Abstract
Laboratory-guided methods have the potential to provide robust mortality prediction for acute respiratory distress syndrome (ARDS) which could improve timely intervention. The objective of this investigation was to predict mortality using regression and machine learning techniques using biomarkers linked to ARDS pathophysiology, including matrix metalloproteinase-3 (MMP-3) and club cell-secretory protein-16 (CC16), and general inflammatory biomarkers. 89 adult patients from the "Aerosolized β₂-agonist for treatment of acute lung injury" (ALTA) trial were randomly separated into training (n = 53), validation (n = 8), and test (n = 28) sets. Logistic regression and supervised machine learning (ML) models were developed. In total, 20 ICU predictors including baseline characteristics (age, sex, APACHE III score, sepsis, vasoactive agent, PaO(2)/FiO(2)) and baseline and serial biomarkers were included. The primary outcome was area under the receiver operating characteristic (AUROC) for 90 day mortality. Random Forest, Support Vector Machine (SVM), and XGBoost achieved AUROCs of 0.917, 0.705, and 0.955, respectively. Stepwise regression achieved an AUROC of 0.508. For the highest performing model (XGBoost), MMP3-based variables were the most important features. ML had high predictive ability for 90-day mortality, and MMP-3 demonstrates moderate-to-high feature importance in ML models. These findings support using pathophysiology-derived biomarkers in ML models for ARDS prediction.