This retrospective cohort study included 109 inpatients who developed heart failure following non-cardiac surgery and 380 inpatients who remained free of heart failure during the perioperative period. All participants were admitted to the Sixth Medical Center of the Chinese PLA General Hospital between October 8, 2013, and September 18, 2024, and underwent at least one non-cardiac surgical procedure during hospitalization. New-onset postoperative heart failure is defined as: heart failure that occurs for the first time after surgery during the index hospitalization in patients without pre-existing chronic heart failure prior to the operation. For external validation, an independent cohort of 10,492 patients—with no overlap with the primary dataset—was obtained from the First Medical Center of the Chinese PLA General Hospital. Inclusion criteria were: (1) age ≥18 years; (2) hospital stay exceeding 2 days with at least one documented non-cardiac surgery; (3) no prior history of cardiac surgery, including valve replacement, coronary artery bypass grafting, or percutaneous coronary intervention; (4) for patients having multiple non-cardiac surgeries, only the first procedure was included. Using SQL, we applied exclusion criteria based on ICD-10 codes (I50.105 for chronic left ventricular dysfunction and I50.908 for chronic heart failure) from preoperative, admission, and historical diagnoses. The final positive cohort consisted solely of patients whom clinicians had assigned a general I50.x diagnosis for heart failure. The patient selection process is illustrated in Figure 1.

Clinical variables were retrospectively extracted from the Oracle database of the Sixth Medical Center of the Chinese PLA General Hospital using Structured Query Language (SQL). The collected variables included: (1) Demographics: age, sex, and ethnicity; (2) Comorbidities: atrial fibrillation, acute coronary syndrome, malignant arrhythmias, primary cardiomyopathies, and others; (3) Laboratory tests: hemoglobin, white blood cell count, platelets count, blood urea nitrogen, serum creatinine, glucose, among others; (4) Vital signs: heart rate, respiratory rate, systolic blood pressure, and diastolic blood pressure. A total of 137 clinical features were initially extracted. The diagnostic criteria for postoperative new-onset heart failure were based on the ICD-10 codes specified in the《2022 ESC Guidelines on cardiovascular assessment and management of patients undergoing non-cardiac surgery》. These include: I50.0 Congestive heart failure (including I50.000 Congestive heart failure, I50.001 Right heart failure, I50.002 Combined heart failure); I50.1 Left ventricular failure (including I50.100 Left ventricular failure, I50.101 Acute left ventricular failure, I50.102 Left atrial failure, I50.103 Left heart failure with pulmonary edema, I50.104 Cardiac asthma); I50.9 Heart failure, unspecified (including I50.900 Heart failure, I50.906 Myocardial impairment, I50.907 Acute heart failure); and I51.7 Cardiomegaly (including I51.700 × 009 Cardiac enlargement) (10, 11). As a retrospective study, the requirement for informed consent was waived.

The raw dataset underwent the following preprocessing steps: (1) surgical risk stratification was conducted according to the 2022 ESC Guidelines on Cardiovascular Assessment and Management of Patients Undergoing Non-Cardiac Surgery. All surgical and interventional procedures performed during hospitalization were categorized into three risk levels (11); (2) features with more than 30% missing values were excluded, reducing total number of features from 137 to 69; (3) categorical variables were converted into numerical format using label encoding; (4) missing values were filled using median imputation via the Simple Imputer module in Python’s scikit-learn library. (5) Outliers were identified based on interquartile range (IQR) method [Q1–1.5 × IQR, Q3 + 1.5 × IQR] [Q1–1.5 × IQR, Q3 + 1.5 × IQR] [Q1–1.5 × IQR, Q3 + 1.5 × IQR] and clinical judgment, and were replaced with the median value of the respective feature.

The dataset was randomly divided into a training set (80%) and a testing set (20%). The training set included 391 patients (87 postoperative heart failure cases, 304 controls), and the testing set comprised 98 patients (22 postoperative heart failure cases, 76 controls). To address class imbalance in the training set, a hybrid method combining random under sampling and Synthetic Minority Oversampling Technique (SMOTE) was used: the negative-to-positive ratio was reduced to 70% by under sampling, followed by Synthetic Minority Oversampling Technique to upsample the minority class. Eight binary classification algorithms were developed: Naive Bayes (NB), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Logistic Regression (LR), Decision Tree (DT), AdaBoost, XG Boost, and Random Forest (RF). Five-fold cross-validation was employed to enhance model robustness and reduce overfitting. Model performance was evaluated using accuracy, sensitivity, specificity, precision, F1 score (harmonic mean of precision and recall), and area under the receiver operating characteristic curve. The model with the best overall performance was selected for further analysis. Due to the imbalanced distribution of negative and positive samples, the negative samples in the original model were undersampled, and the model was retrained. The model’s calibration was assessed by calculating the Brier score and plotting calibration curves on both internal and external validation datasets. Additionally, we conducted Decision Curve Analysis (DCA) to demonstrate its net clinical benefit and plotted a Clinical Impact Curve (CIC) to determine the optimal threshold probability. This comprehensive visual analysis enabled a rigorous evaluation of the model’s clinical utility and the identification of the most effective decision threshold for practical application.

To improve model interpretability and reduce computational complexity, the most predictive features were identified. Initially, 43 features with an importance score greater than 0.001 were selected based on model-derived feature importance rankings. Next, Recursive Feature Elimination (RFE) with five-fold cross-validation was applied to determine the optimal subset of features. Finally, Pearson correlation analysis was performed to eliminate redundant variables; among highly correlated pairs (correlation coefficient r > 0.6), only the most informative feature was retained.

To evaluate the generalizability, both internal and external validation were conducted. Internal validation was performed using the held-out testing set comprising 98 patients (22 postoperative heart failure cases and 76 controls). External validation utilized an independent dataset of 5,585 patients from the First Medical Center of the Chinese PLA General Hospital, including 1.536 postoperative heart failure cases and 4,049 controls. This external cohort was used to assess the model’s robustness and applicability across different clinical settings and patient populations. Performance metrics from both validation sets were compared to verify the model’s reliability and clinical utility.

To enhance model transparency and clinical trust, SHAP were used to interpret prediction results. SHAP values quantify each feature’s contribution to the model output at both global and individual levels, allowing clinicians to understand how specific variables influence risk estimates. This interpretability supports personalized clinical decision-making and facilitates integration of the model into routine practice.

All analyses were conducted using Python 3.8.3. Continuous variables were presented as means ± standard deviations, and categorical variables as frequencies and percentages. Between-group comparisons (postoperative heart failure vs. non-postoperative heart failure) were performed using the student’s t-test for continuous variables and the chi-square test for categorical variables. Pearson correlation was used to assess associations between continuous variables. A two-sided p-value < 0.05 was considered statistically significant.

Compared to the non-heart failure group, patients in the HF group had a significantly higher proportion of males (60.6% vs. 48.2%, p < 0.05), an older mean age (73.9 vs. 53.8 years, p < 0.001), and a greater rate of high-risk surgeries (27.5% vs. 16.3%, p < 0.05). Additionally, the HF group had a higher prevalence of comorbidities, including acute coronary syndrome, acute kidney injury, acute exacerbation of chronic obstructive pulmonary disease (AECOPD), atrial fibrillation, chronic kidney disease, coronary artery disease, diabetes or ketoacidosis, myocardial infarction, cor pulmonale, and severe anemia (all p < 0.05).

Significant physiological and laboratory differences were also observed: the HF group had lower diastolic blood pressure (70.3 vs. 74.8 mmHg, p < 0.001) but higher pulse (89.1 vs. 75.3 bpm, p < 0.001) and respiratory rates (21.1 vs. 19.1 breaths/min, p < 0.001). Furthermore, they exhibited significantly lower levels of absolute lymphocyte count, albumin, serum calcium, cholesterol, estimated glomerular filtration rate, high-density lipoprotein cholesterol, mean corpuscular hemoglobin concentration, platelet count, serum phosphorus, and total protein, alongside elevated levels of absolute neutrophil count, alkaline phosphatase, activated partial thromboplastin time, blood glucose, direct bilirubin, fibrinogen, gamma-glutamyl transferase, international normalized ratio, neutrophil percentage, neutrophil-to-lymphocyte ratio, prothrombin time, serum creatinine, total bilirubin, and white blood cell count (all p < 0.05), as summarized in Table 1.

Eight binary classification models were developed using the training cohort, based on the following machine learning algorithms: Naive Bayes, K-Nearest Neighbors, Support Vector Machine, Logistic Regression, Decision Tree, AdaBoost, XG Boost, and random forest. Model performance was assessed using five-fold cross-validation, and the average metrics across folds are summarized in Table 2. The corresponding ROC curves are presented in Figure 2A. Among all models, the random forest classifier demonstrated the best overall performance, achieving the highest area under the ROC curve of 0.919, along with an accuracy of 0.849, sensitivity of 0.806, specificity of 0.879, and precision of 0.824. Based on these results, the random forest model was selected as the final predictive model for further validation and interpretation.

Initially, 137 features were collected. After excluding those with over 30% missing data, 69 features remained. Based on importance scores from the random forest model, 43 features with a value greater than 0.001 were retained (Supplementary Figure S1). Recursive feature elimination with five-fold cross-validation was then used to evaluate model performance across different feature subsets (Figure 2B; Supplementary Table S1). Performance metrics—including AUROC, accuracy, specificity, and precision—plateaued after the top 10 features, while sensitivity and F1 score declined. To this end, a comprehensive assessment of the model’s performance at various decision thresholds is provided in Supplementary Table S2. As shown in Supplementary Figure S2, strong correlations were observed between albumin and total protein, among absolute lymphocyte count, absolute neutrophil count, neutrophil percentage, and NLR, and between international normalized ratio (INR) and prothrombin time. As the Neutrophil-to-Lymphocyte Ratio (NLR) is calculated from the absolute neutrophil count and absolute lymphocyte count, these two individual components were removed from the model to avoid multicollinearity. Consequently, albumin, NLR, eGFR, and INR were retained for the final model. Ultimately, 10 features were selected for the final model: age, albumin, neutrophil-to-lymphocyte ratio (NLR), blood glucose, international normalized ratio (INR), pulse rate, mean corpuscular hemoglobin concentration (MCHC), serum creatinine, estimated glomerular filtration rate (eGFR), and diastolic blood pressure. The model’s generalizability was assessed using an external validation cohort, with its calibration performance across various risk deciles showcased in Supplementary Table S3. Furthermore, the Precision-Recall curves for both the internal and external validation cohorts are provided in Supplementary Figures S3 and S4, respectively. To elucidate the interactions between the final predictive features and their collective impact on the model’s output, we performed a SHAP analysis, with the resulting interaction value matrix displayed in Supplementary Figure S5.

The internal validation set comprised 98 samples, comprising 22 positive and 76 negative cases. The final random forest model demonstrated strong performance in this cohort (Figures 3A,B), achieving an accuracy of 0.857, sensitivity of 0.864, specificity of 0.855, precision of 0.633, F1 score of 0.731, and an AUROC of 0.923. Of the 76 negative cases, 65 were correctly classified, while 22 of the 19 positive cases were correctly identified. In the external validation cohort, consisting of 5,585 samples (1,536 positive and 4,049 negative cases), the model achieved an accuracy of 0.619, sensitivity of 0.813, specificity of 0.810, and an AUROC of 0.878 (Figure 3C). A total of 3,280 out of 4,049 negative cases and 1,248 out of 1,536 positive cases were correctly predicted (Figure 3D).

To comprehensively evaluate the clinical value of our predictive model, we performed calibration, Decision Curve Analysis (DCA), and Clinical Impact Curve (CIC) analysis on the Random Forest (RF) model.

As shown in Figure 4A, the calibration curves for both the internal test set and external validation set (the x-axis represents the predicted risk of postoperative heart failure, while the y-axis indicates the observed actual incidence) demonstrated good agreement between predictions and observations, with the RF model’s performance (solid line) closely following the ideal diagonal (dashed line). This was supported by Brier scores of 0.093 and 0.123 for internal and external validation, respectively. The DCA (Figure 4B) revealed that across a threshold probability range of 0–1.0, the RF model provided a superior net clinical benefit compared to the strategies of “intervening on all” or “none.” The CIC (Figure 4C) further illustrated the model’s performance across thresholds. In a sample of 1,000 patients, the number of patients identified as high-risk by the model (blue curve) closely approximated the actual number of true positive cases (red curve) as the risk threshold increased.

Based on this comprehensive analysis, a threshold probability of 0.6 was selected as optimal for clinical decision-making. This threshold effectively balances sensitivity and specificity, minimizing both unnecessary interventions from false positives and the risk of missing true positive cases.

To enhance interpretability and identify key factors influencing postoperative heart failure following non-cardiac surgery, SHAP was applied to analyze the random forest model. Feature importance rankings based on SHAP values are shown in Figure 5A, highlighting albumin, age, and estimated glomerular filtration rate as the most influential predictors among the 10 selected features. The mean absolute SHAP values for the 10 predictors are as follows: Age: 0.27, Albumin: 0.18, NLR: 0.16, Blood Glucose: 0.12, INR: 0.08, Pulse Rate: 0.07, MCHC: 0.05, Serum Creatinine: 0.04, eGFR: 0.03, Diastolic Blood Pressure: 0.02. Figure 5B presents the distribution of SHAP values for each feature. The x-axis represents the SHAP value, and dot color indicates the magnitude of the corresponding feature value (red = high, blue = low). A positive SHAP value indicates increased predicted risk of postoperative heart failure, while a negative value suggests a protective effect. The model’s predictions were primarily driven by age, which showed a strong positive correlation with the risk of pHF. This positive association was also shared by NLR, blood glucose, INR, pulse rate, and serum creatinine. In contrast, albumin, MCHC, eGFR, and diastolic blood pressure had a negative correlation, implying a protective effect, as lower values of these features corresponded to a higher pHF risk. Beyond global feature influence, SHAP analysis also enabled interpretation at the individual prediction level. Feature contributions for specific cases are illustrated in Figure 5C (positive case) and Figure 5D (negative case). In the positive case, features such as mean corpuscular hemoglobin concentration, blood glucose, and neutrophil-to-lymphocyte ratio contributed to increased postoperative heart failure risk. Conversely, in the negative case, albumin, age, and estimated glomerular filtration rate were associated with lower risk. The SHAP dependence plots were further used to visualize the impact of individual features on predictions. In these plots, the x-axis represents feature values, and the y-axis indicates SHAP values, with a LOWESS-smoothed line showing the overall trend. As shown in Figure 6, The Random Forest model identified specific inflection points for key predictive features, beyond which the risk of postoperative heart failure notably changed. These critical values were: Age: 68.92 years; Albumin: 36.59 g/L; NLR: 5.18; Serum Creatinine: 104.96 μmol/L; eGFR: 82.64 mL/min/1.73 m²; Blood Glucose: 5.85 mmol/L; Pulse Rate: 86.53 beats/min; INR: 1.04; Diastolic Blood Pressure: 72 mmHg; and MCHC: 330.52 g/L. It is noteworthy that these values represent the physiological ranges at which the model identifies a steep increase in the risk of postoperative heart failure; however, they themselves are not clinically optimized decision thresholds.

To enhance clinical utility, the final model (AUROC: 0.923) was optimized and deployed as a user-friendly, web-based application.¹ This streamlined tool enables clinicians to input recent preoperative laboratory values for rapid estimation of postoperative heart failure risk. In addition, the application provides SHAP-based visual explanations of individual predictions, offering clinicians transparent insights into how each feature contributes to the risk estimate and supporting informed clinical decision-making.