Prediction of intubation risk in patients with alcohol withdrawal using explainable machine learning

Abstract Alcohol withdrawal is a common and potentially life-threatening condition that can rapidly progress to respiratory failure requiring intubation. Clinicians must make timely decisions about monitoring intensity, treatment strategy, equipment needs, and escalation of care, yet existing bedside tools primarily assess withdrawal severity rather than functional outcomes such as intubation risk. In this study, we use explainable machine learning methods to predict intubation in patients with alcohol withdrawal at clinically relevant decision points. In this retrospective study of adult inpatient encounters from a large academic health system, we developed and evaluated machine learning models to predict intubation at hospital admission and 24 h after admission. Ensemble tree-based methods provided the strongest discrimination and the best-calibrated probability estimates, while other model families achieved higher sensitivity at the cost of specificity. At admission, we observed strong overall performance, with an AUC of 0.863 (95% CI 0.838–0.887). At the 24-hour mark, retaining historical information led to consistent improvements relative to using current physiological data alone. Early risk reflected underlying comorbidity and prior clinical history, while later risk was shaped by evolving physiological instability and increasing sedative requirements. These findings support explainable, time-aware intubation risk stratification to inform monitoring and escalation of care.