Machine learning in ARDS: an intensivist’s guide to artificial intelligence applications

Acute Respiratory Distress Syndrome (ARDS) is a heterogeneous, life-threatening condition with persistent diagnostic uncertainty, limited causal treatments, and high mortality in intensive care. ARDS management remains largely supportive, and conventional scores and trial designs have struggled to account for its biological and clinical heterogeneity. This narrative review summarizes current and emerging applications of Artificial Intelligence (AI) in ARDS and mechanical ventilation. It focuses on how Machine Learning (ML) can refine early risk prediction, diagnosis, phenotyping, management and outcome prediction, while outlining the strengths, weaknesses and nuances of these applications. ML models using electronic health records, imaging, physiological waveforms and omics data show strong performance for predicting ARDS onset, enabling early diagnosis, optimising management and forecasting outcomes. These models show performance equivalent to and often outperform traditional guidelines and scores. However, most of these models remain limited to the research setting and show limited clinical adoption. Most studies are retrospective, single-center and lack rigorous external validation, limiting generalizability and real-world impact. Additional challenges include data quality and bias, poor calibration and scarce decision-curve analyses, limited interpretability, and the absence of prospective trials demonstrating that AI-guided strategies improve patient-centered outcomes. ML has substantial potential to advance precision medicine in ARDS by enabling predictive and prognostic enrichment and supporting personalized ventilatory care. Its safe and equitable deployment, however, requires standardized methodology, transparent reporting, multicenter prospective validation, and clinician-led governance to ensure that AI augments rather than replaces expert clinical judgment.