AI-Driven Prediction of ICU Admission After Left Heart Catheterization: A Machine Learning Approach

This study developed an AI-driven predictive model for ICU admission following Left Heart Catheterization (LHC) using the MOVER database (1,779 procedures, 1,591 patients, 2017-2022). A novel three-stage hybrid feature selection methodology-combining mutual information filtering, recursive feature elimination with cross-validation (RFECV), and Lasso L1 regularization-reduced dimensionality from 48 to 12 features (75% reduction). Eight machine learning algorithms were evaluated on ADASYN-balanced data with 10-fold stratified cross-validation, with the optimized CatBoost model achieving exceptional performance: AUROC 0.973, accuracy 93.2%, F1-score 93.2%, MCC 0.897, and false negative rate 7.14%. SHAP analysis identified disch_disp_home routine (SHAP value: 1.023) as the dominant predictor, followed by device_foley and med_beta_blocker. The three-tier risk stratification system demonstrated strong calibration with actual ICU rates of 3.1% (low risk), 9.1% (moderate risk), and 92.9% (high risk). This clinically deployable model addresses critical gaps in perioperative AI through rigorous feature selection, transparent interpretability via SHAP/LIME analysis, and reproducible methodology using open-access data, facilitating evidence-based ICU resource planning and preoperative decision support.