Data-Driven Prediction of Hospital-Acquired Infections in Greek ICUs: Comparative Analysis of Machine Learning and Traditional Techniques

Hospital-acquired infections (HAIs) constitute a critical challenge in intensive care units (ICUs). In Greece, data limitations due to incomplete electronic record implementation hinder advanced risk management. This study aims to compare the predictive performance of machine learning algorithms versus traditional techniques for HAI risk stratification in non-digitized ICU settings. We analyzed 1500 anonymized ICU patient records from two general hospitals in Athens, comparing Logistic Regression against LASSO, Random Forest, and Gradient Boosting Machine (GBM), while unsupervised K-means clustering identified latent patient profiles. Invasive devices were found to be the strongest predictors; GBM identified urinary catheter use as the dominant factor (Relative Influence Score: 36.66), and all models showed high performance convergence (AUC ≈ 0.85 to 0.919). K-means clustering (Silhouette width 0.47) revealed co-occurring risk factors, such as smoking and alcohol (r = 0.60). Combining machine learning with classical statistics enhances risk stratification, offering superior granularity in risk ranking to support targeted device stewardship and individualized prevention strategies in ICUs.