Interpretable artificial intelligence model for predicting heart failure severity after acute myocardial infarction

Abstract

Background: Heart failure (HF) after acute myocardial infarction (AMI) is a leading cause of mortality and morbidity worldwide. Accurate prediction and early identification of HF severity are crucial for initiating preventive measures and optimizing treatment strategies. This study aimed to develop an interpretable artificial intelligence (AI) model for HF severity prediction using multidimensional clinical data.

Methods: This study included data from 1574 AMI patients, including medical history, clinical features, physiological parameters, laboratory test, coronary angiography and echocardiography results. Both deep learning (TabNet, Multi-Layer Perceptron) and machine learning (Random Forest, XGboost) models were employed in constructing model. Additionally, the Shapley Additive Explanation (SHAP) method was used to elucidate clinical factors importance and enhance model interpretability. A web platform ( https://prediction-killip-gby.streamlit.app/ ) was also developed to facilitate clinical application.

Results: Among the models, TabNet demonstrated the best performance, achieving an AUROC of 0.827 for KILLIP four-class classification and 0.831 for KILLIP binary classification. Key clinical factors such as GRACE score, NT-pro BNP, and TIMI score were highly correlated with KILLIP classification, aligning with established clinical knowledge.

Conclusions: By leveraging easily accessible multidimensional data, this model enables accurate early prediction and personalized diagnosis of HF risk and severity following AMI. It supports early clinical intervention and improves patient outcomes, offering significant clinical application value.

Clinical trial number: Not applicable.

Keywords: Acute myocardial infarction; Artificial intelligence; Deep learning; Heart failure.

Fig. 1

Flow chart of the study design

Fig. 2

ROC Curves of machine learning and deep learning models. Fivefold cross-validation was performed in all the 1574 patients. (A). ROC Curves for four-class KILIIP classification (B). ROC Curves for binary KILLIP classification

Fig. 3

Feature importance by the SHAP method for the Tabnet model. (A) SHAP summary bar plot derived from 1574 patients. (B) SHAP summary dot plot for KILLIP 1 classification (1005 patients). The colors of the dots represent the actual feature values for each patient, with red indicating higher values and blue indicating lower values. Dots are stacked vertically to represent density

Fig. 4

Global model explanation by the SHAP method for the TabNet model. SHAP dependence plot for KILLIP 1 classification (1005 patients). Each dot represents a patient and shows how a single feature affects the model’s output. SHAP values greater than zero push the decision toward the “KILLIP 1” class

Fig. 5

Local model explanation by the SHAP method for the TabNet model. (A1–D1, A2-D2, A3-D3, A4-D4) represent prediction result plots for randomly selected patients from each KILLIP class 1 through 4. The raw data for each patient is presented in Appendix Table 1

Fig. 6

Force SHAP value plot for the test set (315 patients). Each patient is represented along the x-axis, while the contributions of features are shown on the y-axis. A larger red area for an individual patient indicates a higher probability of the prediction being classified as “KILLIP 1.”

Fig. 7

The web platform for KILLIP classification prediction model