An ensemble machine learning-based risk stratification tool for 30-day mortality prediction in critically ill cardiovascular patients

Abstract

Background

Early mortality prediction in critically ill patients with cardiovascular disease remains challenging. This study aimed to develop and validate an ensemble machine learning (ML) model to predict 30-day mortality, comparing its performance with conventional severity scores and interrogating the incremental prognostic value of stress hyperglycemia ratio (SHR).

Methods

A retrospective cohort of 1,595 ICU patients with cardiovascular disease combined with diabetes (2008–2022) was analyzed. SHR was calculated as admission glucose divided by estimated average glucose (eAG) from HbA1c. Six ML models (eXtreme Gradient Boosting [XGBoost], Decision Tree [DT], Random Forest [RF], Artificial Neural Network [ANN], Logistic Regression [LR], and Support Vector Machine [SVM]) were trained on 80% of the data, with the top three performers combined into an ensemble model. Model performance was evaluated using area under the curve (AUC), precision-recall, calibration, and clinical utility metrics.

Results

The 30-day mortality rate was 10.8% in the entire cohort (n = 173). The ensemble model demonstrated superior predictive performance with an AUC of 0.912 (95% CI: 0.888–0.936), outperforming both individual ML models (XGBoost, AUC = 0.903) and traditional scoring systems (APS III/SOFA/SAPS II AUCs ≤ 0.742; all P < 0.001). The top six important predictors included anti-hypertensives, aspirin, blood urea nitrogen (BUN), white blood cell (WBC), age, and red blood cell (RBC), with the Shapley Additive Explanations analysis revealing clinically meaningful patterns: a nonlinear risk escalation for age, linear risk increases with rising BUN and bilirubin levels, a protective effect associated with higher RBC counts, and both low and high WBC levels linked to increased early death risk. While SHR significantly improved the performance of traditional scoring systems (e.g., increasing SOFA AUC from 0.741 to 0.757, P = 0.010), its addition to the ensemble model provided limited incremental benefit (ΔAUC = - 0.032, P = 0.094). External validation in an independent cohort (n = 307) confirmed the model’s robustness (AUC = 0.891, 95% CI: 0.864–0.917), with decision curve analysis demonstrating superior clinical utility across a wide range of risk thresholds.

Conclusions

The ensemble ML model outperformed conventional prognostic tools in predicting 30-day mortality, with SHR augmenting traditional tools but not the ensemble ML model. This approach offers a reliable, interpretable framework for risk stratification in high-risk cardiovascular patients.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12933-025-02911-5.

Introduction

Cardiovascular disease remains the leading cause of death worldwide, accounting for nearly 18 million deaths annually and imposing a massive burden on healthcare systems [1]. Despite advances in prevention and treatment, the global cardiovascular disease epidemic continues to escalate, driven by aging populations, rising metabolic risk factors, and socioeconomic disparities [1, 2]. Recent projections estimate that between 2025 and 2050, the crude mortality rate from cardiovascular disease will increase by 73.4%, with total deaths expected to surge from 20.5 million in 2025 to 35.6 million by 2050 [1].

In addition, diabetes mellitus has emerged as a global epidemic, affecting over 500 million adults worldwide, with projections indicating a rise to 693 million by 2045 [3]. The disease poses a significant public health burden due to its association with life-threatening complications [4, 5], particularly cardiovascular disease. Among diabetic patients, cardiovascular disease is also the leading cause of mortality [6], approximately accounting for 50% of deaths. Studies indicate that diabetic individuals face a 2- to 4-fold increased risk of developing cardiovascular disease compared to non-diabetic populations and the risk grows with worsening control of blood glucose [7], with more than 20% of diabetic patients suffering from concurrent cardiovascular conditions [8]. Patients with both diabetes and cardiovascular disease face a sevenfold elevated risk of mortality compared to those without cardiovascular disease [9]. The synergistic effects of hyperglycemia, insulin resistance, and systemic inflammation exacerbate vascular damage, leading to adverse outcomes such as myocardial infarction, stroke, and heart failure [7]. Given the high prevalence and mortality risk, early identification of at-risk diabetic cardiovascular patients is crucial for improving clinical outcomes.

Current research has identified multiple risk factors contributing to mortality in diabetic or cardiovascular patients, such as glycated haemoglobin A1c [10], prognostic nutritional index [11], renal dysfunction, and hematological abnormalities [12, 13]. Among these, the stress hyperglycemia ratio (SHR), a marker of acute glycemic dysregulation, has gained attention for its favorable association with adverse diabetic [14] or cardiovascular outcomes [15–20]. However, despite the known prognostic value of SHR, existing risk prediction models—such as Sequential Organ Failure Assessment (SOFA), Simplified Acute Physiology Score II (SAPS II), and Charlson Comorbidity Index—were not specifically designed for this high-risk subgroup and often lack accuracy in predicting short-term mortality. Traditional scoring systems fail to incorporate comprehensive metabolic variables and machine learning (ML)-driven insights, limiting their clinical utility. In addition, while ML-based predictive models have been developed for cardiovascular populations, such as patients treated with cardiac surgery [21], heart failure with hypertension [22], acute myocardial infarction [23], critically ill atrial fibrillation [24, 25], atherosclerotic cardiovascular disease [26], coronary heart disease [27–29], and life-threatening ventricular arrhythmias [30], none have been optimized for the intersection of cardiovascular and diabetic populations, leaving a critical gap in personalized risk assessment. Furthermore, the predictive performance of these models remains limited, and most lack external validation, thus their generalizability requires further verification.

Therefore, to address this unmet need, this study hypothesized that an interpretable ML ensemble model incorporating multidimensional clinical features could outperform conventional risk scores in predicting 30-day mortality among critically ill diabetic cardiovascular patients. This study aimed to (1) develop and validate an ML-based predictive model tailored to this population, (2) identify key mortality predictors through advanced feature analysis, and (3) evaluate the incremental prognostic value of SHR within the ensemble framework. By leveraging explainable artificial intelligence techniques, we further sought to provide clinically actionable insights into risk stratification and early intervention strategies.

Patients and methods

Study design and patient selection

This retrospective observational cohort study analyzed de-identified clinical data from adult patients ( ≧ 18 years) admitted to the intensive care unit (ICU) of Beth Israel Deaconess Medical Center between 2008 and 2022. The dataset comprised comprehensive clinical records, including demographics, laboratory results, medication histories, and mortality outcomes [31]. Access to the database was granted after completing mandatory training, including the National Institutes of Health certification and the Collaborative Institutional Training Initiative program.

Patients with a primary diagnosis of cardiovascular disease, as classified by ICD-9 and ICD-10 codes (detailed in Supplementary Materials Table S1), were initially screened. Exclusion criteria consisted of: (1) discharged or died within 24 h of ICU admission, (2) missing admission glucose or HbA1c measurements necessary for stress hyperglycemia ratio (SHR) calculation; (3) incomplete mortality follow-up data; and (4) absence of diabetes mellitus. After applying these criteria, 1,595 patients were eligible for analysis. The study cohort was randomly divided into a derivation set (80% of patients) for model development and an internal validation set (20%). Specifically, we employed train_test_split in Python (test_size = 0.2, stratified by outcome, random_state = 0) to create mutually exclusive derivation (80%) and internal validation (20%) sets, confirming no patient overlap between sets. Additionally, external validation was performed using an independent patient cohort (n = 307) from the Chinese PLA General Hospital. The selection process and study design are illustrated in Fig. 1. This study was approved by the Institutional Review Board of the Chinese PLA General Hospital. To comply with ethical standards, all personally identifiable information was anonymized prior to analysis. The research adhered to the principles of the Declaration of Helsinki and followed the STROCSS reporting guidelines [32, 33].

Fig. 1.

Study design and patient selection flowchart. It illustrates the patient inclusion and exclusion process, derivation and validation cohort splits, and external validation cohort selection

Collection of clinical baseline characteristics

Clinical baseline characteristics were extracted and categorized into demographic, comorbidity, treatment, and laboratory parameters. Demographic data included gender, age, height, and weight. Comorbidities and clinical conditions comprised smoking status, acute kidney injury (AKI), sepsis, congestive heart failure, cerebrovascular disease, dementia, chronic pulmonary disease, and invasive ventilation use. Medication history included statins, aspirin, anti-hypertensives, and glucocorticoids. Laboratory measurements encompassed albumin, total bilirubin, white blood cell (WBC) count, red cell distribution width (RDW), platelet count, red blood cell (RBD) count, creatinine, blood urea nitrogen (BUN), total calcium, international normalized ratio (INR), glucose, potassium, sodium, hemoglobin, Hemoglobin A1c (HbA1c), and stress hyperglycemia ratio (SHR). The proportion of missing values for each variable is detailed in Supplementary Materials Table S2. All variables had a missingness rate below 30%, and 26 out of 32 variables (81.3%) had no missing data. For imputation, we employed k-nearest neighbors (k = 5) due to its ability to preserve complex relationships between variables while being robust to outliers—particularly advantageous feature for machine learning applications. This method generates imputations based on similar patient profiles rather than simple mean/median substitution, thereby maintaining the underlying data structure. In addition, we incorporated 20 auxiliary variables to strengthen the imputation model and mitigate potential bias.

Calculation of SHR

To quantify acute glycemic stress, SHR was derived from the ratio of admission blood glucose to estimated average glucose (eAG), the latter reflecting chronic glycemic control based on HbA1c levels. This approach has been validated in prior studies as a reliable marker of acute hyperglycemic response [34, 35]. SHR was computed as:

1

where eAG was determined using the following formula:

2

To rigorously assess the impact of SHR on predictive performance, this study conducted comparative analyses by systematically evaluating its inclusion or exclusion in both our machine learning-based prediction model and conventional severity scores (e.g., SOFA, SAPS II, Charlson, and OASIS). This approach allowed us to determine whether SHR contributed meaningfully to the predictive accuracy of these models for 30-day mortality risk stratification.

Definition of primary outcome

The primary outcome, early death, was defined as all-cause mortality within 30 days of hospital admission, with mortality data rigorously extracted from both hospital records and linked Social Security Administration (SSA) death records via the Medical Information Mart for Intensive Care IV (MIMIC-IV) database to ensure comprehensive capture of in-hospital and post-discharge events. This 30-day endpoint was selected for its clinical relevance in acute care outcomes, and any discrepancies between data sources were resolved by prioritizing the earliest recorded death date; patients without 30-day follow-up data were censored in secondary analyses to mitigate attrition bias.

Feature selection and data engineering

To identify robust predictive features while mitigating overfitting, we employed least absolute shrinkage and selection operator (LASSO) regression method, which performs simultaneous variable selection and regularization by shrinking irrelevant coefficients to zero [36, 37], particularly advantageous for high-dimensional clinical datasets. The selected non-zero-coefficient predictors were consistently used across all ML models to ensure comparability. The dataset was randomly divided into training (80%) and internal validation (20%) groups with stratified sampling to maintain outcome distribution, and next SMOTE-Tomek resampling was performed in both groups to address class imbalance through combined Synthetic Minority Oversampling Technique with Tomek Links Undersampling [38]. To elaborate, in our study, the original dataset exhibited significant class imbalance, with 30-day mortality (the minority class) occurring in 10.8% of cases (173/1,595), while survivors (the majority class) accounted for 89.2%. To ensure balanced representation, we first stratified the cohort into training (n = 1,276; 10.8% mortality) and internal validation (n = 319; 11.0% mortality) sets at an 8:2 ratio while preserving outcome proportions. Subsequently, we applied SMOTE to adjust the class distribution to an equal 50:50 ratio in both subsets, resulting in a training set of 2,252 samples (1,126 events and 1,126 non-events) and a validation set of 550 samples (275 events and 275 non-events). This balanced approach was chosen to mitigate classifier bias toward the majority class, enhance model sensitivity to the clinically critical mortality outcome, and align with established practices in clinical prediction. Additionally, the preprocessing pipeline employed a systematic approach to handle both categorical and continuous variables. For categorical variables, we implemented a dedicated pipeline comprising variable selection followed by binary encoding. Continuous variables were processed through a separate pipeline involving feature selection and standardization using z-score normalization. These parallel pipelines were integrated through a Feature Union operation, which concatenated the transformed outputs while preserving the original variable structure. The final preprocessing step generated a unified feature matrix suitable for ML algorithms, with categorical variables represented as binary indicators and continuous variables scaled to zero mean and unit variance.

Model development and hyperparameter optimization

We implemented six distinct ML algorithms: eXtreme Gradient Boosting (XGBoost), Decision Trees (DT), Random Forests (RF), Artificial Neural Networks (ANN), Logistic Regression (LR), and Support Vector Machines (SVM). Hyperparameter optimization was conducted through an exhaustive strategy combining grid search with randomized search (100 iterations, 5-fold cross-validation), using the receiver operating characteristic area under the curve (ROC-AUC) as the primary optimization metric. The search space encompassed algorithm-specific parameters including learning rate (0.01–0.3), tree depth (3–15), regularization terms (L1/L2 norms), and kernel coefficients, systematically evaluated through repeated stratified k-fold validation (n = 100 shuffles) to ensure robust generalization. In addition, learning curve analysis was conducted across different training-test set ratios to evaluate cross-validation score convergence and identify potential overfitting or underfitting scenarios.

Model performance quantification

A comprehensive evaluation framework was employed to assess model performance across multiple dimensions. Discriminative ability was measured using ROC-AUC with 100 bootstrap resamples for confidence interval estimation and discrimination slope. Standard classification metrics—including accuracy, precision, recall, and F1-score—were derived from confusion matrices to evaluate predictive performance. Probabilistic calibration was assessed using Brier score (measuring prediction accuracy) and log loss (evaluating probabilistic prediction quality), alongside calibration slope and intercept analysis to verify the alignment between predicted probabilities and observed outcomes. The formulas for Brier score and log loss are outlined as follows:

3

4

where = Total number of samples, = Predicted probability of the -th sample belonging to the positive class (0 ≤≤ 1), and = True label of the -th sample ( where 0 = negative class and 1 = positive class). To avoid numerical instability when or apply probability smoothing (e.g., clip to [ε,1−ε]).

Further validation included precision-recall curve analysis to assess model performance at varying risk stratification levels, probability density curve evaluation to examine predicted risk distributions across outcome classes, and clinical decision curve analysis to quantify clinical utility. The decision curve analysis framework calculated net benefit across varying decision thresholds, balancing true positives against false positives to identify optimal cutoffs for clinical decision-making.

Ensemble model development and comparative performance analysis

In this study, we developed an ensemble model based on a comprehensive scoring system that incorporates ten evaluation metrics [39, 40]: accuracy, precision, recall, F1 score, AUC, brier score, log loss, intercept-in-large, calibration slope, and discrimination slope. For each of the six ML models assessed, a maximum score of 6 points was allocated for optimal performance in each metric, with subsequent scores of 5 and so forth down to 0 for the least favorable outcomes. The total model score represents the cumulative points from these ten metrics, with higher scores indicating better predictive performance. Based on the total scores, we ranked the models and selected the top three for the construction of the ensemble model. The ensemble model was implemented using the VotingClassifier, which combines the predictions of the chosen models: LR, DT, and XGBoost. The soft voting method was employed to enhance prediction accuracy. Subsequently, the ensemble model was trained using the training dataset. In the ensemble model, particularly in the context of a Voting Classifier, the formulation of the ensemble model can be expressed as follows:

5

where ​is the final prediction outcome of the ensemble model, represents the class label, is the number of base models participating in the voting process, is the weight of the -th base model (in the case of soft voting, the weights can be the probability predictions of the models), denotes the predicted probability of the -th base model for class In the soft voting framework, this formula aggregates the predicted probabilities from each base model, weighted according to their significance, and selects the class with the maximum probabilistic outcome as the final decision. This ensemble approach effectively combines the strengths of multiple models, thereby enhancing overall predictive performance.

Our analysis of prediction performance included the calculation of the same ten evaluation metrics for the ensemble model, followed by the visualization of the ROC-AUC curve, probability density curve, calibration curve, and clinical decision curve. These results were compared with those of the base estimators to provide a comprehensive evaluation of the ensemble’s performance relative to its constituent models. Additionally, we assessed the predictive performance of the ensemble model in comparison to existing traditional scoring systems, including Acute Physiology Score III (APS III), SAPS II, SOFA, Charlson Index, and Oxford Acute Severity of Illness Score (OASIS). This comparison employed the DeLong test to statistically analyze the differences in predictive efficacy between the ensemble model and traditional scores. Furthermore, to investigate the impact of the variable SHR in the models, we conducted an analysis of the ensemble with and without the SHR variable to examine how this exclusion affects the model’s prediction capabilities. This thorough approach highlights the effectiveness of the ensemble model while providing insights into the significance of SHR in enhancing predictive accuracy.

Feature importance

To identify the critical factors influencing early death and improve interpretability, this study performed an extensive analysis of feature importance among all the ML algorithms applied in this research. The importance rankings for the features were obtained through the inherent metrics of each algorithm. Moreover, we employed the Shapley Additive Explanations (SHAP) method where suitable, providing an enriched understanding of the influence of individual features. The SHAP value () for the feature is calculated as follows:

6

In this equation, represents the complete set of features, while denotes any subset of features that does not include The term indicates the model’s prediction based solely on the features in subset and indicates the factorial of the size of subset size, ensuring equitable treatment of permutations in weight assignments.

SHAP is rooted in cooperative game theory and offers a robust mathematical framework for attributing feature significance. This method calculates the incremental contribution of each feature to the model’s predictions by assessing how the model’s output changes with the inclusion of that feature, focusing on the log odds ratios while considering all potential feature interactions. The SHAP values thus serve as a comprehensive metric that quantifies the influence of each feature against a baseline expectation, allowing for direct comparisons of feature significance. For each ML model analyzed, we generated variable importance rankings. By aggregating the importance scores of individual features across all six models, we established an overall ranking for each feature that reflects its importance within the ensemble. As a result, features that rank higher are considered more impactful, underlining their essential roles in shaping the outcome prediction and enhancing the interpretability of the models.

Statistical analysis

Continuous variables were expressed as mean ± standard deviation (SD), while categorical variables were summarized as frequencies and percentages. Group differences were evaluated using Student’s t-tests or Mann-Whitney U tests (for non-normally distributed continuous variables) and χ² tests (for categorical variables). Multicollinearity among model features was assessed using Spearman’s rank correlation, with an absolute coefficient (|r| >0.8) indicating significant collinearity. To further examine the association between model features and early mortality, restricted cubic spline (RCS) analysis was employed. Receiver operating characteristic (ROC) curve comparisons between models were conducted using DeLong’s test. All ML implementations, including hyperparameter optimization, were performed in Python (v3.9.7) with scikit-learn (v1.2.2), whereas traditional statistical analyses were conducted in R (v4.1.2). The complete code for data processing, modeling, and validation is uploaded to GitHub (accessible at https://github.com/Starxueshu/30_day-mortality-prediction-among-cardiovascular-patients/tree/main). A two-sided P-value of less than 0.05 was considered statistically significant for all analyses.

Results

Baseline clinical characteristics

The study included 1,595 patients (66.5% male, mean age 68.9 ± 11.2 years) with cardiovascular disease and diabetes (Table 1), demonstrating a high comorbidity burden: 84.1% had AKI, 57.9% congestive heart failure, 38.2% sepsis, and 59.1% required invasive ventilation. Most patients received aspirin (92.1%) and anti-hypertensives (93.2%), while statin use was lower (20.9%). Laboratory findings revealed hyperglycemia (glucose 204.9 ± 111.8 mg/dL; HbA1c 7.75 ± 1.93%), renal impairment (creatinine 1.77 ± 1.80 mg/dL; BUN 29.8 ± 19.9 mg/dL), anemia (hemoglobin 11.7 ± 2.2 g/dL), and elevated inflammatory markers (WBC 10.68 ± 8.12 × 10¹²/L), consistent with a metabolically unstable, critically ill population.

Table 1.

Baseline clinical characteristics among patients with cardiovascular disease and diabetes

Characteristics	Overall	Early death		P
		No	Yes
n	1595	1422	173
Gender (male/female, %)	1061/534 (66.5/33.5)	947/475 (66.6/33.4)	114/59 (65.9/34.1)	0.921
Age (years, mean (SD))	68.92 (11.18)	68.34 (11.00)	73.70 (11.49)	< 0.001
Height (cm, mean (SD))	170.61 (9.32)	170.69 (9.43)	170.01 (8.39)	0.366
Weight (kg, mean (SD))	86.66 (20.83)	86.90 (20.53)	84.66 (23.06)	0.181
Smoker (no/yes, %)	1515/80 (95.0/5.0)	1347/75 (94.7/5.3)	168/5 (97.1/2.9)	0.241
AKI (no/yes, %)	254/1341 (15.9/84.1)	243/1179 (17.1/82.9)	11/162 (6.4/93.6)	< 0.001
Sepsis (no/yes, %)	986/609 (61.8/38.2)	915/507 (64.3/35.7)	71/102 (41.0/59.0)	< 0.001
Congestive heart failure (no/yes, %)	671/924 (42.1/57.9)	628/794 (44.2/55.8)	43/130 (24.9/75.1)	< 0.001
Cerebrovascular disease (no/yes, %)	1266/329 (79.4/20.6)	1166/256 (82.0/18.0)	100/73 (57.8/42.2)	< 0.001
Dementia (no/yes, %)	1553/42 (97.4/2.6)	1391/31 (97.8/2.2)	162/11 (93.6/6.4)	0.003
Chronic pulmonary disease (no/yes, %)	1223/372 (76.7/23.3)	1101/321 (77.4/22.6)	122/51 (70.5/29.5)	0.053
Invasive ventilation (no/yes, %)	652/943 (40.9/59.1)	599/823 (42.1/57.9)	53/120 (30.6/69.4)	0.005
Statins (no/yes, %)	1262/333 (79.1/20.9)	1112/310 (78.2/21.8)	150/23 (86.7/13.3)	0.012
Aspirin (no/yes, %)	126/1469 (7.9/92.1)	97/1325 (6.8/93.2)	29/144 (16.8/83.2)	< 0.001
Anti-hypertensives (no/yes, %)	108/1487 (6.8/93.2)	74/1348 (5.2/94.8)	34/139 (19.7/80.3)	< 0.001
Glucocorticoids (no/yes, %)	1412/183 (88.5/11.5)	1276/146 (89.7/10.3)	136/37 (78.6/21.4)	< 0.001
Albumin (g/dL, mean (SD))	3.58 (0.48)	3.62 (0.46)	3.30 (0.54)	< 0.001
Total bilirubin (mg/dL, mean (SD))	0.65 (0.56)	0.63 (0.50)	0.83 (0.88)	< 0.001
WBC (×1012/L, mean (SD))	10.68 (8.12)	10.28 (7.80)	14.01 (9.81)	< 0.001
RDW (%, mean (SD))	14.37 (1.96)	14.28 (1.90)	15.13 (2.27)	< 0.001
Platelet (×109/L, mean (SD))	224.28 (84.65)	225.69 (85.03)	212.65 (80.74)	0.056
RBC (×1012/L, mean (SD))	3.97 (0.75)	4.01 (0.74)	3.61 (0.78)	< 0.001
Creatinine (mg/dL, mean (SD))	1.77 (1.80)	1.68 (1.67)	2.55 (2.53)	< 0.001
BUN (mg/dL, mean (SD))	29.83 (19.89)	28.21 (18.25)	43.08 (26.74)	< 0.001
Total calcium (mg/dL, mean (SD))	8.82 (0.67)	8.83 (0.64)	8.74 (0.84)	0.118
INR (mean (SD))	1.27 (0.50)	1.25 (0.48)	1.46 (0.64)	< 0.001
Glucose (mg/dL, mean (SD))	204.90 (111.84)	203.12 (112.44)	219.57 (105.96)	0.068
Potassium (mEq/L, mean (SD))	4.31 (0.68)	4.28 (0.64)	4.58 (0.87)	< 0.001
Sodium (mEq/L, mean (SD))	137.63 (4.35)	137.71 (4.29)	136.97 (4.76)	0.032
Hemoglobin (g/dL, mean (SD))	11.71 (2.21)	11.83 (2.17)	10.65 (2.24)	< 0.001
HbA1c (%, mean (SD))	7.75 (1.93)	7.79 (1.93)	7.46 (1.83)	0.032
SHR (mean (SD))	1.19 (0.56)	1.17 (0.55)	1.36 (0.66)	< 0.001

SD, Standard Deviation; AKI, Acute Kidney Injury; WBC, White Blood Cell; RDW, Red Cell Distribution Width; RBC, Red Blood Cell count; BUN, Blood Urea Nitrogen; INR, International Normalized Ratio; HbA1c, Hemoglobin A1c; SHR, Stress Hyperglycemia Ratio.

Risk factors associated with early mortality

Subgroup analysis stratified by early mortality status revealed significant differences in multiple clinical parameters (Table 1). In detail, patients with early death were older (73.70 ± 11.49 vs. 68.34 ± 11.00 years, P < 0.001) and had higher rates of AKI (93.6% vs. 82.9%, P < 0.001), sepsis (59.0% vs. 35.7%, P < 0.001), congestive heart failure (75.1% vs. 55.8%, P < 0.001), cerebrovascular disease (42.2% vs. 18.0%, P < 0.001), and dementia (6.4% vs. 2.2%, P = 0.003). Significant treatment differences included lower statin (13.3% vs. 21.8%, P = 0.012) and glucocorticoid use (21.4% vs. 10.3%, P < 0.001), as well as reduced aspirin (83.2% vs. 93.2%, P < 0.001) and anti-hypertensive adherence (80.3% vs. 94.8%, P < 0.001). Laboratory abnormalities associated with early mortality included lower albumin (3.30 ± 0.54 vs. 3.62 ± 0.46 g/dL, P < 0.001), higher bilirubin (0.83 ± 0.88 vs. 0.63 ± 0.50 mg/dL, P < 0.001), leukocytosis (14.01 ± 9.81 vs. 10.28 ± 7.80 × 10¹²/L, P < 0.001), elevated RDW (15.13% ± 2.27% vs. 14.28% ± 1.90%, P < 0.001), anemia (10.65 ± 2.24 vs. 11.83 ± 2.17 g/dL, P < 0.001), worse renal function (creatinine 2.55 ± 2.53 vs. 1.68 ± 1.67 mg/dL, P < 0.001; BUN 43.08 ± 26.74 vs. 28.21 ± 18.25 mg/dL, P < 0.001), prolonged INR (1.46 ± 0.64 vs. 1.25 ± 0.48, P < 0.001), and higher SHR (1.36 ± 0.66 vs. 1.17 ± 0.55, P < 0.001). No significant differences were found in gender (P = 0.921), height (P = 0.366), weight (P = 0.181), smoking status (P = 0.241), platelet count (P = 0.056), total calcium (P = 0.118), or glucose levels (P = 0.068).

Identification of model features

The LASSO regression analysis identified 14 variables with non-zero coefficients for inclusion in the final predictive model (Supplementary Materials Figs. S1, S2): demographic factors (age, LASSO coefficient = 0.018), comorbidities (sepsis = 0.151, cerebrovascular disease = 0.708), interventions (invasive ventilation = 0.144, aspirin = 0.052, anti-hypertensives = 0.742), and laboratory parameters (total bilirubin = 0.055, WBC = 0.012, RDW = 0.001, RBC = -0.041, BUN = 0.016, INR = 0.027, potassium = 0.083) (Supplementary Materials Table S3). Notably, anti-hypertensives and cerebrovascular disease demonstrated the strongest positive associations, while RBC exhibited the most substantial negative association, suggesting that higher RBC levels were a protective factor against early death. This selection underscores the importance of acute clinical conditions, cardiovascular risk factors, and key laboratory markers in predicting the outcome of interest. A correlation matrix analysis of these 14 variables revealed that all correlation coefficients were ≦ 0.38 (Supplementary Materials Fig. S3 and Table S4), indicating no significant multicollinearity among the selected features. To further elucidate the relationship between these variables and early death, RCS analyses were performed. The results demonstrated that age was significantly associated with early death in a nonlinear manner: early death risk increased gradually at younger ages but rose sharply with advancing age (Fig. 2A). Total bilirubin exhibited a significant positive linear relationship with early death risk (Fig. 2B). WBC showed a nonlinear association, with both low and high WBC levels both linked to increased early death risk (Fig. 2C). RDW primarily displayed a positive linear correlation with early death risk (Fig. 2D), whereas RBC demonstrated a negative linear correlation (Fig. 2E). BUN was positively and linearly associated with early death risk (Fig. 2F). INR exhibited a nonlinear relationship, with early death risk initially increasing with rising INR but plateauing at higher levels (Supplementary Materials Fig. S4). Potassium was predominantly associated with early death risk in a positive linear fashion (Supplementary Materials Fig. S5).

Fig. 2.

Restricted cubic spline (RCS) analysis of key predictors for early mortality. A Nonlinear association between age and mortality risk; B Linear relationship between total bilirubin and mortality; C U-shaped association of WBC count with mortality; D Positive linear correlation of RDW with mortality; E Negative linear correlation of RBC count with mortality; F Positive linear association of BUN with mortality

Prediction performance for the ML models

Six ML models were trained, and their best-performing hyperparameters are provided in Supplementary Materials Table S5. The comparative analysis of six ML models revealed XGBoost as the optimal predictive algorithm, demonstrating superior performance across multiple evaluation metrics. As shown in Table 2, XGBoost achieved the highest discriminative ability (AUC = 0.903) (Fig. 3), precision (0.876), and the lowest error rates (Brier score = 0.146, log loss = 0.445) (Fig. 4), complemented by favorable calibration (slope = 1.048) (Supplementary Materials Fig. S6) and the highest overall sum score (51) (Fig. 5).

Table 2.

Prediction performance for all machine learning models

Metrics	Models
	LR	SVM	RF	ANN	DT	XGBoost
Accuracy	0.760	0.756	0.716	0.715	0.807	0.785
Precision	0.770	0.760	0.772	0.766	0.777	0.876
Recall	0.742	0.749	0.615	0.618	0.862	0.665
F1 score	0.756	0.755	0.684	0.684	0.817	0.756
AUC	0.848	0.837	0.831	0.806	0.858	0.903
Brier score	0.162	0.166	0.179	0.219	0.150	0.146
Log loss	0.490	0.504	0.535	0.821	0.471	0.445
Intercept-in-large	0.009	0.011	0.294	1.129	-0.200	0.877
Calibration slope	1.114	1.113	1.688	0.307	1.380	1.048
Discrimination slope	0.334	0.314	0.231	0.394	0.346	0.460
Sum score	40	33	20	17	49	51

LR, Logistic Regression; SVM, Support Vector Machine; RF, Random Forest; ANN, Artificial Neural Network; DT, Decision Tree; XGBoost, eXtreme Gradient Boosting Machine; AUC, Area Under the Curve.

Fig. 3.

Receiver operating characteristic (ROC) curves of six ML models after applying 100 bootstraps. XGBoost achieved the highest AUC (0.903), followed by RF, ANN, DT, LR, and SVM

Fig. 4.

Error metric comparison across ML models. A Accuracy; B Precision; C Recall; D F1 score; E Brier score; F Log loss

Fig. 5.

Sum score ranking of ML models. XGBoost (51 points) outperformed other models (DT: 49, LR: 40, SVM: 33, RF: 20, and ANN: 17) across the 10-evaluation metrics

The probability density curves further validated the superior performance of the ML models, particularly highlighting XGBoost, DT, RF, and ANN as having the most distinct separation between predicted probability distributions for positive and negative outcomes (Fig. 6). XGBoost demonstrated the clearest bimodal distribution with minimal overlap between classes, explaining its exceptional discrimination slope (0.460) (Supplementary Materials Fig. S7). In addition, clinical decision curve analysis consistently showed greater net benefit for XGBoost across various threshold probabilities, confirming its enhanced clinical utility for decision-making (Fig. 7). Furthermore, precision-recall curve analysis reinforced XGBoost’s superiority, particularly in maintaining high precision across varying recall levels (Fig. 8), suggesting robust performance in real-world clinical scenarios where class imbalance is common. While DT exhibited the highest recall (0.862) and competitive accuracy (0.807), its relatively lower precision (0.777) compared to XGBoost indicated a greater tendency for false positives. Traditional models including LR and SVM demonstrated balanced but unexceptional performance, whereas RF and ANN showed suboptimal calibration and recall. In addition, based on the scoring system in Supplementary Materials Table S6, XGBoost achieved the highest cumulative score (51), significantly outperforming other models such as RF (20) and ANN (17), further validating its robustness. However, certain prediction performance metrics of XGBoost, including recall (0.665), F1-score (0.756), and intercept-in-large (0.877), still require further improvement.

Fig. 6.

Probability density curves for predicted outcomes. A LR; B SVM; C RF; D ANN; E DT; F XGBoost. XGBoost, DT, RF, and ANN showed distinct separation between mortality (positive) and survival (negative) classes

Fig. 7.

Clinical decision curve analysis (DCA) for ML models. Net benefit of each model across risk thresholds, with XGBoost offering the highest clinical utility

Fig. 8.

Precision-recall curves for model performance comparison. XGBoost maintained high precision across recall levels, while DT achieved higher recall at lower precision

Feature importance in the ML models

The relative importance of predictive features across the six ML models is summarized in Supplementary Materials Table S7. Anti-hypertensives consistently ranked as the most influential variable, achieving the highest cumulative ranks (sum rank = 12), followed by aspirin (sum rank = 20) and BUN (sum rank = 23). Notably, anti-hypertensives were ranked first in LR, SVM, and XGBoost, while INR emerged as the top feature in RF despite its lower performance in other models (Fig. 9). Age and RBC also demonstrated considerable importance, particularly in ANN, where RBC was the leading predictor. In contrast, RDW and CVD were the least influential features across most models, with the lowest cumulative ranks (sum ranks = 65 and 61, respectively). These findings highlight varying feature contributions depending on the modeling approach, with anti-hypertensives, aspirin, BUN, WBC, age, and RBC consistently identified as the six most important predictors (Supplementary Materials Table S8).

Fig. 9.

Feature importance rankings across ML models. A LR; B SVM; C RF; D ANN; E DT; F XGBoost. Anti-hypertensives, aspirin, and BUN were top predictors in most models

Prediction performance for the ensemble ML models

To further enhance predictive performance, particularly for improving XGBoost’s recall (0.665) and F1 score (0.756), this study developed an ensemble model incorporating the top three performing individual ML algorithms (LR, DT, and XGBoost). The ensemble model demonstrated superior performance across the majority of metrics compared to its constituent models (Supplementary Materials Table S9), achieving an accuracy of 0.845, precision of 0.849, recall of 0.840, and F1 score of 0.845. Most notably, it showed exceptional discriminative ability with an AUC of 0.912 (95% CI: 0.888–0.936) (Supplementary Materials Fig. S8), representing an obvious improvement over the best individual model (XGBoost, AUC = 0.903). The ensemble also exhibited enhanced reliability with lower error rates (Brier score = 0.130; log loss = 0.420) and favorable calibration (Intercept-in-large = 0.144) (Supplementary Materials Fig. S9) compared to individual models. While maintaining the strong discrimination slope (0.380) characteristic of its component models (Supplementary Materials Fig. S10), the ensemble approach successfully balanced the high precision of XGBoost with the superior recall of DT, resulting in more robust overall performance. Overall, based on the scoring system in Supplementary Materials Table S10, the ensemble model achieved the highest total score (32), outperforming all individual models (LR: 17, DT: 25, XGBoost: 26). It ranked first in key metrics such as accuracy, F1 score, AUC, and error-related measures (Brier score and log loss), demonstrating its well-rounded superiority (Supplementary Materials Fig. S11).

In addition, the probability density curves provided critical insights into the ensemble model’s superior performance characteristics, demonstrating optimal separation between predicted probability distributions for positive and negative outcomes (Supplementary Materials Fig. S12). These curves revealed that the ensemble model successfully preserved XGBoost’s exceptional ability to identify true negatives (evidenced by a distinct left-shifted peak in the density plot for negative cases) while simultaneously incorporating DT’s strength in correctly classifying true positives (manifested as a right-shifted peak for positive cases with minimal overlap). This dual advantage resulted in a bimodal distribution with clearer separation between outcome classes compared to any individual model, explaining the ensemble’s enhanced discriminative performance (AUC 0.912). Additionally, clinical decision curve analysis confirmed the ensemble model’s superior net benefit across the entire range of clinically relevant threshold probabilities (Supplementary Materials Fig. S13), significantly outperforming all individual models in clinical utility (Supplementary Materials Fig. S14). Precision-recall curve analysis particularly highlighted the ensemble’s balanced improvement (Supplementary Materials Fig. S15), where it achieved both higher precision (0.849) than DT (0.777) and better recall (0.840) than XGBoost (0.665), effectively addressing the precision-recall trade-off that constrained individual models.

Comparative performance of the ensemble model versus traditional scoring systems

The ensemble ML model demonstrated significantly superior discriminative performance compared to conventional clinical scoring systems. As shown in the analysis (Supplementary Materials Fig. S16), the ensemble model achieved an AUC of 0.912 (95% CI: 0.888–0.936), substantially outperforming all traditional risk assessment tools, including APS III (AUC: 0.742, 95% CI: 0.705–0.778), SAPS II (0.719, 95% CI: 0.679–0.758), SOFA (0.741, 95% CI: 0.702–0.780), Charlson Comorbidity Index (0.735, 95% CI: 0.698–0.772), and OASIS (0.642, 95% CI: 0.597–0.687). DeLong’s test confirmed that the ensemble model’s AUC was significantly higher than all traditional scoring systems (all P < 0.001), highlighting its enhanced predictive accuracy for clinical outcomes. These results suggest that the ensemble ML approach not only improves upon individual ML models but also surpasses established risk stratification tools, offering a more reliable method for prognosis in clinical practice.

Comparative performance of the ensemble model with and without SHR

Although the LASSO regression method did not initially select SHR as a key predictive variable, several studies have demonstrated that SHR was significantly associated with mortality among patients with cardiovascular disease [15, 18, 41] or diabetes [14]. Hence, we systematically compared the performance of our ensemble model and its base estimators (LR, DT, and XGBoost) before and after incorporating SHR (Supplementary Materials Table S11). After incorporating SHR, the ensemble model’s accuracy declined slightly to 0.762, with modest reductions in precision (0.816), recall (0.676), and F1-score (0.740). Notably, its AUC remained high (0.880), though lower than the SHR-excluded version (ΔAUC: −0.032, Delong test: P = 0.094, Supplementary Materials Fig. S17), while calibration metrics were preserved (Brier score: 0.153; log loss: 0.473). These results suggest that SHR’s inclusion did not significantly enhance prediction performance, although it provided slightly improvements in certain contexts (e.g., calibration slope increased from 1.904 to 1.536 for the ensemble model).

However, more notably, SHR significantly enhanced traditional scoring systems, with AUC improvements reaching statistical significance for SAPS II (0.719→0.733, Delong test: P = 0.037, Supplementary Materials Fig. S18), SOFA (0.741→0.757, Delong test: P = 0.010, Supplementary Materials Fig. S19), Charlson Index (0.735→0.753, Delong test: P = 0.006, Supplementary Materials Fig. S20), and OASIS (0.642→0.666, Delong test: P = 0.018, Supplementary Materials Fig. S21), but not significance for APSIII (0.742→0.747, Delong test: P = 0.316, Supplementary Materials Fig. S22). These findings collectively suggest that while SHR’s value in our ensemble model was limited, it demonstrates meaningful clinical utility by augmenting established risk assessment tools, particularly for comorbidity evaluation and organ failure scoring. The differential impact across systems may reflect SHR’s varying relevance to different pathophysiological domains.

External validation

External validation of the AI application was performed in a cohort of 307 critically ill patients with cardiovascular disease and diabetes, and the baseline characteristics of these patients are summarized in the Supplementary Materials Table S12. In this cohort, the incidence of early death was 9.8%. The model development and external validation cohorts shared similarities but also exhibit notable differences in baseline characteristics. For example, the validation cohort showed lower rates of chronic pulmonary disease (11.7% vs. 23.3%), statin (13.0% vs. 20.9%), glucocorticoid use (6.8% vs. 11.5%), and a higher AKI incidence (90.6% vs. 84.1%) and smoking status (9.4% vs. 5.0%). Despite these variations, the ensemble model also demonstrated robust performance in external validation, with an accuracy of 0.784, a precision of 0.825, and a recall of 0.720, with an F1-score of 0.769, underscoring its generalizability across diverse clinical settings. The model’s discriminative ability was further supported by an AUC of 0.891 (95% CI: 0.864–0.917) (Supplementary Materials Fig. S23) and a discrimination slope of 0.363 (Supplementary Materials Fig. S24), indicating strong classification performance. The Brier score (0.143) and log loss (0.448) reflected well-calibrated probabilistic predictions, with calibration curve analysis revealing a slope of 1.541 and an intercept of 0.279 (Supplementary Materials Fig. S25). For clinical utility, the clinical decision curve analysis (Supplementary Materials Fig. S26) showed superior net benefit across a wide range of risk thresholds, supporting its potential for decision-making. The precision-recall curve (Supplementary Materials Fig. S27) highlighted a balanced trade-off between positive predictive value and sensitivity, while the probability density curve (Supplementary Materials Fig. S28) revealed clear separation between predicted probabilities of early death and survivor classes, underscoring the model’s reliability in risk stratification.

Discussion

Principal findings

This study developed and validated an interpretable ML ensemble model that outperformed both individual ML algorithms and conventional scoring systems in predicting 30-day mortality among critically ill cardiovascular patients with diabetes. Key predictors emerged as anti-hypertensive medications, aspirin use, BUN levels, WBC counts, patient age, and RBC levels, with SHAP analysis providing crucial clinical insights—age showed a nonlinear mortality risk progression, BUN and bilirubin exhibited linear risk relationships, higher RBC counts demonstrated protective effects, and WBC levels revealed a U-shaped mortality pattern where both extremes increased risk. While SHR enhanced traditional scoring systems, its contribution to the ensemble model proved limited, though external validation confirmed robust performance across institutions and decision curve analysis demonstrated practical clinical utility at various risk thresholds.

Risk factors for mortality in diabetic cardiovascular disease

Prior research has extensively explored risk factors for mortality in patients with diabetes or cardiovascular disease. For example, Chen et al. [12] identified the red cell distribution width-to-albumin ratio (RAR) as a significant predictor of 30-day, 90-day, and 1-year mortality in ICU patients with both coronary heart disease and diabetes. Their analysis of 3,416 patients revealed a linear relationship between RAR levels and long-term mortality risk, positioning RAR as a potential prognostic marker for this high-risk population. Our findings align with this observation, as we also detected a positive linear correlation between RDW and early mortality risk. Beyond hematological markers, sex differences in mortality have also been investigated. While one study reported that women with diabetes face a higher relative risk of cardiovascular mortality than men—particularly at younger ages, even after adjusting for traditional risk factors like blood pressure and cholesterol [42]—our analysis of diabetic patients with established cardiovascular disease showed no significant sex-based disparity in early mortality risk. This discrepancy may stem from differences in study populations or the stage of cardiovascular disease in our cohort.

Lifestyle factors, such as physical inactivity, further modulate mortality risk in diabetic patients. Dai et al. [43] demonstrated that prolonged sedentary behavior was linked to higher all-cause and cardiovascular mortality in adults with diabetes who did not meet recommended physical activity levels, whereas no such association existed in sufficiently active individuals. This suggests that regular exercise may counteract the detrimental effects of sedentary behavior in this population. Anthropometric measures have also been studied as mortality predictors. Woolcott et al. [44] analyzed over 46,000 U.S. adults and found that relative fat mass (RFM) correlated more strongly with diabetes-related mortality than BMI or waist circumference, though all three parameters showed similar associations with cardiovascular disease and all-cause mortality. Importantly, most prior studies focused on either diabetes or cardiovascular disease in isolation, leaving the validity of these risk factors uncertain in patients with comorbid diabetic cardiovascular disease. Our study specifically addressed this gap by examining ICU patients with both conditions, uncovering several physiological indicators tied to early mortality. These included an age-dependent risk escalation (gradual in younger patients but sharp in older cohorts), a positive linear association with total bilirubin and BUN, a U-shaped relationship with WBC count, a negative linear correlation with RBC count, and a nonlinear trend for INR (risk plateauing at higher levels).

Role of SHR in mortality prediction

The SHR has been widely implicated in mortality risk across various clinical settings. Ding et al. [14], for instance, reported a U-shaped relationship between SHR and all-cause mortality in diabetic and prediabetic patients, indicating that both low and high SHR values may portend poorer outcomes. Similar associations have been observed in other high-risk populations, including those with cardiovascular-kidney-metabolic syndrome [45], acute myocardial infarction [15], chronic kidney disease [46], coronary artery disease [16–18], and atrial fibrillation [19]. Previously, our team also demonstrated that SHR was significantly associated with early mortality after propensity score matching analysis [20]. Despite this robust evidence, our study yielded unexpected results regarding SHR’s predictive utility. During the variable selection via LASSO regression, SHR was not among the 14 variables retained in our final model. Moreover, its inclusion in the ensemble model failed to enhance predictive performance, implying that the selected variables (e.g., RDW, bilirubin, WBC) may more comprehensively capture the pathophysiological drivers of early mortality in patients with diabetic cardiovascular disease. Notably, SHR did improve the accuracy of traditional risk scores, such as SOFA, SAPS II, Charlson, and OASIS, underscoring its context-dependent value. This dichotomy highlights several important considerations: (1) traditional scoring systems may benefit more from explicit inclusion of glycemic variability markers due to their simpler linear architectures, while (2) advanced machine learning algorithms can implicitly extract equivalent prognostic information through nonlinear combinations of other clinical features. Furthermore, this also suggests that the features selected by LASSO in the present have already comprehensively captured the prognostic information relevant to 30-day mortality, rendering SHR redundant in this context. The robustness of the selected variables was further supported by their strong biological plausibility and alignment with established pathophysiological mechanisms in critically ill cardiovascular patients. Thus, while SHR has demonstrated prognostic value in traditional risk scores and other studies [15–20], its exclusion here likely reflects the ensemble model’s ability to distill equivalent or superior predictive signals from other clinically meaningful features. These findings provide valuable insights into the evolving paradigm of risk stratification, where machine learning approaches may reduce reliance on individual biomarkers by identifying complex, higher-order interactions among clinical variables.

Challenges in current mortality prediction models for diabetic cardiovascular disease

Accurate mortality prediction in diabetic cardiovascular disease remains challenging, as existing models exhibit significant limitations in both generalizability and predictive performance for early adverse outcomes. Current risk scores, while useful in certain contexts, often fail to account for the complex interplay between metabolic and cardiovascular pathophysiology in this high-risk population. For instance, Pagano et al. [47] found that the UKPDS-OM2 model, while effective in predicting stroke and ischemic heart disease events in diabetic patients, systematically overestimated all-cause mortality and myocardial infarction risk in Italian and Dutch cohorts. Similarly, the ENFORCE model demonstrated strong real-world performance for 6-year mortality prediction in diabetes (C-statistics 0.75–0.80) [48], but its accuracy declined markedly (C-statistic 0.68) when applied to the tightly controlled ACCORD trial population. These discrepancies highlight the limitations of conventional prediction tools, which were neither designed for early mortality risk stratification nor optimized using advanced computational approaches.

Recent advances in ML have enabled more sophisticated mortality prediction in cardiovascular subpopulations. For example, ML models have been developed for all-cause mortality prediction in cardiac surgery patients [21], as well as studies focusing on heart failure with hypertension [22], acute myocardial infarction [23], critically ill atrial fibrillation patients [24, 25], atherosclerotic cardiovascular disease [26], hypertensive participants [49], coronary heart disease [27–29], Takotsubo syndrome [50], life-threatening ventricular arrhythmias [30] (Table 3). This study further synthesized recent three-year literature on ML-based cardiovascular mortality prediction and demonstrated that our ensemble model offers three key advantages: (1) superior predictive performance (AUC: 0.912) for 30-day mortality in cardiovascular-diabetic patients, exceeding most prior studies (AUC range: 0.741–0.905); (2) incorporation of external validation—unlike most existing models—ensuring clinical generalizability; and (3) specific targeting of cardiovascular-diabetic mortality, a high-risk yet underrepresented population in earlier research, thereby bridging a critical prognostic gap.

Table 3.

Summary of machine learning models predicting mortality in cardiovascular disease

Authors	Published year	Study design	Patients	Sample size	Optimal ML model	Outcome	Prediction performance (AUC)	External validation
Pei et al. [21]	2025	Retrospective	Patients undergoing cardiac surgery	3,848	Naive Bayes	All-cause mortality	In-hospital mortality (AUC: 0.794) and 360-day mortality (AUC: 0.741).	No
Li et al. [34]	2025	Retrospective	Patients with both acute critical illness and pre- existing CHF	913	ANN	28-day hospital mortality	In-hospital, 28-day all-cause mortality (AUC:0.801)	No
Yan et al. [45]	2024	Retrospective	Patients with cardiovascular–kidney–metabolic syndrome	9647	LightGBM	All-cause mortality	AUC: 0.863	No
Chen et al. [25]	2024	Retrospective	Patients with atrial fibrillation	9826	LightGBM	30-day all-cause mortality after ICU admission	AUC: 0.780	Yes
Wang et al. [26]	2025	Retrospective	Patients with atherosclerotic cardiovascular disease	2807	RF	28-day all-cause mortality	AUC: 0.804	No
Li et al. [49]	2025	Retrospective	Hypertensive participants	9432	XGBoost	All-cause mortality and cardiovascular mortality	AUC: 0.823	No
Cheng et al. [27]	2024	Prospective	Elderly coronary heart disease patients with anemia	509	Ensemble	1-year overall mortality	AUC: 0.828	Yes
Ke et al. [28]	2022	Retrospective	Patients with acute coronary syndrome	6482	RF	In-hospital mortality	AUC: 0.913	No
Li et al. [29]	2023	Prospective	Older patients with coronary artery disease combined with impaired glucose tolerance or diabetes mellitus	451	GBM	1-year overall mortality	AUC: 0.836	No
Filippo et al. [50]	2023	Retrospective	Patients with Takotsubo syndrome	4519	Ensemble	In-hospital mortality	AUC: 0.890	Yes
Li et al. [30]	2023	Retrospective	Patients with life-threatening ventricular arrhythmias	5991	CatBoost	In-hospital mortality	AUC: 0.905	Yes
Our study	2025	Retrospective	Cardiovascular diabetic disease	1595	Ensemble	30-day mortality	AUC: 0.912	Yes

ML, Machine Learning; AUC, Area Under the Curve; CHF, Chronic Heart Failure; ANN, Artificial Neural Network; RF, Random Forest; LightGBM, Light Gradient Boosting Machine; GBM, Gradient Boosting Machine; CatBoost, Categorical Boosting.

However, these models target specific cardiovascular conditions rather than the intersection of diabetes and cardiovascular disease—a distinct clinical entity with unique risk profiles. Our study addresses this critical gap by developing an ensemble ML model specifically tailored to diabetic cardiovascular disease. By incorporating multidimensional clinical and biochemical variables, our model achieved exceptional discrimination (AUC 0.912) with robust external validation performance (AUC 0.891). This represents a significant advancement over existing tools, offering clinicians a precise, disease-specific prognostic instrument to guide early intervention in this vulnerable population. Moreover, the model’s strong generalizability across diverse settings underscores its potential for real-world implementation, bridging the gap between research and clinical decision-making for patients with diabetic cardiovascular disease.

Clinical implication and future studies

The ensemble ML model developed in this study demonstrated significant clinical potential by providing a highly accurate and interpretable tool for predicting 30-day mortality in critically ill diabetic cardiovascular patients. Unlike traditional scoring systems, which showed limited discriminative ability, the ensemble model integrated key predictors—including 14 model features such as anti-hypertensives, aspirin, BUN, WBC, age, and RBC—while also incorporating SHR to enhance conventional risk assessment. The model’s superior performance (AUC 0.912) and external validation (AUC 0.891) suggest its potential for real-world implementation in critical care setting to guide clinical decision-making. Notably, the SHAP analysis provided clinically meaningful insights, such as the nonlinear risk associated with age and the protective effect of higher RBC counts, which could inform individualized patient management.

Despite these strengths, several areas warrant further investigation. First, while SHR improved traditional scores, its marginal benefit in the ensemble model suggests that its predictive value may be context-dependent. Future studies should explore whether dynamic SHR measurements or alternative glycemic variability indices enhance predictive accuracy. Second, prospective validation in diverse, multi-center cohorts is needed to assess generalizability across different patient populations and healthcare settings. Third, integrating real-time electronic health record data with the ensemble model could facilitate dynamic risk prediction, enabling timely interventions. Interventional studies should evaluate whether risk stratification using this model translates into improved patient outcomes, such as targeted glycemic control or early escalation of care in high-risk individuals. Lastly, building upon our team’s established expertise in developing clinically applicable AI tools [38] - including our previously validated web application for postoperative mortality prediction - we are currently developing an intuitive interface for this ensemble model to enhance both transparency and clinical utility. This implementation will incorporate interactive visualization of individualized risk factors and provide actionable clinical decision support, bridging the gap between algorithmic prediction and bedside application. By addressing these gaps, future research can further refine and optimize ML-driven mortality prediction in this high-risk population.

Limitations

The present study developed a high-performing ensemble ML model that demonstrated superior predictive accuracy compared to traditional scoring systems, while maintaining interpretability through SHAP analysis. However, the retrospective design of the study introduces potential biases in data collection, including unmeasured confounders such as variations in ICU management protocols and undocumented comorbidities. In addition, the model’s complexity may limit its clinical adoption, suggesting a need for hybrid approaches that balance predictive power with user-friendly interfaces. While the model was rigorously validated in an independent cohort, its generalizability to diverse healthcare settings remains uncertain, highlighting the importance of future multicenter prospective studies. Although SHR improved traditional scores, its limited incremental benefit in the ensemble model indicates that additional dynamic biomarkers may be helpful to better capture acute metabolic stress. The identified predictors aligned well with clinical knowledge, but the absence of novel biomarkers in our dataset suggests opportunities for future research incorporating multi-omics data to further refine risk stratification. Notably, further analysis of this study dataset revealed that there were 10 cases (0.6%) of fracture trauma. Although the proportion of fracture trauma in the 30-day mortality group (1.2%) was twice that in the survivor group (0.6%), the difference was not statistically significant. Thus, future research should further investigate the impact of trauma on 30-day mortality in cardiovascular patients. Lastly, while our current model demonstrates robust performance, future studies may also benefit from incorporating advanced techniques like later temporal attention [51] for improved temporal data analysis and singular pooling [52] for more efficient feature aggregation, potentially enhancing both predictive accuracy and clinical applicability. Hence, moving forward, efforts should focus on prospective validation in diverse populations, integration of real-time physiological data, development of clinically actionable decision-support tools, and introduction of more advanced algorithms to translate this advanced predictive model into improved patient outcomes in clinic.

Conclusions

The ensemble ML model provides a clinically actionable, high-accuracy tool for 30-day mortality prediction in high-risk diabetic cardiovascular patients, outperforming conventional scores while maintaining interpretability. Although SHR augmented traditional systems, its marginal utility in the ensemble suggests that advanced ML architectures may inherently capture glycemic stress through other correlated features. Future work should focus on real-world deployment and assessing the model’s impact on clinical decision-making and patient outcomes.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

All authors took part in designing and writing the manuscript. All authors read and approved the final manuscript.

Funding

None.

Data availability

The datasets of the current study in the model development cohort are available at Medical Information Mart for Intensive Care IV database (version 3.1, [https://mimic.mit.edu/](https:/mimic.mit.edu) ). The data of external validation cohort can be available after request from the corresponding authors.

Declarations

Ethical approval and consent to participate

The Medical Information Mart for Intensive Care IV database project received approval from the Institutional Review Boards of both the Massachusetts Institute of Technology and Beth Israel Deaconess Medical Center. Patient information was anonymized, thereby obviating the need for informed consent from individual patients for this study. This study of external validation was approved by the Institutional Review Board of the Chinese PLA General Hospital. To comply with ethical standards, all personally identifiable information was anonymized prior to analysis.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.