Predicting failure of extubation and non-invasive respiratory support in critically ill patients: clinical complexity, limitations of traditional indices, and machine learning perspectives.

Failure of non-invasive respiratory support represents a major and unresolved clinical challenge in critically ill patients, both after extubation and during non-invasive ventilation or high-flow oxygen therapy used as primary treatment. Delayed or inappropriate escalation to invasive mechanical ventilation is associated with adverse outcomes, underscoring the need for reliable tools to timely identify patients at risk of failure. Traditional bedside indices provide simplicity and immediate applicability but show limited predictive accuracy and poor generalizability across heterogeneous clinical settings. In recent years, predictive models based on machine learning have been developed to improve risk stratification by integrating multiple clinical and physiological variables. Although early models demonstrated improved discrimination compared with conventional indices, their clinical impact has been constrained by retrospective designs, heterogeneous outcome definitions, limited external validation, and poor interpretability-limitations that collectively prevent meaningful cross-study comparisons and hinder safe clinical translation. More recent approaches have shifted toward dynamic and time-dependent models that incorporate early physiological trajectories and treatment response, offering predictions that are more consistent with real-world clinical decision-making processes. Nevertheless, high predictive accuracy alone does not ensure clinical usefulness in the absence of transparent decision support and integration into clinical workflows. Overall, current evidence suggests that predictive models should not replace clinical judgment but rather support timely, contextualized decisions by identifying modifiable risks. Future research should prioritize standardized outcome definitions, prospective multicenter validation, and the development of interpretable, workflow-integrated tools to enable safe and effective clinical translation.