This retrospective modeling study covered modeling, validation, and interpretation. First, the predictive models were preliminarily developed using seven ML algorithms. Then the model performance was fully assessed on the validation set through the area under the curve (AUC) with 95% CI, calibration curve and expected calibration error (ECE), recall, F1-score, positive predictive value (PPV), negative predictive value (NPV), and accuracy, based on which the optimal model was identified. Finally, the optimal model's prediction mechanism and key features were explanatorily analyzed using SHAP and LIME techniques, and an open Web-based prediction tool was deployed.

Data were acquired using electronic medical records from a retrospective cohort of AIHD patients who underwent PCI at the Affiliated Guangzhou Hospital of TCM of Guangzhou University of Chinese Medicine from January 2023 to May 2025. This study was approved by the Institutional Review Board of the hospital (No. 2024NK60), and informed consent was waived because of its retrospective nature.

Following the 10 EPV principle, the minimum number of positive events required for the training set was estimated based on the number of predictor variables included (17).

AIHD patients who underwent PCI at the Affiliated Guangzhou Hospital of TCM of Guangzhou University of Chinese Medicine from January 2023 to May 2025 were included. Inclusion criteria: (a) patients aged >18 years; and (b) patients meeting the diagnostic criteria for AIHD in the current guidelines (18), with clinical symptoms, characteristic electrocardiographic changes, and elevated levels of cardiac biomarkers at the same time, specifically as follows: (i) ST-segment elevation myocardial infarction (STEMI): cardiac troponin (cTn) > 99th percentile upper limit of normal (ULN) or creatine kinase-myocardial band (CK-MB) > 99th percentile ULN, with electrocardiogram (ECG) showing upward convex ST-segment elevation, accompanied by one or more of persistent ischemic chest pain, echocardiogram showing segmental wall motion abnormalities, and coronary angiogram showing abnormalities; (ii) Non-ST-segment elevation myocardial infarction: cTn >99th percentile ULN or CK-MB > 99th percentile ULN, accompanied by one or more of persistent ischemic chest pain, ECG showing new ST-segment depression or flattened/inverted T waves, echocardiogram showing segmental wall motion abnormalities, and coronary angiogram showing abnormalities; (iii) Unstable angina (UA): negative cTn, ischemic chest pain, ECG showing transient ST-segment depression or flattened/inverted T waves, rarely with ST-segment elevation (vasospastic angina). Exclusion criteria: (a) patients with a history of HF; (b) patients who died during hospitalization or with a length of stay <48 h; (c) patients who experienced major clinical events recently (in the last month): major surgery, severe trauma, shock, active infections, or systemic inflammation; (d) patients with severe immune dysfunction; and (e) patients with active malignancy or hematologic malignancy.

The primary outcome of this study was the incidence of in-hospital HF in patients following PCI. The diagnosis of HF was based on the 2022 American College of Cardiology, American Heart Association, and Heart Failure Society of America Guidelines for the Management of Heart Failure (19), specifically as follows: (1) Clinical symptoms and signs: typical symptoms include dyspnea at rest or during physical activity, orthopnea, paroxysmal nocturnal dyspnea, fatigue, and decreased exercise tolerance. Signs include pulmonary crackles, peripheral edema, distension of jugular vein, and positive hepatojugular reflux. (2) Cardiac imaging (primarily echocardiography) revealing structural or functional abnormalities of the heart [e.g., reduced left ventricular ejection fraction (LVEF), cardiac chamber enlargement]. (3) Natriuretic peptide levels: B-type natriuretic peptide (BNP) > 35 pg/mL or N-terminal proB-type natriuretic peptide (NT-proBNP) > 125 pg/mL.

Based on the available literature and clinical practice, 118 features related to the risk of HF were extracted from the electronic medical records, covering demographics, comorbidities, vital signs, laboratory indicators, and angiography findings. They were collected within 24 h post-admission or during PCI. Each variable was independently checked by two researchers and confirmed to have been collected at a time point earlier than the first in-hospital diagnosis of HF for the corresponding patient before it was included in the model analysis. A complete list of features is shown in Supplementary Table S1.

Categorical variables were described by frequencies (percentages), and compared between groups by the chi-square test or Fisher's exact test. Continuous variables of normal distribution were presented as mean ± standard deviation, and compared between groups by the independent-samples t-test; continuous variables of non-normal distribution were presented as median (IQR), and compared between groups by the Mann–Whitney U-test. In the handling of outliers, extreme values were adjusted using two-sided 1% Winsorization (capping method) based on clinical significance and statistical distribution characteristics. Specifically, observations below the 1st percentile were raised to the 1st percentile level, while those above the 99th percentile were lowered to the 99th percentile level. This approach was designed to preserve the potential biological significance of extreme values while effectively reducing their excessive influence on the statistical model. Features with less than 30% missing values were processed by employing the Multiple Imputation by Chained Equations (MICE) method with the mice package in R. To adapt the model to potential nonlinear relationships, a random forest approach was adopted for imputation. Five imputed datasets (m = 5) were generated with a fixed random seed (123) to ensure reproducibility of the results. Features were then screened using L1 regularization based on LASSO regression. The optimal regularization parameter (λ_min) was determined through cross-validation, and key predictive features were selected based on this parameter. Subsequently, for the selected features, Pearson correlation coefficients were used to assess multicollinearity among the variables. Features with high multicollinearity (|r| ≥ 0.8) were excluded to ensure unbiased estimation of model parameters. After the above preprocessing, the retained predictor variables were utilized to establish the predictive model.

Then the samples were randomized into a training set (70%) and a testing set (30%). Seven ML algorithms were used: SVM, KNN, XGBoost, DT, NB, RF, and LR. The models were trained on the training set, and the model performance was assessed on the validation set from the discrimination (AUC with 95% CI), calibration (calibration curve and ECE), and classification (accuracy, recall, F1-score, PPV, and NPV). Finally, based on the assessment results, the optimal model was selected. To thoroughly evaluate the robustness of the stochastic optimal model and mitigate potential bias arising from the limited sample size, the Bootstrap method was employed for internal validation. By performing 1,000 Bootstrap resampling iterations on the dataset, the model was trained using Bootstrap samples in each iteration. Model performance was comprehensively evaluated by calculating the average area under the curve (AUC) and its 95% CI, ECE, recall, F1-score, PPV, and NPV based on out-of-bag (OOB) samples.

In addition, the model's interpretability was enhanced using the SHAP technique. First, we displayed the overall contribution of features to the prediction using SHAP summary plots, and analyzed the relation between individual features and predictions using SHAP dependence plots. Finally, the contribution of important features to the individual prediction was visualized using SHAP force plots and LIME.To operationalize the model, the best-performing algorithm was packaged into an interactive web application using Python's Streamlit framework; this deployment serves both to validate the model's reproducibility in a practical environment and to provide an accessible interface for instant, on-demand predictions.

R 4.4.3 and Python 3.12.7 were utilized for data analysis, modeling, and validation. Statistical significance was assumed at P < 0.05.

Based on the eligibility criteria, 232 AIHD patients were initially screened, of whom six were excluded due to multiple PCI procedures in our hospital, four were excluded due to in-hospital death, one was excluded due to a length of stay <48 h, and 18 were excluded due to a history of surgical procedures, severe infections, and major illnesses. Ultimately, the study cohort encompassed 203 patients (Figure 1).

Following the 10 EPV principle, 12 predictor variables were included: age, AST/ALT, cystatin C, myoglobin, number of diseased vessels, heart rate (HR), absolute neutrophil count, direct bilirubin, lactate dehydrogenase (LDH), mean platelet volume (MPV), platelet count (PLT), and presence or absence of triple vessel disease, with HF as the primary outcome metric. Of these variables, continuous indicators (e.g., age, HR, and myoglobin) each corresponded to one β coefficient, and the dichotomous indicator (presence or absence of triple vessel disease) corresponded to one β coefficient. Therefore, twelve β coefficients were involved in the model, and at least 120 positive outcome events were required for constructing the training set. However, only 40 (<120) positive outcome events were involved in the 142 training samples in this study, suggesting an insufficient sample size.

All patients were assigned to an HF group (n = 55) and a non-HF group (n = 148). The HF group had a significantly higher age of patients (P < 0.001) and a significantly lower proportion of smokers (P = 0.035) than the non-HF group. In the HF group, significant hemodynamic deterioration occurred: elevation of HR (P = 0.005), and significant decreases in systolic and diastolic blood pressure (P < 0.05); coronary artery lesions became significantly more severe: a higher proportion of triple vessel disease (P = 0.019), a higher prevalence of left main coronary artery lesions (P = 0.004), a significant aggravation of stenosis (percentage of reduction of lumen diameter) in the proximal (P = 0.006) and middle (P = 0.001) segments of the anterior descending branch, and a significant difference in the distribution of culprit vessels (P = 0.013). Moreover, the levels of markers of myocardial injury significantly rose in the HF group, and significant inflammatory responses also occurred, as manifested by elevation of neutrophil count, monocyte count (MONO), and D-dimer, accompanied by a decreased lymphocyte percentage. In addition, the HF group exhibited more severe renal impairment and hepatic dysfunction, as well as electrolyte disturbance (hypocalcemia and hyperphosphatemia) (Table 1).

We extracted 118 clinical features within 24 h post-admission and during PCI. Of these, 20 features were excluded due to a data missing rate >30% (Supplementary Table S2). Subsequently, the remaining features were handled by MICE, and features in which hemoglobin was automatically identified as perfect collinearity were excluded. Finally, 97 clinical features were included for modeling (Supplementary Table S3). To identify characteristic variables associated with the risk of in-hospital HF in AIHD, LASSO regression was conducted. The optimal regularization parameter (λ_min) was determined by cross-validation, based on which predictor variables were screened. Ultimately, 12 predictor variables significantly associated with prognosis were screened (Supplementary Figure S1). Spearman's correlation coefficient analysis revealed that all predictor variables exhibited coefficients below 0.8, effectively preventing multicollinearity (Supplementary Figure S2). Bases on these 12 features, seven ML models were constructed for predicting the HF risk: XGBoost, RF, NB, LR, SVM, KNN, and DT. Detailed specifications and hyperparameter settings for each ML algorithm are provided in Supplementary Table S4.

The AUC with 95% CI for each model is illustrated in Figure 2: Both RF (AUC 0.70; 95% CI 0.53–0.84) and XGBoost models (AUC 0.72; 95% CI 0.56–0.86) performed well in the AUC. The calibration curve and ECE for each model are illustrated in Figure 3: Both RF and NB models displayed high consistency between the predicted and actual risks, suggesting their good calibration performance. To assess the model performance more objectively, the recall, F1-score, PPV, and NPV on the validation set were further calculated and ranked. The mean rank of each model was used as its final overall rank (Supplementary Table S5), and the results were visualized by radar charts (Figure 4). To sum up, the RF model demonstrated optimal overall performance: AUC 0.70 (95% CI 0.53–0.84), ECE 0.08, Recall 0.33, F1-score 0.42, accuracy 0.77, PPV 0.56, and NPV 0.81 on the validation set. Therefore, the RF model was selected for the subsequent prediction.

To further assess the robustness of the model with a small sample size, 1,000 bootstrap replications were performed for internal validation. The validation results showed that the adjusted AUC of the model was 0.73 (95% CI: 0.63–0.83), with ECE = 0.11, recall = 0.27, F1-score = 0.35, accuracy = 0.74, PPV = 0.55, and NPV = 0.77 (Table 2). Compared to the results from the original test set, Bootstrap validation not only confirmed the discriminatory capability of the model (AUC > 0.70) but also provided a more precise performance estimate (with the width of CI reduced from 0.31 to 0.20). The optimism of the model was −0.03, indicating good robustness of the original model. (Supplementary Figure S3).

The feature importance was assessed by SHAP values for the RF model on the validation cohort. The top 12 features based on the SHAP values (Figure 5A) and their positive (an increased risk) or negative (a decreased risk) contribution to the model's predictive efficacy (the SHAP summary plot in Figure 5B) were illustrated. The results revealed that the elevation of age, AST/ALT, cystatin C, myoglobin, number of diseased vessels, HR, absolute neutrophil count, direct bilirubin, LDH, and MPV, and the presence of triple vessel disease all made positive contributions, i.e., they increased the risk of HF. A decrease in PLT was also identified as a feature increasing the HF risk. To more deeply investigate how key features influence the probability of HF predicted by the RF model, a SHAP dependence plot was drawn for the six features (age, HR, AST/ALT, LDH, MONO#, and PLT) (Figure 6). The elevation of age, HR, AST/ALT, LDH, and MONO# was closely linked to an increased risk of HF, whereas the elevation of PLT was associated with a decreased risk.

To further explore the application pattern of the RF model in specific patients, the HF risk was predicted for four patients (two HF and two non-HF) from the validation cohort, and visualized using the SHAP force plot (Figure 7). For the first HF patient, the predicted probability of HF was 0.62, and the main risk factors for HF were HR (111 bpm) and MPV (10.6 fL). For the second HF patient, the predicted probability of HF was 0.56, and the main risk factors were age (80 years) and the number of culprit vessels (4). For the two non-HF patients, the predicted probability of HF was 0.52 and 0.51, and the main protective factors were MONO (0.24 × 10⁹/L) and age (53 years), respectively.

Additionally, the HF risk was also predicted for two patients by the LIME technique (Figure 8). For the first HF patient, the predicted probability of HF was 0.67, and the main risk factors for HF were age (81 years) and HR (94 bpm), while the protective factors were the absence of triple vessel disease and cystatin C (0.97 mg/L). For the second HF patient, the predicted probability of HF was 0.66, and the main risk factors for HF were age (81 years) and AST/ALT (5.58), while the protective factor was PLT (290 × 10⁹/L). For the two non-HF patients, the predicted probability of HF was 0.30 and 0.10, and the main protective factors were HR (58 bpm) and MONO (0.39 × 10⁹/L), respectively.

To promote the reproducibility and clinical application, the final prediction model has been deployed as an open-access Web-based tool freely accessible at https://hspifojdhapwpa7wwwadks.streamlit.app.