Enhancing prognostic accuracy in sepsis-induced cardiomyopathy: a machine learning approach

Abstract

Background

Sepsis-induced cardiomyopathy (SIMD) is a severe yet potentially reversible complication of sepsis, characterized by myocardial dysfunction and associated with high short-term mortality. Conventional scoring systems and traditional statistical models inadequately capture the complex pathophysiology of SIMD, highlighting the need for robust prognostic tools.

Methods

We retrospectively analyzed 1068 adult SIMD patients from the MIMIC-IV database, of whom 236 (22.1%) died within 28 days of ICU discharge. Candidate predictors were screened using Boruta, Least Absolute Shrinkage and Selection Operator (LASSO), and Recursive Feature Elimination with Cross-Validation (RFECV). Eight machine learning algorithms were developed and compared. Model performance was evaluated with receiver operating characteristic (ROC) curves, calibration curves, decision curve analysis, confusion matrices, and Kolmogorov–Smirnov (K-S) statistics. Model interpretability was assessed with SHapley Additive exPlanations (SHAP).

Results

Seven independent predictors were identified: Acute Physiology Score III (APS III), age, Charlson Comorbidity Index (CCI), cerebrovascular disease, alkaline phosphatase (ALP), lactate, and creatine kinase-MB (CK-MB). Logistic regression achieved consistent discrimination, with AUC values of 0.80 (95% CI 0.79–0.82) in the training set, 0.80 (95% CI 0.77–0.84) in the validation set, and 0.88 (95% CI 0.81–0.94) in the test set. Model accuracies were 70.0%, 67.0%, and 79.0%, respectively, with sensitivities ranging from 0.76–0.82 and specificities from 0.65–0.79. Negative predictive values (NPV) remained high (0.91–0.93), while positive predictive values (PPV) were moderate (0.38–0.52). The K-S statistic indicated strong discrimination (0.46, 0.45, and 0.52 across cohorts). SHAP analysis confirmed APS III (≈0.11), age (≈0.05), and ALP (≈0.03) as the most influential predictors, with CCI (≈0.02) and CK-MB (≈0.01) contributing modest but clinically relevant effects.

Conclusion

We established and validated a parsimonious logistic regression model with robust discrimination (AUC up to 0.88) and calibration (K-S > 0.45). The model underscores acute illness severity, aging, and hepatic dysfunction as principal determinants of short-term mortality in SIMD, offering valuable support for early risk stratification in critical care.

Supplementary Information

The online version contains supplementary material available at 10.1186/s40001-025-03354-0.

Introduction

Sepsis is a life-threatening condition caused by a dysregulated host response to infection, frequently resulting in multiple organ failure [1]. It affects approximately 48.9 million individuals and contributes to 11 million deaths worldwide each year. In the U.S., sepsis is responsible for more than one-third of all in-hospital mortality. Sepsis-induced cardiomyopathy (SIMD), first described by Parker et al. in 1984, is a common complication characterized by reversible cardiac dysfunction, with reported prevalence ranging from 10 to 70% due to inconsistent diagnostic criteria [2, 3]. Clinical manifestations include impaired myocardial contractility, electrophysiological abnormalities, cardiac remodeling, refractory hypotension during fluid resuscitation, and diminished responsiveness to vasoactive agents [3]. Although SIMD is potentially reversible, it remains associated with considerable mortality—35% during hospitalization and 51% within one year [4]. Diagnosing SIMD is challenging owing to the lack of standardized criteria, often requiring the exclusion of alternative diagnoses, such as acute coronary syndrome or pre-existing cardiomyopathies. The pathophysiology involves a complex interplay of inflammatory cytokines, nitric oxide, mitochondrial dysfunction, calcium dysregulation, autonomic disturbance, and myocardial edema [3]. These mechanisms collectively contribute to the development and progression of SIMD. Echocardiography serves as a key diagnostic tool but has inherent limitations; for instance, left ventricular ejection fraction may reflect vascular resistance rather than intrinsic contractility, potentially leading to misdiagnosis and suboptimal treatment. Although more specific diagnostic modalities for SIMD exist, they often suffer from insufficient data integration, hampering effective prevention [5, 6].

Accurate assessment of disease severity is essential for optimizing therapeutic strategies in SIMD, yet the absence of dedicated severity and prognostic scores complicates evaluation and outcome prediction in critically ill patients. Conventional scoring systems such as the Sequential Organ Failure Assessment (SOFA) and Acute Physiology and Chronic Health Evaluation II (APACHE II) evaluate multisystem organ function but inadequately capture the dynamic cardiac impairment specific to SIMD, relying predominantly on clinical signs and basic laboratory values rather than elucidating the underlying complex mechanisms of myocardial dysfunction [7, 8]. This underscores the urgent need for robust prognostic tools incorporating multidimensional biomarkers to facilitate real-time risk assessment, personalized therapy, and data-informed clinical decision-making in acute care settings.

Traditional prediction models such as logistic regression and Cox proportional hazards regression exhibit limited performance in forecasting outcomes among ICU patients, including short-term mortality, due to their dependence on linearity assumptions and inability to accommodate complex interactions [9]. They are poorly suited to incorporate the multifactorial risk profile of SIMD, which encompasses inflammatory markers, hemodynamic variables, and comorbid conditions [3, 10]. Consequently, these models often yield suboptimal predictive accuracy. Machine learning plays an increasingly vital role in medical prognosis, particularly in SIMD research and ICU risk prediction [9–11]. It demonstrates superior capability in predicting sepsis-related outcomes by managing high-dimensional data and capturing nonlinear relationships. By integrating diverse clinical data sources, ML models frequently exceed the performance of conventional statistical approaches. Deep learning further enhances predictive accuracy by modeling temporal sequences, thereby improving diagnostic precision, enabling personalized treatment, and supporting real-time monitoring. Empirical evidence indicates that machine learning outperforms traditional models in sepsis prognosis by detecting complex patterns, with algorithms such as support vector machines, random forests, and neural networks consistently demonstrating improved predictive performance [12]. Tseng et al. [13] employed multiple ML models—including logistic regression, SVM, random forest, XGBoost, and an ensemble of RF and XGBoost—to predict acute kidney injury following cardiac surgery, enabling early intervention. Kalimouttou et al. applied machine learning to implement the Surviving Sepsis Campaign bundles, significantly reducing mortality among 42,735 patients through six targeted interventions [14]. Bomrah et al. [15] systematically reviewed 29 studies involving 1,147,202 patients and found that feature selection methods such as random forest and XGBoost improved model accuracy, while deep learning provided additional insights for enhancing sepsis prognosis.

The rapid advancement of SIMD is fueled by inflammatory cascades and hemodynamic instability, with significant variability in patient outcomes. Traditional methods struggle to capture SIMD’s dynamic nature, complicating severity assessment. In contrast, machine learning excels at analyzing complex data, identifying new biomarkers, and understanding clinical interactions, enhancing prognostic accuracy and supporting clinical decisions. Improved medical informatics infrastructure allows machine learning to update predictive models more effectively than traditional approaches. Therefore, creating a big data–driven predictive model for sepsis-induced cardiomyopathy is timely and crucial.

Materials and methods

Source of the data

This study relied on precise data collection and thorough analysis using the Medical Information Mart for Intensive Care (MIMIC-IV) 2.0 database, a highly authoritative resource developed at MIT. This database includes diverse data, particularly on critically ill patients in the ICU at Beth Israel Deaconess Medical Center [16]. The MIMIC-IV database is a robust clinical resource containing data from over 200,000 emergency visits and 70,000 ICU admissions. It offers core clinical data, standardized interpretations with ICD-9/ICD-10 codes, and high-frequency physiological data verified by algorithms and nurses. All personal health information is deidentified per HIPAA standards, enabling secondary analysis without individual consent. This infrastructure supports real-time monitoring, treatment decisions, and long-term outcome studies through its unified data architecture [17, 18]. Researcher LX adhered to protocols to access and extract data from the MIMIC-IV database, using certification number 46830776 for compliance. The study followed the TRIPOD guidelines to ensure quality and reliability [19].

Population selection criteria and definitions

The inclusion criteria for this study were as follows: (1) aged ≥ 18 years and first ICU admission at first hospitalization; (2) fulfilled the Sepsis 3.0 criteria; (3) length of stay in the ICU > 24 h but < 100 days; and (4) diagnosis of SIMD (defined as cTnT > 0.01 ng/mL within 24 h of ICU admission according to criteria and database limitations) [18, 20–23].

Data extraction

We utilized Navicat Premium 15.2 and structured query language (SQL) to extract data across multiple dimensions. The dataset included patient demographics (age at admission, sex, time of death, length of hospitalization, ICU stay duration); clinical information within the first 24 h of ICU admission (respiratory rate [RR], blood pressure [BP], heart rate [HR], oxygen saturation, body temperature); comorbidities indexed by the Charlson Comorbidity Index (e.g., hypertension, diabetes mellitus, history of myocardial infarction, coronary artery disease, renal disease, severe liver disease, malignancy); laboratory results from the first day of ICU admission (white blood cell count [WBC], platelet count [PLT], troponin T [cTnT], lactate [Lac], serum creatinine [Scr]); coagulation indicators (partial thromboplastin time [PTT], prothrombin time [PT]); severity scores (SOFA, Acute Physiology Score III [APS III]); and initial treatments (use of vasoactive medications, renal replacement therapy [RRT], invasive mechanical ventilation [IMV]). Variables with > 20% missing data were excluded from the analysis. For variables with ≤ 20% missing data, missing values were imputed using a Random Forest–based imputation method, which has been shown to provide robust performance for mixed clinical data [24].See the Supplementary Table S1 for details (Fig. 1).

Fig. 1.

Flowchart of patient selection from the MIMIC-IV 2.0 database. From 45,706 septic patients in MIMIC-IV 2.0, those with cardiac comorbidities, missing cTnT data, or cTnT < 0.01 ng/mL were excluded, leaving 3202 with SIMD. After excluding short/long ICU stays and repeat admissions, 1068 patients were included, divided into the 28-day survival group (n = 832) and mortality group (n = 236). MIMIC Medical Information Mart for Intensive Care IV, SIMD sepsis-induced cardiomyopathy, ACS acute coronary syndrome, CHF chronic heart failure, severe tachyarrhythmias include supraventricular tachycardia, ventricular tachycardia, and ventricular fibrillation

Outcome

The primary outcome was 28-day mortality after ICU discharge.

Statistical analyses

Before analyzing continuous variables, normality was assessed using the Shapiro–Wilk test. For variables not normally distributed, such as APACHE III and SOFA scores, the Mann–Whitney U test compared group differences. Categorical variables (e.g., sex, comorbidities) were analyzed using the chi-square test. Patients were categorized into survival and nonsurvival groups based on 28-day post-ICU discharge status.

Feature selection proceeded as follows: the Boruta algorithm first assessed feature importance using Z scores; variables retained by Boruta were then evaluated with LASSO regression, keeping only predictors with non-zero coefficients while optimizing mean squared error (MSE) and standard errors; finally, the intersection of LASSO-selected variables with those identified via recursive feature elimination with cross-validation (RFECV) was taken. The factors identified by all three methods were used to develop the prognostic model [25].

Model performance was evaluated using area under the receiver operating characteristic curve (AUC), optimal cutoff, accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1 score on the training, validation, and test sets. ROC curves were plotted for visualization.

To ensure reliability, the dataset was divided into training and validation sets in a 7:3 ratio during model selection. In multiclass settings, separate validation partitioning was omitted to (1) estimate generalization via k-fold cross-validation (e.g., average AUC and variance) and (2) improve computational efficiency by integrating validation into training. For final model construction, data were split into training, validation, and test sets in a 7:1.5:1.5 ratio.

Eight algorithms were considered: logistic regression (LR), Extreme Gradient Boosting (XGBoost), random forest (RF), AdaBoost, Gaussian Naive Bayes (GNB), Complement Naive Bayes (CNB), support vector machine (SVM), and k-nearest neighbors (KNN). Variables were selected using the joint Boruta–LASSO–RFECV workflow. Ten-fold nested cross-validation ensured stability, and a grid search optimized hyperparameters to maximize AUC–ROC. Performance metrics included AUC–ROC, accuracy, sensitivity, specificity, PPV, NPV, F1 score, and Cohen’s kappa. After selecting the optimal model, prescreened variables were used to fit the final model. Visualization included calibration curves, decision curves, confusion matrices, and Kolmogorov–Smirnov (K–S) plots. SHapley Additive exPlanations (SHAP) analyses generated heatmaps, global summary plots, force plots, and parallel-coordinate plots to quantify feature contributions and interpret decision mechanisms.

All statistical analyses were performed using R (version 3.6.3) and Python (version 3.7). The following packages were used: XGBoost (version 2.0.1), scikit-learn (version 1.1.3), and SHAP (version 0.43.0). A two-sided p value < 0.05 was considered statistically significant. All statistical tests were two-sided, and a p value < 0.05 was considered statistically significant. Reported p values and 95% confidence intervals were adjusted for multiplicity using the Holm–Bonferroni method (Fig. 2).

Fig. 2.

Workflow of predicting 28-day mortality in SIMD using the MIMIC-IV database. After exclusions, 1068 patients were divided into survival (n = 832) and mortality (n = 236) groups. Clinical features were screened by Boruta, LASSO, and RFECV, then tested across multiple classifiers, with logistic regression showing the best performance. SHAP analysis was applied for model interpretation. MIMIC-IV Medical Information Mart for Intensive Care IV, ICU intensive care unit, HR heart rate, RR respiratory rate, SBP systolic blood pressure, DBP systolic blood pressure, MBP mean blood pressure, WBC white blood cell, PLT platelet, CK-MB creatine kinase-muscle brain. CBD cerebrovascular disease, CKD chronic kidney disease, LASSO, least absolute shrinkAge and selection operator, XGBoost extreme gradient boosting; logistic, logistic regression, AdaBoost adaptive boosting; GNB, Gaussian naive Bayes, CNB complementive naive Bayes, SVM support vector machine, KNN nearest neighbor. Created at https://BioRender.com

Results

From the MIMIC-IV database, a total of 45,706 records of sepsis patients were reviewed, among which 3,202 met the SIMD criteria. After exclusion, 1,068 patients were analyzed. The incidence of SIMD was 2.34%.They were divided into survivor and nonsurvivor groups based on 28-day post-ICU discharge survival. Table 1 compares variables between these groups.

Table 1.

Baseline characteristics of patients stratified by 28-day outcome

Variables	Total(n = 1068)	28d-SG	28d-MG	P
N	1068	832	236
Age(years old) Median (Q1, Q3)	68.00 (57.00;78.25)	67.00 (56.00;77.00)a	73.00 (62.75;81.25)ab	< 0.001
Gender n (%)				0.801
Female	357 (33.4%)	276 (33.2%)	81 (34.3%)
Male	711 (66.6%)	556 (66.8%)	155 (65.7%)
Vital Signs on ICU Admission, Median (Q1, Q3)
HR,(bmp)	105.00(90.00;122.00)	102.00(89.00;120.00)a	113.00(96.00;127.00)ab	< 0.001
SBP,(mmHg)	86.00 (77.00;94.00)	86.00 (78.50;95.00)a	83.00 (72.00;92.50)ab	< 0.001
DBP,(mmHg)	46.00 (39.00;52.00)	46.00 (40.00;52.00)a	44.00 (35.00;52.00)ab	0.008
MBP,(mmHg)	57.00 (50.00;63.75)	58.00 (51.00;64.00)a	54.00 (47.00;63.00)ab	0.001
RR,(bpm)	29.00 (25.00;33.00)	28.00 (24.00;33.00)a	30.50 (26.00;35.00)ab	< 0.001
Body temperature,(°C)	37.22 (36.90;37.83)	37.28 (36.94;37.83)	37.11 (36.83;37.80)	0.112
SpO2,(%)	93.00 (90.00;95.00)	93.00 (90.00;95.00)a	92.00 (88.00;95.00)ab	0.036
Laboratory parameters on ICU admission, Median (Q1, Q3)
WBC count,(103/μL)	13.30 (9.40;18.55)	13.30 (9.30;18.40)	13.70 (9.70;19.42)	0.361
Hematocrit	30.90 (26.50;35.70)	30.96 (26.60;35.73)	30.65 (25.40;35.70)	0.348
Hb,(g/dL)	10.20 (8.60;11.72)	10.20 (8.70;11.80)	10.10 (8.20;11.60)	0.060
PLT(103/μL)	160.50(114.00;216.25)	160.00(117.75;216.00)	162.00(107.00;220.25)	0.222
pH	7.33 (7.29;7.37)	7.33 (7.30;7.37)	7.33 (7.24;7.37)	0.102
PO2,(mmHg)	102.05 (81.00;124.60)	102.69 (82.00;125.72)a	100.00 (74.00;119.16)ab	0.029
PCO2,(mmHg)	45.00 (39.00;51.00)	45.54 (40.00;50.77)a	42.00 (35.62;52.16)ab	0.002
Aniongap,(mmol/L)	17.00 (14.00;20.00)	16.00 (14.00;19.00)a	19.00 (16.00;23.00)ab	< 0.001
Lactate,(mmol/L)	2.20 (1.70;2.70)	2.12 (1.66;2.69)a	2.20 (1.77;2.80)ab	0.024
BicarbonTate	24.00 (21.00;27.00)	25.00 (22.00;28.00)a	23.00 (19.75;26.25)ab	< 0.001
Albumin,(g/dL)	3.19 (2.90;3.40)	3.20 (2.95;3.40)a	3.10 (2.70;3.40)ab	0.005
ALT,(U/L)	36.83 (25.26;68.02)	35.53 (26.00;60.00)a	43.00 (22.00;118.25)ab	0.019
AST,(U/L)	55.00 (41.00;92.00)	54.55 (41.00;77.08)a	66.00 (40.75;251.50)ab	< 0.001
ALP,(U/L)	8.61 (6.80;9.80)	8.27 (6.69;9.63)a	8.61 (7.10;10.50)ab	0.001
Na+,(mmol/L)	140.00(137.00;142.00)	140.00(137.00;142.00)	140.00(136.00;143.00)	0.958
K+,(mmol/L)	4.60 (4.10;5.10)	4.50 (4.10;5.10)a	4.75 (4.10;5.50)ab	0.001
BUN,(mg/dL)	30.00 (19.00;47.00)	27.00 (18.00;42.00)a	40.00 (26.00;63.25)ab	< 0.001
Creatinine(mg/dL)	1.40 (1.00;2.30)	1.30 (0.98;2.00)a	2.00 (1.28;3.00)ab	< 0.001
TBil(μmol/L)	1.02 (0.61;1.64)	1.00 (0.63;1.60)a	1.10 (0.60;2.40)ab	0.012
Ck-MB(U/L)	7.00 (4.00;15.00)	6.64 (4.00;13.59)a	8.75 (4.70;18.00)ab	< 0.001
Ck-CPK(IU/L)	234.61(124.00;658.36)	225.98(123.18;605.97)a	260.00(131.62;1096.88)ab	0.036
LDH(U/L)	334.50(278.00;430.12)	330.01(275.98;386.28)a	374.50(283.75;664.75)ab	< 0.001
Glucose,(mmol/L)	168.00(134.00;216.00)	167.00(133.00;214.00)	175.50(136.00;234.00)	0.117
Fibrinogen,(mg/dL)	251.94 (238.00;259.34)	252.24(238.65;259.37)	251.56 (235.95;259.18)	0.761
PT,(S)	14.28 (12.70;17.00)	14.00 (12.60;16.40)a	15.10 (13.20;19.62)ab	< 0.001
PTT,(S)	30.50 (26.80;35.10)	30.10 (26.70;34.40)a	31.50 (27.35;38.10)ab	0.001
INR	1.30 (1.10;1.60)	1.30 (1.10;1.50)a	1.40 (1.20;1.80)ab	< 0.001
Critical assessment on admission, Median (Q1, Q3)
APS III	52(38.00,73.00)	47.00(36.00,64.00)a	77.00(57.00,102.00)ab	< 0.001
SOFA	6.00(4.00,10.00)	6.00(4.00,9.00)a	9.00(6.00,13.00)ab	< 0.001
GCS	14.00 (10.00;15.00)	14.00 (11.00;15.00)a	11.00 (5.50;14.00)ab	< 0.001
Comorbidity, n (%)
CCI	7.00 (5.00;8.00)	6.00 (5.00;8.00)a	8.00 (6.00;9.00)ab	< 0.001
CBD	127 (11.9%)	79 (9.5%)a	48 (20.3%)ab	< 0.001
CPD	317 (29.7%)	244 (29.3%)	73 (30.9%)	0.692
Diabetes with cc	110 (10.3%)	81 (9.7%)	29 (12.3%)	0.309
Renal disease	353 (33.1%)	255 (30.7%)a	98 (41.5%)ab	0.002
Malignancy	121 (11.3%)	87 (10.5%)	34 (14.4%)	0.116
Advanced life support,n(%)
RRT	136(12.7%)	83(10.0%)a	53(22.5%)ab	< 0.001
IMV	964(90.3%)	746(89.7%)	218(92.8%)	0.265
Vasoactive_Agent	631(59.1%)	467(56.1%)a	164(69.5%)ab	< 0.001
Antibiotic	1032(96.6%)	809(97.2%)	223(94.5%)	0.063
Length of Stay (LOS), Median (Q1, Q3)
Los of Hospital (day)	9.94(5.96,16.81)	10.04(6.12,16.94)a	9.73(4.84,15.86)ab	0.035
Los of ICU (day)	3.46(1.83,6.98)	3.13(1.71,5.95)a	5.41(2.27,9.81) ab	< 0.001

Values are presented as median (Q1, Q3) or n (%), as appropriate. P values were calculated using the Mann–Whitney U test for continuous variables and the chi-square test for categorical variables. P values were adjusted for multiplicity using the Holm–Bonferroni method. Superscript a indicates a significant difference compared with the Total group; superscript b indicates a significant difference when compared with the 28d-SG group

HR Heart Rate; SBP Systolic Blood Pressure; DBP Diastolic Blood Pressure; MBP Mean Blood Pressure; RR Respiratory Rate; SpO₂ Peripheral Oxygen Saturation; WBC White Blood Cell Count; Hb Hemoglobin; PLT Platelet Count; PO₂ Partial Pressure of Oxygen in Arterial Blood; PCO₂ Partial Pressure of Carbon Dioxide in Arterial Blood; ALT Alanine Aminotransferase; AST Aspartate Aminotransferase; ALP Alkaline Phosphatase; Na⁺, Sodium; K⁺ Potassium; BUN Blood Urea Nitrogen; TBil Total Bilirubin; CK-MB Creatine Kinase-MB; CK-CPK Creatine Phosphokinase; LDH Lactate Dehydrogenase; PT Prothrombin Time; PTT Partial Thromboplastin Time; INR International Normalized Ratio; APS III Acute Physiology Score III; SOFA Sequential Organ Failure Assessment; GCS Glasgow Coma Scale; CCI Charlson Comorbidity Index; CBD Cerebrovascular Disease; CPD Chronic Pulmonary Disease; RRT renal replacement therapy; IMV Invasive Mechanical Ventilation; LOS Length of Stay

Baseline characteristics

A total of 1,068 SIMD patients were enrolled, with 236 (22.1%) in the 28-day mortality group and 832 (77.9%) survivors. The median age was 68 years, and 66.6% were male. Significant differences (p < 0.05) between groups were observed for body temperature, WBC count, Hct, Hb, PLT, serum Na +, glucose, fibrinogen, pre-existing conditions (COPD, diabetes complications, malignancy), and the use of antibiotics and IMV in the ICU. In the 28-day mortality group, 22.5% received renal replacement therapy (RRT) and 69.5% received vasoactive agents—both higher than in the survivor group, where 10.0% received RRT and 56.1% received vasoactive agents (p < 0.05). The mortality group had a shorter hospital stay [9.73 (4.84, 15.86) days] but a longer ICU stay [5.41 (2.27, 9.81) days] compared with the survivor group [hospital stay: 10.04 (6.12, 16.94) days; ICU stay: 3.13 (1.71, 5.95) days]. Baseline characteristics are detailed in Table 1 by 28-day survival status.

Predictor screening

A total of 1,068 SIMD patients were divided into a training group (n = 908) and a validation group (n = 321) in a 7:3 ratio, with no significant differences between the two groups (Table 2). From an initial pool of 49 clinical and laboratory variables (Supplementary Material 1), three feature selection methods were applied: Boruta retained 23 variables, LASSO-CV retained 8, and RFECV retained 24. By integrating their results, seven consistent predictors—APS III, Age, CBD, CCI, Lactate, ALP, and CK-MB—were identified for model construction. A graphical summary (Fig. 3E) illustrates the overlaps and differences among methods, and Spearman correlation analysis confirmed no significant multicollinearity (|r|< 0.8). This transparent multi-method selection pipeline strengthens both the robustness and reproducibility of our model development [25–27].

Table 2.

Training set and test set variability analysis

Variables	Total	Validation	Training	P
N	1068	321	747
Age(years old) Median (Q1, Q3)	68.00 (57.00, 78.25)	67.00 (57.00, 79.00)	68.00 (58.00, 78.00)	0.985
Gender n (%)				0.506
Female	357 (33.4%)	112 (34.9%)	245 (32.8%)
Male	711 (66.6%)	209 (65.1%)	502 (67.2%)
Vital Signs on ICU Admission, Median (Q1, Q3)
HR,(bmp)	105.00 (90.00, 122.00)	105.00 (89.00, 124.00)	104.00 (91.00, 121.00)	0.944
SBP,(mmHg)	86.00 (77.00, 94.00)	86.00 (78.00, 93.88)	85.00 (77.00, 95.00)	0.719
DBP,(mmHg)	46.00 (39.00, 52.00)	45.00 (38.25, 51.00)	46.00 (39.00, 52.00)	0.288
MBP,(mmHg)	57.00 (50.00, 63.75)	57.00 (50.00, 63.00)	57.00 (50.12, 64.00)	0.760
RR,(bpm)	29.00 (25.00, 33.00)	29.00 (25.00, 33.00)	29.00 (25.00, 33.00)	0.846
Body temperature,(°C)	37.22 (36.90, 37.83)	37.22 (36.94, 37.88)	37.22 (36.89, 37.83)	0.757
SpO2,(%)	93.00 (90.00, 95.00)	93.00 (90.00, 95.00)	92.00 (90.00, 94.75)	0.133
Laboratory parameters on ICU admission, Median (Q1, Q3)
WBC count,(103/μL)	13.30 (9.40, 18.55)	13.40 (9.50, 18.80)	13.30 (9.30, 18.50)	0.751
Hematocrit	30.90 (26.50, 35.70)	30.60 (26.30, 34.90)	31.10 (26.50, 35.95)	0.352
Hb,(g/dL)	10.20 (8.60, 11.72)	10.20 (8.60, 11.60)	10.20 (8.60, 11.80)	0.970
PLT(103/μL)	160.50 (114.00, 216.25)	164.65 (114.00, 217.00)	159.00 (114.00, 216.00)	0.728
pH	7.33 (7.29, 7.37)	7.33 (7.30, 7.36)	7.33 (7.29, 7.37)	0.843
PO2,(mmHg)	102.05 (81.00, 124.60)	102.00 (84.00, 124.02)	102.18 (79.37, 124.75)	0.518
PCO2,(mmHg)	45.00 (39.00, 51.00)	45.00 (39.00, 51.00)	45.00 (39.00, 51.00)	0.678
Aniongap,(mmol/L)	17.00 (14.00, 20.00)	17.00 (14.00, 20.00)	17.00 (14.00, 20.00)	0.715
Lactate,(mmol/L)	2.20 (1.70, 3.20)	2.13 (1.69, 3.13)	2.20 (1.70, 3.20)	0.512
BicarbonTate	24.00 (21.00, 27.00)	24.00 (22.00, 27.00)	24.00 (21.00, 28.00)	0.801
Albumin,(g/dL)	3.19 (2.90, 3.40)	3.19 (2.96, 3.40)	3.19 (2.90, 3.40)	0.514
ALT,(U/L)	36.83 (25.26, 68.02)	39.00 (25.28, 72.00)	36.00 (25.32, 67.16)	0.328
AST,(U/L)	55.00 (41.00, 92.00)	55.39 (42.21, 95.26)	54.76 (39.50, 91.60)
ALP,(U/L)	8.61 (6.80, 11.00)	8.60 (6.99, 11.00)	8.62 (6.69, 11.00)	0.670
Na+,(mmol/L)	140.00 (137.00, 142.00)	140.00 (137.00, 142.00)	140.00 (137.00, 142.00)	0.569
K+,(mmol/L)	4.60 (4.10, 5.10)	4.50 (4.20, 5.10)	4.60 (4.10, 5.15)	0.939
BUN,(mg/dL)	30.00 (19.00, 47.00)	30.00 (19.00, 46.00)	29.00 (19.00, 47.50)	0.922
Creatinine(mg/dL)	1.40 (1.00, 2.30)	1.40 (1.00, 2.20)	1.40 (1.00, 2.30)	0.899
TBil(μmol/L)	1.02 (0.61, 1.64)	1.00 (0.60, 1.70)	1.02 (0.66, 1.60)	0.944
Ck-MB(U/L)	7.00 (4.00, 15.00)	6.74 (4.00, 14.48)	7.00 (4.01, 15.14)	0.606
Ck-CPK(IU/L)	234.61 (124.00, 658.36)	222.77 (131.90, 590.17)	237.19 (122.66, 675.99)	0.630
LDH(U/L)	334.50 (278.00, 430.12)	340.21 (278.68, 407.17)	332.11 (277.70, 438.26)	0.819
Glucose,(mmol/L)	168.00 (134.00, 216.00)	165.50 (132.00, 214.00)	169.00 (135.00, 217.00)	0.613
Fibrinogen,(mg/dL)	251.94 (238.00, 259.34)	253.79 (238.43, 259.87)	251.46 (237.99, 259.03)	0.188
PT,(S)	14.28 (12.70, 17.00)	14.20 (12.80, 16.80)	14.30 (12.60, 17.15)	0.736
PTT,(S)	30.50 (26.80, 35.10)	30.10 (26.60, 35.50)	30.80 (26.90, 34.90)	0.822
INR	1.30 (1.10, 1.60)	1.30 (1.10, 1.50)	1.30 (1.10, 1.60)	0.770
Critical assessment on admission, Median (Q1, Q3)
APS III	52.00 (38.00, 73.25)	52.00 (39.00, 70.00)	52.00 (38.00, 74.00)	0.483
SOFA	6.00 (4.00, 10.00)	7.00 (4.00, 10.00)	6.00 (4.00, 10.00)	0.406
GCS	14.00 (10.00, 15.00)	14.00 (10.00, 15.00)	14.00 (10.00, 15.00)	0.678
Comorbidity, n (%)
CCI	7.00 (5.00, 8.00)	7.00 (5.00, 8.00)	7.00 (5.00, 8.00)	0.566
CBD	127 (11.9%)	38 (11.8%)	89 (11.9%)	0.972
CPD	317 (29.7%)	87 (27.1%)	230 (30.7%)	0.227
Diabetes with cc	110 (10.3%)	34 (10.6%)	76 (10.2%)	0.837
Renal disease	353 (33.1%)	105 (32.7%)	248 (33.2%)	0.876
Malignancy	121 (11.3%)	34 (10.6%)	87 (11.7%)	0.618
Advanced life support,n(%)
RRT	136 (12.7%)	33 (10.3%)	103 (13.8%)	0.115
IMV	964 (90.3%)	283 (88.2%)	681 (91.2%)	0.129
Vasoactive_Agent	631 (59.1%)	197 (61.4%)	434 (58.1%)	0.319
Antibiotic	1032 (96.6%)	312 (97.2%)	720 (96.4%)	0.501
Length of Stay (LOS), Median (Q1, Q3)
Los of Hospital (day)	9.94 (5.96, 16.81)	9.22 (5.80, 16.97)	10.12 (6.08, 16.75)	0.406
Los of ICU (day)	3.45 (1.83, 6.99)	3.44 (1.85, 7.00)	3.46 (1.83, 6.97)	0.901

Values are presented as median (Q1, Q3) or n (%), as appropriate. P values were calculated using the Mann–Whitney U test for continuous variables and the chi-square test for categorical variables

HR Heart Rate; SBP Systolic Blood Pressure; DBP Diastolic Blood Pressure; MBP Mean Blood Pressure; RR, Respiratory Rate; SpO₂, Peripheral Oxygen Saturation; WBC White Blood Cell Count; Hb Hemoglobin; PLT Platelet Count; PO₂ Partial Pressure of Oxygen in Arterial Blood; PCO₂ Partial Pressure of Carbon Dioxide in Arterial Blood; ALT Alanine Aminotransferase; AST Aspartate Aminotransferase; ALP Alkaline Phosphatase; Na⁺ Sodium; K⁺ Potassium; BUN Blood Urea Nitrogen; TBil Total Bilirubin; CK-MB Creatine Kinase-MB; CK-CPK Creatine Phosphokinase; LDH Lactate Dehydrogenase; PT Prothrombin Time; PTT Partial Thromboplastin Time; INR International Normalized Ratio; APS III Acute Physiology Score III; SOFA Sequential Organ Failure Assessment; GCS Glasgow Coma Scale; CCI Charlson Comorbidity Index; CBD Cerebrovascular Disease; CPD Chronic Pulmonary Disease; RRT renal replacement therapy; IMV Invasive Mechanical Ventilation; LOS Length of Stay

Fig. 3.

Feature selection process for predicting 28-day mortality in patients with SIMD. A LASSO coefficient profiles: The trajectories of coefficients shrink toward zero with increasing penalty parameter (λ), indicating variable selection. B LASSO cross-validation curve: Optimal λ value determined via tenfold cross-validation, minimizing binomial deviance. C Boruta feature importance ranking: Box plots showing confirmed, tentative, and rejected variables, highlighting features with significant predictive value. D RFECV performance curve: Recursive feature elimination with cross-validation identified the optimal number of features based on model accuracy. E Integration of feature selection methods: Venn-type network visualization of key predictors identified by Boruta, LASSO, and RFECV, including APSIII, lactate, ALP, age, CKMB, CCI, and CBD. F Correlation heatmap of predictors. HR, heart rate; SBP, systolic blood pressure; DBP systolic blood pressure, MBP mean blood pressure, RR respiratory rate, WBC white blood cell, Hb hemoglobin, PLT platelet; ALT alanine aminotransferase, AST aspartate aminotransferase, ALP alkaline phosphatase, BUN blood urea nitrogen, TBil total bilirubin, CK-MB creatine kinase-muscle brain, CK-CPK creatine kinase; LDH, lactate dehydrogenase, PT prothrombin time, PTT partial thromboplastin time, INR international normalized ratio, APS III acute physiology and chronic health evaluation scoring system III, APACHE III; SOFA Sequential Organ Failure Assessment; GCS Glasgow Coma Scale, CCI Charlson Comorbidity Index, CBD cerebrovascular disease, CPD chronic pulmonary disease, Diabetes with cc diabetes with complications, RRT renal replacement therapy, IMV invasive mechanical ventilation

Model performance

We trained eight models (XGBoost, logistic regression, random forest, AdaBoost, GaussianNB, ComplementNB, SVC, and KNN), each with five runs. On the training set, random forest achieved the best performance (AUC = 0.98, accuracy = 0.97), while XGBoost and KNN also performed well (AUC = 0.89 and 0.91, respectively). In contrast, GNB and CNB performed poorly (AUC = 0.79 and 0.73) (Fig. 4A). On the validation set, random forest declined markedly (AUC = 0.79, sensitivity = 0.28), suggesting overfitting, whereas XGBoost, logistic regression, and AdaBoost demonstrated more stable generalization (AUC = 0.79–0.82). KNN performed worst (AUC = 0.70) (Fig. 4B). Decision curve analysis confirmed clinical utility for logistic regression and random forest, while calibration curves indicated reliable probability estimation for XGBoost and random forest (Fig. 4C–D). On the test set, random forest yielded the highest average precision (AP = 0.96), whereas logistic regression achieved an AP of 0.60 (Fig. 4F). Considering stability and clinical applicability, the logistic regression model was ultimately selected (Supplementary Material 2) (Figs. 4, 5).

Fig. 4.

Model performance evaluation for predicting 28-day mortality in SIMD. A ROC curves (Training set): Discrimination ability of multiple classifiers, including XGBoost, logistic regression, random forest, AdaBoost, GNB, CNB, SVM, and KNN. B ROC curves (Validation set): External performance comparison of classifiers in the validation cohort. C Decision curve analysis (Validation set): Net clinical benefit across a range of threshold probabilities, showing logistic regression with superior overall utility. D Calibration curves (Validation set): Agreement between predicted and observed outcomes, indicating the goodness of fit for each model. E Precision–recall (PR) curves (Training set): Trade-off between precision and recall in the training cohort, highlighting robustness of logistic regression. F PR curves (Validation set): Precision–recall relationship in the validation cohort, further confirming the comparative performance of classifiers

Fig. 5.

Comprehensive performance metrics of machine learning models for predicting 28-day mortality in SIMD. A Training set: Radar-like circular plot summarizing key performance metrics (AUC, accuracy, sensitivity, specificity, F1 score, positive predictive value, and negative predictive value) across different classifiers. B Validation set: External comparison of the same performance indicators in the validation cohort, demonstrating logistic regression with balanced and robust predictive ability across multiple dimensions. AUC area under the ROC curve; KNN k-nearest neighbors, SVM support vector machine, CNB complementive naive Bayes, GNB Gaussian naive Bayes, AB adaptive boosting, RF random ferrost, Log logistic regression, XGB extreme gradient boosting

Logistic regression model

The dataset (N = 1,068) was randomly partitioned into training (70%, n = 748), validation (15%, n = 160), and test (15%, n = 160) subsets via stratified sampling, as detailed in Supplementary Material 3 [28]. Seven prognostic features were identified using the LASSO, Boruta, and RFECV algorithms, and a logistic regression model was developed with parameters optimized through nested tenfold cross-validation.

The model demonstrated stable and reliable discrimination across all subsets, with AUC values of 0.80 (95% CI 0.79–0.82) in the training cohort, 0.80 (95% CI 0.77–0.84) in the validation cohort, and 0.88 (95% CI 0.81–0.94) in the test cohort. Accuracy ranged from 0.67 to 0.79, with balanced sensitivity (0.76–0.82) and specificity (0.65–0.79), indicating robust performance and generalizability. Confusion matrix analysis revealed consistently high-negative predictive values (0.91–0.93), underscoring the model’s reliability in excluding adverse outcomes and minimizing false negatives, whereas positive predictive values remained moderate (0.38–0.52), reflecting class imbalance but supporting conservative classification in critical care contexts. The Kolmogorov–Smirnov (K–S) statistic further confirmed the model’s discriminatory capacity, with values of 0.46, 0.45, and 0.52 in the training, validation, and test cohorts, respectively, demonstrating clear separation between survivors and non-survivors and good probability calibration (Table 3, Figs. 6, 7).

Table 3.

Performance of the logistic regression model in the training, validation, and test sets

Indicators	Training Set	Validation Set	Test Set
AUC(95%CI)	0.80(0.79–0.82)	0.80(0.77–0.84)	0.88 (0.81–0.94)
Cut-off(95%CI)	0.19(0.16–0.21)	0.19(0.16–0.21)	0.25
Accuracy(95%CI)	0.70(0.67–0.73)	0.67(0.64–0.71)	0.79
Sensitivity(95%CI)	0.82(0.76–0.87)	0.76(0.65–0.87)	0.78
Specificity (95%CI)	0.67(0.61–0.72)	0.65(0.60–0.71)	0.79
PPV(95%CI)	0.42(0.38–0.45)	0.38(0.35–0.41)	0.52
NPV(95%CI)	0.93(0.92–0.94)	0.91(0.88–0.94)	0.93
F1 score(95%CI)	0.55(0.53–0.56)	0.50(0.46–0.55)	0.62
Kappa(95%CI)	0.36(0.33–0.38)	0.30(0.24–0.36)	0.48

Values are presented as point estimates with 95% confidence intervals (95% CI)

AUC Area Under the Curve, PPV Positive Predictive Value, NPV Negative Predictive Value; CI Confidence Interval

Fig. 6.

Performance evaluation of the logistic regression model for predicting 28-day mortality in SIMD. A ROC curve (Training set): Discriminative ability of the logistic regression model in the training cohort. B ROC curve (Validation set): External validation of model discrimination across multiple folds. C ROC curve (Test set): Predictive performance of the model in the independent test cohort. D Calibration curve (Test set): Agreement between predicted and observed probabilities, showing good calibration. E Decision curve analysis (Test set): Clinical utility of the logistic regression model across different threshold probabilities, indicating positive net benefit. F Learning curve: Stability and robustness of model performance with increasing training samples, showing minimal risk of overfitting.

Fig. 7.

Classification performance of the logistic regression model for predicting 28-day mortality in SIMD. A Confusion matrix (Training set): Distribution of true positives, true negatives, false positives, and false negatives in the training cohort. B Confusion matrix (Test set): Classification outcomes in the independent test set, reflecting generalization ability. C Kolmogorov–Smirnov (KS) curve (Test set): Separation between cumulative distributions of positive and negative classes, with a KS statistic of 0.65 at threshold 0.20, indicating strong discriminative power

Collectively, these findings highlight that the logistic regression model not only achieved satisfactory predictive accuracy but also provided well-calibrated probability estimates, reinforcing its potential clinical utility for early risk stratification in critically ill patients.

SHAP-based interpretability analysis

SHAP-based interpretability analysis provided comprehensive insights into the logistic regression model for predicting 28-day mortality in patients with septic cardiomyopathy (Fig. 8).At the global level, the summary heatmap (Fig. 8A) demonstrated substantial heterogeneity in feature contributions, with APS III, age, and ALP identified as the most influential predictors. The beeswarm plot (Fig. 8B) further revealed that when APS III ≥ 90, age ≥ 70 years, and ALP ≥ 150 U/L, median SHAP values exceeded |0.2| and shifted toward positive contributions, consistently increasing predicted mortality risk. These thresholds were derived from SHAP dependence plots at turning points where median SHAP values crossed zero or surpassed the effect threshold (|SHAP|≥ 0.2), and they largely aligned with clinically recognized cutoffs (e.g., ≥ 70 years for elderly status). In contrast, higher Charlson Comorbidity Index (CCI ≥ 10) and lower CK-MB (< 5 ng/mL) were generally associated with negative SHAP values, indicating relative protective effects. Notably, although CCI typically reflects poor long-term outcomes, its short-term impact was overshadowed by acute severity markers (e.g., APS III), thus appearing protective in the model. Similarly, lower CK-MB suggested limited myocardial injury, consistent with reduced predicted risk.The ranking of mean absolute SHAP values (Fig. 8C) confirmed APS III (≈ 0.11) as the dominant predictor, followed by age (≈ 0.05) and ALP (≈ 0.03), whereas CCI (≈ 0.02), CK-MB (≈ 0.01), total bilirubin (≈ 0.005), and lactate (≈ 0.002) contributed less prominently.

Fig.8.

SHAP-based interpretability analysis of the logistic regression model for predicting 28-day mortality in SIMD. A SHAP heatmap: Patient-level heterogeneity of variable contributions, highlighting APSIII, age, and CBD as major drivers of mortality risk. B SHAP beeswarm plot: Distribution of SHAP values for each predictor, showing APSIII, age, and ALP with the strongest positive influence, while CCI and CKMB contributed less. C Mean absolute SHAP values: Ranking of variable importance confirmed APSIII as the most influential predictor, followed by age and CBD. D SHAP force plots (individual patients): (1) Example case where elevated ALP (20.5), advanced age (84 years), and lower APSIII (49.0) shifted baseline mortality risk moderately, partially offset by CCI = 6.0. (2) Example case where high APSIII (103.0) and CBD = 1.0 strongly increased mortality risk despite high CCI (13.0). APSIII Acute Physiology and Chronic Health Evaluation III, ALP Alkaline Phosphatase, CCI Charlson Comorbidity Index, CBD Cerebrovascular disease; CK-MB Creatine Kinase MB

At the individual level, the SHAP force plots (Fig. 8D) demonstrated how combined features shifted the baseline mortality risk. In one representative case, the baseline risk increased from 0.12 to 0.22, primarily driven by elevated ALP (20.8 U/L), advanced age (84 years), and APS III (49.0), partially counterbalanced by a CCI score of 6. In another case, the risk rose from 0.15 to 0.39, largely attributable to APS III = 103.0, CK-MB = 9.0 ng/mL, and bilirubin = 1.0 mg/dL. Although a high CCI (13.0) exerted a modest protective contribution (≈0.02), this was outweighed by the stronger positive effects of APS III (≈0.11) and organ dysfunction markers, resulting in a substantial net increase in predicted risk. Collectively, Fig. 8 highlights that acute physiological derangement (APS III), advanced age, and hepatic dysfunction markers (ALP, bilirubin) are the principal drivers of short-term mortality in septic cardiomyopathy, whereas comorbidity burden (CCI) and myocardial injury markers (CK-MB) show context-dependent but secondary prognostic effects.

To facilitate external validation and potential iterative refinement, we have made the current version of the model available through a web-based platform (https://www.xsmartanalysis.com/model/list/predict/model/html?mid=27647&symbol=81Eo75VX6QR9XG886912). Meanwhile, we are actively collecting local clinical data, with future work aiming to integrate such data and investigate adaptive strategies, including RL-based approaches, to further enhance the model’s generalizability and clinical applicability across diverse healthcare settings.

Discussion

This study developed a three-stage feature selection framework that integrated Boruta, LASSO, and RFECV to identify seven independent predictors from 27 variables, including APS III, Age, and lactate [29]. Boruta’s random forest–based method effectively captured nonlinear associations, LASSO’s L1 penalty reduced multicollinearity [30], and RFECV iteratively eliminated uninformative features to minimize overfitting. The optimized logistic regression model achieved a small generalization error (ΔAUC = 2.7% between training and test sets), clearly outperforming the overfitting-prone random forest (ΔAUC = 23%). Seven clinically relevant predictors—APS III age, CBD, CCI, ALP, lactate, and CK-MB—were ultimately retained, forming the basis of the final model. SHAP analysis confirmed their pathophysiological relevance, with APS III (mean SHAP = 0.11) as the dominant contributor, followed by age (≈0.05) and ALP (≈0.03), whereas CBD, CCI, lactate, and CK-MB contributed more modest effects (≤ 0.03). The logistic model with L2 regularization (C = 1 × 10⁻⁴) yielded a Brier score of 0.13, striking a balance between accuracy and computational efficiency, and was consistent with the optimal framework summarized by Gupta et al. for sepsis prognostic modeling [31]. Decision curve analysis further demonstrated significant net benefit across a wide range of threshold probabilities (15–90%). When compared with nonlinear algorithms such as XGBoost (training AUC = 0.89), the logistic model was computationally lighter, more interpretable, and thus more suitable for real-time integration into electronic health record systems.

Gao et al. [32] applied XGBoost and random forest algorithms to predict ICU mortality in sepsis patients, reporting AUC values of 0.79–0.85, thereby highlighting the strong discriminative capacity of nonlinear models. Likewise, Fan et al. [24] developed an interpretable machine learning model for sepsis-associated acute kidney injury, which demonstrated robust performance and underscored the translational importance of interpretability. In comparison, our logistic regression model achieved relatively higher predictive performance (AUC up to 0.88, K–S > 0.45) while maintaining lower computational cost and greater interpretability, making it more suitable for real-time implementation in EHR systems. Nevertheless, its positive predictive value remained moderate (0.38–0.52), indicating the need for strategies to address class imbalance and improve positive case recognition. Future studies should aim to preserve the simplicity and interpretability of the model while exploring ensemble learning and dynamic time-series methods to enhance predictive comprehensiveness and clinical applicability.

High APS III scores are strongly linked to worse outcomes, including mortality, making it valuable for patient prognostication and resource management. Unlike other scoring systems such as SOFA and SAPS, APS III encompasses a broader range of variables, enhancing predictive accuracy in complex conditions [33, 34]. Future research should explore mechanistic links between APS III components and disease processes to advance precision medicine in this high-risk population.

Our multivariate screening further identified ALP, CK-MB, and lactate as key indicators associated with SIMD, offering mechanistic insights into their roles in systemic inflammation and myocardial injury. CK-MB, a recognized marker of myocardial damage, was strongly associated with SIMD. Prior research has linked microRNA-378a-3p dysregulation to elevated CK-MB in sepsis patients, suggesting that CK-MB may form part of a novel diagnostic panel for early cardiac dysfunction [35]. In addition, neogambogic acid was shown to attenuate myocardial injury in septic mice by reducing CK-MB, highlighting its dual potential as a biomarker and therapeutic indicator [36, 37].

ALP elevation in septicemia may reflect impaired hepatobiliary excretion and bile reflux. Its positive correlation with C-reactive protein (CRP) and neutrophil-to-lymphocyte ratio underscores its association with infection-related inflammation [38]. Proximity of ALP to the endotoxin receptor CD14 suggests a role in host defense. Elevated ALP has been linked to liver injury and infection-related mortality, supporting its value as a prognostic marker [39]. Experimental studies also show alkaline phosphatase (AP) exerts protective effects by detoxifying lipopolysaccharide (LPS) and extracellular ATP, converting ATP into anti-inflammatory adenosine, and thereby attenuating inflammation [40]. Recombinant human tissue nonspecific alkaline phosphatase (rhTNAP) improved survival and reduced LPS-driven inflammation in murine sepsis models [41]. AP also modulates intestinal barrier integrity in colitis, reducing bacterial translocation [42]. Collectively, these findings suggest ALP is not only a biomarker but also a promising therapeutic candidate for SIMD, warranting further clinical trials.

Lactate levels and clearance are widely recognized as key prognostic indicators in sepsis and septic shock, with direct implications for SIMD [3–5]. Lactate serves as both a substrate for myocardial oxidation and a marker of systemic hypoperfusion [43]. Elevated lactate at ICU admission has been strongly associated with SIMD, particularly in younger patients, those with prior heart failure, or positive blood cultures [44]. Mechanistically, ALP elevation reflects bile acid–induced myocardial fibrosis via TGR5 receptor activation, while excessive lactate disrupts mitochondrial function through CK-MB–mediated pathways [45].

Age profoundly influences SIMD prognosis. In elderly patients, cardiovascular aging, immune senescence, and comorbidity burden exacerbate sepsis-induced myocardial dysfunction. The pathophysiology of SIMD involves inflammation, mitochondrial dysfunction, and impaired energy metabolism; these effects are amplified in older adults by age-related metabolic decline and dysregulated inflammatory responses [3, 46]. Emerging therapies targeting these pathways show promise: melatonin has been reported to mitigate septic cardiac dysfunction via mitochondrial modulation [47], while inhibition of toll-like receptor–mediated pyroptosis may represent a novel strategy, especially for elderly patients with heightened inflammatory reactivity [48].

The Charlson Comorbidity Index (CCI) remains a valuable predictor of mortality risk, stratifying patients by comorbidity burden [49]. In sepsis, higher CCI scores worsen prognosis, reflecting the additive effect of chronic disease on acute illness [46]. Understanding interactions between specific comorbidities and SIMD may enhance precision in risk stratification and therapy.

Cerebrovascular disease (CBD) was defined according to the Charlson Comorbidity Index and encompassed a history of previous or current stroke (with or without sequelae), transient ischemic attack (TIA), cerebral hemorrhage, cerebral infarction, and other non-traumatic cerebrovascular events. CBD represents another important comorbidity, as patients with stroke or TIA frequently exhibit cardiac abnormalities, even in the absence of pre-existing heart disease. For example, a clinical study identified frequent ECG changes and myocardial alterations following CVD events [50], suggesting cerebrovascular disease can exacerbate cardiac dysfunction in SIMD. Management of sepsis patients with CVD requires special consideration, particularly fluid therapy, to avoid fluid overload in those with concurrent heart failure [51]. This underscores the need for personalized treatment strategies in SIMD patients with cerebrovascular comorbidities.

Limitations

This study has several limitations. First, SIMD was diagnosed solely based on cTnT levels (> 0.01 ng/mL), which may have excluded patients with impaired myocardial function but normal cTnT, potentially reducing diagnostic accuracy. Second, the model was developed and validated exclusively using the single-center MIMIC-IV database. Although authoritative, its specific demographics and treatment regimens may limit generalizability. Differences in sepsis management and cardiac evaluation across centers were not addressed, highlighting the need for external multicenter validation. Nevertheless, our inclusion and exclusion criteria were consistent with prior studies [18, 20–23].

Third, the model relied on static data from the first 24 h of ICU admission, without incorporating dynamic disease progression. Since SIMD evolves rapidly, early recognition is critical. Future studies should integrate longitudinal data, including daily laboratory and hemodynamic measures, and apply time-series modeling to improve predictive accuracy.

Finally, while SHAP analysis identified clinically relevant predictors (e.g., APS III, age, lactate), the underlying biological pathways remain incompletely defined. For instance, elevated ALP may reflect hepatic dysfunction or systemic inflammation, but its direct causal link to myocardial injury warrants experimental confirmation. To enhance clinical utility, future research should integrate multicenter datasets, echocardiography, ambulatory monitoring, and multi-omics analyses, while linking predictive modeling with mechanistic studies. Embedding real-time early warning systems into electronic health records (EHRs) may ultimately enable precision risk stratification and personalized therapy in SIMD.

Conclusion

This study developed a 28-day mortality prediction model for SIMD by integrating Boruta, LASSO, and RFECV with diverse clinical and biochemical indicators. Seven key predictors were identified, with APS III emerging as the most significant. The logistic regression model demonstrated consistent performance, achieving AUC values of 0.80 in the training cohort, 0.80 in the validation cohort, and 0.88 in the test cohort, thereby confirming its robustness and clinical utility. SHAP analysis underscored the central role of inflammation, metabolic derangements, and organ dysfunction in SIMD prognosis. Limitations include the single-center design and reliance on cTnT for diagnosis, which may underestimate true SIMD prevalence. Future research should focus on external multicenter validation, incorporation of echocardiographic and multi-omics data, and exploration of novel biomarkers to improve diagnostic accuracy and guide individualized treatment strategies.

Author contributions

Xiang Li: study design, manuscript drafting. Huixin Cheng: data proofreading, formal analysis. Dina Ainiwaer, Xinxin Du, Yang Chunbo, Ilzati Aizezi: data collection, processing, and figure preparation. Yi Wang, Xiangyou Yu, Zhan Sun: manuscript review and editing. Zhan Sun, Yi Wang: supervision. All authors read and approved the final manuscript.

Funding

This work was supported by the Innovation Team of Tian Shan in Xinjiang Uygur Autonomous Region (Grant No. 2024D14013), the National Natural Science Foundation of China (Grant No. 82460372), the Natural Science Foundation of Xinjiang Uygur Autonomous Region (Grant No. 2022D01E53 and 2022D01C477).

Data availability

The dataset analyzed in this study was derived from the MIMIC-IV database, a well-established and authoritative resource in clinical informatics research. In accordance with database policies, raw data cannot be openly shared due to privacy and ethical restrictions. Researchers interested in accessing the data must complete the official MIMIC application process, which requires submission of project details, research objectives, and confirmation of ethical compliance. Upon approval, access will be granted for the specified research purpose. For inquiries related to the dataset specific to this study, requests may be directed to the corresponding author, Zhan Sun. Data sharing will follow MIMIC’s ethical regulations and institutional guidelines to ensure appropriate use and protection of patient information.

Declarations

Ethics approval and consent to participate

This study was conducted in accordance with the principles of the Declaration of Helsinki. All data were obtained from the de-identified, HIPAA-compliant MIMIC-IV database; therefore, no personal health information was included. In compliance with institutional review board (IRB) requirements, the study protocol was submitted in writing and received official approval.

Consent for publication

All authors have reviewed and approved the final version of this manuscript and consent to its publication. The work has not been published previously and is not under consideration for publication elsewhere.

Competing interests

The authors declare no competing interests.