Machine learning-based prediction of sepsis-induced myocardial injury: external validation and SHAP interpretation

Abstract

Background

Sepsis-induced myocardial injury (SIMI) is a common complication in sepsis patients with poor prognosis. Consequently, its early accurate prediction is crucial for optimizing clinical management.

Methods

We collected data on 14,208 patients with sepsis from the MIMIC IV database, and used the Bootstrap Sampling to handle unbalanced data. 70% of sepsis patients were selected randomly from the MIMIC-IV database, and feature selection was performed using a Boruta algorithm. Six machine learning prediction models (Decision Tree, GBDT, AdaBoost, Logistic Regression, Random Forest, CatBoost) were constructed. The remaining 30% of patients were utilized to validate the model’s accuracy. Additionally, local data from the Intensive Care Unit (ICU) of Zunyi First People’s Hospital were employed for further model validation. Model performance was assessed using AUC, decision curve analysis, PR curves, calibration plots, and SHapley Additive Explanations (SHAP).

Results

A total of 1882 sepsis patients were enrolled, with a SIMI risk prediction of 43.3%. Univariate analysis identified significant differences in multiple indicators between SIMI and non-SIMI groups (p < 0.05). Among six machine learning models, the CatBoost model exhibited the optimal performance, with an accuracy of 0.8690, AUROC of 0.9215, F1-score of 0.8452, and MCC of 0.7368. Five-fold cross-validation confirmed its stable performance (mean AUROC = 0.9183), and external validation showed good generalizability (AUROC = 0.7567, 95% CI: 0.6944–0.8191). SHAP analysis revealed that CK-MB, D-dimer, and NT-proBNP were the top three risk factors for SIMI.

Conclusions

The CatBoost model constructed in this study has high accuracy, stability, and generalizability for predicting SIMI risk prediction in sepsis patients. SHAP analysis confirms that CK-MB, D-dimer, and NT-proBNP stand out as the primary predictive factors for sepsis-induced myocardial injury (SIMI) risk. This model could assist clinicians in early identification of high-risk patients and guide targeted interventions, thereby improving patient prognosis.

Clinical trial number

Not applicable.

Introduction

Sepsis is a life-threatening syndrome characterized by a dysregulated host response to infection, leading to acute organ dysfunction and potentially fatal outcomes [1, 2]. As one of the most vulnerable organs during sepsis, the heart is frequently involved, resulting in sepsis-induced myocardial injury (SIMI)—a common complication defined by ventricular dilatation, reduced systolic function, biventricular dysfunction, and impaired volume responsiveness [3], Clinical studies have reported that the risk prediction of SIMI could reach up to 70% among sepsis patients, and those with concurrent myocardial injury faced a 2–3 fold higher mortality risk compared to those without [4]. This profound impact on prognosis underscores the critical need for early identification and targeted intervention in high-risk individuals.

The pathogenesis of SIMI is complex and multifactorial, involving excessive cardiac inflammatory responses, oxidative stress, myocardial cell death, impaired myocardial energy metabolism, and mitochondrial dysfunction [5]. Despite advances in clinical management, there is currently no valid scoring system for predicting the occurrence of sepsis-induced myocardial injury (SIMI) [6–8]. These tools often fail to capture the nonlinear interactions between clinical variables and the heterogeneous nature of sepsis-induced organ injury, highlighting the unmet need for more precise, personalized risk stratification models to optimize clinical decision-making.

In recent years, machine learning (ML) has emerged as a powerful tool in medical research, particularly in disease prediction and prognosis [9]. Unlike conventional statistical methods, ML algorithms excel at handling high-dimensional, nonlinear clinical data, enabling the discovery of complex relationships that may be overlooked by traditional approaches [10, 11], Furthermore, the integration of explainable artificial intelligence (XAI) techniques, such as SHapley Additive exPlanations (SHAP), has addressed a key limitation of ML models—lack of interpretability—by quantifying the contribution of individual features to each prediction. This transparency fosters clinician trust, facilitates the translation of ML models into clinical practice, and supports the development of personalized treatment strategies tailored to individual patient characteristics [12].

To address the limitations of existing predictive tools for SIMI, this multicenter study aimed to develop and externally validate an interpretable ML model for predicting the risk prediction of SIMI in sepsis patients. By leveraging data from the MIMIC-IV database and a local clinical cohort, we compared the performance of multiple ML algorithms to identify the optimal predictive model. Additionally, we employed SHAP analysis to elucidate the key predictors of SIMI and their relative contributions, providing clinically actionable insights into the model’s decision-making process. The ultimate goal of this study is to provide a reliable, interpretable tool that can assist clinicians in early risk stratification of SIMI, thereby guiding timely interventions and improving patient outcomes.

Methods

Data source

This retrospective cohort study was conducted using data from two independent databases: the Medical Information Mart for Intensive Care IV (MIMIC-IV) and a local institutional database (Department of Critical Care Medicine, Zunyi First People’s Hospital). The MIMIC-IV database was utilized for model development and internal validation [13], which is a large-scale, deidentified clinical database encompassing comprehensive health records of 380,000 + sepsis patients admitted to the Beth Israel Deaconess Medical Center (BIDMC), Boston, Massachusetts, USA, between 2008 and 2019. This database is jointly maintained by BIDMC and the Massachusetts Institute of Technology (MIT), Cambridge, Massachusetts, and provides high-quality, structured data covering demographic characteristics, vital signs, laboratory test results, medication administration, procedural records, and outcomes. For external validation, retrospective data were collected from 325 sepsis patients admitted to the Department of Critical Care Medicine, Zunyi First People’s Hospital (Zunyi, China) between January 2019 and December 2024.

Ethical approval for MIMIC-IV was obtained from the Institutional Review Boards (IRB) of Beth Israel Deaconess Medical Center (Boston, MA, USA) and the Massachusetts Institute of Technology (Cambridge, MA, USA). The local database complies with the Declaration of Helsinki and was approved by the Ethics Committee of Zunyi First People’s Hospital (approval number: 2025-1-14). All data were fully deidentified to remove protected health information (PHI) in compliance with the Health Insurance Portability and Accountability Act (HIPAA) guidelines (for MIMIC-IV) and local ethical regulations (for the local database), thus individual informed consent was waived for both databases. Access to MIMIC-IV was restricted to investigators who completed the Collaborative Institutional Training Initiative (CITI) program in human subjects research protection. Author Botao Li obtained official access authorization (Certification Number: 13878778) and was responsible for data extraction, cleaning, and validation. For the local database, data extraction was conducted independently by two investigators using a standardized form; discrepancies were resolved via consultation with senior clinician Min Xu to ensure accuracy. All extracted data were encrypted and stored on a password-protected server in adherence to data security regulations.

Study population

Inclusion criteria

Patients with sepsis meeting the Sepsis-3.0 criteria [14], age ≥ 18 years.

Exclusion criteria

Severe heart failure (LVEF < 30% or grade IV cardiac function) and recent acute myocardial infarction, patients with malignant tumors, patients expected to die within 24 h, patients with bleeding disorders, chronic obstructive pulmonary disease, chronic kidney disease, severe liver function impairment, data missing. Patients with a history of coronary heart disease (CHD) combined with acute myocardial infarction, unstable angina, or acute cardiac ischemia within 3 months before enrollment were strictly excluded (Only patients with stable CHD without acute cardiac events were included in the final analysis).

Diagnostic criteria

For outcome definition, the maximum cTnT value was determined from 24 h after ICU admission throughout the entire ICU stay to ensure strict temporal separation between predictors and the outcome. Sepsis-induced myocardial injury (SIMI) was diagnosed based on the 99th percentile upper reference limit (URL) of cTnT at each study site [15, 16]. For the MIMIC-IV database, the 99th percentile of the cTnT URL was 0.1 ng/mL; therefore, SIMI at this site was defined as a cTnT value > 0.1 ng/mL [17]. For Zunyi First People’s Hospital, the 99th percentile of the cTnT URL was 0.03 ng/mL; thus, SIMI at this site was defined as a cTnT value > 0.03 ng/mL. During external validation, the local site-specific threshold was used as the ground truth to evaluate model generalizability.

Data collection

Data of patients admitted to the intensive care unit (ICU) within 24 h were extracted from the MIMIC-IV database using Navicat Premium (Version 17.0) via structured query language (SQL) [18]. This local database derived from the electronic medical record (EMR) system of the Department of Critical Care Medicine, Zunyi First People’s Hospital. No interventions were performed during the data collection process, ensuring the integrity of the observational nature of the study. Two databases incorporated six key categories of characteristics measured on the 24 h of ICU admission: (1) Demographics: age, gender; (2) Comorbidities: hypertension, diabetes mellitus, coronary heart disease (CHD), cerebral infarction; (3) Scale scores: Acute Physiology and Chronic Health Evaluation II (APACHE II); (4) Treatments: use of mechanical ventilation, continuous renal replacement therapy (CRRT), antibiotic agents, and vasoactive agents; (5) Laboratory indicators: white blood cells (WBC), neutrophils, hemoglobin, platelets, alanine transaminase (ALT), aspartate transaminase (AST), total bilirubin, creatinine, calcium, potassium, phosphorus, lactate, magnesium, etc. ICU length of stay (ICU LOS) was collected as a final statistical variable after ICU discharge. For outcome definition, the maximum cTnT value was determined from 24 h after ICU admission throughout the entire ICU stay, as cTnT is the gold-standard biomarker for defining SIMI outcome., before the diagnosis of SIMI was confirmed. SIMI was defined using cTnT measured after the 24-h predictor window to ensure strict temporal separation. To reduce bias caused by missing data, variables with more than 20% missing values were excluded, while the remaining variables were processed using the multiple imputation by chained equations (MICE) method. Patients were divided into two groups based on the primary endpoint: the sepsis-induced myocardial injury (SIMI) group and the non-SIMI group.

Establishment and validation of the prediction models

The Boruta algorithm is a feature selection technique designed to identify the most influential features affecting a target variable within a dataset [19]. It works by generating g selection technique designed to identify the most influential features affecting a target variable within a datasety. glucose. For vscore for each feature. A feature is deemed important if its Z-score significantly exceeds the maximum Z-score of the shadow features over multiple independent tests.

Following the application of the Synthetic Minority Over-sampling Technique (SMOTE) to mitigate class imbalance inherent in the dataset, the MIMIC-IV database was partitioned into a 70% training subset and a 30% internal validation subset for predictive model development. To address potential class imbalance in model training, SMOTE was applied only to the training set and strictly nested within each cross-validation fold to prevent data leakage. No resampling was performed on the internal test set or external validation cohort. Complemented by five-fold cross-validation to ensure rigorous assessment of model robustness and mitigate overfitting. While the local cohort from Zunyi First People’s Hospital served as the external validation set. Missing data were handled using multiple imputation by chained equations (MICE). Variables with a missing rate > 20% were directly excluded. For variables with a missing rate < 20%, MICE was performed with 5 imputation cycles. Linear regression was used for continuous variables, and logistic regression was used for categorical variables. All clinical baseline variables were included in the imputation prediction model to ensure imputation accuracy and reliability. Six machine learning models (Decision Tree, GBDT, AdaBoost, Logistic Regression, RF, CatBoost) were constructed, with hyperparameters optimized through grid search and five-fold cross-validation, followed by internal cross-validation to assess stability using metrics including AUROC, sensitivity, specificity, DCA, and calibration curves, external validation to confirm generalizability, and SHAP analysis to quantify individual feature contributions and enhance clinical interpretability of the optimal model’s decision-making process.

Statistical analysis

Data preprocessing, statistical analysis, and visualization were performed using Python 3.7 and R 4.4.2. DecisionLinnc was only used for auxiliary automated statistical verification, and all core modeling and analysis results were based on Python and R to ensure the reproducibility of our findings. For continuous variables, the data are expressed as the mean ± standard deviation when normally distributed. For non-normal distributions, the data were presented as median(Q1-Q3). Categorical variables were expressed as counts and percentages. Between-group comparisons of continuous variables were performed using the independent samples t-test (for normal distributions) or Wilcoxon rank-sum test (for non-normal distributions), whereas categorical variables were analyzed with the chi-square test or Fisher’s exact test, as appropriate.

Despite the robust predictive capacity of machine learning (ML) models, their inherent “black-box” nature and limited interpretability hinder clinical translation. To address this limitation, the SHapley Additive exPlanations (SHAP) method was employed to decode the underlying. Decision-making mechanisms of the developed ML models. By calculating Shapley values for each input feature, SHAP quantifies the marginal contribution of individual variables to each prediction, generating intuitive visualizations that illustrate both the direction and magnitude of each factor’s influence on outcomes. This enhanced transparency fosters clinician trust in the model’s predictions and facilitates its integration into clinical practice. Statistical significance was defined as a two-tailed P-value < 0.05. All statistical analyses and model interpretations were performed using R software (version 4.4.2) and Python (version 3.7.2).

Results

Patients

The inclusion and exclusion criteria for patient selection were illustrated in Fig. 1. This study enrolled a total of 1882 patients from the MIMIC-IV database and 251 patients from the ICU of Zunyi First People’s Hospital. Based on whether cardiac troponin levels were higher than the 99th percentile of the cardiac troponin T (cTnT) upper reference limit (URL), the patients were categorized into two groups: “SIMI group” and “Non-SIMI group” (Fig. 1).

Fig. 1.

Patients recruitment flowchart

Baseline characteristics

In the MIMIC-IV database, the risk prediction of sepsis-induced myocardial injury (SIMI) was 43.3% (n = 815). Univariate analysis showed that there were significant differences between the two groups in the following variables (p < 0.05): age, hypertension (HTN), diabetes mellitus (DM), coronary heart disease (CHD), Glucocorticoid use, low molecular weight heparin (LMWH) use, Vasopressor use, Acute Physiology and Chronic Health Evaluation II (APACHE II) score, left ventricular ejection fraction (LVEF), hospital mortality, and various physiological parameters, including Neutrophil, Hemoglobin, red cell distribution width - standard deviation (RDW-SD), red cell distribution width - coefficient of variation (RDW-CV), Magnesium, Phosphate, alanine aminotransferase (ALT), aspartate aminotransferase (AST), Albumin, creatine kinase (CK), creatine kinase-MB isoenzyme (CK-MB), N-terminal pro-B-type Natriuretic Peptide (NT-proBNP), D-dimer, Fibrinogen, and Lactate (p < 0.05) (Table 1).

Table 1.

Comparison of baseline characteristics between the non-SIMI group and SIMI group

Variable	Overall(n = 1882)	Non-SIMI group(n = 1067)	SIMI group(n = 815)	p-value
Age(years)	64.48 ± 17.51	65.35 ± 17.40	63.35 ± 17.60	0.014
Male, n(%)	1,060.00 (56.32%)	618.00 (57.92%)	442.00 (54.23%)	0.110
HTN, n(%)	1,057.00 (56.16%)	621.00 (58.20%)	436.00 (53.50%)	0.042
CI, n(%)	185.00 (9.83%)	112.00 (10.50%)	73.00 (8.96%)	0.266
DM, n(%)	467.00 (24.81%)	282.00 (26.43%)	185.00 (22.70%)	0.063
CHD, n(%)	522.00 (27.74%)	216.00 (20.24%)	306.00 (37.55%)	< 0.001
WBC(×10⁹/L)	12.20 (8.50–17.10)	11.80 (8.10–16.00)	13.00 (8.90–18.50)	0.007
Neutrophil Count (×10⁹/L)	9.55 (6.16–14.28)	9.11 (5.92–13.61)	10.06 (6.64–15.35)	< 0.001
Lymphocyte Count (×10⁹/L)	1.03 (0.60–1.54)	1.02 (0.59–1.54)	1.03 (0.60–1.54)	0.593
RBC(×10¹²/L)	3.69 ± 0.85	3.65 ± 0.82	3.73 ± 0.89	0.060
Hemoglobin(g/dL)	11.09 ± 2.45	10.98 ± 2.33	11.22 ± 2.58	0.035
RDW-SD(fL)	33.58 ± 7.28	33.25 ± 6.90	34.01 ± 7.73	0.028
RDW-CV(%)	14.98 ± 2.33	15.11 ± 2.41	14.81 ± 2.22	0.007
Platelet(×10⁹/L)	206.26 ± 111.11	209.48 ± 116.83	202.05 ± 103.07	0.144
Calcium(mmol/L)	1.10 ± 0.13	1.10 ± 0.11	1.11 ± 0.15	0.444
Magnesium(mmol/L)	1.95 ± 0.46	1.94 ± 0.46	1.98 ± 0.47	0.045
Phospate(mmol/L)	3.73 ± 1.63	3.62 ± 1.49	3.86 ± 1.79	0.002
ALT(U/L)	36.00 (19.00–80.00)	33.00 (17.00–62.00)	43.00 (22.00–107.00)	< 0.001
AST(U/L)	55.00 (30.00–120.00)	46.00 (26.00–95.00)	69.00 (37.00–164.00)	< 0.001
Total Bilirubin(mg/dL)	0.60 (0.40–1.20)	0.60 (0.40–1.30)	0.60 (0.40–1.20)	0.103
Albumin(g/dL)	2.90 (2.50–3.40)	2.90 (2.40–3.40)	3.00 (2.50–3.40)	0.006
Creatinine(mg/dL)	1.29 ± 0.86	1.27 ± 0.89	1.31 ± 0.82	0.368
CK(U/L)	209.00 (76.00–603.00)	155.00 (60.00–510.00)	273.00 (109.00–757.00)	0.004
CK-MB(ng/mL)	6.00 (3.00–13.00)	4.00 (2.00–8.00)	9.00 (4.00–23.00)	< 0.001
NT-proBNP(pg/mL)	3,368.00 (1,180.00–9,472.00)	2,371.00 (868.00–5,432.00)	5,432.00 (1,726.00–12,404.00)	< 0.001
D-dimer(ng/ml)	4785.00 (2,323.00–6,435.00)	3897.00 (1,811.00–5,302.00)	5368.00 (3,897.00–7,325.00)	< 0.001
Fibrinogen(mg/dl)	386.14 ± 194.97	410.53 ± 195.13	354.22 ± 190.21	< 0.001
Lactate(mmol/L)	2.00 (1.30–3.20)	1.90 (1.30–2.90)	2.10 (1.40–3.70)	< 0.001
LVEF(%)	54.54 ± 11.03	57.72 ± 9.04	50.38 ± 11.98	< 0.001
APACHE II Score	19.42 ± 7.49	18.95 ± 7.19	20.04 ± 7.83	0.002
LMWH Use, n(%)	130.00 (6.91%)	89.00 (8.34%)	41.00 (5.03%)	0.005
Vasopressor Use, n(%)	1,196.00 (63.55%)	625.00 (58.58%)	571.00 (70.06%)	< 0.001
Glucocorticoid Use, n(%)	505.00 (26.83%)	306.00 (28.68%)	199.00 (24.42%)	0.039
Antibiotic therapy, n(%)	1,745.00 (92.72%)	994.00 (93.16%)	751.00 (92.15%)	0.403
CRRT, n(%)	140.00 (7.44%)	69.00 (6.47%)	71.00 (8.71%)	0.066
ICU LOS(days)	3.82 (1.94–9.64)	3.84 (1.94–10.16)	3.78 (1.93–8.75)	0.249
Hospital Mortality(%)	483.00 (25.66%)	232.00 (21.74%)	251.00 (30.80%)	< 0.001

HTN (Hypertension), CI ( Cerebrovascular Disease, ), DM (Diabetes Mellitus); CHD (Coronary Heart Disease); WBC (White Blood Cell count); RBC (Red Blood Cell count); RDW-SD (Red Cell Distribution Width - Standard Deviation); RDW-CV (Red Cell Distribution Width - Coefficient of Variation, ); ALT (Alanine Aminotransferase); AST (Aspartate Aminotransferase); CK (Creatine Kinase, ); CK-MB (Creatine Kinase-MB isoenzyme); NT-proBNP (N-terminal pro-B-type Natriuretic Peptide); LVEF (Left Ventricular Ejection Fraction), APACHE II Score (Acute Physiology and Chronic Health Evaluation II score), LMWH use (Low Molecular Weight Heparin use, ); CRRT (Continuous Renal Replacement Therapy); ICU LOS(ICU Length of Stay)

Feature selection

Feature selection via the Boruta algorithm identified 39 Z-score-ranked variables most closely linked to SIMI occurrence, including the top 10 variables: CHD, ALT, AST, CK, CK-MB, D-dimer, fibrinogen, lactate, LVEF, and NT-proBNP. Additionally, to further optimize the predictive performance of the model and based on the clinical experience of two senior physicians. APACHE II score, Total Bilirubin, Hemoglobin, and RDW-SD were also selected (Fig. 2).

Fig. 2.

Feature selection based on the Boruta algorithm. The horizontal axis denotes the name of each variable, while the vertical axis illustrates the Z-value corresponding to each variable. The box plot displays the Z-values for every variable throughout the model computation. HTN (Hypertension), CI (Cerebrovascular Disease), DM (Diabetes Mellitus); CHD (Coronary Heart Disease); WBC (White Blood Cell count); RBC (Red Blood Cell count); RDW-SD (Red Cell Distribution Width - Standard Deviation); RDW-CV (Red Cell Distribution Width - Coefficient of Variation); ALT (Alanine Aminotransferase); AST (Aspartate Aminotransferase); CK (Creatine Kinase); CK-MB (Creatine Kinase-MB isoenzyme); NT-proBNP (N-terminal pro-B-type Natriuretic Peptide); LVEF (Left Ventricular Ejection Fraction), APACHE II Score (Acute Physiology and Chronic Health Evaluation II score), LMWH use (Low Molecular Weight Heparin use); CRRT (Continuous Renal Replacement Therapy); ICU LOS(ICU Length of Stay)

Model performance comparisons

Among all models, the CatBoost model demonstrated optimal overall performance, outperforming other models across multiple core metrics. It achieved the highest accuracy (0.8690 vs. mean 0.7858) and AUROC (0.9215), substantially exceeding the second-best GBDT (AUROC = 0.8576) and the mean AUROC (0.8350). Notably, CatBoost yielded the highest F1-score (0.8452) and MCC (0.7368, robust for imbalanced datasets with SIMI prevalence 0.4496), significantly outperforming the second-ranked AdaBoost (MCC = 0.6445) and the mean MCC (0.5702). CatBoost also showed excellent precision (0.9018) and specificity (0.9293), trailing only Random Forest. It maintained high recall (0.7953) and low FNR (0.2047), ensuring reliable positive case identification. Further highlighting CatBoost’s superiority in SIMI prediction (Table 2).

Table 2.

Performance metrics of different machine learning models for outcome prediction

ModelName	Accuracy	Prevalence	Recall	F1-Score	MCC	AUROC	Presicion	Specificity	FNR
DesicionTreeTEST	0.7327	0.4496	0.6024	0.6696	0.4578	0.7977	0.7537	0.8392	0.3976
GBDTTEST	0.8124	0.4496	0.8189	0.7969	0.6236	0.8576	0.7761	0.8071	0.1811
AdaBoostTEST	0.8230	0.4496	0.8268	0.8077	0.6445	0.8234	0.7895	0.8199	0.1732
LogisticTEST	0.7062	0.4496	0.4921	0.6010	0.4105	0.7820	0.7716	0.8810	0.5079
RFTEST	0.7717	0.4496	0.5945	0.7007	0.5479	0.8279	0.8531	0.9164	0.4055
CatBoostTEST	0.8690	0.4496	0.7953	0.8452	0.7368	0.9215	0.9018	0.9293	0.2047
mean_scores	0.7858	0.4496	0.6883	0.7368	0.5702	0.8350	0.8076	0.8655	0.3117

MCC(Matthews Correlation Coefficient): AUROC(Area Under ROC Curve): FNR(False Negative Rate)

Comprehensive performance evaluations further confirmed the superiority of the CatBoost model: it had the largest AUROC (0.9215), significantly exceeding other models (Fig. 3A); the model provided the highest net benefit across a wide range of risk thresholds, with an optimal threshold of 0.446 (Fig. 3B); it exhibited the highest AUC-PR (0.901), reflecting superior balance between precision and recall (Fig. 3C); and its predicted probabilities were most consistent with actual outcomes (calibration error = 0.102), outperforming other models (Fig. 3D).

Fig. 3.

Comprehensive performance evaluation of machine learning models for SIMI outcome prediction. (A) ROC curves; (B) Decision curves; (C) Precision-recall curves; (D) Calibration curves

Internal validation

To ensure the robustness and generalizability of the CatBoost model for predicting sepsis-induced myocardial injury (SIMI) outcomes, the mean accuracy of the CatBoost model across 5 folds was 0.8527, ranging from 0.8191 to 0.8856. For discriminative ability, the mean area under the receiver operating characteristic curve (AUROC) was 0.9183, with fold-specific values between 0.9031 and 0.9285. In terms of class-balanced performance, the mean Matthews Correlation Coefficient (MCC) was 0.6994 (range: 0.6345–0.7597), and the mean F1-score was 0.8206 (range: 0.7821–0.8532).Regarding sensitivity (recall) and specificity, the model achieved a mean recall of 0.7894 (range: 0.7262–0.8446) and a mean specificity of 0.9033 (range: 0.8686–0.9282), with a mean false negative rate (FNR) of 0.2106 (range: 0.1554–0.2738). The prevalence of SIMI across the 5 folds was relatively consistent, with a mean of 0.4287 (range: 0.3723–0.4814) (Table 3).

Table 3.

Cross-validation performance metrics of the CatBoost model for SIMI outcome prediction

ModelName	Accuracy	Prevalence	Recall	F1-Score	MCC	AUROC	Presicion	Specificity	FNR
CatBoost 1TEST	0.8511	0.4814	0.7680	0.8323	0.7081	0.9261	0.9085	0.9282	0.2320
CatBoost 2TEST	0.8511	0.3723	0.8214	0.8042	0.6845	0.9102	0.7877	0.8686	0.1786
CatBoost 3TEST	0.8564	0.4495	0.7870	0.8313	0.7103	0.9285	0.8808	0.9130	0.2130
CatBoost 4TEST	0.8856	0.3936	0.8446	0.8532	0.7597	0.9238	0.8621	0.9123	0.1554
CatBoost 5TEST	0.8191	0.4468	0.7262	0.7821	0.6345	0.9031	0.8472	0.8942	0.2738
mean_scores	0.8527	0.4287	0.7894	0.8206	0.6994	0.9183	0.8573	0.9033	0.2106

MCC (Matthews Correlation Coefficient): AUROC (Area Under the Receiver Operating Characteristic Curve): FNR (False Negative Rate)

The internal Cross-Validation visualizations supported these findings: AUROC values were consistent across all folds (0.9031–0.9285) with minimal fluctuations (Fig. 4A); the model consistently provided higher net benefit than other strategies across risk thresholds (Fig. 4B); AUC-PR performance remained stable across folds, indicating reliable precision-recall balance (Fig. 4C); and the model’s predicted probabilities aligned closely with actual outcomes in each fold, confirming robust calibration (Fig. 4D).

Fig. 4.

Cross-validation performance of the CatBoost model for SAE outcome prediction. (A) ROC curves; (B) Decision curves; (C) Precision-recall curves; (D) Calibration curves. ROC (Receiver Operating Characteristic); DCA (Decision Curve Analysis); PR (Precision-Recall)

External validation

To further verify the generalizability of the model for SIMI outcome prediction, we conducted external validation using an independent cohort. The model achieved an accuracy of 0.6972, a prevalence of 0.3904, and a recall of 0.4588. The F1-score was 0.5309, and the MCC was 0.3375, indicating a moderate correlation between predicted and actual outcomes. In terms of discriminative ability, the AUROC was 0.7567. Precision was 0.6719, specificity was 0.8827 (Table 4).

Table 4.

Presents the performance metrics of the CatBoost model in the external validation cohort

ModelName	Accuracy	Prevalence	Recall	F1-Score	MCC	AUROC	Presicion	Specificity	FNR
CatBoostTEST	0.6972	0.3904	0.4388	0.5309	0.3375	0.7567	0.6719	0.8627	0.5612

The model exhibited discriminative ability in external validation, with an area under the ROC curve (AUROC) of 0.7567 and a 95% confidence interval (95% CI) of 0.6944–0.8191. Additionally, the optimal threshold probability for SIMI prediction was determined to be 0.429, corresponding to a sensitivity of 0.712 and a specificity of 0.714 (Fig. 5).

Fig. 5.

The ROC curve for the external validation cohort. ROC (Receiver Operating Characteristic)

Interpretability analysis

To interpret the CatBoost model’s decision-making process and identify clinically meaningful predictors of sepsis-induced myocardial injury (SIMI), we performed SHapley Additive exPlanations (SHAP) analysis. The SHAP values quantify the direction and magnitude of feature impacts, with positive values increasing SIMI risk and negative values reducing it. Results indicated that creatine kinase-MB (CK-MB), D-dimer, and N-terminal pro-B-type natriuretic peptide (NT-proBNP) were the top three risk factors for SIMI (Fig. 6).

Fig. 6.

SHAP Analysis of Feature Contribution in sepsis-induced myocardial injury Prediction. SHAP (SHapley Additive Explanations); CK-MB (Creatine Kinase-MB Isoenzyme); NT-proBNP(N-terminal Pro-B-type Natriuretic Peptide); LVEF (Left Ventricular Ejection Fraction); CHD (Coronary Heart Diseases)

Discussion

Sepsis-induced myocardial injury (SIMI) is a life-threatening complication of sepsis, characterized by complex pathogenesis and high mortality [20]. Its early identification and precise risk stratification have long been key challenges in critical care medicine [21]. Traditional prognostic evaluation tools such as APACHE II and SOFA scores, while widely used in clinical practice, are constrained by their linear modeling frameworks and inadequate capture of the nonlinear interactions between multiple clinical variables in sepsis [22]. This limitation often results in insufficient accuracy in predicting SIMI, failing to meet the clinical demand for personalized risk assessment. Against this backdrop, the application of machine learning algorithms in medical prediction models has opened up new avenues for addressing this unmet need, leveraging their inherent advantages in handling high-dimensional, heterogeneous clinical data.

In this study, six mainstream machine learning models were constructed and compared, with the CatBoost model demonstrating superior comprehensive performance, consistent with prior research. The model achieved an AUROC of 0.9215 in the training set, accompanied by high accuracy (0.8690), F1-score (0.8452), and MCC (0.7368). These metrics not only outperformed traditional logistic regression but also surpassed other ensemble learning models such as GBDT and AdaBoost. The superiority of the CatBoost model can be attributed to its unique algorithmic design: it employs ordered boosting to mitigate prediction bias caused by categorical feature ordering, effectively handles missing values without relying on complex preprocessing, and adaptively captures nonlinear relationships between clinical indicators and SIMI occurrence. Notably, the 5-fold cross-validation yielded a mean AUROC of 0.9183, with minimal fluctuations across folds, confirming the model’s robust stability and resistance to overfitting. In external validation using an independent cohort from Zunyi First People’s Hospital, the model maintained an AUROC of 0.7567 (95% CI: 0.6944–0.8191), the model exhibited moderate and stable generalizability in the external population, and we further analyzed the underlying reasons for the slight decline in predictive performance across different cohorts. The slight performance reduction in external validation may be partly attributed to differences in cTnT detection systems and reference thresholds between cohorts, which reflects real-world clinical heterogeneity rather than model deficiency.

The interpretability of machine learning models remains a critical barrier to their clinical adoption, as the “black-box” nature often hinders clinician trust [23, 24]. To address this, SHAP analysis was employed to decode the decision-making mechanism of the CatBoost model, quantifying the direction and magnitude of each feature’s contribution to SIMI prediction. The results identified CK-MB, D-dimer, and NT-proBNP as the top three predictive factors, which aligns with current clinical understanding and pathological mechanisms of SIMI. CK-MB, a specific marker of myocardial cell injury, is released into the bloodstream following damage to myocardial cell membranes. In sepsis, excessive inflammatory responses and oxidative stress directly induce myocardial cell necrosis, leading to elevated CK-MB levels—consistent with its status as a direct indicator of SIMI [25, 26]. D-dimer, a product of fibrinolysis, reflects the activation of the coagulation system [27]. Sepsis-induced systemic inflammatory response syndrome triggers coagulopathy, characterized by microthrombosis formation in myocardial microvasculature. This impairs myocardial perfusion, exacerbates ischemia-hypoxia, and ultimately contributes to myocardial injury, explaining the positive correlation between D-dimer elevation and SIMI risk. NT-proBNP, primarily secreted by ventricular myocytes in response to increased wall tension, is a sensitive indicator of cardiac function. Sepsis-induced hemodynamic disturbances and inflammatory mediator-induced myocardial contractile dysfunction led to elevated NT-proBNP levels, serving as an early warning signal for impending myocardial injury [28]. Beyond the core predictors (CK-MB, D-dimer, NT-proBNP), CHD, LVEF, lactate, liver enzymes, albumin, APACHE II score, and RDW-SD also contribute to SIMI risk stratification, reflecting the multi-factorial nature of the condition. Collectively, these findings not only validate the clinical relevance of the model’s predictions but also provide a biological rationale for the identified risk factors, enhancing the model’s credibility among clinicians.

Compared with existing SIMI prediction studies, this work offers several notable advancements [29]. First, in contrast to most previous studies that relied on logistic regression or single machine learning algorithms, we systematically compared six models and optimized hyperparameters via grid search and cross-validation, ensuring the selection of the most robust predictive framework. Second, the integration of SHAP analysis addressed the interpretability gap of machine learning models, enabling quantitative assessment of each feature’s impact—an advantage over traditional regression models that only provide coefficient estimates without capturing complex variable interactions. Third, the use of two independent datasets (MIMIC-IV and a local clinical cohort) for internal and external validation, respectively, enhances the model’s generalizability and reduces the risk of overfitting to a single data source. Additionally, strict inclusion/exclusion criteria were applied to the study population, excluding patients with pre-existing severe heart failure, acute myocardial infarction, and other conditions that could confound the diagnosis of SIMI, thereby improving the accuracy of the outcome definition.

Despite these strengths, several limitations remain in this study. First, although patients with acute myocardial infarction, pulmonary embolism and decompensated heart failure were strictly excluded, CK-MB, D-dimer and NT-proBNP could still be affected by subclinical cardiac or thrombotic conditions. Additionally, the retrospective design cannot establish causal relationships between risk factors and SIMI; further prospective studies are needed to verify these associations and underlying mechanisms. Second, inter-center heterogeneity in laboratory detection systems and reference ranges represents an inherent limitation of this retrospective multicenter study. Differences in assay reagents, detection platforms, and reference populations between the MIMIC-IV cohort and the local Chinese cohort may introduce systematic variations not only in cardiac troponin T but also in other biochemical indicators. Although the CatBoost algorithm exhibits inherent robustness to feature distribution shifts, and our supplementary sensitivity analysis using three uniform absolute cTnT cut-offs (0.03 ng/mL, 0.065 ng/mL, and 0.1 ng/mL) confirmed that the model’s discriminative performance ranking, core predictor set, and overall conclusion remain stable across different diagnostic thresholds, this residual measurement heterogeneity may still partially attenuate the external validation performance. Future prospective multicenter studies should adopt unified detection protocols and standardized calibration procedures to further improve the generalizability and comparability of the prediction model. Third, dynamic time-series clinical data were not included due to data limitations, while the progression of sepsis and myocardial injury is time-dependent; incorporating dynamic data may further enhance predictive performance. Fourth, some potential confounding variables were unavailable for analysis, which may have partially influenced the model’s predictive ability.

Future research directions could focus on the following aspects: First, expanding the sample size with multi-center, prospective cohorts to optimize the model and validate its performance across diverse clinical settings. Second, integrating dynamic monitoring data to develop a real-time predictive model that adapts to changes in patient condition. Third, exploring the underlying mechanisms of the identified key risk factors in SIMI pathogenesis to identify potential therapeutic targets. Fourth, developing user-friendly clinical decision support tools based on the CatBoost model to facilitate its integration into electronic medical record systems, enabling bedside risk assessment and guiding timely interventions. Finally, conducting cost-effectiveness analyses to evaluate the clinical and economic benefits of implementing the model in routine practice, such as reducing mortality, shortening ICU length of stay, and lowering healthcare costs.

Conclusion

This study successfully constructed and validated a machine learning-based SIMI prediction model, among which the CatBoost model performed optimally with high accuracy, strong discriminative ability, and good generalizability. The identification of key risk factors (CK-MB, D-dimer, NT-proBNP) provides a theoretical basis for understanding the pathogenesis of SIMI. This work not only provides a reliable tool for early risk stratification of SIMI but also deepens our understanding of the clinical and biological determinants of this complication. With further optimization and validation, this model has the potential to be translated into clinical practice, assisting clinicians in making personalized treatment decisions and ultimately improving the prognosis of sepsis patients.

Abbreviations

CI

Cerebral Infarction

CHD

Coronary Heart Disease

WBC

White Blood Cell

ALT

Alanine Aminotransferase

AST

Aspartate Aminotransferase

TBIL

Total Bilirubin

CRRT

Continuous Renal Replacement Therapy

APACHE

Acute Physiology and Chronic Health Evaluation

LOS

Length of Stay

Author contributions

SYP, XML, XRH, ZSZ, WL and ZLW contributed to the conception of the study and data’s collection. MX, YKL and DQW made statistical analysis. BTL，HM and XL completed the writing of the paper. ZRC, WL and FJD were responsible for the revision of the paper. All authors confirmed the final version of the paper.

Funding

This research was financially supported by the National Natural Science Foundation of China (Grant Nos 82260079, 82360079 and 82400468), the Guizhou Provincial Science and Technology Projects (Grant No. Guizhou Science and Technology Cooperation Support Program [2024] General 160).

Data availability

The datasets used during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This study complies with the Declaration of Helsinki and was approved by the Ethics Committee of Zunyi First People’s Hospital (approval number: 2025-1-14).

Consent to participate

Not applicable.

Transparency

The lead author (WL) affirms that the manuscript is an honest, accurate, and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies from the study as planned have been explained.

Competing interests

The authors declare no competing interests.