A Clinically Actionable Risk Score for Ventilator-Associated Pneumonia Using Interpretable Machine Learning

Le résumé fourni par la source

Background Ventilator-associated pneumonia (VAP) remains the most common ICU-acquired infection, with high morbidity and substantial mortality. Diagnosis often is delayed, reducing the potential benefit of timely treatment. Research Question Can a machine learning-based VAP risk score (VRS) enable dynamic daily risk prediction of VAP in patients receiving ventilation for ≥ 48 hours? Study Design and Methods Data were obtained from the prospective Perpetual Observational Study of Ventilator-Associated Pneumonia cohort, which enrolled 3,571 adults receiving mechanical ventilation for ≥ 48 hours across 25 European ICUs between 2022 and 2024. VAP was defined according to US Food and Drug Administration criteria. Thirty-eight candidate predictors were selected based on clinical relevance. Several machine learning algorithms were trained, and a reduced model was used to derive a continuous VRS ranging from 0 to 100. This score was stratified into 3 risk categories (low, medium, and high), each associated with increasing VAP prevalence. Sensitivity analysis included the use of physician-diagnosed VAP and evaluation of the predictive performance. External validation was carried out with data from 4 geographically distinct ICUs. Results The reduced model used 5 clinical variables collected at intubation (Sequential Organ Failure Assessment score, Acute Physiology and Chronic Health Evaluation II, Charlson Comorbidity Index, age, and number of VAP prevention strategies applied) and the ongoing duration of mechanical ventilation, enabling a daily updated VRS. In the external validation cohort (n = 556), VAP prevalence was 5.5% (95% CI, 3.4%-8.7%) in the low-risk (n = 18/326), 16.8% (95% CI, 11.9%-23.0%) in the medium-risk (n = 32/191), and 61.5% (95% CI, 44.7%-76.2%) in the high-risk (n = 24/39) VRS category. Likelihood ratios for VAP were 0.4, 1.3, and 10.4, respectively, confirming meaningful separation. The VRS achieved a precision recall area under the receiver operating characteristic curve of 0.75 (95% CI, 0.71-0.79). Similar results were observed for sensitivity analysis. Interpretation Our results show that the VRS allows daily stratification of VAP risk in patients receiving ventilation for ≥ 48 hours using 6 clinical variables. It provides a reproducible and externally validated tool that may help clinical decision-making. Clinical Trial Registration ClinicalTrials.gov; No.: NCT05719259; URL: www.clinicaltrials.gov.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.