70]	0.141
BUN, mg/dL	23.00 [15.00, 39.00]	20.00 [13.00, 33.00]	21.00 [14.00, 36.00]	24.00 [14.00, 40.00]	28.00 [17.00, 49.00]	<0.001
Potassium, mEq/L	4.20 [3.70, 4.70]	4.10 [3.70, 4.60]	4.10 [3.70, 4.60]	4.20 [3.70, 4.70]	4.30 [3.80, 4.80]	0.001
Sodium, mEq/L	139.00 [136.00, 142.00]	140.00 [137.00, 142.00]	139.00 [136.00, 142.00]	139.00 [135.00, 142.00]	138.00 [134.00, 141.00]	<0.001
Glucose, mg/dL	135.00 [109.00, 178.50]	133.00 [107.00, 173.00]	134.00 [107.00, 180.00]	139.00 [110.00, 186.00]	137.00 [109.00, 176.00]	0.194
Creatinine, mg/dL	1.10 [0.80, 1.80]	1.00 [0.70, 1.60]	1.10 [0.80, 1.70]	1.20 [0.80, 1.90]	1.30 [0.80, 2.20]	<0.001
AST, U/L	50.00 [27.00, 123.50]	53.00 [27.00, 133.00]	46.00 [26.00, 102.00]	50.00 [27.00, 131.00]	51.00 [29.00, 124.00]	0.074
WBC, K/uL	12.50 [8.70, 17.50]	11.10 [7.60, 15.20]	11.60 [8.00, 15.80]	13.10 [9.70, 18.50]	14.50 [10.40, 20.42]	<0.001
Hemoglobin, g/dL	10.40 [8.80, 12.20]	10.00 [8.50, 11.70]	10.60 [9.00, 12.40]	10.60 [9.00, 12.40]	10.60 [8.80, 12.50]	<0.001
Lactate, mmol/L	1.80 [1.20, 2.90]	2.00 [1.30, 3.20]	1.70 [1.20, 2.70]	1.80 [1.20, 2.90]	1.80 [1.20, 2.90]	<0.001
Platelet, K/uL	179.00 [122.00, 243.00]	160.00 [113.00, 223.00]	183.00 [125.00, 251.00]	194.00 [131.00, 252.00]	186.00 [126.00, 244.25]	<0.001
PO2, mmHg	91.00 [53.00, 174.00]	127.00 [65.00, 280.00]	91.00 [55.00, 163.00]	84.00 [50.00, 151.00]	80.00 [47.75, 135.00]	<0.001
PCO2, mmHg	41.00 [35.00, 48.00]	40.00 [35.00, 47.00]	41.00 [35.00, 48.00]	41.00 [35.00, 49.00]	41.00 [35.00, 50.00]	0.139
MLR (M [IQR])	0.62 [0.37, 1.06]	0.24 [0.17, 0.31]	0.49 [0.43, 0.55]	0.80 [0.70, 0.93]	1.55 [1.25, 2.20]	<0.001
Comorbidity, n(%)
Myocardial infarct	685 (19.8)	146 (16.9)	159 (18.4)	181 (20.9)	199 (22.9)	0.008
Heart failure	1212 (35.0)	292 (33.8)	284 (32.8)	293 (33.9)	343 (39.5)	0.014
Peripheral vascular disease	467 (13.5)	132 (15.3)	123 (14.2)	102 (11.8)	110 (12.7)	0.148
Cerebrovascular disease	686 (19.8)	153 (17.7)	204 (23.6)	174 (20.1)	155 (17.9)	0.006
Chronic pulmonary disease	1003 (29.0)	241 (27.9)	248 (28.7)	255 (29.5)	259 (29.8)	0.806
Diabetes	1125 (32.5)	285 (32.9)	281 (32.5)	271 (31.3)	288 (33.2)	0.849
Renal disease	828 (23.9)	181 (20.9)	186 (21.5)	211 (24.4)	250 (28.8)	<0.001
Severe liver disease	377 (10.9)	66 (7.6)	82 (9.5)	116 (13.4)	113 (13.0)	<0.001
Severity of illness scores (M [IQR])
CCI	5.00 [3.00, 7.00]	5.00 [3.00, 6.00]	5.00 [3.00, 7.00]	5.00 [3.00, 7.00]	5.00 [3.00, 7.00]	<0.001
SOFA score	3.00 [2.00, 5.00]	4.00 [2.00, 5.00]	3.00 [2.00, 5.00]	3.00 [2.00, 5.00]	4.00 [2.00, 5.25]	<0.001
GCS	15.00 [13.00, 15.00]	15.00 [13.00, 15.00]	15.00 [13.00, 15.00]	15.00 [13.00, 15.00]	15.00 [13.00, 15.00]	0.109
SAPS II	41.00 [33.00, 52.00]	41.00 [33.00, 50.00]	40.00 [31.00, 50.00]	42.00 [33.00, 53.00]	44.00 [36.00, 54.00]	<0.001
Intervention
Vasopressor, n (%)	2398 (69.2)	646 (74.7)	565 (65.3)	563 (65.1)	624 (71.9)	<0.001
Midazolam, n (%)	1305 (37.7)	326 (37.7)	333 (38.5)	341 (39.4)	305 (35.1)
MV, n (%)	2632 (76.0)	677 (78.3)	644 (74.5)	656 (75.8)	655 (75.5)	0.293
CRRT, n (%)	505 (14.6)	122 (14.1)	90 (10.4)	126 (14.6)	167 (19.2)	<0.001
Input amount, mL (M [IQR])	4529.56 [2545.66, 7192.72]	5406.50 [3402.32, 7465.14]	4324.17 [2488.67, 6825.20]	4145.50 [2474.26, 7191.67]	4085.44 [2225.07, 7051.81]	<0.001
Output amount, mL (M [IQR])	1990.00 [1111.50, 3072.50]	2270.00 [1436.00, 3235.00]	1993.00 [1200.00, 3037.00]	1902.00 [1094.00, 3107.00]	1691.00 [894.75, 2817.50]	<0.001

MLR: Q1 (MLR < 0.37), Q2 (0.37 ≤ MLR <0.62), Q3 (0.62 ≤ MLR <1.05), and Q4 (1.05 ≤ MLR <11.93)

Abbreviations: MAP: mean arterial pressure; SpO₂: oxygen saturation; PTT: partial thromboplastin time; INR: international normalized ratio; BUN: blood urea nitrogen; AST: aspartate transaminase; WBC: white blood cells; PO₂: partial pressure of oxygen; PCO₂: partial pressure of carbon dioxide; MLR: monocyte-to-lymphocyte ratio; CCI: Charlson comorbidity index; SOFA: sequential organ failure assessment; GCS: glasgow coma scale; SAPS II: simplified acute physiology score II; MV: mechanical ventilation; CRRT: continuous renal replacement therapy. p-values less than 0.05 are shown in bold

Patients were stratified into four groups based on MLR quartile: Q1 (MLR < 0.37), Q2 (0.37 ≤ MLR <0.62), Q3 (0.62 ≤ MLR <1.05), and Q4 (1.05 ≤ MLR <11.93). Groups Q1, Q2, and Q3 each consisted of 865 patients, while Q4 included 868 SAD patients. Compared with other quartiles, patients in Q4 exhibited a higher heart rate, BUN, creatinine, WBC count, potassium, anion gap, and PTT, as well as higher SAPS II score and MLR values. Conversely, lower levels of SpO₂, sodium, bicarbonate, and PaO₂ were observed in Q4 patients. Additionally, patients in Q4 were more likely to have a history of myocardial infarction, HF, renal disease, and severe liver disease; a greater proportion of them received vasopressor use and CRRT, with the lowest total fluid input and output volume among the four groups.

The 28-day all-cause mortality rates in Q1, Q2, Q3, and Q4 were 135 (15.61%), 155 (17.92%), 199 (23.01%), and 259 (29.84%), respectively. Table 2 presents the comparisons of characteristics between the 28-day survival and non-survival cohorts. Patients in the non-survival cohort were significantly older, had a higher proportion of males, a higher prevalence of comorbidities, and higher disease severity scores (SOFA score and SAPS II score). Laboratory tests revealed that the non-survival cohort had elevated INR, PTT, anion gap, BUN, creatinine, AST, WBC count, and lactate, as well as a markedly higher MLR compared with the survival cohort.

Table 2.

Baseline characteristics according to 28-day all-cause mortality

Variables	Overall (n = 3463)	Survival (n = 2715)	Non-survival (n = 748)	P
Demographic characteristics
Male, n(%)	1983 (57.3)	1583 (58.3)	400 (53.5)	0.02
Race (White), n(%)	2063 (59.6)	1639 (60.4)	424 (56.7)	0.076
Age, years (M [IQR])	66.54 [54.61, 77.06]	64.66 [53.00, 75.03]	73.01 [60.40, 82.96]	<0.001
Weight, kg (M [IQR])	85.00 [69.85, 102.20]	86.10 [70.90, 102.65]	81.40 [66.97, 99.80]	<0.001
Vital signs (M [IQR])
MAP,mmHg	81.00 [70.00, 94.00]	82.00 [70.00, 94.00]	79.00 [68.00, 93.00]	0.001
Respiratory rate	20.00 [16.00, 24.00]	19.00 [16.00, 24.00]	21.00 [17.00, 25.00]	<0.001
Temperature, °C	36.83 [36.50, 37.22]	36.89 [36.50, 37.28]	36.67 [36.44, 37.06]	<0.001
SPO2, %	98.00 [95.00, 100.00]	98.00 [95.00, 100.00]	98.00 [94.00, 100.00]	<0.001
Heart rate, beats/min	90.00 [78.00, 106.50]	90.00 [78.00, 106.00]	92.00 [78.00, 108.00]	0.12
Laboratory tests (M [IQR])
INR	1.30 [1.10, 1.70]	1.30 [1.10, 1.60]	1.40 [1.20, 2.10]	<0.001
PTT, seconds	31.30 [27.50, 39.30]	30.70 [27.10, 37.60]	34.40 [28.78, 47.23]	<0.001
Anion gap, mmol/L	15.00 [13.00, 19.00]	15.00 [13.00, 18.00]	16.00 [14.00, 20.00]	<0.001
Bicarbonate, mmol/L	22.00 [19.00, 25.00]	22.00 [19.00, 25.00]	21.00 [18.00, 24.00]	<0.001
Calcium, mg/dL	8.20 [7.60, 8.70]	8.20 [7.70, 8.70]	8.20 [7.60, 8.80]	0.729
BUN, mg/dL	23.00 [15.00, 39.00]	21.00 [14.00, 35.00]	32.00 [19.00, 52.00]	<0.001
Potassium, mEq/L	4.20 [3.70, 4.70]	4.10 [3.70, 4.70]	4.30 [3.80, 4.80]	0.001
Sodium, mEq/L	139.00 [136.00, 142.00]	139.00 [136.00, 142.00]	138.00 [135.00, 142.00]	0.006
Glucose, mg/dL	135.00 [109.00, 178.50]	134.00 [109.00, 177.00]	139.50 [108.00, 186.25]	0.162
Creatinine, mg/dL	1.10 [0.80, 1.80]	1.10 [0.80, 1.70]	1.40 [0.90, 2.40]	<0.001
AST, U/L	50.00 [27.00, 123.50]	48.00 [27.00, 113.00]	57.00 [30.00, 151.25]	<0.001
WBC, K/uL	12.50 [8.70, 17.50]	12.30 [8.60, 17.00]	12.95 [9.40, 18.90]	<0.001
Hemoglobin, g/dL	10.40 [8.80, 12.20]	10.50 [8.90, 12.40]	9.90 [8.40, 11.80]	<0.001
Lactate, mmol/L	1.80 [1.20, 2.90]	1.70 [1.20, 2.80]	2.10 [1.40, 3.50]	<0.001
Platelet, mmol/L	179.00 [122.00, 243.00]	181.00 [124.00, 243.00]	170.50 [114.00, 245.25]	0.11
PO2, mmol/L	91.00 [53.00, 174.00]	97.00 [58.00, 184.50]	74.00 [45.00, 135.00]	<0.001
PCO2, mmol/L	41.00 [35.00, 48.00]	41.00 [35.00, 48.00]	41.00 [34.00, 48.25]	0.325
MLR	0.62 [0.37, 1.06]	0.59 [0.35, 1.00]	0.79 [0.46, 1.33]	<0.001
Comorbidity n(%)
Myocardial infarct	685 (19.8)	483 (17.8)	202 (27.0)	<0.001
Heart failure	1212 (35.0)	876 (32.3)	336 (44.9)	<0.001
Peripheral vascular disease	467 (13.5)	343 (12.6)	124 (16.6)	0.006
Cerebrovascular disease	686 (19.8)	510 (18.8)	176 (23.5)	0.005
Chronic pulmonary disease	1003 (29.0)	758 (27.9)	245 (32.8)	0.011
Diabetes	1125 (32.5)	879 (32.4)	246 (32.9)	0.825
Renal disease	828 (23.9)	600 (22.1)	228 (30.5)	<0.001
Severe liver disease	377 (10.9)	238 (8.8)	139 (18.6)	<0.001
Severity of illness scores (M [IQR])
Charlson comorbidity index	5.00 [3.00, 7.00]	4.00 [3.00, 6.00]	6.00 [4.00, 8.00]	<0.001
SOFA score	3.00 [2.00, 5.00]	3.00 [2.00, 5.00]	4.00 [3.00, 6.00]	<0.001
GCS	15.00 [13.00, 15.00]	15.00 [13.00, 15.00]	15.00 [13.00, 15.00]	0.526
SAPS II	41.00 [33.00, 52.00]	40.00 [32.00, 49.00]	49.00 [40.00, 58.00]	<0.001
Interventions
Vasopressor, n (%)	2398 (69.2)	1793 (66.0)	605 (80.9)	<0.001
Midazolam, n (%)	1305 (37.7)	1001 (36.9)	304 (40.6)	0.065
MV, n (%)	2632 (76.0)	2051 (75.5)	581 (77.7)	0.246
CRRT, n (%)	505 (14.6)	309 (11.4)	196 (26.2)	<0.001
Input amount, mL (M [IQR])	4529.56 [2545.66, 7192.72]	4807.29 [2741.88, 7354.10]	3719.48 [1994.38, 6117.56]	<0.001
Output amount, mL (M [IQR])	1990.00 [1111.50, 3072.50]	2136.00 [1295.00, 3260.00]	1345.00 [665.75, 2329.50]	<0.001

Abbreviations: MAP: mean arterial pressure; SpO₂: oxygen saturation; PTT: partial thromboplastin time; INR: international normalized ratio; BUN: blood urea nitrogen; AST: aspartate transaminase; WBC: white blood cells; PO₂: partial pressure of oxygen; PCO₂: partial pressure of carbon dioxide; MLR: monocyte-to-lymphocyte ratio; SOFA: sequential organ failure assessment; GCS: glasgow coma scale; SAPS II: simplified acute physiology score II; MV: mechanical ventilation; CRRT: continuous renal replacement therapy. p-values less than 0.05 are shown in bold

Association between MLR and 28-day all-cause mortality risk

The test for the proportional hazards assumption of Cox regression was performed (p > 0.05), confirming that the proportional hazards assumption of the Cox model was satisfied. Table 3 presents the association between MLR and 28-day all-cause mortality risk. In both univariate (HR: 1.22, 95% CI: 1.16–1.28) and multivariate (HR: 1.08, 95% CI: 1.02–1.15) Cox regression models, elevated MLR levels were significantly associated with an increased risk of 28-day all-cause mortality.

Table 3.

The association between MLR and the risk of 28-day all-cause mortality

Variables	Model 1 HR (95% CI)	P	Model 2 HR (95% CI)	P	Model 3 HR (95% CI)	P
MLR levels Continuous	1.22 (1.16, 1.28)	<0.001	1.20 (1.14, 1.26)	<0.001	1.08 (1.02, 1.15)	0.013
MLR levels Categorize
Q1	Ref		Ref		Ref
Q2	1.17 (0.93, 1.48)	0.171	1.19 (0.94, 1.50)	0.139	1.09 (0.86, 1.38)	0.473
Q3	1.56 (1.25, 1.94)	<0.001	1.57 (1.26, 1.95)	<0.001	1.24 (0.99, 1.56)	0.063
Q4	2.11 (1.71, 2.60)	<0.001	2.05 (1.66, 2.52)	<0.001	1.44 (1.15, 1.80)	0.001
P for Trend	1.73 (1.51, 1.99)	<0.001	1.68 (1.47, 1.93)	<0.001	1.30 (1.12, 1.51)	<0.001

HR: hazard ratio; CI: confidence interval; all-cause mortality: all-cause mortality; MLR: monocyte-to-lymphocyte ratio

MLR:Q1 (MLR < 0.37), Q2 (0.37 ≤ MLR <0.62), Q3 (0.62 ≤ MLR <1.05), and Q4 (1.05 ≤ MLR <11.93).

Model 1 was the unadjusted model;

Model 2 adjusted for gender, age, weight, and race based on Model 1;

Model 3 adjusted for SOFA score, MAP, respiratory rate, temperature, SPO₂, INR, PTT, anion gap, BUN, potassium, sodium, AST, WBC, Hb, lactate, PO₂, myocardial infarction, HF, peripheral vascular disease, cerebrovascular disease, chronic pulmonary disease, renal disease, severe liver disease, CCI, vasopressor use, CRRT, fluid input, fluid output volume and based on Model 2.

Bold values represent significant p values

Further analyses using RCS revealed a significant non-linear positive correlation between MLR and 28-day all-cause mortality risk (P overall = 0.0038, P non-linearity = 0.0226) (Fig. 2). When MLR was categorized into quartiles (Q1–Q4), the highest MLR quartile (Q4) was associated with a heightened risk of 28-day all-cause mortality compared to Q1, as evidenced by both univariate (HR: 2.11, 95% CI: 1.71–2.60) and multivariate (HR: 1.44, 95% CI: 1.15–1.80) models.

Fig. 2.

RCS analysis of 28-day all-cause mortality. Curves represent estimated adjusted hazard ratios, and shaded ribbons represent 95% confidence intervals. The horizontal dashed line represents a hazard ratio of 1.0. HR hazard ratio, CI confidence interval

K- M survival curves (Fig. 3) stratified by the four MLR quartiles demonstrated a significantly higher 28-day all-cause mortality risk in the Q4 group (log-rank p < 0.001). Additionally, Supplementary Table S3 shows that the variance inflation factor (VIF) values of all continuous variables were below 5, indicating the absence of multicollinearity in the regression models. ROC curve analyses revealed that the predictive capacity of MLR for 28-day all-cause mortality remained moderate (AUC = 0.60) (Table S4).

Fig. 3.

28-day KM survival curve. KM curves showing the survival rates at 28 days for each quartile. MLR: Q1 (MLR < 0.37), Q2 (0.37 ≤ MLR < 0.62), Q3 (0.62 ≤ MLR < 1.05), and Q4 (1.05 ≤ MLR < 11.93)

Subgroup analysis

Figure 4 presents the results of a subgroup analysis. In subgroups stratified by age > 60 years, gender, HF, cerebrovascular disease, chronic pulmonary disease, renal disease, absence of severe liver disease, the Q4 group consistently conferred an elevated risk of 28-day all-cause mortality (HRs > 1, all p < 0.05).

Fig. 4.

Subgroup forest plot for 28-day all-cause mortality. The same adjustments as in model 3 were applied. MLR: Q1 (MLR < 0.37), Q2 (0.37 ≤ MLR < 0.62), Q3 (0.62 ≤ MLR < 1.05), and Q4 (1.05 ≤ MLR < 11.93)

In contrast, no statistically significant association between Q4 group and 28-day all-cause mortality was observed in the subgroups of patients aged ≤ 60 years, those with severe liver disease. Furthermore, no significant interaction effects were detected between MLR and any of the stratified subgroups (all interaction p > 0.05).

Machine learning-based models predicting the 28-day all-cause mortality risk

Feature selection

During feature selection, a total of 46 variables in the training set were included as candidate feature. Based on the weight ranking of feature variables and the C-index of the training and test sets (Fig. 5), 11 feature variables were ultimately selected for model construction. These features included age, fluid output, SAPS II, CCI, fluid input, CRRT, temperature, PTT, MLR, calcium, and severe liver disease. Table S5 presents the weight ranking of all feature variables in the model, among which MLR exhibited a relatively high weight of 0.004547.

Fig. 5.

C-index vs. Number of selected features in training and test sets. The vertical dotted line indicates that the number of selected feature variables is 11, at which point the training set and test set achieve the relatively optimal Concordance index

Internal validation of models

Using the selected features, five ML survival models were constructed. Hyperparameter tuning was optimized via 5-fold cross-validation and grid search; the hyperparameter search spaces and selected best parameters for each survival model are detailed in Table S6.

Comparison of the td-AUC curves of each model (Fig. 6) revealed that the MLR-included Cox PH Survival (AUC = 0.75), Gradient Boosting Survival (AUC = 0.75), RSF (AUC = 0.76), and Extra Survival Trees (AUC = 0.74) all demonstrated excellent predictive efficacy. Figure 7 displays the ROC curve at the median survival time point (12 days): the AUC of Cox PH Survival was 0.75, Fast Kernel Survival SVM was 0.52, Gradient Boosting Survival was 0.74, RSF was 0.76, and Extra Survival Trees was 0.73. Except for the Fast Kernel Survival SVM model, the other four MLR-based survival ML models also exhibited superior 28-day all-cause mortality predictive efficacy at the median survival time point of SAD patients.

Fig. 6.

Summary of time-dependent area under the curve (td-AUC) performance for five survival models predicting 28-day all-cause mortality

Fig. 7.

Summary of ROC curve at the median survival time point (12 days) performance for five survival models

Figure 8 summarizes the C-indices of the five survival ML models across the training and test cohorts. The Cox PH model achieved C-indices of 0.73 in both datasets. The Fast Kernel Survival SVM exhibited a C-index of 1.0 in the training set but only 0.48 in the test set, indicating substantial overfitting. The Gradient Boosting Survival model attained C-indices of 0.76 (training) and 0.73 (test), whereas the RSF model yielded values of 0.94 (training) and 0.74 (test). The Extra Survival Trees model demonstrated C-indices of 0.87 (training) and 0.73 (test). With the exception of the Fast Kernel Survival SVM, the remaining four MLR-based survival models showed strong and well-balanced performance across both datasets, reflecting robust predictive capacity. Notably, the RSF model achieved the highest C-index of 0.74 on the test set. Given that interpretability and practical applicability in clinical settings are equally crucial as predictive performance, the RSF model was selected as the optimal model for risk stratification and further analysis.

Fig. 8.

The C-indexes of the five survival machine learning models on the training and test sets

External validation of models

Baseline characteristics of patients in the MIMIC-IV and eICU-CRD cohorts are summarized in Table S7. In brief, the 368 patients with SAD in the eICU-CRD cohort presented with lower severity-of-illness scores, a greater proportion of White individuals, a reduced comorbidity burden, and fewer therapeutic interventions. As illustrated in Fig. 9, the top-performing model (RSF) was subsequently assessed in the eICU-CRD external validation cohort. The RSF demonstrated satisfactory discriminative performance for in-hospital mortality, achieving an AUC of 0.67. Collectively, these results substantiate its potential clinical utility and confirm its generalizability within a geographically and institutionally distinct multicentre population.

Fig. 9.

External validation of the RSF model in the eICU-CRD cohort; (A) the C-indexes of the RSF model on the training, test and external sets; (B) the ROC curve at the median survival time point (6 days)

Model explainability

As shown in Fig. 10, the SHAP summary plot (A) and SHAP beeswarm plot (B) illustrate the magnitude and direction of each variable’s influence on model predictions. Age stands out as the most influential prognostic factor across most clinical scenarios, with advanced age correlating with an elevated risk of 28-day all-cause mortality. Specifically, age, SAPS II, PTT, CCI, MLR, CRRT, and severe liver disease exhibited positive SHAP contributions. In contrast, liquid output, liquid input, temperature, and calcium demonstrated negative SHAP contributions. In these plots, the red line represents a higher predicted mortality risk, while the blue line denotes a lower predicted risk. These plots decompose the cumulative impact of key features on the model’s prediction, thereby providing actionable insights to inform personalized care planning based on the optimal predictive model.

Fig. 10.

(A) SHapley Additive exPlanations (SHAP) summary plot shows the top features contributing to the random survival forest model’s prediction of 28-day all-cause mortality. Features are positioned along the y axis based on importance. (B) SHAP beeswarm plot shows the location of each feature on the y-axis ranks its importance according to the model prediction in descending order. For each feature, one dot represents a single patient and the dot’s color is an indicator of that feature value, where blue represents the lower value and red represents the higher value. A dot’s position along the x-axis (ie, the actual SHAP value) illustrates the impact that the feature had on the model’s output (positively or negatively) for that specific patient. Mathematically, this corresponds to the (logarithm of the) likelihood of 28-day all-cause mortality relative across patients (ie, a patient with a higher SHAP value has a higher likelihood of 28-day all-cause mortality relative to a patient with a lower SHAP value)

Discussion

This study comprehensively elucidates the relationship between the MLR and 28-day all-cause mortality in patients with SAD, as well as its potential predictive utility. We demonstrated that elevated MLR levels are associated with an increased risk of 28-day all-cause mortality. RCS analysis further revealed a non-linear relationship between MLR and 28-day all-cause mortality. After adjusting for covariates, these findings remained consistent across nearly all subgroups, indicating the robustness of the study results. Taking the 28 day survival time and 28 day survival status of SAD patients as the target variables, tree-based algorithms were employed to screen for feature variables associated with the target variables and calculate their weights. By incorporating MLR as a core feature into various survival ML algorithms, RSF performed best with AUC = 0.76 in internal validation and AUC = 0.67 in external validation, showing that these predictive models exhibited favorable performance. This indicates that MLR plays a pivotal role in the present study and is significantly associated with the research objectives, highlighting the clinical value of monitoring MLR at the time of ICU admission for 28 day risk stratification in SAD patients.

The MLR, calculated as the ratio of monocytes to lymphocytes, effectively captures the relative changes in two immune-related cell populations and integrates distinct immunological pathways. As crucial components of the innate immune system, monocytes and lymphocytes are closely involved in the pathophysiological processes of the organism. Specifically, monocytes participate in immune responses through phagocytosis, neutrophil recruitment, release of reactive oxygen species and inflammatory cytokines, and antigen presentation [24]. Monocytes and macrophages promote the inflammatory cascade, and elevated monocyte levels indicate exacerbated inflammation. Peripheral monocytes can serve as a source of matrix metalloprotein-9 and aggravate brain damage [25]. Lymphocytes, as the core of immune responses, are produced by lymphoid organs and primarily reside in lymph fluid, serving as key cellular components of the body’s immune defense [26]. Lymphocytes are thought to have neuroprotective and neurofunctional effects [27]. A reduction in lymphocytes may lead to deleterious inflammatory states and adverse clinical outcomes. In recent years, as a peripheral blood-derived inflammatory marker, MLR has emerged as a potential biomarker offering a more precise approach for disease assessment and monitoring. Previous studies have confirmed that a higher MLR is associated with adverse clinical outcomes in patients with HF, stroke, and coronary artery disease [28–30]. Additionally, MLR is correlated with overall survival in cancer patients (HR: 2.229, 95% CI: 1.260–3.944, p = 0.006), and exhibits convenient and cost-effective utility in predicting survival and recurrence in patients with central non-small cell lung cancer undergoing lobectomy [31].

Our study is consistent with previous research. Multivariate Cox regression analysis confirmed that an elevated MLR significantly increases the risk of 28-day all-cause mortality in patients with SAD, with each 1-unit increase in MLR corresponding to a 2–15% elevated risk. RCS curves revealed a non-linear positive association between MLR and the HR. These findings suggest that excessive inflammatory activation is a critical warning sign for short-term mortality in SAD patients, and maintaining immune homeostasis within a moderate range may serve as an effective intervention to improve the prognosis of SAD patients. To further validate this association, SAD patients were stratified into four groups based on MLR quartiles. The Q4 group (1.05 ≤ MLR <11.93) was associated with a 1.44-fold increased risk of 28-day all-cause mortality, indicating that the risk of 28-day all-cause mortality in SAD patients further escalates with increasing MLR. Subgroup analyses were performed to explore the relationship between MLR and 28-day all-cause mortality risk across different subgroups and the stability of this association. The results demonstrated that the predictive capacity of MLR for 28-day all-cause mortality risk was most prominent in patients aged > 60 years, as well as those with HF, cerebrovascular disease, chronic pulmonary disease, and renal disease. This indirectly suggests that MLR tends to exhibit high sensitivity as a crucial predictor of 28-day all-cause mortality in elderly patients and those with major comorbidities. This aligns with evidence that inflammatory indices most effectively capture delirium-related mortality risk in immunologically vulnerable states such as aging and chronic inflammation [32]. By monitoring MLR, clinicians can identify high-risk patients earlier and implement timely interventions to reduce the risk of 28-day all-cause mortality and improve patient clinical outcomes.

Furthermore, the association between the MLR and 28-day all-cause mortality was not statistically significant in the severe liver disease subgroup (HR: 1.25, 95% CI: 0.7–2.24). This finding may be attributed to the following key factors: Firstly, as a critical immune organ of the human body, the liver plays a pivotal role in immune regulation. Severe liver disease impairs hepatic immune regulatory function, disrupting monocyte/lymphocyte homeostasis—such as inducing lymphocyte apoptosis and dysregulating monocyte function [33, 34]. In this context, comorbidities associated with severe liver disease may become dominant predictors of mortality, thereby masking the prognostic value of MLR. Secondly, the small sample size of the severe liver disease subgroup may limit statistical power to detect potential true associations. Additionally, selection bias inherent in single-center studies may result in unrepresentative samples. Therefore, for patients with SAD complicated by severe liver disease, further research with larger sample sizes may be warranted to validate the relationship between MLR and 28-day all-cause mortality risk.

Building on the robust correlation between the MLR and 28-day all-cause mortality, we further evaluated the predictive performance of MLR for 28-day all-cause mortality risk. The results indicated that MLR alone had limited efficacy in predicting 28-day all-cause mortality in SAD patients (AUC = 0.60). This limitation may be attributed to the following factors: systemic inflammatory markers such as MLR indirectly reflect neuroinflammation but lack specificity for key neural events (e.g., blood-brain barrier (BBB) disruption or microglial activation), and single-time-point measurements fail to capture the inflammatory trajectory essential for SAD progression. Therefore, there is a need to develop clinically applicable survival prediction models with superior performance by integrating MLR with other relevant features. Unlike previous studies that only considered 28 day survival status [35, 36], our study incorporated both 28 day survival time and survival status as target variables, enabling a more comprehensive characterization of 28-day all-cause mortality. Tree-based random forest algorithms were used for iterative feature selection to identify and quantify the weights of variables associated with 28-day all-cause mortality. The feature selection results demonstrated that MLR had a relatively high weight in predicting 28-day all-cause mortality, and survival prediction models excluding MLR showed a significant reduction in predictive efficacy for 28-day all-cause mortality. The td-AUC curves and median survival time ROC curves revealed that the RSF model, constructed with MLR as the core feature, exhibited better and more stable predictive performance for 28-day all-cause mortality both during the follow-up period and at the median survival time. Feature importance plots further validated that MLR was a critical variable for predicting 28-day all-cause mortality in SAD patients. SHAP plots also confirmed that an increase in MLR was associated with adverse prognosis in SAD patients. These findings are consistent with the results of multivariate Cox regression analysis, further confirming that MLR is an important risk factor for 28-day all-cause mortality in SAD patients.

While the underlying mechanisms through which elevated MLR augments the risk of 28-day all-cause mortality in patients with SAD remain incompletely elucidated, the present findings offer potential mechanistic insights, which are discussed as follows: First, sepsis is a well-recognized precipitant of BBB dysfunction [37]. As a critical surrogate marker of systemic inflammatory progression, elevated MLR reflects heightened peripheral inflammatory responses in the context of sepsis. Following BBB disruption, peripherally activated immune cells and proinflammatory mediators can infiltrate the central nervous system (CNS), thereby amplifying neuroinflammatory cascades and inducing neuronal injury—pathological processes that may exacerbate delirium severity and compromise survival outcomes [38]. Second, neuroinflammation constitutes a core pathophysiological driver of SAD development and progression, and MLR may indirectly mirror the severity of neuroinflammatory responses. Neuroinflammation is closely associated with an increased risk of encephalopathy progression and is predominantly characterized by the activation of microglia and astrocytes [39]. Upon activation, these glial cells release a spectrum of proinflammatory mediators within the CNS, including TNF-α, IL-1β, and IL-6, which initiate and propagate neuroinflammatory reactions [40]. These inflammatory cytokines can further impair neuronal function, induce oxidative stress and neuronal apoptosis, and ultimately precipitate cognitive dysfunction—key features of SAD that may contribute to poor survival [41, 42]. Third, MLR serves as a marker of immune and inflammatory dysregulation in SAD. Such dysregulation is characterized by excessive production of proinflammatory cytokines and aberrant leukocyte activation, which can promote microvascular thrombosis, reduce renal perfusion and oxygen delivery, and trigger dysfunction of vital organs [9]. Meanwhile, the level of circulating extracellular vesicles (EVs) is elevated in sepsis patients. Vascular cell adhesion molecule 1 overexpressed on these EVs interacts with integrin subunit α4 on the surface of monocytes, thereby activating the Nuclear factor - κB (NF-κB) signaling pathway. This activation regulates monocyte differentiation, targets monocytes, and reprograms them into proinflammatory macrophages, ultimately exacerbating sepsis-associated acute respiratory distress syndrome [43]. These pathological alterations may culminate in multiple organ dysfunction syndrome, exacerbate septic complications, and ultimately lead to patient death.

Compared with previous studies, our research has several advantages. First, the sufficient sample size and strict control of confounding variables enhance the representativeness and reliability of our findings. We not only used Cox regression models to explore the association between the MLR and 28-day all-cause mortality but also employed survival ML algorithms to further validate the prominent role of MLR in all-cause mortality and establish survival prediction models. This approach avoids the limitations of relying solely on traditional Cox regression, combining Cox regression with survival ML algorithms to more comprehensively capture linear, non-linear, and complex relationships. Among the constructed models, the RSF, a survival analysis method capable of handling highly correlated variables without strict data distribution assumptions was identified as the optimal model [44]. Second, MLR is derived from routine complete blood count tests, which are easily accessible, cost-effective, and practical for clinical application, thereby holding broad potential for widespread use. Third, we also conducted an external validation. Models trained in the MIMIC-IV dataset were directly applied to the eICU-CRD cohort, which encompasses multiple medical centers. The models maintained relatively favorable discriminative performance, with an AUC greater than 0.65, in an independent population. These results suggest that MLR retains transportability across institutions and workflows and can complement traditional scores for early risk stratification. Finally, this study incorporated time-to-event data, avoiding the risks associated with censored data that arise when models are built solely based on outcome status. This confers greater potential value in clinical decision-making, particularly in the early identification and management of elderly SAD patients and those with comorbidities, highlighting the important clinical significance of MLR.

However, this study also has several limitations that warrant further investigation in future research. Firstly, as a retrospective cohort study, although we have strived to balance the impact of confounding factors on the study to the greatest extent possible, the results may still be subject to the influence of potential confounding factors. Although external validation was undertaken in the eICU-CRD cohort, 28-day follow-up data were unavailable. Consequently, in-hospital mortality was employed as the endpoint for external validation. Owing to the divergent follow-up structures between the training and external validation set, the construction of td-AUC for the external cohort was not feasible. Nevertheless, we assessed model performance using the C-index and the AUC at the median survival time. While the model maintained acceptable discriminative capacity (AUC > 0.65), the multicentre external validation cohort exhibited a censoring distribution distinct from that of the training dataset, population heterogeneity, regional clinical practice variations, and database-specific data collection biases which may partially explain the observed decrement in AUC. Thus, prospective studies are required to verify the present findings. Secondly, MLR data were only collected on the first day of ICU admission, which may introduce potential bias. Future studies should conduct longitudinal assessments of MLR trajectory changes to gain a more detailed understanding of the dynamic relationship between these changes and 28-day all-cause mortality. Thirdly, based on previous studies and a clearly defined time threshold for the onset of delirium, we maximized the consistency of the diagnosis of SAD with the diagnostic criteria. However, the MIMIC-IV and eICU-CRD databases lack detailed clinical data to distinguish the specific etiologies of delirium in the enrolled patients. Moreover, the progression of delirium is a multifaceted and complex process influenced by numerous biological and environmental factors, which constitutes an inherent limitation of studies based on retrospective databases. Finally, our study lacked time-dependent versions of traditional classification metrics (e.g., accuracy, sensitivity, specificity, calibration assessment, and decision-analytic evaluation). Although the mathematical foundations for these metrics have been established, their integration with widely used survival analysis library (such as scikit-survival) remains limited. The application of these metrics would require specialized methodological development, which is beyond the scope of this study. Future research should explore their practical implementation.

Conclusions Elevated MLR is independently associated with 28-day all-cause mortality in patients with SAD. Survival machine learning models established based on MLR can effectively predict 28-day all-cause mortality in SAD patients.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Abbreviations

ICUs

Intensive care units

SAD

Sepsis-associated delirium

SAE

Sepsis-associated encephalopathy

IL

Interleukin

TNF-α

Tumour necrosis factor-alpha

MLR

Monocyte-to-lymphocyte ratio

ML

Machine learning

MIMIC-IV

Medical Information Mart for Intensive Care-IV

eICU-CRD

eICU Collaborative Research Database

SOFA

Sequential organ failure assessment

CAM-ICU

Confusion assessment method for the Intensive Care Unit

MAP

Mean arterial pressure

SpO₂

Oxygen saturation

PO₂

Partial pressure of oxygen

PCO₂

Partial pressure of carbon dioxide

BUN

Blood urea nitrogen

HCO₃^-

Bicarbonate

AST

Aspartate transaminase

PTT

Partial thromboplastin time

INR

International normalized ratio

HF

Heart failure

CCI

Charlson comorbidity index

GCS

Glasgow Coma Scale

SAPS II

Simplified acute physiology score II

CRRT

Continuous renal replacement therapy

K-M

Kaplan-Meier

HR

Hazard ratio

95% CI 95%

Confidence interval

RCS

Restricted cubic spline

VIF

Variance inflation factor

ROC

Receiver operating characteristic

RSF

Random Survival Forest

AUC

Area under the curve

C-index

Concordance index

SHAP Shapley

Additive exPlanations

BBB

Blood-brain barrier

EVs

Extracellular vesicles

NF-κB

Nuclear factor - κB

Author contributions

Youhui Zhao and Hong Zhu contributed to the study concept and study design. Youhui Zhao and Jun Jiang performed the statistical analysis and data interpretation. Kun Yang, Hong Zhu, and Jianghui Pu were responsible for the quality control of the data and algorithms. Jinyou Yang, Aifei Na, Jun Ding, and Qiuyi Lu performed the literature research and data extraction. Kun Yang and Hong Zhu critically reviewed, edited, and approved the manuscript. Youhui Zhao and Jun Jiang contributed equally to this work and are co-first authors. All authors contributed to the writing of the manuscript and approved the final version.

Funding

This work was supported by the National Nature Science Foundation of China Grants (82260371), Yunnan health training project of high level talents (L-2024008), and Yunnan Revitalization Talent Support Program.

Data availability

The datasets generated and/or analyzed during the current study are available in the MIMIC-IV and eICU-CRD databases. Upon reasonable request, the corresponding author can provide all detailed information on statistical analysis code and structured data.

Declarations

Ethics approval and consent to participate

The data used in this study were obtained from the publicly available MIMIC (Medical Information Mart for Intensive Care) and eICU-CRD. The use of the MIMIC database was approved by the Institutional Review Boards (IRBs) of Beth Israel Deaconess Medical Center and the Massachusetts Institute of Technology. The use of the eICU-CRD database was approved by the IRBs of the Massachusetts Institute of Technology and Philips Healthcare. Both databases contain de-identified patient information in compliance with the Health Insurance Portability and Accountability Act (HIPAA). As all data are fully anonymized and publicly accessible, the requirement for written informed consent was waived by the corresponding IRBs. Given the anonymized nature and standardized structure of the database, additional ethical approval was not required for the present study. All methods were carried out in accordance with relevant guidelines and regulations stated in the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.