Toward precision risk stratification: predicting in-hospital death in sepsis patients with MRSA bacteremia

Abstract

Patients with sepsis and concurrent methicillin-resistant Staphylococcus aureus (MRSA) bloodstream infection face a substantial risk of mortality. This study aimed to develop and validate a pragmatic, interpretable prediction model for in-hospital mortality in this high-risk population. We conducted a retrospective, single-center cohort study including 1,605 eligible patients, who were randomly divided into a training set (n = 1,124) and a validation set (n = 481). A rigorous, multi-stage feature selection pipeline—integrating univariate analysis, multi-model machine learning importance assessment, the Boruta algorithm, and stability selection via LASSO—was applied to 64 candidate variables. This process identified seven parsimonious, clinically accessible predictors: Glasgow Coma Scale score, minimum pH, maximum blood urea nitrogen, minimum white blood cell count, minimum platelet count, maximum lactate level, and the presence of pneumonia. Among six compared machine learning algorithms, logistic regression was selected for its optimal balance of performance and inherent interpretability. The final model demonstrated strong discriminative ability, with an area under the receiver operating characteristic curve (AUC) of 0.861 (95% CI: 0.840–0.882) in the training set and 0.844 (95% CI: 0.811–0.877) in the independent validation set, and showed good calibration (Hosmer-Lemeshow test p = 0.274). Decision curve analysis confirmed superior clinical net benefit across a wide range of risk thresholds. The model maintained robust and equitable performance across key patient subgroups and exhibited stability in extensive sensitivity analyses, including multiple imputation and bootstrap internal validation. This interpretable, seven-variable logistic regression model provides a clinically actionable tool for early mortality risk stratification, potentially supporting timely and tailored intervention strategies for sepsis patients with MRSA bacteremia.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12879-026-12584-4.

Introduction

Sepsis and septic shock represent a monumental global health challenge, characterized by dysregulated host response to infection leading to life-threatening organ dysfunction [1]. They are leading causes of mortality in intensive care units (ICUs) worldwide, contributing to substantial healthcare expenditures and long-term morbidity among survivors [2]. This burden is exponentially compounded by the concurrent crisis of antimicrobial resistance (AMR). The Global Burden of Disease Study 2019 estimated that bacterial AMR was directly associated with millions of deaths, underscoring its urgency as a paramount public health threat [3]. Among the most formidable antibiotic-resistant pathogens is methicillin-resistant Staphylococcus aureus (MRSA), a prevalent cause of both community-onset and healthcare-associated bloodstream infections (BSIs) [4]. Staphylococcus aureus bacteremia (SAB) is a serious infectious disease syndrome with a notoriously high attributable mortality [5]. Recent systematic analyses estimate the all-cause mortality of SAB at approximately 18% at one month and 27% at three months, with MRSA infection often, though not universally, associated with worse outcomes compared to methicillin-susceptible S. aureus (MSSA) [6]. The latest evidence suggests that while MRSA may increase the risk of intensive care admission, its independent association with in-hospital mortality, when adjusted for confounders like appropriate empiric therapy, may be less pronounced than historically presumed [7]. Nonetheless, SAB, particularly when caused by MRSA, frequently leads to metastatic complications such as infective endocarditis, septic arthritis, osteomyelitis, and abscess formation in more than one-third of cases, profoundly complicating clinical management and prolonging therapy [8]. The pathogenesis is driven by a complex arsenal of virulence factors and immune evasion strategies, making S. aureus a uniquely adaptable and persistent pathogen [9].

The clinical trajectory of sepsis with concurrent MRSA BSI is especially perilous [10]. These patients often present with profound inflammatory response, significant organ dysfunction, and are at high risk for septic shock [11]. Factors such as older age, elevated lactate levels (indicative of tissue hypoperfusion and metabolic stress), and higher sequential organ failure assessment (SOFA) scores have been consistently identified as strong predictors of adverse outcomes in general sepsis and specific BSI populations [12–14]. The presence of MRSA adds layers of complexity regarding initial empiric antibiotic selection, the urgency of achieving source control (e.g., removal of infected devices), and the potential need for longer or combination antimicrobial regimens, all of which can influence survival [15, 16].

Despite decades of research and established guidelines, the management of SAB, especially MRSA-BSI, remains fraught with practice variation and clinical dilemmas. Key controversies persist around the optimal choice and duration of antibiotic therapy, the role of combination regimens for persistent bacteremia, and the precise definition of “complicated” infection [15, 17]. This heterogeneity in clinical presentation and management underscores the critical need for sophisticated, individualized risk stratification tools. Accurate early prognosis is not merely an academic exercise; it directly informs clinical decisions, such as the aggressiveness of diagnostic workup for metastatic foci (e.g., echocardiography), the intensity of monitoring, the timing of antibiotic de-escalation or switch to oral therapy, and discussions with patients and families regarding goals of care.

In response to this need, prognostic modeling in sepsis and bacteremia has evolved significantly. Traditional severity-of-illness scores like SOFA and APACHE II provide a generalized assessment of physiological derangement but are not specific to SAB and may not fully incorporate the unique determinants of outcome in this population [18]. Consequently, research has shifted towards developing syndrome- specific models. For instance, studies have identified risk factors for recurrence in SAB, such as indwelling vascular grafts, nasal MRSA carriage, and specific resistance patterns [19]. In pediatric S. aureus BSI, models have been built to predict sepsis development, highlighting factors like hospital-acquired infection, central venous catheters, underlying comorbidities (e.g., gastrointestinal malformations), and elevated procalcitonin [20]. Other research has leveraged large electronic health record databases to predict AMR patterns or hospital-onset bacteremia using machine learning and even large language models, demonstrating the potential of data-driven approaches [21, 22]. However, a conspicuous and critical gap remains: there is a paucity of robust, validated prediction models specifically designed to estimate the risk of in-hospital mortality for adult patients diagnosed with sepsis (by Sepsis-3 criteria) suffering from concurrent, monomicrobial MRSA bloodstream infection. Existing models either focus on broader sepsis populations without pathogen specification, on SAB outcomes regardless of sepsis severity, or on pediatric cohorts, limiting their direct applicability to this high-stakes clinical niche.

To bridge this gap, the use of large, high-fidelity ICU databases like the Medical Information Mart for Intensive Care (MIMIC) offers a powerful solution. These repositories provide granular, longitudinal clinical data ideal for developing data-intensive predictive models [23]. Coupled with advanced analytic techniques, including rigorous feature selection and machine learning algorithms, they enable the identification of a parsimonious yet highly predictive set of variables from a vast array of potential predictors. This approach balances model performance with clinical interpretability and practicality—a crucial consideration for bedside implementation.

Therefore, this study aimed to develop and validate a pragmatic, interpretable, and accurate prediction model for in-hospital mortality in patients with sepsis and concurrent MRSA BSI. Utilizing data from the MIMIC-IV database, we implemented a rigorous, multi-stage feature selection pipeline. This process integrated univariate filtering, multi-model machine learning importance aggregation, the Boruta all-relevant feature selection algorithm, and stability selection via LASSO regression to distill a core set of robust predictors from an extensive list of candidate variables spanning demographics, comorbidities, vital signs, laboratory values, and severity scores. We then systematically compared the performance of multiple machine learning algorithms, ultimately selecting logistic regression for its optimal blend of discrimination, calibration, stability against overfitting, and inherent clinical interpretability. We hypothesize that this deliberately constructed model, derived from readily available clinical data within the first 24 h of diagnosis, will provide superior individualized risk estimation for this specific patient subgroup compared to generic severity scores (e.g., SOFA, APACHE II), and will serve as a practical decision-support tool.

Methods

Study design, data source, and ethical approval

This was a retrospective cohort study utilizing data from the Medical Information Mart for Intensive Care IV (MIMIC-IV, version 3.2) database. MIMIC-IV is a publicly available, de-identified database containing comprehensive clinical data for patients admitted to the Beth Israel Deaconess Medical Center (Boston, MA, USA) between 2008 and 2019. The study protocol was reviewed and approved by the Ethics Committee of Zhongshan City People’s Hospital. Detailed statements regarding ethical approval, waiver of informed consent, and compliance with the Declaration of Helsinki are provided in the ‘Declarations’ section.

Study population and patient selection

Adult patients (≥ 18 years) diagnosed with sepsis-3 and concurrent monomicrobial Methicillin-resistant Staphylococcus aureus (MRSA) bloodstream infection were eligible for inclusion. Sepsis was defined retrospectively according to the Sepsis-3 criteria, operationalized as follows: (1) “Suspected infection” was identified by the concurrent administration of systemic antibiotics and procurement of a body fluid culture within a 24-hour window anchored to “time zero”; (2) Organ dysfunction was defined as an acute increase in the Sequential Organ Failure Assessment (SOFA) score by ≥ 2 points from baseline. The baseline SOFA was defined as the first score recorded at hospital admission, or the lowest value in the 24 h preceding antibiotic initiation if admission data were unavailable; The time of antibiotic initiation was defined as the first documented administration of a new systemic antibiotic course targeting the suspected infection. (3) The “suspected infection” window and the time point meeting the ΔSOFA ≥ 2 criterion were required to fall within a shared 48-hour period to ensure a valid temporal relationship. MRSA bacteremia was confirmed by at least one positive blood culture for MRSA. To ensure objectivity and alignment with our primary outcome (MRSA bacteremia), we define “time zero” as the timestamp of the first positive blood culture for MRSA. We include only the first eligible hospital encounter for each patient to ensure the independence of observations. The exclusion criteria were: (1) length of hospital stay < 24 h; (2) polymicrobial bloodstream infection; (3) missing data for the primary outcome (in-hospital mortality) or key severity scores (e.g., SOFA, APACHE II); (4) discharge against medical advice. The STROBE-compliant flow diagram was shown in Supplementary Figure S1. The comparing key baseline characteristics and in-hospital mortality between the included cohort and patients excluded at the major filtering step were shown in Supplementary Table 1.

Data extraction, variable definition, and preprocessing

Data were extracted using structured query language (SQL). A total of 64 candidate predictor variables were initially collected, encompassing four domains: Demographics and Comorbidities: Age, sex, and 23 comorbidities defined using ICD-9/10 codes (e.g., congestive heart failure, chronic kidney disease, cancer). Clinical Severity and Vital Signs: Sequential Organ Failure Assessment (SOFA) score, Acute Physiology and Chronic Health Evaluation II (APACHE II) score, Glasgow Coma Scale (GCS) score, and vital signs (heart rate, blood pressure, respiratory rate, temperature, SpO₂). Laboratory Parameters: The minimum and maximum values recorded within the 24-hour window immediately preceding and including the time of the first positive MRSA blood culture (“time zero”) for key parameters: sodium, potassium, blood urea nitrogen (BUN), white blood cell count (WBC), platelet count, pH, lactate, and hemoglobin. Treatment and Microbiological Factors: Use of specific antibiotics (e.g., vancomycin, anti-pseudomonal agents) and co-isolation of other pathogens. Continuous variables were presented as mean (± SD) or median (IQR), and categorical variables as frequencies (percentages). Variables with > 30% missing data were excluded. For remaining variables with minor missingness (< 5% for most), multiple imputation by chained equations (MICE) was performed to create five imputed datasets for sensitivity analysis. The primary model development was based on complete cases for the final predictors.The seven predictors selected for the final model had a near-complete data availability (> 99% in both training and validation sets), ensuring the model was developed and validated on a virtually complete-case basis for these key variables.

Dataset construction and balance assessment

The entire cohort was randomly split into a development set (70%, n = 1,124) for feature selection and model training, and an independent validation set (30%, n = 481) for performance assessment. To quantitatively evaluate the balance between sets, we calculated the Standardized Mean Difference (SMD) for all baseline variables. An SMD < 0.1 was considered indicative of negligible imbalance.

Multi-stage feature selection framework

All steps of the feature selection pipeline described below were conducted strictly within the training set (n = 1,124). The independent validation set (n = 481) was not accessed during this process. To identify a robust and parsimonious set of predictors, we implemented a rigorous, four-stage feature selection pipeline (summarized in Fig. 1). Stage 1: all candidate variables were subjected to univariate logistic regression with in-hospital mortality as the outcome. Variables with a p-value < 0.2 were retained for subsequent analysis. Stage 2: the remaining variables were processed through three independent, data-driven selection methods run in parallel. Six distinct algorithms (Logistic Regression, Support Vector Machine with linear kernel, Random Forest, XGBoost, K-Nearest Neighbors, Multilayer Perceptron) were trained. The normalized importance scores for each variable were averaged across all models. This wrapper algorithm, based on Random Forest with shadow features, was used to identify all features statistically relevant to the outcome, distinguishing them from random noise. Least Absolute Shrinkage and Selection Operator (LASSO) regression was combined with bootstrap resampling (100 iterations). Features consistently selected (non-zero coefficient) across a high percentage of bootstrap samples were deemed stable. Stage 3: a variable was promoted to the candidate longlist if it was selected by at least two of the three methods in Stage 2. The stability of this longlist was further verified via a dedicated bootstrap analysis (100 repetitions) to estimate each variable’s selection frequency. Stage 4: the candidate longlist was reviewed by a panel of clinicians and statisticians. Final inclusion was determined by a panel of clinicians and statisticians who integrated quantitative evidence (selection frequency, effect size) with pre-specified qualitative criteria: (1) clinical actionability (e.g., is the variable modifiable or a target for monitoring?), (2) bedside accessibility within the first 24 h of ICU care, and (3) strong pathophysiological plausibility in the context of sepsis and MRSA infection. Variables with weaker or inconsistent support across these criteria were considered for exclusion.

Fig. 1.

Schematic overview of the multi-stage feature selection pipeline integrating machine learning and statistical methods

Model development, algorithm comparison, and final model building

Using the final predictor set, we trained and tuned the same six machine learning algorithms listed in Stage 2. Hyperparameter tuning was performed via 5-fold cross-validation on the development set, optimizing for the area under the receiver operating characteristic curve (AUC). Model performance was then evaluated on the held-out validation set. The optimal algorithm was selected based on a comprehensive comparison of: (1) Discriminatory Power (AUC with 95% CI); (2) Stability (Absolute difference between training and validation AUC); (3) Statistical Comparison (Pairwise DeLong’s test); and (4) Clinical Applicability (Interpretability, computational simplicity). The selected algorithm was then used to build the final prediction model.

Model evaluation, interpretability, and validation analyses

The performance of the final model was thoroughly assessed: 1) Discrimination and Calibration: Evaluated using the AUC, the receiver operating characteristic (ROC) curve, and calibration plots with the Hosmer-Lemeshow test. The optimal probability cut-off was determined by maximizing Youden’s J index. 2)Clinical Utility: Decision Curve Analysis (DCA) was performed to quantify the net clinical benefit across a range of threshold probabilities. Patients were stratified into risk categories based on model-predicted probabilities. 3) Model Interpretability: For the final logistic regression model, coefficients were exponentiated to generate odds ratios (ORs) with 95% confidence intervals (CIs). SHapley Additive exPlanations (SHAP) analysis was conducted to quantify individual feature contributions and explore pairwise interaction effects. 4)Robustness and Validation: Internal Validation: A bootstrap resampling procedure with 1,000 replicates was conducted for optimism correction, yielding a bias-corrected AUC. Sensitivity Analyses: These included (a) re-fitting the model on multiply imputed datasets, (b) sequential removal of the least important features, (c) evaluating performance using different probability thresholds, and (d) assessing performance across key clinical subgroups (e.g., by age, sex, SOFA score) to evaluate fairness.

Statistical software

All data analyses, including data preprocessing, feature selection, model development, and statistical validation, were performed using R software (version 4.3.0, R Foundation for Statistical Computing). A two-sided p-value < 0.05 was considered statistically significant.

Results

Cohort characteristics and dataset balance assessment

Baseline characteristics of the overall cohort and by outcome are detailed in Table 1 and Supplementary Table 2, respectively. Key findings are summarized below. A total of 1,605 eligible patients with sepsis and concurrent MRSA bloodstream infection were included in this study. All patients were randomly allocated into a training set (n = 1,124) and an independent validation set (n = 481). Baseline characteristics, including demographics, vital signs, severity scores (e.g., APACHE II, SOFA, GCS), comorbidities, and the vast majority of laboratory parameters, were well-balanced between the training and validation sets, with no significant differences (all p > 0.05, Table 1). Significant differences were only observed in a few variables, including heart rate, congestive heart failure, coronary artery disease, daptomycin use, and Escherichia coli infection (all p < 0.05).

Table 1.

Baseline characteristics: training vs. Validation datasets

Variable	Training (n = 1124)	Validation (n = 481)	P-value
spo2			0.510
Mean (SD)	97.04 (1.83)	96.97 (1.84)
Median (Q1, Q3)	97.30 (95.80, 98.50)	97.20 (95.70, 98.40)
Weight			0.402
Mean (SD)	80.45 (20.23)	81.39 (20.60)
Median (Q1, Q3)	79.05 (66.45, 93.00)	80.70 (67.70, 93.20)
APACHEII_score			0.583
Mean (SD)	15.61 (2.50)	15.68 (2.47)
Median (Q1, Q3)	15.80 (13.60, 17.70)	15.90 (13.60, 17.70)
APS3_score			0.816
Mean (SD)	2.01 (0.95)	2.02 (0.90)
Median (Q1, Q3)	2.00 (1.00, 3.00)	2.00 (1.00, 3.00)
gcs_score			0.977
Mean (SD)	4.13 (0.36)	4.13 (0.34)
Median (Q1, Q3)	4.20 (3.90, 4.50)	4.20 (3.90, 4.50)
heart_rate			0.018
Mean (SD)	69.37 (12.32)	67.85 (11.59)
Median (Q1, Q3)	68.45 (60.05, 77.65)	66.60 (60.00, 75.50)
SBP			0.683
Mean (SD)	114.40 (14.39)	114.72 (14.33)
Median (Q1, Q3)	112.35 (103.90, 123.50)	112.80 (104.10, 124.10)
DBP			0.550
Mean (SD)	61.62 (9.55)	61.31 (9.49)
Median (Q1, Q3)	60.80 (55.20, 67.75)	60.70 (54.70, 67.60)
respiratory_rate			0.983
Mean (SD)	19.47 (3.82)	19.48 (3.67)
Median (Q1, Q3)	19.10 (16.60, 22.00)	19.00 (16.80, 21.60)
Temperature			0.093
Mean (SD)	36.89 (0.46)	36.85 (0.50)
Median (Q1, Q3)	36.80 (36.60, 37.10)	36.80 (36.50, 37.10)
Gender			0.231
0	654.0 (58.2%)	296.0 (61.5%)
1	470.0 (41.8%)	185.0 (38.5%)
Age			0.911
Mean (SD)	62.64 (16.48)	62.74 (16.30)
Median (Q1, Q3)	64.00 (52.00, 75.00)	63.00 (52.00, 75.00)
sofa_score			0.813
Mean (SD)	4.99 (1.11)	4.98 (1.08)
Median (Q1, Q3)	5.00 (4.00, 6.00)	5.00 (4.00, 6.00)
myocardial_infarct			0.069
0	1,000.0 (89.0%)	443.0 (92.1%)
1	124.0 (11.0%)	38.0 (7.9%)
congestive_heart_failure			0.032
0	692.0 (61.6%)	324.0 (67.4%)
1	432.0 (38.4%)	157.0 (32.6%)
peripheral_vascular_disease			0.338
0	1,056.0 (94.0%)	445.0 (92.5%)
1	68.0 (6.0%)	36.0 (7.5%)
cerebrovascular_disease			0.239
0	971.0 (86.4%)	404.0 (84.0%)
1	153.0 (13.6%)	77.0 (16.0%)
chronic_pulmonary_disease			0.249
0	789.0 (70.2%)	323.0 (67.2%)
1	335.0 (29.8%)	158.0 (32.8%)
rheumatic_disease			0.974
0	1,077.0 (95.8%)	460.0 (95.6%)
1	47.0 (4.2%)	21.0 (4.4%)
peptic_ulcer_disease			0.708
0	1,088.0 (96.8%)	468.0 (97.3%)
1	36.0 (3.2%)	13.0 (2.7%)
mild_liver_disease			0.538
0	979.0 (87.1%)	425.0 (88.4%)
1	145.0 (12.9%)	56.0 (11.6%)
severe_liver_disease			0.549
0	1,064.0 (94.7%)	451.0 (93.8%)
1	60.0 (5.3%)	30.0 (6.2%)
diabetes			0.737
0	680.0 (60.5%)	296.0 (61.5%)
1	444.0 (39.5%)	185.0 (38.5%)
paraplegia			0.732
0	1,081.0 (96.2%)	465.0 (96.7%)
1	43.0 (3.8%)	16.0 (3.3%)
renal_disease			0.744
0	624.0 (55.5%)	272.0 (56.5%)
1	500.0 (44.5%)	209.0 (43.5%)
hypertension			0.460
0	390.0 (34.7%)	157.0 (32.6%)
1	734.0 (65.3%)	324.0 (67.4%)
cancer			0.622
0	1,011.0 (89.9%)	428.0 (89.0%)
1	113.0 (10.1%)	53.0 (11.0%)
atrial_fibrillation			0.685
0	957.0 (85.1%)	414.0 (86.1%)
1	167.0 (14.9%)	67.0 (13.9%)
chronic_kidney_disease			0.818
0	777.0 (69.1%)	336.0 (69.9%)
1	347.0 (30.9%)	145.0 (30.1%)
previous_venous_thromboembolism			0.537
0	1,040.0 (92.5%)	440.0 (91.5%)
1	84.0 (7.5%)	41.0 (8.5%)
pulmonary_embolism			0.917
0	1,094.0 (97.3%)	467.0 (97.1%)
1	30.0 (2.7%)	14.0 (2.9%)
pneumonia			0.154
0	783.0 (69.7%)	317.0 (65.9%)
1	341.0 (30.3%)	164.0 (34.1%)
stroke			0.984
0	1,073.0 (95.5%)	460.0 (95.6%)
1	51.0 (4.5%)	21.0 (4.4%)
coronary_heart_disease			0.016
0	716.0 (63.7%)	337.0 (70.1%)
1	408.0 (36.3%)	144.0 (29.9%)
mechanical_ventilation			> 0.999
0	2.0 (0.2%)	1.0 (0.2%)
1	1,122.0 (99.8%)	480.0 (99.8%)
sodium_min			0.230
Mean (SD)	133.69 (4.31)	133.40 (4.43)
Median (Q1, Q3)	134.00 (131.00, 137.00)	134.00 (131.00, 136.00)
sodium_max			0.890
Mean (SD)	143.24 (4.68)	143.21 (4.64)
Median (Q1, Q3)	143.00 (140.00, 146.00)	143.00 (140.00, 146.00)
potassium_min			0.770
Mean (SD)	3.43 (0.43)	3.42 (0.45)
Median (Q1, Q3)	3.40 (3.10, 3.70)	3.40 (3.10, 3.70)
potassium_max			0.811
Mean (SD)	4.94 (0.66)	4.93 (0.67)
Median (Q1, Q3)	4.90 (4.50, 5.40)	4.80 (4.50, 5.30)
bun_max			0.088
Mean (SD)	26.59 (16.15)	28.08 (15.93)
Median (Q1, Q3)	21.90 (14.10, 36.20)	23.90 (15.10, 39.10)
wbc_min			0.256
Mean (SD)	6.70 (2.80)	6.53 (2.76)
Median (Q1, Q3)	6.50 (4.70, 8.50)	6.20 (4.60, 8.20)
wbc_max			0.367
Mean (SD)	16.72 (7.11)	17.09 (7.53)
Median (Q1, Q3)	15.70 (11.60, 21.25)	15.90 (11.50, 21.70)
rbc			0.619
Mean (SD)	3.25 (0.50)	3.27 (0.52)
Median (Q1, Q3)	3.18 (2.87, 3.57)	3.20 (2.89, 3.60)
hemoglobin_min			0.926
Mean (SD)	30.92 (1.03)	30.93 (1.10)
Median (Q1, Q3)	30.50 (30.10, 31.30)	30.50 (30.10, 31.30)
platelet_min			0.421
Mean (SD)	161.30 (83.76)	165.07 (87.21)
Median (Q1, Q3)	154.00 (100.00, 217.00)	154.00 (99.00, 226.00)
ph_min			0.442
Mean (SD)	7.31 (0.10)	7.31 (0.09)
Median (Q1, Q3)	7.32 (7.25, 7.38)	7.32 (7.25, 7.37)
lactate_max			0.915
Mean (SD)	2.24 (1.12)	2.25 (1.20)
Median (Q1, Q3)	2.00 (1.40, 2.90)	2.00 (1.30, 2.90)
rrt			0.886
0	1,117.0 (99.4%)	479.0 (99.6%)
1	7.0 (0.6%)	2.0 (0.4%)
vancomycin			0.646
0	175.0 (15.6%)	80.0 (16.6%)
1	949.0 (84.4%)	401.0 (83.4%)
linezolid			0.898
0	996.0 (88.6%)	428.0 (89.0%)
1	128.0 (11.4%)	53.0 (11.0%)
daptomycin			0.005
0	984.0 (87.5%)	445.0 (92.5%)
1	140.0 (12.5%)	36.0 (7.5%)
ceftaroline			> 0.999
0	1,083.0 (96.4%)	464.0 (96.5%)
1	41.0 (3.6%)	17.0 (3.5%)
aztreonam			> 0.999
0	1,099.0 (97.8%)	471.0 (97.9%)
1	25.0 (2.2%)	10.0 (2.1%)
meropenem			0.367
0	941.0 (83.7%)	412.0 (85.7%)
1	183.0 (16.3%)	69.0 (14.3%)
imipenem			> 0.999
0	1,112.0 (98.9%)	476.0 (99.0%)
1	12.0 (1.1%)	5.0 (1.0%)
piperacillin			0.198
0	802.0 (71.4%)	359.0 (74.6%)
1	322.0 (28.6%)	122.0 (25.4%)
other_staphylococcus			> 0.999
0	1,078.0 (95.9%)	462.0 (96.0%)
1	46.0 (4.1%)	19.0 (4.0%)
streptococcus			0.559
0	1,068.0 (95.0%)	461.0 (95.8%)
1	56.0 (5.0%)	20.0 (4.2%)
enterococcus			0.659
0	1,065.0 (94.8%)	459.0 (95.4%)
1	59.0 (5.2%)	22.0 (4.6%)
ecoli			0.021
0	1,068.0 (95.0%)	442.0 (91.9%)
1	56.0 (5.0%)	39.0 (8.1%)
klebsiella			0.292
0	1,091.0 (97.1%)	472.0 (98.1%)
1	33.0 (2.9%)	9.0 (1.9%)
pseudomonas			0.379
0	1,051.0 (93.5%)	456.0 (94.8%)
1	73.0 (6.5%)	25.0 (5.2%)
clostridium			0.537
0	1,073.0 (95.5%)	455.0 (94.6%)
1	51.0 (4.5%)	26.0 (5.4%)
candida			0.704
0	1,108.0 (98.6%)	476.0 (99.0%)
1	16.0 (1.4%)	5.0 (1.0%)
outcome			0.792
0	953.0 (84.8%)	411.0 (85.4%)
1	171.0 (15.2%)	70.0 (14.6%)

To further quantify the balance between datasets, we calculated the Standardized Mean Difference (SMD) for all variables. As shown in Supplementary Figure S2, the vast majority of variables (52/58, 89.7%) had an SMD of less than 0.1, indicating negligible differences. Only six variables had an SMD between 0.1 and 0.2 (small difference), and no variable exceeded an SMD of 0.2, confirming that the two datasets were comparable for model development and validation.

Within the entire cohort, univariate analysis between the non-survivors group (n = 241) and the survivors group (n = 1,364) revealed that non-survivors had significantly higher disease severity scores (APACHE II, SOFA), lower GCS scores, a higher prevalence of comorbidities (e.g., congestive heart failure, liver disease, renal disease, cancer, atrial fibrillation, pneumonia, and stroke), and more deranged laboratory parameters (e.g., potassium, blood urea nitrogen, white blood cell count, lactate) (all p < 0.05, Supplementary Table 2). These findings consistently indicated that non-survivors presented with more severe illness and greater inflammatory and organ dysfunction, justifying the inclusion of these variables in subsequent predictive modeling.

Multi-stage feature selection and determination of the final predictor set

The identification of core predictors followed a rigorous, multi-stage pipeline summarized in Fig. 1. Beginning with 64candidate variables, univariate logistic regression (P < 0.2) retained 42 for subsequent analysis (Supplementary Figure S3). These variables were then subjected to three parallel, data-driven selection methods: a multi-model machine learning importance assessment (Supplementary Figure S4), the Boruta all-relevant feature filter (Supplementary Figure S5), and stability selection via LASSO regression with bootstrap resampling (Supplementary Figure S6). The union of results from these methods yielded a candidate longlist of 10 variables. The selection robustness of these 10 variables was further confirmed through a dedicated bootstrap stability analysis (Supplementary Figure S7). This longlist was then evaluated by integrating data-driven evidence strength (e.g., selection frequency, importance scores) with clinical accessibility, pathophysiological relevance, and literature support. Following this comprehensive evaluation and clinical expert deliberation, three variables were excluded due to weaker or less consistent evidence across the criteria, resulting in the final set of seven robust and clinically applicable predictors: gcs_score, ph_min, bun_max, wbc_min, platelet_min, lactate_max, and pneumonia.

Machine learning model comparison and rationale for logistic regression selection

We systematically compared the performance of six machine learning models to identify the most suitable algorithm for our clinical prediction task. As shown in Fig. 2A, all models demonstrated good discriminatory ability in the independent validation set. The Support Vector Machine (SVM) with a linear kernel achieved the highest AUC (0.847, 95% CI: 0.815–0.879), closely followed by the Logistic Regression model (AUC = 0.844, 95% CI: 0.811–0.877). Model stability analysis (Fig. 2B), assessed by the difference between training and testing AUC, revealed that SVM Linear (AUC drop = 0.013), Logistic Regression (AUC drop = 0.017), and Multilayer Perceptron (MLP, AUC drop = 0.024) showed minimal overfitting. In contrast, Random Forest (drop = 0.171), XGBoost (drop = 0.073), and K-Nearest Neighbors (KNN, drop = 0.239) exhibited substantial performance degradation, indicative of overfitting. Pairwise DeLong tests (Fig. 2C) indicated no statistically significant difference in AUC between Logistic Regression and SVM Linear (p = 0.415), MLP (p = 0.714), Random Forest (p = 0.287), or XGBoost (p = 0.052), except for KNN (p < 0.001). A comprehensive multi-dimensional radar chart (Fig. 2D) further visualized that Logistic Regression offered an ideal balance between predictive performance, model stability, interpretability, and computational simplicity.

Fig. 2.

Comparative performance of six machine learning algorithms for model selection. (A) Bar plot of area under the receiver operating characteristic curve (AUC) in the independent validation set, ranked by discriminative ability. (B) Model stability assessed by the absolute difference in AUC between training and validation sets. (C) Pairwise statistical comparison of AUCs between logistic regression and other algorithms using DeLong’s test (dashed line indicates p = 0.05). (D) Radar chart integrating three normalized dimensions: training AUC, validation AUC, and stability, providing a comprehensive view of each algorithm’s trade-offs

Therefore, considering its optimal trade-off between predictive performance, model stability, and, most importantly, inherent clinical interpretability—which facilitates trust, understanding, and the direct translation of odds ratios into clinical insight—Logistic Regression was selected as the final modeling algorithm.

Development, performance, and validation of the logistic regression prediction model

Using the seven selected predictors, we developed and validated a logistic regression model for in-hospital mortality risk. The model demonstrated excellent performance (Fig. 3). The ROC curve (Fig. 3A) showed robust discriminatory power, with an AUC of 0.861 (95% CI: 0.840–0.882) in the training set and 0.844 (95% CI: 0.811–0.877) in the independent validation set. Calibration analysis (Fig. 3B) revealed a strong agreement between predicted probabilities and observed event rates (Hosmer-Lemeshow test p = 0.274). A radar chart comparing key performance metrics (Accuracy, F1-Score, Sensitivity, Specificity, PPV, NPV) between the training and validation sets (Fig. 3C) confirmed the model’s excellent generalization capability without significant performance degradation. The predicted probability density plot (Fig. 3D) clearly distinguished between survivors and non-survivors. The optimal decision threshold, determined by maximizing Youden’s index, was 0.135, yielding a sensitivity of 0.927 and a specificity of 0.386, making it particularly suitable for clinical screening purposes where high sensitivity is paramount. At this threshold (0.135), the model also achieved a positive predictive value (PPV) of 96.2% and a negative predictive value (NPV) of 35.6% in the validation set. This performance profile suggests the model is excellent for ‘ruling out’ low-risk patients (high sensitivity), albeit at the cost of a higher false-positive rate (low specificity), which should be considered in the context of subsequent diagnostic or therapeutic resource allocation. To facilitate clinical translation and provide flexibility in decision-making, we conducted a comprehensive threshold analysis. The performance metrics (sensitivity, specificity, PPV, and NPV) across multiple clinically relevant risk thresholds (e.g., 0.05, 0.10, 0.135, 0.20, 0.30, 0.40, 0.50) are presented in Supplementary Table 3. This analysis allows clinicians to understand the trade-offs between sensitivity and specificity and select a threshold appropriate for their specific clinical context (e.g., prioritizing high sensitivity for screening or higher specificity when resources are constrained).

Fig. 3.

Performance evaluation of the logistic regression model in training and validation sets. (A) Receiver operating characteristic (ROC) curves for training and validation sets. (B) Calibration plot for the validation set, showing agreement between predicted probabilities and observed event rates. (C) Radar chart summarizing key performance metrics (accuracy, F1-score, sensitivity, specificity, positive predictive value, negative predictive value) in training and validation sets. (D) Density distribution of predicted probabilities for survivors and non-survivors, illustrating model separation

Interpretability and clinical utility of the prediction model

The logistic regression model provided direct and clinically meaningful insights into risk factors (Fig. 4). The coefficient forest plot (Fig. 4A) revealed both protective and risk factors: Glasgow Coma Scale score (gcs_score) and minimum platelet count (platelet_min) were significant protective factors (negative coefficients), whereas maximum lactate level (lactate_max) and maximum blood urea nitrogen (bun_max) were significant risk factors (positive coefficients). The odds ratio (OR) forest plot (Fig. 4B) quantified these effects. For instance, each unit increase in lactate_max was associated with an 84% increase in the odds of mortality (OR = 1.84, 95% CI: 1.32–2.58). A comprehensive feature importance bar chart, ranked by absolute coefficient value and annotated with ORs (Fig. 4C), visually distinguished risk factors (red) from protective factors (blue), with ph_min and gcs_score identified as the strongest predictors.

Fig. 4.

Interpretability and clinical utility of the logistic regression prediction model. (A) Coefficient forest plot showing raw logistic regression coefficients with 95% confidence intervals. The vertical line at zero indicates no effect; positive values denote risk factors, negative values denote protective factors. (B) Odds ratio (OR) forest plot with 95% confidence intervals, derived from exponentiated coefficients. The vertical line at 1 indicates no effect. (C) Feature importance bar chart ranked by absolute coefficient magnitude, annotated with ORs. Red bars indicate risk factors, blue bars indicate protective factors. (D) Performance metrics (sensitivity, specificity, PPV, NPV) across different classification thresholds. The red dashed line indicates the optimal threshold determined by maximizing Youden’s index

To evaluate the model’s practical value in clinical decision-making, we performed Decision Curve Analysis (DCA, Supplementary Figure S8A). Across a wide range of threshold probabilities (0.1–0.7), using the model to guide interventions provided a higher net benefit than the strategies of “treating all” or “treating none.” The clinical impact curve (Supplementary Figure S8B) further demonstrated that the number of patients identified as high-risk by the model closely aligned with the actual number of true-positive cases across thresholds. Finally, risk stratification (Supplementary Figure S8D) using model-predicted probabilities effectively categorized patients into four distinct risk tiers (Low, Moderate, High, Very High), with observed mortality rates increasing significantly from 6.5% to 88.9% (p for trend < 0.001), underscoring the model’s powerful risk discrimination capability.

To translate the model’s risk stratification into concrete bedside actions, a predicted probability exceeding a pre-defined threshold (e.g., 0.135) could trigger a structured clinical pathway. This may include: (1) escalation of monitoring (e.g., to hourly vital signs and serial lactate measurements); (2) automatic consultation request to a sepsis response team or infectious disease service; (3) urgent re-evaluation for and removal of potential infection sources (e.g., intravascular catheters); and (4) prioritization of intensive care unit evaluation if not already in an ICU. The specific threshold and actions can be adapted based on institutional resources and the desired balance between sensitivity and specificity.

Robustness and stability validation of the final model

We conducted extensive internal and sensitivity analyses to rigorously assess the model’s robustness and stability.

Internal Bootstrap Validation: A 1000-replicate bootstrap resampling analysis was performed (Fig. 5). The distribution of the apparent AUC was approximately normal, with a mean of 0.865 (Fig. 5A). The mean optimism (difference between apparent and test performance) was small at 0.007 (Fig. 5B). After optimism correction, the adjusted AUC was 0.858 (Fig. 5C), indicating minimal over-optimism. The calibration slope after bootstrap correction was close to the ideal value of 1 (Fig. 5D). A strong correlation was observed between bootstrap AUC estimates and the original model AUC (r = 0.96, p < 0.001, Fig. 5E), and boxplots of key performance metrics confirmed their stability across resamples (Fig. 5F).

Fig. 5.

Internal validation and robustness assessment via bootstrap resampling (1,000 replicates). (A) Histogram of bootstrap AUC distribution, with red dashed line indicating the mean. (B) Distribution of optimism (difference between apparent and test performance). (C) Comparison of model performance metrics across bootstrap samples. (D) Calibration curve after bootstrap correction, showing predicted versus observed probabilities. (E) Scatter plot of bootstrap AUC estimates against original model AUC, with Pearson correlation coefficient. (F) Boxplots of key performance metrics across bootstrap replicates

Multi-scenario Sensitivity Analysis: We evaluated model performance under various challenging scenarios (Fig. 6). The baseline full model achieved an AUC of 0.844. Performance remained robust when sequentially removing the least important features (defined by the smallest absolute coefficient), with AUCs of 0.826 and 0.789 after removing one and two features, respectively (Fig. 6A, “Feature Removal”). When different classification thresholds were applied to the baseline model’s predicted probabilities, the balanced accuracy varied, reaching 0.701, 0.656, and 0.642 at thresholds of 0.3, 0.4, and 0.5, respectively (Fig. 6A, “Prediction Threshold”). Using only subsets of the most important features (e.g., top 3–5 features ranked by absolute coefficient) yielded AUCs between 0.757 and 0.789 (Fig. 6A, “Feature Selection”). The standardized coefficients and their statistical significance for all seven features are displayed in Fig. 6B, reaffirming ph_min (β = -5.78, p < 0.001) and gcs_score (β = -2.30, p < 0.001) as the dominant predictors. The model demonstrated good calibration in the validation set (Brier Score = 0.097, Fig. 6C). A summary of the final model’s comprehensive performance metrics is presented in Fig. 6D.

Fig. 6.

Sensitivity analysis and comprehensive performance of the logistic regression model. (A) Heatmap of AUC under different sensitivity scenarios: prediction threshold adjustment, feature selection, and feature removal. (B) Feature importance plot showing standardized regression coefficients and statistical significance (***p < 0.001, **p < 0.01, *p < 0.05). (C) Calibration curve in the validation set with Brier score. (D) Summary of model performance metrics in the validation set: AUC, accuracy, sensitivity, specificity, PPV, NPV, and F1-score

Multiple Imputation Sensitivity Analysis: To assess the impact of missing data handling, the model was re-developed and validated on four independently generated multiply imputed datasets. All key performance metrics showed minimal variation across these datasets: validation set AUC ranged from 0.842 to 0.857 (mean 0.851 ± 0.006), accuracy from 0.852 to 0.865 (0.858 ± 0.005), and F1-Score from 0.916 to 0.923 (0.919 ± 0.003) (Supplementary Figure S9A). The coefficient of variation for all metrics was below 0.01 (Supplementary Figure S9B), confirming the model’s robustness to the specific method of imputation.

Advanced interpretability via SHAP interaction analysis

To uncover non-linear and interaction effects among predictors, we performed a SHapley Additive exPlanations (SHAP) interaction analysis on the final model. The SHAP interaction network (Fig. 7A) identified 19 significant pairwise interaction combinations among the seven features. Node size represents feature importance (consistent with coefficient rankings), with blue and red nodes denoting protective and risk factors, respectively. Edge thickness quantifies interaction strength. The strongest interaction was observed between gcs_score and ph_min (interaction strength = 0.2763). A bar chart ranking all significant interaction pairs by strength is provided in Fig. 7B, offering novel insights into the synergistic pathophysiology underlying mortality risk in this patient population.

Fig. 7.

SHAP-based interaction analysis of predictive features. (A) SHAP interaction network visualizing pairwise interactions among the seven predictors. Node size reflects feature importance (based on absolute logistic regression coefficients). Blue nodes denote protective factors, red nodes denote risk factors. Edge thickness indicates interaction strength. (B) Bar chart ranking the top 10 strongest feature interaction pairs by interaction strength (range 0–1), derived from SHAP interaction values

Model fairness and subgroup analysis

To ensure equitable predictive performance across diverse patient subgroups, we conducted a comprehensive fairness analysis (Fig. 8). The model maintained good and stable discriminatory ability (AUC > 0.82) across all eight pre-defined subgroups based on age, gender, SOFA score, and comorbidity burden, with an overall AUC of 0.844 (Fig. 8A). Calibration curves were excellent and nearly identical for male and female patients (DeLong test for AUC difference, p = 0.275) and across different age groups (p = 0.587) (Fig. 8B and C). Furthermore, the distribution of predicted mortality probabilities logically increased with higher disease severity (SOFA score) and greater comorbidity burden (Fig. 8D-F). These results confirm that the predictive model performs robustly and fairly across key demographic and clinical subgroups.

Fig. 8.

Fairness and subgroup analysis of the mortality prediction model. (A) Forest plot of AUC across eight predefined patient subgroups, with error bars indicating 95% confidence intervals. The red dashed line represents overall AUC. (B) Calibration curves stratified by sex, with DeLong’s test p-value for AUC difference between males and females. (C) Calibration curves stratified by age (< 65 vs. ≥65 years), with DeLong’s test p-value. (D) Distribution of predicted mortality probabilities by SOFA score (< 6 vs. ≥6). (E) Distribution of predicted mortality probabilities by APACHE II score (< 15 vs. ≥15). (F) Distribution of predicted mortality probabilities by comorbidity burden (number of comorbidities)

Head-to-head comparison with general severity scores

To explicitly evaluate the added discriminative and calibration value of our custom model beyond established general severity scores, we performed a direct comparison between our final logistic regression model and the SOFA and APACHE II scores on the independent validation set (n = 481). The comprehensive performance metrics are summarized in Supplementary Table 4.

Our custom model demonstrated significantly superior discriminatory power, with an AUC of 0.844 (95% CI: 0.799–0.890), compared to the SOFA score (AUC = 0.602, 95% CI: 0.523–0.681; DeLong’s test p < 0.001) and the APACHE II score (AUC = 0.503, 95% CI: 0.427–0.578; DeLong’s test p < 0.001). Regarding calibration, our model showed good agreement between predicted and observed outcomes (Hosmer-Lemeshow test p = 0.274, Brier Score = 0.097). In contrast, both the SOFA and APACHE II scores exhibited significant miscalibration (Hosmer-Lemeshow test p < 0.001 for both), as reflected in their substantially higher Brier scores (0.403 and 0.346, respectively). At its optimal threshold (0.135), our model also achieved better sensitivity (0.829) and specificity (0.745) than either general score at the conventional threshold of 0.5.

Discussion

This study successfully developed and validated a machine learning model for predicting in-hospital mortality risk in patients with sepsis and concurrent MRSA bloodstream infection. Through a rigorous, multi-stage feature selection process integrating statistical filtering, multi-model ensemble evaluation, and stability testing, we identified seven core predictive variables: Glasgow Coma Scale (GCS) score, minimum pH, maximum blood urea nitrogen (BUN), minimum white blood cell (WBC) count, minimum platelet count, maximum lactate level, and presence of pneumonia. The logistic regression model demonstrated excellent discriminative ability (AUC = 0.844) and good calibration in an independent validation set, with decision curve analysis confirming its broad clinical net benefit. Particularly important, SHAP interaction analysis revealed complex non-linear relationships among predictors, notably a strong synergistic effect between GCS score and minimum pH, offering a new perspective for understanding the mortality risk mechanisms in this population. The value of this study lies not only in providing a ready-to-use risk assessment tool but also in employing a systematic data science approach to uncover the key, parsimonious, and actionable pathophysiological axes driving clinical outcomes in this specific high-risk group.

A core contribution of this work is its advancement beyond the traditional logic of nomogram model construction, representing a methodological iteration and deepening. Although our team has previously developed prognostic nomogram models using similar data (critically ill patients from the MIMIC database), the two studies differ fundamentally in objectives, methods, and outputs. Traditional nomogram models typically focus on visualizing established risk factors and their weights, providing an intuitive graphical tool to visualize results from multivariable logistic regression, facilitating manual calculation of individual risk probabilities by clinicians. However, the construction of such models often relies on researchers’ prior knowledge or relatively simple univariate/multivariable regression screening; the process for determining the variable set may lack systematicity, and the models have limited capacity to capture complex interactions and non-linear relationships among variables. In contrast, this study employed a proactive, systematic multi-stage feature selection framework. Starting from a “large pool” of 64 broad candidate variables, this framework objectively identifies the minimum subset of variables with the greatest information value and robustness for predicting the highly specific outcome of “sepsis with concurrent MRSA bloodstream infection” through a machine learning-driven “screening-aggregation-stabilization” process. This approach reduces dependency on prior assumptions and enhances the objectivity of discovering new associations or confirming core ones. The final set of seven variables emerged as the most parsimonious and robust solution through this stringent computational process. Therefore, this study can be viewed as an evolution from “constructing an easy-to-use risk calculator (nomogram)” to “systematically discovering and validating the optimal set of predictive variables for this specific clinical problem.” The latter provides a more solid and reliable scientific foundation for the former—we first answered the fundamental question, “Which variables are truly the most important?” and the subsequently generated logistic regression model (easily convertible to a nomogram) is the direct application of that answer.

These seven core variables do not exist in isolation; they collectively and systematically map the critical pathway from initial insult to irreversible multiple organ failure in patients with sepsis and MRSA bacteremia. Impaired consciousness (low GCS score) and severe acidosis (low pH) were identified as the most important contributors in the model, and SHAP analysis revealed their strongest positive interaction. This strongly suggests that neurological decompensation and systemic metabolic collapse are not independent events but rather mutually exacerbating vicious cycles. Severe acidosis and inflammatory storms damage the blood-brain barrier and neuronal function, while the failure of central regulation (e.g., vasomotor dysfunction, dysregulated endocrine stress response) further aggravates peripheral circulatory failure and cellular metabolic disturbance. This finding aligns with the conclusion of Cheng et al. that neurologic involvement is a critical determinant of poor outcomes in severe infections [24]. However, our work, through interaction effect analysis, further couples it with metabolic status, revealing the extreme danger of their synergistic action, where the risk far exceeds simple summation. The elevated lactate level, a classic marker of tissue hypoperfusion and hypoxia, had its predictive value reaffirmed in our model. This is entirely consistent with the series of findings by Liu et al. and Lee et al., who confirmed the prognostic utility of lactate in various critically ill cohorts [25, 26]. Our model places this classic indicator in a more specific context and quantifies its covariance with other core indicators (e.g., pH).

Sequential damage to the renal and hematopoietic/coagulation systems constitutes another key axis driving mortality risk. The inclusion of maximum BUN and minimum platelet count precisely points to progressive renal dysfunction and systemic coagulation/inflammatory system disturbance. In the torrent of sepsis, rapidly rising BUN often heralds the onset of acute kidney injury or the persistent deterioration of pre-renal hypoperfusion, both of which are established poor prognostic factors in SAB [27]. Thrombocytopenia is a sensitive output signal of sepsis’ complex pathological network, potentially stemming from bone marrow suppression, increased consumption in disseminated intravascular coagulation (DIC), or immune-mediated destruction. Mitra et al. and Claushuis et al. emphasized the strong association between thrombocytopenia and adverse outcomes in sepsis, linking it to endothelial activation and immunothrombosis [28, 29]. The inclusion of minimum WBC count warrants deeper discussion. It challenges the traditional notion that sepsis is always accompanied by leukocytosis, potentially identifying a subset of patients entering a state of “immune paralysis” or bone marrow failure. This immunoparalytic phenotype is associated with Persistent Inflammation, Immunosuppression, and Catabolism Syndrome (PICS) and is a significant cause of later uncontrolled infection and multi-organ failure, receiving increasing attention in recent years [30]. Finally, the inclusion of pneumonia as the sole primary infection site feature in the model may reveal a specific and fatal high-risk subtype. The lungs, as an infection source, due to their vast microvascular surface area and direct alveolar-capillary gas exchange interface, may be more prone to triggering uncontrolled systemic inflammatory response and acute respiratory distress syndrome once hematogenous seeding occurs. This could also be related to MRSA’s own virulence characteristics, such as its propensity to cause necrotizing pneumonia and cavity formation [31].

The molecular mechanistic interpretation of this study could be further strengthened by integrating insights from recent research on sepsis molecular subtyping and metabolic heterogeneity. Several multi-omics integration studies have revealed significant immunometabolic reprogramming and molecular phenotypic heterogeneity in sepsis patients. For example, Wang et al. integrated large-scale cohorts, single-cell transcriptomics, and proteomics to identify a high-risk sepsis subtype characterized by severe CD4⁺ T cell exhaustion. They elucidated that monocyte-derived matrix metalloproteinase-9 (MMP9) drives immune dysfunction by disrupting T cell receptor signaling and calcium homeostasis. This provides a direct molecular pathway explanation for the association between metabolic stress indicators, such as elevated lactate and acidosis, and poor prognosis (PMID: 41306770) [32]. Another study employed single-cell sequencing combined with multiple machine learning algorithms to identify LILRA5 as a diagnostic marker for sepsis and revealed its association with monocyte inflammatory phenotypes, supporting the prognostic value of abnormal white blood cell counts from the perspective of immune cell differentiation (PMID: 36920635) [33]. Additionally, Ning et al. utilized single-cell trajectory analysis and multi-model machine learning to construct a monocyte differentiation-related prognostic signature and validated the diagnostic and prognostic value of genes such as EVL, highlighting the impact of molecular-level information on immune cell differentiation states on clinical outcomes (PMID: 36941583) [34]. Collectively, these studies indicate that the molecular heterogeneity of sepsis is profoundly reflected in the dysregulation of immune cell function, differentiation states, and metabolic pathways. The predictors incorporated in our model—such as low pH (acidosis), high lactate (metabolic stress), low platelet count (coagulation/inflammatory dysregulation), and low white blood cell count (immune paralysis)—represent the integration of these underlying molecular and immune phenotypes into routine clinical indicators. Therefore, our prediction model not only offers clinical utility but also resonates with findings from cutting-edge multi-omics research, thereby strengthening its scientific foundation for precision risk stratification in sepsis at a molecular mechanistic level.

The head-to-head comparison revealing superior performance of our parsimonious model over SOFA and APACHE II scores (Section “Head-to-head comparison with general severity scores”) warrants careful interpretation. This finding does not diminish the well-established utility of these general severity scores in broader ICU populations. Rather, it suggests that for the highly specific phenotype of “sepsis with concurrent MRSA bacteremia,” the predictive signal is powerfully concentrated within a few key physiological axes: neurological integrity (GCS), acid-base homeostasis (pH, lactate), renal function (BUN), and hematological/inflammatory response (platelets, WBC). Composite scores like SOFA, while containing some of these elements (e.g., platelets, GCS), also include components (e.g., bilirubin for liver) that may contribute less discriminative power in this specific context and whose weights are not optimized for this outcome. Our data-driven approach effectively distilled this core signal, removing redundancy and re-weighting components, thereby achieving more efficient prediction for this niche population. This underscores the value of developing syndrome- and pathogen-specific models alongside general tools. This line of thinking resonates with the work of Cohen et al., who compared clinical large language models to traditional machine learning for AMR prediction, aiming to mine predictive signals from diverse data representations [35].

Placing this study within a broader academic context further clarifies its unique positioning and innovative value. Previous extensive research has explored the prognosis of sepsis or the outcomes of SAB from different angles. For example, investigations into the burden of AMR in specific regions, such as the Eastern Mediterranean, highlight the sustained threat posed by resistant pathogens like MRSA [36]. This underscores the importance of developing targeted management tools. The study by Jeon et al. on persistent MRSA bacteremia explored treatment patterns and found limited impact of antimicrobial switching on mortality, pointing to the complexity of managing advanced disease [37]. This complements our study, which focuses on early risk stratification at presentation. Similarly, the bibliometric analysis by Lin et al. maps the global research landscape of MRSA bacteremia, identifying AI and machine learning as emerging hotspots [38]. Our work directly contributes to this trend by applying such methods to a focused clinical question. Furthermore, while pediatric studies like that of Sun et al. have developed predictive models for sepsis in children with S. aureus BSI, our model targets the distinct adult population and the more definitive endpoint of in-hospital death [20]. Compared to these important studies, the core innovation of this work lies in being the first to intersect the two strict conditions of “sepsis meeting Sepsis-3 criteria” and “microbiologically confirmed monomicrobial MRSA bloodstream infection,” specifically constructing a dedicated prediction model with in-hospital mortality as the endpoint for this high-risk subgroup characterized by complex management and high resource consumption. This allows our tool to more directly answer the core question clinicians face at the moment of diagnosis: “What is the mortality risk for this patient with MRSA hematogenous sepsis?”

From a clinical translation perspective, this model holds immediate application and deep integration potential. First, its high sensitivity makes it an excellent “screening tool” for effectively identifying nearly all high-risk patients (low missed diagnosis rate), which is crucial for initiating early intensive interventions. Second, the model’s parsimony (only 7 routine variables) and the interpretability of the logistic regression algorithm make it easy for clinicians to understand and trust. The regression coefficient and odds ratio for each variable provide clear risk quantification. Third, the model can be seamlessly integrated into electronic health record systems. The moment a patient is diagnosed with sepsis and has a positive MRSA blood culture, the system can automatically capture relevant data from the first 24 h, calculate and display the mortality risk probability and the main contributing risk factors in real-time (achievable via SHAP values), providing immediate, quantitative support for bedside decision-making. Finally, the model’s success conversely enlightens the standardization of clinical assessment: for patients with MRSA hematogenous sepsis, the initial assessment must systematically and quantitatively focus on neurological status, acid-base balance, renal function, coagulation profile, and a clear infection source, forming a structured assessment template.

While the current study demonstrates robust internal validation, the generalizability of our model requires further confirmation in external populations. As a logical next step, future research should prioritize evaluating the model’s transportability by applying it to independent critical care cohorts with comparable data granularity. Subsequently, prospective validation in a multi-center setting will be essential to assess its real-time predictive accuracy and clinical impact before any implementation can be considered. These steps are crucial to establish the model’s robustness and utility beyond the derivation context of the present study.

Certainly, this study has limitations. The foremost limitation of this study is the lack of external validation. The model was developed and validated using data from a single-center, retrospective database (MIMIC-IV). Despite comprehensive internal validation, the absence of testing in an independent cohort from a different healthcare setting significantly constrains the generalizability of our findings and their immediate clinical applicability. Additionally, the retrospective application of Sepsis-3 criteria presents inherent methodological challenges. Precisely defining the ‘time zero’ of sepsis onset is difficult. Our operational definition (anchoring on the first positive MRSA blood culture and extracting data from the preceding 24 h) aimed to capture early physiological derangement but may not fully align with the exact clinical onset identified prospectively. Thirdly, the inherent characteristics of the MIMIC-IV database limited our ability to incorporate certain potentially important predictors, such as detailed antimicrobial pharmacokinetic/pharmacodynamic parameters, the exact timing and completeness of source control, and pathogen molecular typing or specific virulence genes. This information might further optimize the model’s predictive performance and mechanistic interpretation [39]. Fourthly, the model did not incorporate several important treatment-related variables known to influence outcomes, such as the timing of appropriate antibiotic therapy or source control. This was a deliberate choice to ensure the model’s applicability at the point of initial assessment, estimating the patient’s inherent physiological risk distinct from subsequent therapeutic confounders. However, this decision also underscores a more fundamental methodological challenge in prognostic modeling of acute infections. Variables like “antibiotic administration within the first 24 hours,” often considered a treatment factor, can also act as proxies for initial disease severity and the treating team’s responsiveness. Their strong association with outcomes in univariate analysis (e.g., vancomycin use, Supplementary Table 2) likely reflects this confounding. Thus, even if included, such variables would conflate treatment effect with baseline risk, making causal interpretation difficult. Our approach prioritizes the estimation of a treatable physiological risk signature at presentation, acknowledging that the absolute predictive value may be limited in scenarios where post-presentation therapeutic factors become paramount. Finally, although our sample size and event per variable ratio meet general guidelines, the multi-stage feature selection process, which initially evaluated a larger set of variables using complex algorithms, inherently carries a risk of feature selection bias and overfitting. A specific limitation is that we did not employ nested cross-validation during feature selection, which is considered a best practice to obtain an unbiased performance estimate when the same data is used for both selection and model evaluation. The extensive internal validation (bootstrap, sensitivity analyses) mitigates but does not eliminate this concern.

In summary, this study, through a systematic machine learning feature selection method, identified and validated a parsimonious model composed of seven easily obtainable clinical variables for accurately predicting in-hospital mortality risk in patients with sepsis and concurrent MRSA bloodstream infection. This research transcends the traditional paradigm of nomogram model construction, striving to objectively discover the optimal set of predictive variables from data, thereby providing a more solid scientific foundation for the development of clinical tools. The cluster of predictors revealed by the model profoundly delineates the core pathophysiological pathways leading to multi-organ failure and death in this patient group. This study fills the gap for a dedicated prognostic assessment tool for this specific high-risk population. In the future, through prospective external validation and intelligent integration into clinical workflows, this model has the potential to become an important decision-support unit assisting clinicians in achieving early, precise risk identification and individualized treatment strategy formulation.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AMR

Antimicrobial Resistance

APACHE II

Acute Physiology and Chronic Health Evaluation II

AUC

Area Under the Curve

BSI

Bloodstream Infection

DCA

Decision Curve Analysis

GCS

Glasgow Coma Scale

LASSO

Least Absolute Shrinkage and Selection Operator

MIMIC-IV

Medical Information Mart for Intensive Care IV

MRSA

Methicillin-resistant Staphylococcus aureus

ROC

Receiver Operating Characteristic

SAB

Staphylococcus aureus Bacteremia

SHAP

SHapley Additive exPlanations

SOFA

Sequential Organ Failure Assessment

Author contributions

All authors contributed to the study conception and design. The research was designed by W. Zhao, and L. Chen. Data collection and analysis were performed by Y. Huang, and L. Chen. The first draft of the manuscript was written by Y. Huang, W. Zhao and L. Chen and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Funding

The authors declare that no funds, grants, or other support were received during the preparation of this manuscript.

Data availability

The dataset(s) supporting the conclusions of this article is(are) included within the article (and its additional file(s)).The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This retrospective study was conducted using data from the publicly available, de-identified Medical Information Mart for Intensive Care IV (MIMIC-IV) database. The establishment and release of MIMIC-IV were approved by the Institutional Review Boards (IRBs) of the Massachusetts Institute of Technology (MIT) and Beth Israel Deaconess Medical Center (BIDMC), with a waiver of informed consent due to retrospective data collection and full de-identification. The secondary analysis protocol for the present study was reviewed and approved by the Ethics Committee of Zhongshan City People’s Hospital. All procedures performed in this study involving human data were in accordance with the ethical standards of the Declaration of Helsinki. Given the retrospective nature of this study and the use of fully anonymized data, the requirement for written informed consent from participants was formally waived by the Ethics Committee of Zhongshan City People’s Hospital.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.