Interpretable machine learning model for predicting the ICU readmission of patients after coronary intervention

Abstract

To develop and validate an interpretable machine learning (ML) model integrated with inflammatory indicators for predicting ICU readmission in patients after coronary intervention (PCI/CABG). We used retrospective data from MIMIC-IV (training, 2196 patients) and MIMIC-III (external validation, 1717 patients). Patients aged > 18 years with complete hematological data and < 40% missing data were included in the study. LASSO regression selected 20 clinical variables; nine ML algorithms (e.g., RF, XGBoost) were compared. The optimal model was evaluated using AUC, calibration curves, DCA, and CIC, with SHAP for interpretability. In MIMIC-IV, 674 (30.7%) patients had multiple ICU admissions, and 440 (25.6%) in MIMIC-III. RF outperformed the other models (internal AUC = 0.85, external AUC = 0.706), with balanced accuracy (0.772), sensitivity (0.748), and specificity (0.795). The key predictors included MLR, maximum creatinine levels, invasive procedures, and so on. The model was optimized for 12 variables and deployed online. The RF model achieves moderate predictive performance for ICU readmission risk of post-coronary intervention patients within the same single tertiary hospital across different periods, and SHAP improves model interpretability to facilitate screening high-risk patients. SHAP enhances interpretability, aiding in the early identification of high-risk patients and improving outcomes.