Answers are grounded in this paper and cite their sources.
Ventilator-associated pneumonia (VAP) affects up to 20% of critically ill patients, prolonging hospital stay, increasing costs and contributing to mortality, yet its diagnosis remains challenging. Most machine-learning models for VAP rely on administrative codes or microbiological surrogates and provide a single static prediction. We conducted a retrospective single-centre study of 2,817 adults receiving invasive mechanical ventilation for at least 48 h (2014–2025), 334 of whom developed VAP, using physician-documented diagnoses with exact onset dates and hourly electronic health record data. Twelve algorithms, multiple feature-selection strategies and three prediction horizons (1, 24 and 48 h before onset) were compared. The final CatBoost model, based on 18 routinely collected variables, predicted VAP 24 h before clinical diagnosis with an area under the ROC curve (AUC) of 0.86 and a sensitivity of 0.76 in the held-out test set. SHapley Additive exPlanations (SHAP) identified antibiotic administration, maximum positive end-expiratory pressure, maximum temperature, maximum fraction of inspired oxygen and purulent secretions as the main predictors. The model has been deployed in silent mode, generating hourly, explainable risk estimates with full traceability. External and prospective validation are needed before its clinical impact can be established.
No takes yet. Share an insight, caveat, or question.
Berrueta et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: