Systemic immune-inflammation (SII) index as a novel prognostic biomarker in critically ill patients with sepsis: analysis of the MIMIC-IV cohort and predictive modeling based on machine learning

Abstract

Background

Sepsis remains a leading cause of mortality in intensive care units (ICUs), highlighting the need for reliable and accessible prognostic biomarkers. The systemic immune-inflammation (SII) index, integrating neutrophils, platelets, and lymphocytes, has shown prognostic value in various diseases but remains understudied in sepsis.

Methods

We conducted a retrospective cohort study including 4,001 sepsis patients from the MIMIC-IV database. SII was calculated as (neutrophil × platelet)/lymphocyte and log-transformed due to non-normal distribution. Tertiles of log-SII were evaluated. Multivariable Cox proportional hazards models, restricted cubic splines (RCS), and Kaplan-Meier analyses were used to assess the relationship between log-SII and 30-/90-day mortality. Five machine learning (ML) models, including random forest (RF), logistic regression (LR), extreme gradient boosting (XGBoost), support vector machine (SVM), and multilayer perception (MLP) were constructed for mortality prediction. Feature contributions were further interpreted using Shapley additive explanation (SHAP) values.

Results

Patients in the low log-SII (Q1) had significantly higher 30-day (hazard ratio [HR] 1.36, 95% confidence interval [CI] 1.11–1.67) and 90-day mortality (HR 1.35, 95% CI 1.10–1.65) compared to Q2. An initial non-linear association between log-SII and mortality was observed in unadjusted and minimally adjusted models; however, this non-linear relationship was attenuated and no longer significant in the fully adjusted model (Model 3), suggesting that the apparent non-linearity may be explained by confounding factors. Subgroup analyses confirmed consistent results across most strata. Among ML models, LR demonstrated the highest discriminative performance (area under the receiver operating characteristic curve [AUROC] = 0.770, 95%CI: 0.752–0.788). SHAP analysis identified SII as a key predictor of mortality risk.

Conclusion

Low SII on ICU admission is independently associated with increased short- and long-term mortality in sepsis. As a simple and cost-effective biomarker, SII may support early risk stratification. Prospective validation and investigation of its dynamic changes are warranted.

Clinical trial number

Not applicable.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12879-026-12703-1.

Introduction

Sepsis is a life-threatening syndrome caused by a dysregulated host response to infection, leading to systemic inflammation, immune dysfunction, and multiple organ failure [1]. Despite advances in critical care, sepsis remains one of the most common causes of mortality in intensive care units (ICUs) worldwide, with one-third of sepsis-related deaths occurring within the first 48 h of ICU admission [2, 3]. This high and early mortality underscores the urgent need for reliable tools to identify high-risk patients and guide early clinical interventions.

Conventional prognostic tools based on systemic manifestations, organ dysfunction scores, or microbiological data are limited by suboptimal accuracy and insufficient real-time applicability [4]. Given the pivotal role of immune dysregulation in sepsis, biomarkers that capture inflammatory and immunologic status are gaining attention as candidates for risk stratification. Among these, the systemic immune-inflammation (SII) index, calculated as (neutrophil × platelet)/lymphocyte, offers a composite marker reflecting both innate and adaptive immune responses [5].

Initially developed in oncology and cardiovascular disease, SII has been further validated as a prognostic indicator in anemia, ischemic stroke, and sarcopenia [6–8]. The biological rationale is grounded in the interplay between neutrophils, lymphocytes, and platelets. Neutrophils, key mediators of innate immunity, contribute to excessive inflammation in sepsis, releasing reactive oxygen species (ROS) and neutrophil extracellular traps (NETs), which can exacerbate tissue damage and organ dysfunction [9]. Conversely, lymphocyte depletion, driven by accelerated apoptosis and bone marrow suppression, is a hallmark of immunosuppression and is strongly associated with poor prognosis [10]. Lymphocytes, including T cells, B cells, and NK cells, play a critical role in adaptive immunity. In sepsis, the loss of these cells, particularly CD4 + T cells, impairs pathogen clearance and increases susceptibility to secondary infections [11]. Platelets, beyond their role in coagulation, also interact with immune cells and endothelial barriers, influencing the progression of sepsis through mechanisms such as immunothrombosis and cytokine release [12]. In sepsis, this triad is frequently disrupted, suggesting that SII may serve as an integrated marker of immune-inflammatory imbalance. However, despite its clinical availability and theoretical relevance, the prognostic value of SII on critically ill septic patients remains poorly defined.

In this study, we leveraged the Medical Information Mart for Intensive Care (MIMIC-IV) database to evaluate the prognostic utility of SII in ICU patients with sepsis. We aimed to characterize the relationship between SII and short- and long-term mortality, and to integrate SII into predictive models using machine learning approaches. Additionally, we applied model explainability techniques to identify key predictors and enhance clinical interpretability.

Methods

Study population

This retrospective study was conducted using data from the Medical Information Mart for Intensive Care IV (MIMIC-IV), a publicly available and widely validated critical care database maintained by the Massachusetts Institute of Technology [13]. This database contains comprehensive clinical information from patients admitted to the intensive care units (ICUs) of Beth Israel Deaconess Medical Center. Data regarding the initial ICU admissions of patients aged 18 years or older were collected for this study. Patients with sepsis were identified based on the International Classification of Diseases, 9th and 10th Revisions, including ICD-9 codes: 995.91 (septic shock), 995.92 (severe sepsis), and 038.x (septicemia/bacteremia); ICD-10 codes: A40.x (streptococcal sepsis) and A41 (including A41.9, sepsis, unspecified organism) [14]. These ICD codes were extracted from either the primary or secondary diagnosis fields, without restriction to primary diagnosis, to ensure a representative and generalizable cohort. Exclusion criteria included: (1) age under 18 years; (2) ICU stay less than 3 h; (3) multiple ICU or hospital admissions; (4) missing values for neutrophil, platelet, or lymphocyte on the first day of ICU admission; (5) missing outcome data. Finally, a total of 4001 patients with sepsis were enrolled in this study (Fig. 1).

Fig. 1.

Flow chart of patients from MIMIC to study

Study outcomes

The primary outcomes were 30-day and 90-day all-cause mortality. Mortality status was determined based on discharge data and follow-up records available with the MIMIC-IV database.

Data extraction

Data were extracted using Structured Query Language (SQL) in a PostgreSQL environment. Baseline characteristics collected within the 24 h of ICU admission, including (1) demographic characteristics: age, sex, and weight; (2) laboratory test results: neutrophil, platelet, lymphocyte, red blood cell (RBC), white blood cell (WBC), hemoglobin, hematocrit, serum creatinine (Scr), blood urea nitrogen (BUN), anion gap, bicarbonate, chloride, sodium, potassium, and glucose; (3) severity of illness scores: Sequential Organ Failure Assessment (SOFA) score, the Acute Physiology Score III (APSIII), the Simplified Acute Physiology Score II (SAPSII), Oxford Acute Severity of Illness Score (OASIS), and Glasgow Coma Scale (GCS), and Charlson Comorbidity Index (CCI); (4) comorbidities: heart failure, arterial fibrillation, diabetes, acute myocardial infarction (AMI), chronic kidney disease (CKD), chronic obstructive pulmonary disease (COPD), hypertension, acute kidney injury (AKI), acute respiratory distress syndrome (ARDS), stroke, and septic shock; (5) vital signs: heart rate, systolic blood pressure (SBP), diastolic blood pressure (DBP), mean blood pressure (MBP), and respiratory rate; (6) clinical outcomes: length of hospital and ICU stay, in-hospital mortality, in-ICU mortality. Chronic comorbidities were identified based on ICD-9 and ICD-10 diagnosis codes recorded at the time of ICU admission, and do not include acute conditions or complications arising during the hospital stay. The systemic immune-inflammation (SII) index was calculated using formula: (neutrophil × platelet)/lymphocyte. All three cell types were quantified using their absolute counts derived from the first 24-hour ICU laboratory data in MIMI-IV to ensure the accuracy and validity of the SII calculation. In this study, variables with more than 20% missing data were excluded. Remaining missing data were imputed using the missForest package in R, a non-parametric random forest-based algorithm, and extreme values in the data were identified based on predefined physiological ranges and handled appropriately [15, 16].

Machine learning

The machine learning workflow of this study followed the systematic framework proposed in previous study, which comprises data preprocessing, feature engineering, and model training [17]. Candidate variables for model construction were selected based on the intersection of features identified by the least absolute shrinkage and selection operator (LASSO) regression and Boruta algorithm. This feature selection strategy, coupled with systematic feature engineering, aligns with established machine learning best practices that prioritize both predictive performance and model interpretability [18]. The study cohort (n = 4,001) was randomly divided into a training set (75%, n = 3,000) and an internal testing set (25%, n = 1,001) using a fixed random seed (seed = 4321) to ensure reproducibility. Stratified sampling was employed to maintain the same distribution of the outcome (30-day mortality) in both sets. Five supervised machine learning (ML) models were developed to predict 30-day mortality: random forest (RF), logistic regression (LR), extreme gradient boosting (XGBoost), support vector machine (SVM), and multilayer perception (MLP). RF is an ensemble learning method that combines multiple decision trees to enhance prediction accuracy and control overfitting [19]. LR is a linear classifier suitable for binary outcomes, offering interpretability through estimated coefficients [20]. XGBoost is a gradient-boosting framework widely used for its high efficiency and predictive performance [21]. SVM is effective in high-dimensional spaces and aims to construct an optimal separating hyperplane between calsses [22]. MLP is a type of neural network capable of modeling complex non-linear relationships, suitable for various pattern recognition tasks [23]. Given the class imbalance, we used evaluation metrics that are robust to imbalance, such as AUROC and F1 score, and did not apply any oversampling or undersampling techniques to preserve the original data distribution. For each model, hyperparameter tuning was performed using grid or random search with five-fold cross-validation on the training set [24]. This systematic parameter tuning was employed to accelerate model convergence while maintaining stability [25]. The optimal classification threshold was determined using Youden’s index (sensitivity + specificity − 1), selecting the point with the highest index on the receiver operating characteristic (ROC) curve [26]. Final models were retrained on the full training dataset using the optimal hyperparameters.

Performance evaluation

Model performance was assessed using a comprehensive set of metrics, including accuracy, AUROC, sensitivity, specificity, Youden’s J, F1 Score, and MCC. Calibration was evaluated using calibration plots with a 45-degree reference line representing perfect agreement between predicted and observed probabilities. Model calibration accuracy was further quantified using the Brier score. Clinical utility was assessed using decision curve analysis (DCA), which estimates the net benefit of each prediction model relative to the “treat all” and “treat-none” strategies across the threshold probabilities ranging from 0% to 100%. Model interpretability was addressed using Shapley additive explanation (SHAP) method [27]. Global SHAP analysis were calculated to identify the most influential features across the entire cohort, while local SHAP values provided individualized insights into patient-specific risk drivers.

Statistical analyses

Continuous variables with a normal distribution were presented as mean ± standard deviation (SD) and compared using the Student’s t-test. Categorical variables were summarized as counts (percentages) and compared using Chi-square or Fisher’s exact tests, as appropriate.

Due to the non-normal distribution of SII, a natural logarithm transformation was applied prior to analysis. The log-transformed SII values were then stratified into tertiles, including Q1 (n = 1334, 0.035 ≤ log-SII ≤ 7.331), Q2 (n = 1333, 7.331 < log-SII ≤ 8.294), Q3 (8.294 < log-SII ≤ 11.651). To assess the association between log-SII and mortality, Cox proportional hazards regression models were employed. Hazard ratios (HRs) with 95% confidence intervals (CIs) were calculated using the middle tertile (Q2) as the reference group. A restricted cubic spline (RCS) function with three pre-specified knots was used to explore the relationship continuous log-SII and mortality. The Wald test was applied to evaluate the presence of non-linearity. A threshold-effect model was developed to determine the inflection point of the curve. Both 30-day and 90-day all-cause mortality were analyzed. Kaplan-Meier survival curves were generated for each tertile and compared using the log-rank test. Multivariate Cox regression analyses were performed: Model 1: Not adjusted; Model 2: Adjusted for age, sex; Model 3: Adjusted for the variables in model 2 and further adjusted for RBC, WBC, hemoglobin, hematocrit, Scr, BUN, anion gap, bicarbonate, chloride, sodium, potassium, glucose, SOFA, APSIII, SAPSII, OASIS, GCS, CCI, heart failure, arterial fibrillation, diabetes, AMI, CKD, COPD, hypertension, ARDS, stroke, and septic shock. Finally, to ensure the robustness of our findings, subgroup analyses were conducted based on clinically relevant strata, including sex, age, severity scores (e.g., SOFA, SAPSII), and key comorbidities. All statistical tests were two-tailed, with a significance level set at P < 0.05. Analyses were performed using R software (version 4.2.2).

Results

Baseline characteristics of study population

A total of 4,001 patients with sepsis were identified from the MIMIC-IV database, including 1797 (44.9%) female and 2204 (55.1%) male. The mean age was 67.03 ± 16.20 years. Baseline characteristics were stratified according to tertiles of log-transformed systemic immune-inflammation index (log-SII) and are presented in Table 1. Compared to patients in the higher log-SII group, those in the low-SII group were generally younger and had lower levels of neutrophil, platelet, hematocrit, hemoglobin, RBC, WBC. They also had a lower prevalence of hypertension but presented with higher sodium level and higher SOFA and APSIII scores. In terms of clinical outcomes, patients in lower log-SII group had significantly higher in-hospital and ICU mortality rates, as well as longer hospital and ICU stays. As shown in Table 2, non-survivors exhibited lower mean SII values compared to survivors (4063.71 ± 5883.2 vs. 4367.51 ± 5714.92). The lower log-SII (Q1) accounted for the largest proportion among non-survivors (39.53%) and the smallest among survivors (30.87%).

Table 1.

Baseline characteristics of patients categorized by log-transformed SII tertiles

Log-transformed SII tertile	Total (n = 4001)	Q1(0.035–7.331) (n = 1334)	Q2(7.331–8.294) (n = 1333)	Q3(8.294–11.651) (n = 1334)	p.value
Demographic characteristics
Age (years)	67.03 ± 16.20	65.24 ± 16.41	67.09 ± 16.32	68.75 ± 15.67	< 0.0001
Sex					0.13
Female	1797(44.91)	588(44.08)	580(43.51)	629(47.15)
Male	2204(55.09)	746(55.92)	753(56.49)	705(52.85)
Weight	83.15 ± 26.72	83.63 ± 26.63	84.58 ± 27.44	81.23 ± 25.99	< 0.01
Laboratory tests
RBC	3.41 ± 0.70	3.32 ± 0.75	3.47 ± 0.66	3.45 ± 0.67	< 0.0001
WBC	16.40 ± 14.37	13.55 ± 20.22	15.62 ± 8.08	20.05 ± 11.13	< 0.0001
Hemoglobin	10.04 ± 1.96	9.95 ± 2.04	10.23 ± 1.94	9.95 ± 1.87	< 0.001
Hematocrit	31.11 ± 5.87	30.55 ± 6.16	31.66 ± 5.71	31.13 ± 5.68	< 0.0001
Scr	1.92 ± 1.65	1.93 ± 1.66	1.92 ± 1.64	1.91 ± 1.65	0.95
BUN	37.31 ± 27.56	36.53 ± 27.13	37.68 ± 28.14	37.73 ± 27.42	0.44
Anion gap	16.31 ± 5.65	16.50 ± 5.97	16.34 ± 5.67	16.08 ± 5.29	0.16
Bicarbonate	21.34 ± 4.59	21.26 ± 4.61	21.33 ± 4.56	21.42 ± 4.59	0.67
Chloride	103.85 ± 6.72	104.13 ± 7.23	103.87 ± 6.49	103.54 ± 6.41	0.07
Sodium	138.61 ± 5.69	138.99 ± 5.96	138.67 ± 5.64	138.18 ± 5.44	< 0.001
Potassium	4.44 ± 0.77	4.42 ± 0.77	4.42 ± 0.78	4.47 ± 0.75	0.25
Glucose	162.88 ± 107.80	159.72 ± 81.51	163.91 ± 75.59	165.00 ± 150.02	0.41
Severity of illness scores
SOFA	7.38 ± 4.15	8.41 ± 4.33	7.06 ± 4.18	6.69 ± 3.71	< 0.0001
APSIII	59.32 ± 24.12	61.32 ± 25.64	58.01 ± 23.56	58.64 ± 22.96	< 0.001
SAPSII	44.94 ± 15.91	45.33 ± 16.40	44.10 ± 15.85	45.38 ± 15.46	0.06
OASIS	35.91 ± 9.03	35.84 ± 8.94	35.72 ± 9.22	36.18 ± 8.94	0.39
GCS	13.46 ± 2.96	13.44 ± 3.03	13.42 ± 2.99	13.52 ± 2.86	0.65
CCI	5.67 ± 3.15	5.67 ± 3.22	5.59 ± 3.10	5.76 ± 3.12	0.41
Comorbidities
Heart failure	25(0.62)	11(0.82)	6(0.45)	8(0.60)	0.47
Arterial fibrillation	1330(33.24)	427(32.01)	443(33.23)	460(34.48)	0.40
Diabetes	1278(31.94)	417(31.26)	455(34.13)	406(30.43)	0.10
AMI	6(0.15)	2(0.15)	2(0.15)	2(0.15)	1.00
CKD	964(24.09)	297(22.26)	330(24.76)	337(25.26)	0.15
COPD	584(14.60)	179(13.42)	192(14.40)	213(15.97)	0.17
Hypertension	2496(62.38)	780(58.47)	843(63.24)	873(65.44)	< 0.001
AKI	2633(65.81)	903(67.69)	881(66.09)	849(63.64)	0.09
ARDS	353(8.82)	114(8.55)	116(8.70)	123(9.22)	0.81
Stroke	506(12.65)	179(13.42)	181(13.58)	146(10.94)	0.07
Septic shock	65(1.62)	17(1.27)	26(1.95)	22(1.65)	0.38
Vital signs
Heart_rate	91.46 ± 17.66	91.76 ± 18.11	90.58 ± 17.12	92.02 ± 17.71	0.08
SBP	110.84 ± 13.65	110.60 ± 13.80	111.27 ± 13.77	110.66 ± 13.39	0.37
DBP	61.97 ± 9.86	62.13 ± 9.86	62.22 ± 10.09	61.55 ± 9.62	0.16
MBP	76.04 ± 9.47	76.15 ± 9.43	76.33 ± 9.76	75.65 ± 9.21	0.16
Respiratory_rate	21.55 ± 4.48	21.44 ± 4.65	21.51 ± 4.40	21.70 ± 4.37	0.30
Clinical outcomes
LOS hospital, days	17.21 ± 18.53	19.01 ± 20.63	16.76 ± 17.74	15.86 ± 16.89	< 0.0001
LOS ICU days	6.36 ± 8.74	6.78 ± 9.33	6.37 ± 8.63	5.95 ± 8.20	0.05
ICU mortality	826(20.64)	331(24.81)	262(19.65)	233(17.47)	< 0.0001
Hospital mortality	1141(28.52)	451(33.81)	346(25.96)	344(25.79)	< 0.0001
SII parameters
Neutrophil	13.74 ± 9.64	8.53 ± 7.34	13.39 ± 7.09	19.30 ± 10.80	< 0.0001
Platelet	218.99 ± 128.41	143.84 ± 92.35	219.60 ± 101.18	293.55 ± 139.67	< 0.0001
Lymphocyte	1.52 ± 11.31	2.71 ± 19.52	1.12 ± 0.69	0.72 ± 0.54	< 0.0001
SII	4280.88 ± 5764.31	775.36 ± 434.32	2587.50 ± 703.82	9478.50 ± 7538.04	< 0.0001
Log-transformed SII	7.70 ± 1.35	6.29 ± 1.21	7.82 ± 0.28	8.98 ± 0.55	< 0.0001

RBC, red blood cell; WBC, white blood cell; Scr, serum creatinine; BUN, blood urea nitrogen; SOFA, sequential organ failure assessment; APSIII, acute physiology score III; SAPSII, simplified acute physiological score II; OASIS, oxford acute severity of illness score; GCS, Glasgow coma scale; CCI, Charlson comorbidity index; AMI, acute myocardial infarction; CKD, chronic kidney disease; COPD, chronic obstructive pulmonary disease; AKI, acute kidney injury; ARDS, acute respiratory distress syndrome; SBP, systolic blood pressure; DBP, diastolic blood pressure; MBP, mean blood pressure; LOS, length of stay

Table 2.

Baseline characteristics of the survivors and Non-survivors groups

Categories	Non-survivors (n = 1141)	Survivors (n = 2860)	p.value
Demographic characteristics
Age (years)	69.84 ± 15.06	65.90 ± 16.50	< 0.0001
Sex			0.16
Female	492(43.12)	1305(45.63)
Male	649(56.88)	1555(54.37)
Weight	82.83 ± 25.99	83.27 ± 27.02	0.63
Laboratory tests
RBC	3.33 ± 0.74	3.44 ± 0.68	< 0.0001
WBC	17.53 ± 13.74	15.96 ± 14.60	< 0.01
Hemoglobin	9.95 ± 2.09	10.08 ± 1.90	0.08
Hematocrit	30.92 ± 6.31	31.19 ± 5.69	0.22
Scr	2.35 ± 1.66	1.75 ± 1.61	< 0.0001
BUN	46.66 ± 29.70	33.59 ± 25.74	< 0.0001
Anion gap	19.08 ± 6.92	15.20 ± 4.61	< 0.0001
Bicarbonate	20.19 ± 5.37	21.79 ± 4.15	< 0.0001
Chloride	103.00 ± 7.39	104.18 ± 6.41	< 0.0001
Sodium	138.83 ± 6.23	138.52 ± 5.46	0.14
Potassium	4.66 ± 0.86	4.35 ± 0.71	< 0.0001
Glucose	177.33 ± 102.25	157.11 ± 109.43	< 0.0001
Severity of illness scores
SOFA	9.76 ± 4.31	6.44 ± 3.67	< 0.0001
APSIII	74.56 ± 26.47	53.24 ± 20.12	< 0.0001
SAPSII	54.62 ± 15.73	41.07 ± 14.26	< 0.0001
OASIS	40.53 ± 8.93	34.07 ± 8.39	< 0.0001
GCS	12.86 ± 3.62	13.70 ± 2.61	< 0.0001
CCI	6.51 ± 2.96	5.34 ± 3.16	< 0.0001
Comorbidities
Heart failure	16(1.40)	9(0.31)	< 0.001
Arterial fibrillation	476(41.72)	854(29.86)	< 0.0001
Diabetes	344(30.15)	934(32.66)	0.13
AMI	3(0.26)	3(0.10)	0.48
CKD	326(28.57)	638(22.31)	< 0.0001
COPD	192(16.83)	392(13.71)	0.01
Hypertension	733(64.24)	1763(61.64)	0.13
AKI	910(79.75)	1723(60.24)	< 0.0001
ARDS	156(13.67)	197(6.89)	< 0.0001
Stroke	177(15.51)	329(11.50)	< 0.001
Septic shock	25(2.19)	40(1.40)	0.10
Vital signs
Heart rate	94.44 ± 18.66	90.26 ± 17.10	< 0.0001
SBP	108.37 ± 14.08	111.83 ± 13.36	< 0.0001
DBP	60.00 ± 10.07	62.76 ± 9.66	< 0.0001
MBP	74.26 ± 9.72	76.75 ± 9.28	< 0.0001
Respiratory rate	22.72 ± 4.85	21.08 ± 4.23	< 0.0001
Clinical outcomes
LOS hospital, days	13.36 ± 16.72	18.74 ± 18.99	< 0.0001
LOS ICU days	6.97 ± 8.68	6.12 ± 8.75	< 0.01
ICU mortality	772(67.66)	54(1.89)	< 0.0001
SII parameters
Neutrophil	13.83 ± 10.62	13.71 ± 9.23	0.73
Platelet	197.63 ± 122.59	227.52 ± 129.71	< 0.0001
Lymphocyte	1.63 ± 8.58	1.47 ± 12.24	0.65
SII	4063.71 ± 5883.20	4367.51 ± 5714.92	0.14
Log-transformed SII	7.51 ± 1.51	7.77 ± 1.27	< 0.0001
Q catrgory			< 0.0001
Q1	451(39.53)	883(30.87)
Q2	346(30.32)	987(34.51)
Q3	344(30.15)	990(34.62)

RBC, red blood cell; WBC, white blood cell; Scr, serum creatinine; BUN, blood urea nitrogen; SOFA, sequential organ failure assessment; APSIII, acute physiology score III; SAPSII, simplified acute physiological score II; OASIS, oxford acute severity of illness score; GCS, Glasgow coma scale; CCI, Charlson comorbidity index; AMI, acute myocardial infarction; CKD, chronic kidney disease; COPD, chronic obstructive pulmonary disease; AKI, acute kidney injury; ARDS, acute respiratory distress syndrome; SBP, systolic blood pressure; DBP, diastolic blood pressure; MBP, mean blood pressure; LOS, length of stay

Study outcomes

In the stratified analysis by log-SII tertiles, patients in the lower log-SII (Q1) tertile had higher 30-day and 90-day mortality rates compared to those in middle (Q2) and higher (Q3) tertiles (Table S1). Kaplan-Meier curves further revealed significant differences in both 30-day and 90-day survival probabilities across the three SII groups (Fig. 2).

Fig. 2.

Kaplan-Meier survival analysis for (A) 30-day and (B) 90-day mortality according to log-transformed SII. (log-transformed SII tertile Q1: 0.035–7.331; Q2: 7.331–8.294; Q3: 8.294–11.651)

Relationship between log-SII and clinical outcomes of patients with sepsis

Cox proportional hazards regression models were used to investigate the independent association between log-SII and both 30-day and 90-day mortality. Three models were constructed: Model 1 (unadjusted), Model 2 (adjusted for age, sex), and Model 3 (further adjusted for RBC, WBC, hemoglobin, hematocrit, Scr, BUN, anion gap, bicarbonate, chloride, sodium, potassium, glucose, SOFA, APSIII, SAPSII, OASIS, GCS, CCI, heart failure, arterial fibrillation, diabetes, AMI, CKD, COPD, hypertension, ARDS, stroke, and septic shock). Using the middle (Q2) tertile as the reference group, the hazard ratios (HR) and 95% confidence intervals (CIs) for 30-day mortality were as follows: Q1 (1.47, 1.24–1.75), Q3 (1.03, 0.86–1.24) in Model 1; Q1 (1.54, 1.29–1.83), Q3 (1, 0.84–1.20) in Model 2; Q1 (1.36, 1.11–1.67), Q3 (1.01, 0.82–1.25) in Model 3 (Fig. 3A). Regarding 90-day mortality, the corresponding HRs and 95% CIs were Q1 (1.46, 1.24–1.73), Q3 (1.00, 0.84–1.19) in Model 1; Q1 (1.52, 1.29–1.80), Q3 (0.98, 0.83–1.17) in Model 2; Q1 (1.35, 1.10–1.65), Q3 (0.95, 0.80–1.20) in Model 3 (Fig. 3B). Detailed results are provided in Table S2 and S3. These findings indicate that patients with log-SII ≤ 7.33 had significantly higher risk of both 30-day and 90-day all-cause mortality compared to those with log-SII > 7.33. No significant difference in mortality risk was observed between patients in the middle tertile (Q2, log-SII: 7.33–8.294) and in the higher tertile (Q3, log-SII: 8.294–11.65). Additionally, restricted cubic spline (RCS) analysis revealed a non-linear correlation (p-overall < 0.05, p non-linear < 0.05) between log-SII and both 30-day and 90-day mortality in Model 1 and Model 2, but not in Model 3 (Figure S1).

Fig. 3.

The association between log-transformed SII groups and (A) 30-day, (B) 90-day mortality. Model 1: Unadjusted; Model 2: Adjusted for age, sex; Model 3: further adjusted for WBC, hemoglobin, hematocrit, Scr, BUN, anion gap, bicarbonate, chloride, sodium, potassium, glucose, SOFA, APSIII, SAPSII, OASIS, GCS, CCI, heart failure, arterial fibrillation, diabetes, AMI, CKD, COPD, hypertension, ARDS, stroke, and septic shock. The Q2 group was set as reference group

Subgroup analyses

Subgroup analyses were performed to assess the association between log-SII and both 30-day mortality and 90-day mortality across clinically relevant strata, including sex, age, SAPSII, SOFA, diabetes, stroke, CKD, COPD, and ARDS. Interaction test was used to explore whether the effect of log-SII on mortality was consistent across subgroups. In Fig. 4, patients in the lower of log-SII (Q1) had significantly higher 30-day mortality across nearly all subgroups, with no evidence of interaction effects. These findings suggest a stable association between low log-SII and increased mortality risk. However, the relationship was not significantly in patients with SOFA scores ≤ 3, stroke, or ARDS, indicating that in these subgroups, lower log-SII may have limited prognostic values. There were no significant differences in mortality between the middle (Q2) and higher (Q3) tertiles in any subgroup. Similar results were observed for 90-day mortality (Figure S2). Additionally, when log-SII was analyzed as a continuous variable, an inverse relationship between log-SII and both 30-day and 90-day mortality was observed in most subgroups, again with the exception of patients with SOFA scores ≤ 3, stroke, and ARDS (Figure S3, S4).

Fig. 4.

Subgroup analysis for the 30-day mortality and log-transformed SII tertile in different groups

Selection of features

Common clinical indicators—including weight, laboratory values (hemoglobin, RBC, sodium, bicarbonate, monocyte, platelet, lymphocyte, neutrophil), severity scores (APSIII, GCS, CCI), and the log-SII—were initially selected as candidate variables, with 30-day mortality defined as the outcome for subsequent analysis. Previous RCS analysis revealed a non-linear relationship between log-SII and mortality. To enhance interpretability and reduce unnecessary complexity in model construction, log-SII was categorized into two groups: Q1 as one group, and Q2–Q3 as the other. Lasso regression was then applied to perform variable selection among the 13 candidates. As shown in the coefficient path and cross-validation plots, 10 variables were retained at the optimal penalty parameter (λ = 0.00188), excluding weight, bicarbonate, and lymphocyte (Fig. 5A and B). Receiver operating characteristic (ROC) analysis demonstrated good discriminative performance of the model, with an area under the curve (AUC) of 0.761 (Fig. 5C). Using Youden’s index, the optimal threshold (0.48) achieved a sensitivity of 81.3% and specificity of 58.3%, reflecting balanced classification performance for identifying high-risk patients. Additionally, feature selection using the Boruta algorithm confirmed 12 important predictors (excluding weight), all of which had importance scores significantly higher than those of shadow features (Fig. 5D). The intersection of variables identified by both Lasso and Boruta methods was used as the final feature set for machine learning modeling (Fig. 5E).

Fig. 5.

Feature selection for machine learning modeling. (A) LASSO regression path showing coefficient shrinkage with increasing penalization; (B) The optimal λ value identified by cross-validation; (C) Receiver operating characteristic (ROC) curve with the optimal cutoff point; (D) Feature importance plot derived from the Boruta algorithm; (E) Overlap of selected features from Lasso and Boruta analyses

Construction and evaluation of models

A total of 10 independent variables and one dependent outcome were used to construct five machine learning (ML) models. Comparative analysis showed that logistic regression (LR) achieved the highest discriminative ability, with an area under the receiver operating characteristic curve (AUROC) of 0.770 (95%CI: 0.752–0.788), followed by random forest (RF, 0.740, 95%CI: 0.721–0.759), extreme gradient boosting (XGBoost, 0.724, 95%CI: 0.705–0.743), support vector machine (SVM, 0.724, 95%CI: 0.705–0.743), and multilayer perceptron (MLP, 0.689, 95%CI: 0.669–0.709) (Fig. 6A and B). Model calibration was assessed using calibration plots and Brier scores. Among all models, LR showed the best calibration with the lowest Brier score (0.159), while MLP exhibited the poorest calibration with the highest Brier score (0.197) (Fig. 6C). Notably, SVM and XGBoost showed minor deviations in the high probability range, whereas MLP tended to overestimate risk at lower probability levels. Decision curve analysis (DCA) demonstrated that the LR model provided the greatest net clinical benefit across a threshold probability range of 0–70%, outperforming the default strategies of treating all or treating none (Fig. 6D).

Fig. 6.

Performance comparison of five machine learning models. (A) Receiver operating characteristic (ROC) curves for logistic regression (LR), random forest (RF), XGBoost, support vector machine (SVM), and multilayer perceptron (MLP); (B) Summary of performance metrics for each model; (C) Calibration curves and Brier scores assessing agreement between predicted probabilities and observed outcomes; (D) Decision curve analysis (DCA) comparing net clinical benefit

Model explanation

Feature importance was first evaluated using a permutation-based method, which ranked predictors based on their average impact on model performance. As illustrated in Fig. 7A, APSIII, CCI, and RBC consistently ranked as the top contributors, indicating their strong influence on model output. To investigate individualized prediction behavior, SHapley additive explanations (SHAP) values were calculated for representative patient samples. In the example shown in Fig. 7B, features such as APSIII (+ 0.260), CCI (+ 0.078), RBC (+ 0.046), and monocyte (+ 0.012) contributed positively to the predicted mortality risk, whereas sodium (− 0.001), neutrophil (− 0.004), SII (− 0.009), GCS (− 0.021), hemoglobin (− 0.027), and platelet (-0.03) count exerted a negative impact. Furthermore, a SHAP summary plot was generated to visualize the distribution of feature contributions across the entire cohort (Fig. 7C). This plot revealed both the magnitude and directionality of each feature’s effect on the model, as well as the dependency patterns. Notably, APSIII and CCI exhibited clear monotonic trends, with higher values consistently associated with increased predicted risk.

Fig. 7.

Feature contribution analysis using permutation and SHAP frameworks. (A) Permutation-based feature importance ranking showing the average impact of each predictor on model performance. (B) SHAP value decomposition for an individual case illustrating the direction and magnitude of each feature’s contribution to the predicted outcome. (C) SHAP summary plot displaying the overall distribution of feature effects across the cohort, indicating both the relative importance and directional influence of each variable

Discussion

In this large-scale retrospective study involving 4,001 ICU patients with sepsis, we found that lower systemic immune-inflammation (SII) index levels at first ICU admission were independently associated with higher 30-day and 90-day mortality. Notably, patients in the lowest log-SII (≤ 7.33) tertile exhibited significantly worse survival outcomes compared to those with middle or high log-SII levels, even after rigorous adjustment for demographics, laboratory parameters, severity scores, and comorbidities. Additionally, among five machine learning (ML) models, logistic regression (LR) incorporating SII alongside established clinical predictors achieved strong discriminative performance with an area under the receiver operating characteristic curve (AUROC) of 0.77. These findings underscore the prognostic value of SII as a simple and accessible biomarker for risk stratification in critical ill septic patients.

Sepsis is characterized by a dysregulated immune response to infection, which typically evolves from an initial hyperinflammatory phase to a subsequent immunosuppressive state. It is also associated with significantly higher rates of rehospitalization and substantially increased healthcare expenditures, contributing to a growing global public health burden [28]. Thus, there is an urgent need to identify reliable biomarkers for early intervention in septic patients, particularly those capable of predicting disease progression prior to overt organ dysfunction, which could significantly reduce morbidity and resource utilization. Despite decades of biomarker development, existing indicators often lack sensitivity, specificity, or practicality for real-time risk assessment [29]. In this context, the SII-a composite index integrating neutrophils, platelets, and lymphocytes-may serve as a comprehensive reflection of the host immune-inflammatory status.

Each component of the SII plays a distinct role in sepsis pathophysiology. Neutrophils, the most abundant circulating leukocytes, can timely and effectively sense and eliminate microorganisms [30]. During sepsis, immune activation recruits large numbers of neutrophils to the infection sites, where they engage in phagocytosis, pathogen killing, and cytokines secretion as part of the immune response [31]. However, overactivation of neutrophils may exacerbate tissue damage and propagate systemic inflammation [32]. Recently, a novel neutrophil-based nanorobot emphasize the therapeutic potential of modulating neutrophil activity in sepsis [33]. Likewise, platelets play a crucial role in immune response and inflammation. They interact with leukocytes, endothelial cells, and the innate immune system, contributing to pathogen clearance and immune regulation [34]. Previous study has found that patients with sepsis often presents with thrombocytopenia due to platelet activation and consumption, and it is associated higher mortality, increased risk of multi-organ failure, and prolonged ICU stays, acting as a key prognostic indicator [35]. Lymphocytes, a type pf white blood cell (WBC), are essential for adaptive immunity, and are primarily classified into three types: T cells, B cells, and natural killer cells. Septic patients who died are more likely to manifest severe lymphopenia, which means depletion in absolute lymphocyte count, but the prognostic results are contradictory for subgroup analyses [36]. While individual cell counts offer mechanistic insights, they may inadequately capture the complex interplay of immune suppression and hyperinflammation in sepsis. Previous studies have identified that neutrophil-to-lymphocyte ratio (NLR) and platelet- to-lymphocyte ratio (PLR) as potential biomarkers to predict the sepsis prognosis [37, 38]. However, the prognostic value of integrating all three components-neutrophils, platelets, and lymphocyte-into a single index remains unexplored. The SII consolidates these parameters and has shown prognostic relevance in other diseases such as cancers [39], cardiovascular diseases [40], and metabolic syndrome [41]. Our findings extend this utility to sepsis, revealing a L-type relationship between SII levels and hospital/ICU stays, as well as mortality risk in sepsis patients. Then we found that compared with died patients with sepsis, alive patients with sepsis had higher value of platelet, but no significant difference was found in neutrophil, lymphocyte, and SII. This might attributable to the complexity in sepsis and the non-normal data distribution of SII [42]. Therefore, we applied log-SII in this study, which is significant higher in alive patients with sepsis. We further divided the log-SII values into three subgroups: low log-SII (Q1, 0.035–7.331), middle log-SII (Q2, 7.311–8.294), and high log-SII (Q3, 8.294–11.651). In the survival analysis, we found that patients with higher log-SII value have higher survival probability both in 30-day and 90-day mortality, indicating its protect role in sepsis. However, recent reports have proposed that the higher level of log-SII significantly increased all-cause and hypertension mortality, demonstrating the two-sidedness of SII as a biomarker for different diseases [43]. The paradoxical association between lower SII and increased mortality in sepsis may reflect the transition from hyperinflammatory to immunosuppressive phases. During early sepsis, elevated neutrophils and platelets (high SII) signify intact innate immune mobilization, whereas subsequent lymphocyte depletion and thrombocytopenia (low SII) may indicate bone marrow exhaustion and immune paralysis—a state linked to secondary infections and fatal outcomes [31, 44]. This hypothesis aligns with our subgroup findings: in patients with SOFA scores ≤ 3 (presumed poor organ dysfunction), SII lost predictive power, possibly because immune dysregulation had not yet fully manifested. Conversely, in advanced cases (SOFA scores > 3), low SII effectively captured immunosuppressive trajectories, mirroring post-mortem studies showing lymphocyte apoptosis dominance in non-survivors. Additionally, Contrary to oncology studies where high SII predicts poor prognosis, our inverse association underscores sepsis-specific pathophysiology [45, 46]. In cancer, sustained SII elevation reflects chronic inflammation promoting tumor progression [47]. In sepsis, however, rapid SII decline may signal immunometabolic collapse—a notion supported by the ‘immune clock’ theory proposing that sequential hyperinflammation and immunosuppression determine sepsis mortality [48]. The non-linear relationship between log-SII and mortality, evidenced by restricted cubic spline analysis, suggests a threshold effect. Below a log-SII of ~ 7.33, mortality risk escalates sharply, whereas higher values do not confer additional survival benefits. Subgroup analyses confirmed the robustness of these findings across strata defined by age, sex, and comorbidities. However, SII lost predictive value in patients with low SOFA scores (≤ 3), stroke, or ARDS, which may reflect distinct pathophysiology. For example, preserved organ function (SOFA scores ≤ 3) might mitigate immune dysregulation’s impact, while stroke and ARDS involve unique inflammatory pathways that could overshadow SII’s predictive capacity [49, 50].

Additionally, we conducted lasso regression and Boruta algorithm, and identified 10 key predicators (APSIII, CCI, RBC, and others) that contributed to the SII-based model. Our ML models, particularly LR, demonstrated strong discriminative power, with an AUROC of 0.77, highlighting LR’s superior ability to predict patient outcomes compared to other models (e.g., RF, XGBoost, SVM, and MLP). Furthermore, calibration plots and Brier scores confirmed LR’s robust performance, suggesting it may be the optimal tool for bedside risk prediction in sepsis. The Shapley additive explanation (SHAP) analysis in this study further elucidated the contribution of individual features to model predictions. Features like APSIII and CCI emerged as the most influential, providing valuable insights into the underlying drivers of sepsis outcomes, which is consistent with previous studies [51, 52]. In contrast, features like GCS, hemoglobin, and platelet counts decreased the probability of adverse outcomes. While SII is a valuable biomarker, it should be interpreted within the context of a comprehensive risk assessment framework. These results highlight the dynamic and multifactorial nature of sepsis progression, which needs more comprehensive evaluation through integrated biomarkers rather than relying on single parameters.

Our study has several strengths. First, we leveraged the MIMIC-IV database, which offers a large and representative ICU cohort with high-quality longitudinal data. Second, we systemically evaluated SII using both traditional statistical models and advanced ML approaches, enhancing the interpretability and robustness of our findings. Third, SII is derived from routine blood tests, making it feasible for widespread adoption in ICU settings. However, limitations should be noted. Although we employed rigorous 5-fold cross-validation and hold-out testing, it does not replace the need for external validation in independent populations. SII was calculated using laboratory data from the first 24 h of ICU admission; dynamic changes over time were not captured. Moreover, we used total lymphocyte and neutrophil counts without analyzing subtypes, which may limit granularity in immune characterization. Finally, external validation in independent multicenter cohorts is necessary to confirm the generalizability of the identified SII threshold and its prognostic value across diverse clinical settings.

In conclusion, SII is a readily available, cost-effective biomarker that independently predicts short- and long-term mortality in ICU patients with sepsis. However, its prognostic contribution is modest when considered alongside other established clinical predictors. Incorporating SII into predictive models may enhance risk stratification, but it should be viewed as a supplementary tool rather than a dominant biomarker. Future prospective studies should explore the temporal dynamics of SII and its role in guiding immunomodulatory interventions in sepsis care.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

SII

Systemic immune-inflammation

ICUs

Intensive care units

RCS

Restricted cubic splines

ML

Machine learning

RF

Random forest

LR

Logistic regression

XGBoost

Extreme gradient boosting

SVM

Support vector machine

MLP

Multilayer perception

SHAP

Shapley additive explanation

HR

Hazard ratio

CI

Confidence interval

AUROC

Area under the receiver operating characteristic curve

ROS

Reactive oxygen species

NETs

Neutrophil extracellular traps

MIMIC-IV

The Medical Information Mart for Intensive Care

SQL

Structured Query Language

RBC

Red blood cell

WBC

White blood cell

Scr

Serum creatinine

BUN

Blood urea nitrogen

SOFA

Sequential Organ Failure Assessment

APSIII

The Acute Physiology Score III

GCS

Glasgow Coma Scale

AMI

Acute myocardial infarction

CKD

Chronic kidney disease

COPD

Chronic obstructive pulmonary disease

AKI

Acute kidney injury

ARDS)

Acute respiratory distress syndrome

SBP

Systolic blood pressure

DBP

Diastolic blood pressure

MBP

Mean blood pressure

LASSO

Least absolute shrinkage and selection operator

DCA

Decision curve analysis

SD

Standard deviation

ROC

Receiver operating characteristic

AUC

Area under the curve

NLR

Neutrophil-to-lymphocyte ratio

PLR

Platelet- to-lymphocyte ratio

Author contributions

XDZ and YQX conducted the conception and design of the study. YL extracted the data from the MIMIC-IV database. MMT performed the statistical analysis and interpretation. YQW guided the manuscript review and editing. JHL and SYZ supervised the conceptualization. All authors read and approved the final manuscript.

Funding

This article is supported by 2024 Sichuan Provincial Cadre Health Care Research Project (2024 − 805).

Data availability

The raw data supporting the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

This study adhered to the Declaration of Helsinki. As MIMIC-IV contains anonymized data, patient consent was waived by Beth Israel Deaconess Medical Center (BIDMC), and no further ethical approval was required by the Ethics Committee at Hospital Clinic in Barcelona due to the dataset’s public availability.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.