Artificial intelligence in mechanical ventilation: a narrative review of clinical applications and research gaps.

BACKGROUND AND OBJECTIVE: Artificial intelligence (AI) has been increasingly applied to various aspects of mechanical ventilation (MV) to address the complexity and data-intensive nature of respiratory management in intensive care units (ICUs). Despite rapid growth in algorithm development and frequently reported high predictive performance, the clinical adoption of AI-based tools in MV remains limited. This review aims not only to summarize current AI applications in MV, but more importantly, to critically examine the barriers that have hindered their translation into routine clinical practice. METHODS: We conducted a narrative review of studies published up to June 2026 that developed or evaluated AI applications across five key domains of MV: intubation prediction or assistance, ventilation strategy optimization, detection of patient-ventilator asynchrony (PVA), prediction of weaning or extubation, and prediction of noninvasive ventilation (NIV) failure. Studies were identified through PubMed, Google Scholar, and ResearchGate searches. Methodological quality and risk of bias were assessed using the Prediction model Risk Of Bias ASsessment Tool (PROBAST) tool. KEY CONTENT AND FINDINGS: A total of 104 studies were included. Most AI models demonstrated good to excellent discriminative performance, with reported area under the curve (AUC) values commonly exceeding 0.80. However, the majority of studies were retrospective, single-center, and exploratory in nature, with limited external validation and scarce prospective evaluation. Importantly, few studies assessed the real-world clinical impact of AI-assisted decision-making on patient-centered outcomes, workflow integration, or clinician behavior. CONCLUSIONS: Although AI models for MV frequently achieve high performance in controlled research settings, their translation into routine clinical care remains minimal. The primary challenges are no longer related to algorithmic capability, but rather to methodological limitations, lack of generalizability, insufficient integration into clinical workflows, and absence of outcome-driven evidence. Addressing these translational barriers is essential before AI can meaningfully contribute to safe, effective, and sustainable MV management.