Establishment and Validation of an Interpretable Machine Learning-Based Predictive Model for Risk of Post-PCI In-Hospital Heart Failure in AIHD Patients

Xinying Zhao¹Zhihang Wang¹Qiqi Yang¹Huiqi Liu²Yigen Li^1,3Xi Ye^1*

• ¹The Affiliated Guangzhou Hospital of TCM of Guangzhou University of Chinese Medicine, Guangzhou, China
• ²School of Computer Science, University of Bristol, Bristol BS8 1UB, United Kingdom, Bristol, United Kingdom
• ³National TCM Master Lin Tiandong’s Heritage and Inheritance Studio, The Affiliated Guangzhou Hospital of TCM of Guangzhou University of Chinese Medicine, Guangzhou Guangdong 510130, China, Guangzhou, China

BACKGROUND: This study intends to establish and validate an interpretable machine learning (ML) model based on clinical features for early prediction of the risk of post-percutaneous coronary intervention (PCI) in-hospital heart failure (HF) in patients with acute ischemic heart disease (AIHD). METHODS:This study retrospectively included AIHD patients who underwent PCI at the Affiliated Guangzhou Hospital of TCM of Guangzhou University of Chinese Medicine from January 2023 to May 2025. LASSO regression was utilized for feature screening first, and then seven predictive models for HF risk in AIHD patients were established using ML algorithms. 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, recall, F1-score, positive predictive value, negative predictive value, and accuracy, and internal validation was conducted using the Bootstrap method. In addition, feature importance was evaluated by SHapley Additive exPlanations (SHAP) values, and individualized predictions were explained by Local Interpretable Model-Agnostic Explanations (LIME). RESULTS: Two hundred and three patients with AIHD were ultimately included, of whom 55 (27.1%) experienced in-hospital HF. Of the seven ML models, the random forest (RF) model demonstrated optimal performance on the validation set, with an AUC of 0.70 (95% CI 0.53-0.84) This is a provisional file, not the final typeset article and an accuracy of 0.77; the calibration curve revealed high agreement between predicted and actual risks. Twelve predictive features associated with endpoint events were identified by LASSO regression, and the top five features contributing to the predictive efficacy of the RF model were age, monocyte count, heart rate, platelet count, and mean platelet volume according to the ranking of feature importance. In addition, the contribution of features to the prediction of HF risk was visualized by SHAP summary plots and LIME.Finally, an open Web-based prediction tool was deployed. CONCLUSION: This study developed a random forest model to predict in-hospital HF risk after PCI in AIHD patients. Using SHAP and LIME methods significantly enhanced the model's clinical interpretability. Future research with larger samples is needed to optimize training and validate model generalizability.