Mining Association Rules for ICU Patients: Unveiling Patterns in in-Hospital Mortality Prediction

Forecasting in-hospital mortality for ICU patients is a key challenge in clinical decision-making. This study investigates a novel approach that utilizes Mining Association Rules (MAR) for feature selection to enhance both the precision and interpretability of mortality prediction models. Using a publicly available ICU dataset with heart failure, we performed rigorous data preprocessing, exploratory analysis, and applied synthetic oversampling to handle class imbalance. We mined over 33,000 rules, of which 1,200 were initially evaluated and 26 were ultimately selected based on lift and support criteria. To assess this approach, we trained multiple machine learning (ML) algorithms using the original, MAR-derived, and optimized feature sets. Among them, the Extra Trees Classifier demonstrated the highest predictive performance with the 26 selected features, achieving 97 % accuracy, F1-score, precision, and recall. Furthermore, it registered $\text{RMSE}=0.202, \text{MAE} =0.041$, and $\mathrm{R}^{2} =0.835$, indicating robust and reliable predictions. These results underscore the potential of MAR-guided feature engineering to improve clinical outcome prediction in a clear and effective way.