Explainable machine learning model for early prediction of ICU death in chronic heart failure with pulmonary infection

Abstract

Objective

To develop and validate a machine learning model for predicting ICU mortality in CHF patients with pulmonary infection.

Methods

Clinical data were extracted from the MIMIC-IV database, and ICU mortality within 15 days was defined as the primary endpoint. Patients were stratified into HFrEF, HFmrEF, and HFpEF based on left ventricular ejection fraction. Six models were constructed—logistic regression, random forest, light gradient boosting machine, extreme gradient boosting (XGBoost), multilayer perceptron, and k-nearest neighbor—and optimized using cross-validated grid search. Predictive performance was evaluated using accuracy, precision, recall, F1-score, AUC, Brier score, Youden index, and calibration slope. Model interpretability was assessed using SHAP and LIME. External validation was conducted using the eICU-CRD database.

Results

A total of 1308 patients were included, with an ICU mortality of 15.98%. Among all models, XGBoost achieved the best balance of discrimination and calibration (AUC 0.968; calibration slope 1.070), and retained useful but attenuated discrimination in external testing (AUC 0.718). SHAP analysis identified lactic acid, white blood cell count, prior diagnosis of hypertension, serum potassium, and pre-existing diabetes mellitus as the five most influential predictors of the outcome. Subgroup analyses further demonstrated distinct pathophysiological trajectories associated with mortality: HFrEF predominantly exhibited hemodynamic instability, impaired tissue perfusion, and electrolyte disturbances; in contrast, HFpEF were more frequently associated with chronic metabolic-inflammatory dysregulation and coagulopathy.

Conclusion

The XGBoost-based model provides interpretable prediction of ICU mortality in CHF patients with pulmonary infection. Phenotype-specific interpretability analysis further suggested that mortality risk mechanisms differ between heart failure subtypes, supporting more individualized risk assessment and management in critically ill CHF patients.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12911-026-03682-z.

Introduction

Chronic heart failure (CHF) is a clinical syndrome resulting from diverse structural or functional cardiac abnormalities, leading to impaired ventricular contraction and/or filling. These alterations eventually cause insufficient tissue perfusion and venous congestion. CHF is closely associated with long-standing hypertension, coronary atherosclerotic disease, cardiomyopathies, and valvular disorders. Epidemiological data indicate that approximately seven million individuals in the United States and over 64 million people worldwide are currently affected by heart failure [1, 2]. With the global population aging and cardiovascular risk factors—such as obesity, diabetes, smoking, and hypertension—remaining highly prevalent, the burden of CHF continues to rise, posing a significant challenge to healthcare systems worldwide [3]. Indeed, heart failure is increasingly understood as one manifestation of a broader spectrum of cardiovascular and metabolic diseases that share overlapping risk factors and pathophysiological pathways, including dyslipidemia, insulin resistance, chronic inflammation, and oxidative stress. Notably, even with optimal control of established traditional risk factors, a substantial residual risk persists, with many events occurring in individuals classified as low or intermediate risk by conventional models—an observation that has motivated the search for novel risk markers and more sophisticated, data-driven risk-prediction tools [4].

Pulmonary infection is among the most common hospital-acquired infections in intensive care units (ICUs). Patients with CHF are particularly vulnerable due to impaired hemodynamics, reduced immune competence, prolonged immobilization, and frequent reliance on mechanical ventilation. These infections further exacerbate cardiac dysfunction, may induce a systemic inflammatory response, and contribute to multiorgan impairment, prolonged hospitalization, and increased risk of death [5, 6]. Previous studies have shown that CHF patients with pulmonary infection have a 3–5-fold higher likelihood of dying during ICU admission compared with those without infection [7–9]. Especially in the case of bacterial pneumonia, ICU are often caused by opportunistic, biofilm-forming pathogens that are resistant to antibiotics and disinfectants, such as Pseudomonas aeruginosa. This pathogen possesses intrinsic antimicrobial resistance and can persist on medical devices and environmental surfaces, making eradication difficult and contributing to treatment failure and cross-infection [10]. Early identification of high-risk patients and accurate risk stratification are therefore essential for optimizing antimicrobial therapy, guiding clinical decision-making, and improving outcomes in critically ill CHF patients. However, CHF is a heterogeneous condition. Significant variation exists among patients in terms of cardiac remodeling, hemodynamic status, inflammatory response, and comorbidity burden. Treating CHF as a homogeneous population may obscure important risk characteristics. Based on left ventricular ejection fraction (LVEF), CHF can be classified into heart failure with reduced ejection fraction (HFrEF), heart failure with mildly reduced ejection fraction (HFmrEF), and heart failure with preserved ejection fraction (HFpEF). Understanding differences in infection profiles and prognostic implications across these phenotypes is therefore clinically relevant.

In parallel, advancements in computational power and the increasing availability of large-scale clinical datasets have driven rapid growth in the use of machine learning in medicine. Machine learning–based models have shown potential advantages over traditional statistical tools in screening, diagnostic assistance, risk stratification, and prognosis prediction. In oncology, such models integrate imaging, genomic, and clinical data to enable earlier tumor detection and individualized treatment planning [11–13]. Machine learning has also been successfully applied to predict postoperative complications—including infection [14], organ dysfunction [15], and hospital readmission [16]—and is increasingly used in critical care to anticipate clinical deterioration, guide intervention strategies, and estimate mortality risk [17–19]. More broadly, machine learning is reshaping cardiovascular and cardiometabolic risk modeling, because conventional risk equations capture only a fraction of the perturbations that precede disease. By integrating traditional risk factors with emerging biomarkers and routine clinical data, such models enable individualized, precision risk stratification [20]. In CHF research, machine learning has been employed to predict readmission, short-term mortality, and infectious complications [21–24]. However, most studies have focused on single cohorts or isolated outcomes and have not explored heterogeneity across heart failure subtypes, particularly in the context of pulmonary infection in ICU patients.

Therefore, this study aimed to develop a machine learning–based model to predict ICU mortality in CHF patients complicated by pulmonary infection using comprehensive clinical and laboratory data. Furthermore, subgroup analyses were conducted to identify infection characteristics and risk determinants across different heart failure phenotypes. By enabling timely and individualized risk stratification, this approach may support clinical decision-making and facilitate individualized management for critically ill patients with CHF.

Methods

Data sources

This retrospective study used data from two large, publicly available critical care databases. Model development was conducted using the Medical Information Mart for Intensive Care IV database (MIMIC-IV, version 3.0) [25], which contains comprehensive clinical information from patients admitted to the emergency departments and ICU of Beth Israel Deaconess Medical Center in Boston, Massachusetts. External validation was performed using the eICU Collaborative Research Database (eICU-CRD, version 2.0) [26], which includes ICU data from multiple centers across the United States, providing broader population heterogeneity and healthcare variability. Both databases are fully de-identified. Access and use were approved after completion of required training (certificate No. 12829470).

Study population and variable selection

Patients with a diagnosis of CHF complicated by pulmonary infection were identified using International Classification of Diseases (ICD-9 and ICD-10) codes. Inclusion criteria were: (1) first ICU admission and (2) age ≥ 18 years. Exclusion criteria included: (1) missing key variables, (2) death within 48 hours of ICU admission, and (3) >20% missing data. Based on these criteria, 2,780 eligible patients were initially identified from MIMIC-IV.

To define outcome timing, we analyzed the distribution of death following ICU admission. Approximately 70–80% of deaths occurred within 15 days, with a median of 11.23 days. Therefore, 15-day mortality was selected as the primary outcome to more accurately characterize early mortality among patients with chronic heart failure and pulmonary infection during their ICU stay. The outcome was treated as a binary variable (dead or alive within 15 days). The study flowchart is presented in Fig. 1. This work adhered to the TRIPOD reporting guidelines [27].

Fig. 1.

The patient selects the flowchart and the research design

Sample size estimation

Because this study was retrospective, sample size was determined by available data to maximize precision [28]. Adequacy was assessed using the events-per-variable approach and recent model-based sample size recommendations [29], implemented with the R package “pmsampsize”. Based on previous studies, we assumed an event rate of approximately 20%, an expected C-statistic of 0.90–0.95, and 30–40 candidate predictors [30]. Under these parameters, a minimum of 764–941 subjects was required to ensure sufficient power and model performance, and our sample exceeded this requirement.

Variable extraction

A total of 38 variables were selected based on clinical relevance, previous research, and database availability [31–34], covering demographics, comorbidities, vital signs, disease severity, cardiac function, and laboratory findings:

• Demographics and history: gender, age, chronic pulmonary disease, hypertension, diabetes, chronic renal disease, and type of pneumonia (bacterial vs. non-bacterial);

• Vital signs and severity scoring: systolic/diastolic blood pressure, mean pressure, heart rate, respiratory rate, temperature, pulse oxygen, and SAPS II score;

• Laboratory measurements: base excess (BE), lactic acid (LAC), PH, PaO₂, PaCO₂, glucose, white/red blood cell counts, creatinine, urea, sodium, potassium, calcium, chlorine, platelets (PLT), hemoglobin (HB), prothrombin time (PT), APTT, total bilirubin;

• Cardiac function: LVEF;

• Clinical information: use of norepinephrine, vasoactive agents (including vasopressors and inotropes), and mechanical ventilation status.

• All the rationale for selecting each variable is provided in Supplementary Table S2.

Missing data and preprocessing

All variables were screened for outliers and missing values. Implausible entries (e.g., null or physiologically impossible values) were removed. Variables with more than 20% missingness were excluded, while those with less than 20% missingness were imputed using multiple imputation by chained equations “mice” in R. Continuous variables were standardized using z-score normalization. Binary variables were encoded as 1 (presence) and 0 (absence). All predictor variables were extracted from the first 24 hours after ICU admission.

Handling of class imbalance

The ICU mortality rate was 15.98%, yielding a positive-to-negative ratio of approximately 1:6. To reduce bias toward the majority class and improve model discrimination, class imbalance was addressed using the Synthetic Minority Oversampling Technique combined with Tomek links (SMOTE–Tomek) [35]. This hybrid method simultaneously oversamples the minority class and removes borderline/noisy samples, resulting in improved dataset balance and reduced overfitting.

Model development and evaluation

Six machine learning algorithms were implemented using Python (version 3.10): logistic regression (LR), random forest (RF), Light Gradient Boosting Machine (LightGBM), eXtreme Gradient Boosting (XGBoost), multilayer perceptron (MLP), and k-nearest neighbors (KNN). The dataset was randomly divided into a training set and an independent validation set at a ratio of 70:30. Model development and hyperparameter tuning were conducted exclusively within the training set using 10-fold cross-validation combined with grid search.After optimization, the final model was evaluated on the independent validation set.Model performance was assessed using accuracy, precision, recall, F1-score, area under the receiver operating characteristic curve (AUC), Brier score, Youden index, and calibration slope. The average cross-validation performance within the training set was used for model selection, whereas the final reported performance metrics were obtained from the independent validation set. The 95% confidence intervals were estimated using bootstrap resampling.

Model interpretability

Feature contribution and global model behavior were analyzed using Shapley Additive Explanations (SHAP) [36]. SHAP values quantify each feature’s marginal effect on individual predictions, enabling visualization of key predictors and direction of influence. To enhance clinical interpretability, the Local Interpretable Model-agnostic Explanations (LIME) was additionally applied [37]. LIME approximates local decision boundaries and provides per-patient feature weights, revealing how variable changes influence predicted mortality risk. The combined ues of SHAP and LIME supports transparent prediction and individualized clinical decision-making.

Statistical analysis

Statistical analyses were performed using R (version 4.4.1), SPSS (version 28.0), and Python (version 3.10). Categorical variables were summarized as frequencies and percentages. Continuous variables were expressed as mean ± standard deviation or median (interquartile range), depending on distribution. Group comparisons were conducted using the t-test, chi-square test, or non-parametric tests as appropriate. A two-sided p < 0.05 was considered statistically significant.

Results

A total of 1308 patients were included in the final cohort, with a mean age of 73.93 ± 13.28 years. The most prevalent comorbidities were chronic renal disease (43.65%), diabetes (43.50%), and hypertension (33.87%). As shown in Table 1, survivors and non-survivors differed significantly with respect to history of hypertension, use of norepinephrine, use of vasoactive agents, chronic renal disease, APTT, BE, serum potassium, serum calcium, respiratory rate, serum creatinine, HB, LAC, PH, PLT, urea, PT, SAPS II score, and WBC (all p < 0.05). Overall, 209 patients (15.98%) died within 15 days of ICU admission, including 81 deaths in the HFrEF group (18.41%), 11 in the HFmrEF group (12.79%), and 117 in the HFpEF group (14.96%).

Table 1.

Comparison of various characteristics in the two groups of patients (n = 1308)

Characteristics	Total (n = 1308)	Survivors (n = 1099)	Non-survivors (n = 209)	p-value
Age (years)	73.93 ± 13.28	73.07 ± 13.48	78.45 ± 11.15	<0.001
Gender				0.812
Male	742 (56.73%)	625 (56.87%)	117 (55.98%)
Female	566 (43.27%)	474 (43.13%)	92 (44.02%)
COPD				0.316
Yes	243 (18.58%)	199 (18.11%)	44 (21.05%)
No	1065 (81.42%)	900 (81.89%)	165 (78.95%)
Diabetes				0.770
Yes	569 (43.50%)	480 (43.68%)	89 (42.58%)
No	739 (56.50%)	619 (56.32%)	120 (57.42%)
Hypertension				0.012
Yes	443 (33.87%)	388 (35.30%)	55 (26.32%)
No	865 (66.13%)	711 (64.70%)	154 (73.68%)
Invasive vent				0.281
Yes	625 (47.78%)	518 (47.13%)	107 (51.20%)
No	683 (52.22%)	581 (52.87%)	102 (48.80%)
Norepinephrine				0.028
Yes	416 (31.80%)	336 (30.57%)	80 (38.28%)
No	892 (68.20%)	763 (69.43%)	129 (61.72%)
Pneumonia type				0.053
Bacterial	223 (17.05%)	197 (17.93%)	26 (12.44%)
Other	1085 (82.95%)	902 (82.07%)	183 (87.56%)
Renal disease				0.016
Yes	571 (43.65%)	464 (42.22%)	107 (51.20%)
No	737 (56.65%)	635 (57.78%)	102 (48.80%)
Vasoactive agent				0.003
Yes	574 (43.88%)	463 (42.13%)	111 (53.11%)
No	734 (56.12%)	636 (57.87%)	98 (46.89%)
Lvef (%)				0.194
≤40	439 (33.56%)	358 (32.58%)	81 (38.76%)
41%~49	86 (6.57%)	75 (6.82%)	11 (5.26%)
≥50	783 (59.86%)	666 (60.60%)	117 (55.98%)
Activated Partial Thromboplastin Time (sec)	35.29 (30.24–48.90)	35.16 (29.95–47.60)	37.04 (31.14–55.52)	0.021
Base Excess (mEq/L)	0.83 (−1.80–4.48)	1.00 (−1.50–4.67)	−0.38 (−3.56–3.00)	<0.001
Blood Potassium(mEq/L)	4.04 (3.81–4.34)	4.03 (3.81–4.30)	4.21 (3.84–4.50)	<0.001
Blood Sodium (mEq/L)	139.51 (136.62–142.75)	139.36 (136.64–142.47)	139.95 (136.45–144.03)	0.051
Blood calcium (mmol/L)	8.43 (8.06–8.84)	8.45 (8.09–8.84)	8.29 (7.86–8.81)	0.012
Blood chlorine (mEq/L)	102.59 (98.29–106.27)	102.57 (98.21–106.00)	102.88 (99.03–108.50)	0.024
Body temperature (°F)	98.35 (97.84–98.98)	98.35 (97.83–99.00)	98.36 (97.91–98.85)	0.401
Breathe (breaths/min)	20.41 (17.80–23.56)	20.26 (17.72–23.39)	20.97 (18.27–24.06)	0.019
Creatinine (mg/dL)	1.30 (0.88–2.07)	1.23 (0.87–1.95)	1.72 (1.06–2.83)	<0.001
Diastolic blood pressure (mmhg)	60.19 (53.48–67.61)	60.23 (53.69–67.73)	59.80 (52.43–66.85)	0.376
Glucose (mg/dL)	130.53 (111.00–162.60)	129.74 (111.00–161.50)	133.00 (108.86–167.92)	0.322
Heart Rate (bpm)	85.54 (74.04–97.27)	84.67 (73.85–96.33)	87.17 (75.70–97.27)	0.053
Hemoglobin (g/dl)	9.62 (8.54–10.93)	9.66 (8.63–10.96)	9.34 (8.34–10.65)	0.027
Lactic Acid (mmol/L)	1.46 (1.10–1.98)	1.42 (1.10–1.92)	1.86 (1.27–2.70)	<0.001
Mean pressure (mmhg)	72.96 (66.73–80.23)	73.00 (66.88–80.24)	72.83 (64.56–80.31)	0.285
PH	7.37 (7.31–7.42)	7.39 (7.35–7.43)	7.36 (7.31–7.43)	<0.001
PaCO2 (mmhg)	42.85 (37.33–50.84)	43.00 (37.50–51.00)	42.46 (36.85–49.51)	0.412
PaO2 (mmhg)	93.00 (69.61–122.00)	93.75 (69.60–122.67)	88.50 (69.90–118.97)	0.383
Platelets (K/uL)	211.55 (155.80–291.40)	216.00 (158.29–296.67)	190.00 (141.20–261.07)	<0.001
Prothrombin time (sec)	14.79 (13.14–18.37)	14.66 (13.10–17.87)	15.98 (13.28–19.21)	0.008
Pulse oxygen (%)	96.81 (95.06–98.31)	96.81 (95.04–98.32)	96.88 (95.12–98.27)	0.886
Red blood cells (K/uL)	3.27 (2.94–3.71)	3.27 (2.95 -3.72)	3.25 (2.84–3.68)	0.487
SAPS II (score)	42.00 (35.00–50.00)	41.00 (34.00–49.00)	49.00 (39.50–56.50)	<0.001
Systolic blood pressure (mmhg)	112.51 (103.21–125.44)	112.74 (103.46–125.58)	111.15 (101.34–124.07)	0.250
Total bilirubin (mg/dL)	0.50 (0.35–0.90)	0.50 (0.35–0.85)	0.55 (0.40–0.91)	0.267
Urea (mg/dL)	32.53 (21.81–50.52)	31.18 (20.53–48.33)	42.80 (27.13–64.69)	<0.001
White Blood Cell (K/uL)	10.75 (8.34–13.56)	10.53 (8.20–13.09)	12.63 (9.66–15.91)	<0.001

Model performance

All 38 variables were included for model development. Hyperparameters were optimized using GridSearchCV, and detailed search spaces are provided in Supplementary Table S1. Table 2 summarizes performance metrics with 95% confidence intervals. In terms of individual metrics, KNN achieved the highest accuracy, recall, F1, and Youden index, LR demonstrated the best calibration slope. XGBoost achieved the highest precision and the lowest Brier score, indicating a favorable balance of discrimination and calibration.

Table 2.

Predictive performance of six machine learning models

Measure	LR	RF	LGBM	XGBoost	MLP	KNN
Accurary	0.784 (0.753–0.815)	0.903 (0.879–0.926)	0.897 (0.874–0.921)	0.906 (0.883–0.927)	0.889 (0.865–0.914)	0.914 (0.891–0.935)
Precision	0.795 (0.749–0.839)	0.888 (0.853–0.920)	0.880 (0.842–0.915)	0.894 (0.862–0.926)	0.866 (0.832–0.901)	0.863 (0.827–0.896)
Recall	0.752 (0.708–0.796)	0.917 (0.886–0.945)	0.913 (0.882–0.943)	0.917 (0.886–0.946)	0.916 (0.884–0.945)	0.979 (0.961–0.994)
F1-score	0.773 (0.738–0.807)	0.902 (0.879–0.925)	0.897 (0.873–0.921)	0.906 (0.882–0.929)	0.890 (0.864–0.915)	0.917 (0.896–0.939)
AUC	0.863 (0.832–0.889)	0.968 (0.955–0.978)	0.970 (0.960–0.980)	0.968 (0.956–0.979)	0.947 (0.930–0.963)	0.921 (0.901–0.941)
Brier Score	0.149	0.101	0.072	0.071	0.083	0.110
Youden index	0.566	0.806	0.795	0.812	0.780	0.830
Calibration slope	0.940	1.439	0.742	1.070	1.061	0.864

Predicted probability density plots (Supplementary Figure S1) demonstrated clearer separation between survivors and non-survivors in the LightGBM, XGBoost, and MLP models, indicating stronger discriminative capability. Calibration curves are presented in Supplementary Figure S2. LR, XGBoost, and MLP demonstrated favorable calibration slopes. Given its overall balance between discrimination and calibration, XGBoost was selected as the final model for subsequent interpretability analysis.

Feature interpretability

SHAP analysis (Fig. 2) identified LAC, WBC, history of hypertension, serum potassium, and history of diabetes mellitus as the major predictors of ICU mortality in patients with chronic heart failure complicated by pneumonia. Blue-to-red color gradients indicate increasing variable values, and SHAP values represent each feature’s marginal effect on model output.

Fig. 2.

Importance ranking of SHAP variables. Note: (A) SHAP graph on the training set; (B) SHAP graph on the test set; each line represents a feature, and the abscissa is the SHAP value. Red dots represent higher feature values, and blue dots represent lower feature values

LIME was applied for individualized interpretation (Fig. 3). Green denotes that the feature enhances the model’s propensity to assign the current input to the target category, whereas red indicates that the feature suppresses the model’s likelihood of predicting the target category. In representative non-survivors (Fig. 3A–B), bacterial pneumonia, a history of chronic disease, prolonged PT, and the absence of norepinephrine use contributed to a higher predicted risk of death, whereas a normal white blood cell count and a lower SAPS II score were associated with a reduced risk. For example, in Fig. 3A, the model predicted a mortality probability of death of 0.819, which was consistent with the observed outcome.

Fig. 3.

LIME locally interpretable model agnostic interpretation map. Note: (A) Deceased patients, true-positive cases; (B) Deceased patients, true-positive cases; (C) Surviving patients, true-negative cases; (D) Surviving patients, true-negative cases

Subgroup analysis

Subtype-specific SHAP rankings for HFrEF, HFmrEF, and HFpEF patients are shown in Fig. 4A–F. Supplementary Figure S4 illustrates the common characteristics of the top ten predictors across the three heart failure subtypes in the global model. Supplementary Figure S5 highlights differences in predictor importance between HFrEF and HFpEF, while Supplementary Figure S6 further demonstrates subtype-specific variations in SHAP-based feature contributions. These findings indicate substantial heterogeneity in mortality risk factors between these heart failure phenotypes.

Fig. 4.

Differential feature contributions across LVEF phenotypes within the global prediction model. Note: (A) and (B): LVEF ≤ 40%, (C) and (D): 40%＜LVEF＜50%, (E) and (F) LVEF ≥ 50%

External validation

External validation was performed in 365 patients from the eICU-CRD database, among whom ICU mortality within 15 days was 15.62% (57/365), which not significantly different from the MIMIC-IV cohort (χ² = 0.028, p = 0.867). In the external validation cohort, XGBoost achieved: Accuracy: 0.763; Precision: 0.761; Recall: 0.383; AUC: 0.718.

Discussion

The global burden of CHF complicated by pulmonary infection remains substantial, and this combination has a well-established adverse impact on clinical outcomes [38, 39]. In the present study, ICU mortality within 15 days was 15.98% in the derivation cohort and 15.62% in the external validation cohort. Previous literature often uses 28-day or 30-day mortality to evaluate short-term prognosis in critically ill heart failure populations [40, 41]. However, most of these studies treat CHF patients as a homogeneous group or have focused on specific clinical settings, such as COVID-19 infection, chronic obstructive pulmonary disease, or single subtypes of heart failure (e.g., HFrEF or AHF) [42, 43]. Evidence regarding early mortality risk specifically in CHF patients with pulmonary infection across different heart failure phenotypes remains limited.

One important finding of this study was the early occurrence of mortality events: nearly 80% of deaths occurred within the first 15 days after ICU admission, indicating a rapid deterioration of clinical status in this patient population. This suggests that the therapeutic window may be much narrower than previously recognized. Therefore, we selected 15-day mortality as the primary outcome to better capture acute clinical deterioration associated with critical illness, including hemodynamic instability, respiratory failure, electrolyte disturbances, and dependence on vasopressor therapy, all of which are particularly common among patients with chronic heart failure complicated by pulmonary infection. This outcome measure enables a more accurate assessment of early clinical worsening and reflects the real-time risks most relevant to clinical decision-making and timely intervention.

The interaction between CHF and pulmonary infection is bidirectional and pathophysiologically complex. In CHF, chronic pulmonary congestion, interstitial edema, and disruption of the alveolar–capillary barrier promote microbial colonization and markedly increase infection susceptibility. Reduced perfusion and chronic inflammation further weaken immunity, including impaired phagocytic function and cellular defense [44, 45]. Immobilization, invasive procedures, and mechanical ventilation in the ICU add additional risk. Conversely, pulmonary infection worsens cardiac function through systemic inflammation, hypoxemia, altered preload, increased afterload, and infection-related myocardial depression, leading to rapid hemodynamic instability and acute decompensation [46, 47]. This forms a vicious cycle in which infection accelerates heart failure, and deteriorating heart failure promotes further infection, ultimately increasing mortality in ICU patients [48].

Given this pathophysiological complexity and the lack of early risk-stratification tools tailored to this specific population, developing early mortality prediction models is clinically meaningful. In this study, we applied machine learning algorithms to develop an ICU mortality prediction model and further compared risk patterns across different heart failure phenotypes. By integrating multi-dimensional clinical and laboratory data within the first 24 hours of ICU admission, the proposed approach may support early identification of high-risk patients and guide timely, individualized interventions.

In the present study, LAC, WBC, history of hypertension, serum potassium, and history of diabetes mellitus emerged as the most influential predictors of ICU mortality among CHF patients with pulmonary infection. These findings are consistent with previous evidence highlighting the roles of systemic inflammation, metabolic disturbance, and hemodynamic instability in critically ill heart failure patients [49, 50].

Elevated LAC reflects tissue hypoperfusion and anaerobic metabolism, and is strongly associated with poor outcomes in severe infection and acute decompensated heart failure. WBC indicates the severity of systemic inflammation and infection, which may contribute to organ dysfunction and clinical deterioration. Abnormal serum potassium may lead to fatal arrhythmias and circulatory instability, while diabetes mellitus contributes to immune dysfunction and exacerbated inflammatory responses, increasing susceptibility to severe infection.

Overall, these predictors suggest that mortality in CHF patients with pulmonary infection is driven by the combined effects of inflammation, metabolic and electrolyte disturbances, and impaired cardiovascular reserve. These findings may support early risk stratification and individualized management in the ICU.

Interestingly, SHAP analysis also revealed negative contributions (i.e., protective effects) for several comorbidities, including hypertension, diabetes, and chronic renal disease. This phenomenon may reflect differences in baseline hemodynamics and clinical management rather than a true protective role of the diseases themselves. Since SHAP values measure the association between each feature and the model’s output—not causal effects—this observed pattern is most likely attributable to confounding by indication. CHF patients with acute infection often experience vasodilation and circulatory insufficiency. However, individuals with chronic disease frequently exhibit chronically elevated sympathetic tone, compensatory vasoconstriction, and higher baseline arterial pressure, which may help maintain perfusion during early infection [51]. Moreover, patients with chronic comorbidities are more likely to receive early fluid resuscitation, nutritional support, and hemodynamic optimization. Collectively, these factors may lower short-term mortality. Therefore, this “protective effect” likely represents clinical management advantages and physiological compensation, rather than a beneficial effect of the comorbidities themselves. Future studies should incorporate medication use, comorbidity severity, trertment timing, and prospective validation to better clarify the prognostic of these factors in CHF patients with pulmonary infection.

Heart failure is a highly heterogeneous syndrome, and the influence of individual risk factors may vary across different HF phenotypes. To address this heterogeneity, we further explored predictors of mortality within each LVEF-defined subgroup. As shown in Supplementary Figure S4–S6, the HFmrEF subgroup included only 11 deaths, its subtype-specific SHAP rankings are statistically unstable and should be interpreted as exploratory. Therefore, our discussion primarily focuses on the HFrEF and HFpEF subgroups.

Patients with heart failure with reduced ejection fraction (HFrEF) are characterized by impaired systolic function and diminished cardiac output. In the setting of concomitant pulmonary infection, reduced systemic perfusion and limited cardiovascular reserve may render these patients particularly vulnerable to acute hemodynamic deterioration. In the present study, SHAP analysis demonstrated that serum potassium, SAPS II score, PaO₂, pH, norepinephrine use, renal disease, creatinine, and serum chloride contributed more substantially to mortality prediction in the HFrEF subgroup. These variables collectively reflect circulatory instability, impaired oxygen delivery, electrolyte disturbances, renal dysfunction, and acid–base imbalance. The prominent role of norepinephrine further suggests that progression toward shock and vasopressor dependence may be a critical pathway leading to death in HFrEF patients. Together, these findings indicate that mortality risk in HFrEF may be closely to acute hemodynamic deterioration and multiorgan hypoperfusion during severe infection.

In contrast, HFpEF is increasingly recognized as a systemic syndrome characterized by chronic inflammation, endothelial dysfunction, and multiple cardiometabolic comorbidities. In our study, diabetes mellitus, LAC, WBC, PT, PLT, hypertension, COPD, HB, and gender emerged as the most influential predictors in the HFpEF subgroup. Notably, many of these variables are associated with chronic metabolic dysregulation, inflammatory activation, coagulation abnormalities, and comorbidity burden rather than acute circulatory failure. Diabetes and hypertension are well-established contributors to myocardial stiffening and microvascular dysfunction, whereas COPD may exacerbate systemic hypoxia and inflammatory stress. A potential mechanistic link among several of these predictors is insulin resistance, which is increasingly understood not as a narrowly glycaemic defect but as a multifactorial, systemic state that impairs endothelial function through reduced nitric oxide bioavailability, sustains chronic low-grade inflammation via cytokines such as TNF-α and IL-6, and promotes the dyslipidaemia and atherogenesis underlying cardiometabolic disease [52]. This framework offers a plausible substrate for the predictors dominating the HFpEF subgroup: hyperglycaemia and the diabetic state reflect the metabolic core of insulin resistance, while the elevated white blood cell count and coagulation-related markers (PT and platelet count) are consistent with the inflammation–thrombosis interface that accompanies it. Furthermore, the importance of coagulation-related markers (PT and platelet count) suggests a potential role of inflammation–coagulation interactions in disease progression. Viewed more broadly, the comorbidity-laden HFpEF phenotype aligns with the concept of panvascular disease, in which atherosclerosis is treated as a systemic process spanning multiple arterial beds and in which insulin resistance and hyperglycaemia contribute to endothelial damage, vascular inflammation, and adverse outcomes whose risk compounds across affected territories [53]. These findings imply that mortality in HFpEF patients with pulmonary infection may be more strongly influenced by chronic cardiometabolic vulnerability and systemic inflammatory responses than by abrupt hemodynamic deterioration.

Our phenotype-specific SHAP analysis revealed distinct mortality-related predictors between HFrEF and HFpEF patients. While HFrEF mortality appeared to be primarily associated with hemodynamic instability, impaired tissue perfusion, and electrolyte imbalance, HFpEF mortality was more closely linked to chronic metabolic-inflammatory comorbidities and coagulation abnormalities. These observations suggest that clinical management strategies may need to be tailored according to heart failure phenotype. For HFrEF patients, maintaining circulatory stability, correcting electrolyte abnormalities, and optimizing oxygen delivery may be particularly important. In contrast, for HFpEF patients, greater attention may be warranted toward metabolic control, management of chronic comorbidities, and mitigation of systemic inflammatory and coagulation disturbances. Such phenotype-specific approaches may facilitate more precise risk stratification and individualized treatment in critically ill CHF patients with pulmonary infection.

Overall, this study developed a machine learning-based model to predict ICU mortality in CHF patients with pulmonary infection, and compared six algorithms. The XGBoost model performed robustly. In the external validation cohort (eICU-CRD), the XGBoost model achieved an AUC of 0.718, a precision of 0.761, a recall of 0.383, and an accuracy of 0.763. Although recall was lower than that observedin internal validation, the model maintained acceptable discriminative ability and a relatively high precision, suggesting that its predictive performance remained reasonably robust across independent datasets.

Several factors may explain the reduced recall. First, the distribution of pneumonia subtypes differed substantially between cohorts. Bacterial pneumonia accounted for 17.05% of cases in MIMIC-IV but 82.19% in eICU-CRD, which may have altered the relationships between predictors and outcomes. Second, patients in the eICU cohort exhibited a higher burden of chronic comorbidities and different LVEF phenotype distributions, indicating a case-mix shift between the derivation and validation populations. Third, some laboratory variables in eICU lacked the reference-range information available in MIMIC-IV, potentially introducing measurement heterogeneity and reducing feature consistency across datasets. Finally, the combination of a relatively preserved AUC and high precision with reduced recall suggests that the model retained its ability to rank risk but may require recalibration of the decision threshold when applied to external populations.

Clinically, the model could be integrated into ICU information systems for real-time risk scoring. Physiological and laboratory data can be automatically extracted to identify high-risk patients early. Unlike traditional “black-box” algorithms, SHAP and LIME provide transparent interpretation and quantify the contribution of each predictor. For instance, when rising LAC or reduced LVEF emerge as dominant risk drivers, clinicians may intensify hemodynamic and perfusion monitoring and cardiac support sooner. This interpretability enhances clinical confidence and supports precision management in CHF patients with pulmonary infection.

Limitations

This study has several limitations. First, it was a retrospective analysis, and some variables were subject to missing data or measurement error, which may introduce selection bias. Second, important clinical factors such as medication regimens, fluid management strategies, and infection-control interventions were not included. The absence of these variables limits the model’s ability to capture dynamic changes in clinical status. Third, the present work focused on comparing individual machine-learning algorithms rather than developing integrated ensemble models. Although model fusion may theoretically improve predictive accuracy, ensemble strategies often increase structural complexity and reduce interpretability, thereby limiting usability in clinical settings. Fourth, the external evaluation assessed discrimination exclusively, and the reduced external recall suggests that the model is designed for risk stratification rather than for standalone triage. Finally, external validation was performed using historical datasets. Future research should involve prospective, multi-center studies and incorporate real-time clinical data to support continuous risk monitoring and improve generalizability.

Conclusion

In this study, we developed and externally validated a machine learning-based model to predict ICU mortality in patients with chronic heart failure and pulmonary infection. Among six algorithms, XGBoost provided the optimal balance of discrimination, calibration, and interpretability. SHAP analysis identified lactic acid, WBC, history of hypertension, serum potassium, diabetes mellitus, and related physiological indicators as the strongest predictors of mortality. Subgroup analysis revealed phenotype-specific risk profiles: in HFrEF patients, mortality was more strongly associated with hemodynamic instability, impaired tissue perfusion, and electrolyte imbalance, whereas in HFpEF patients, it was more closely related to chronic metabolic inflammation and coagulation abnormalties.

Overall, XGBoost demonstrated strong internal performance and generalized to an external cohort with attenuated but still useful discrimination, indicating promise for early risk stratification once recalibrated. Future research should include prospective multicenter validation and real-time integration into clinical information systems to enhance bedside application.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Abbreviations

CHF

Chronic heart failure

HFmrEF

Heart Failure with mildly reduced Ejection Fraction

HFpEF

Heart Failure with preserved Ejection Fraction

HFrEF

Heart Failure with Reduced Ejection Fraction

LVEF

Left Ventricular Ejection Fraction

Author contributions

Yihai Zhai: Writing – original draft, Methodology, Formal analysis, Conceptualization, Funding acquisition. Bingbing Wei: Writing – original draft, Methodology, Conceptualization. Danxiu Lan: Writing – original draft, Data curation. Siying Lv: Writing – original draft, Data curation. Liqin Mo: Writing – review & editing, Supervision, Funding acquisition.

Funding

This study was supported by the National Key Clinical Specialty Project, the 2022 Guangxi Medical and Health Key Discipline Construction Project (Guangxi Health Science and Education Development, 2023-1), and the Guangxi Zhuang Autonomous Region Health Commission Self-Funded Research Project (Z-A20240455).

Data availability

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

Declarations

Ethics approval and consent to participate

The MIMIC-IV database and eICU database contains de-identified patient data and has received institutional review board approval with a waiver of informed consent. Therefore, additional ethics approval and individual informed consent were not required for this retrospective database study.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.