Can one-step reinforcement learning guide optimal timing for PEG and tracheostomy in severe TBI? Insights from a 2016-2023 retrospective cohort study at a single academic institution.

Background: Acute management of traumatic brain injury (TBI) presents several challenges in hospital resource planning. While early tracheostomy (trach) and percutaneous endoscopic gastrostomy (PEG) tube placement may improve patient outcomes, the optimal timing and selection criteria for these interventions remain unclear. This study evaluates the impact of PEG and trach timing on key clinical outcomes and applies one-step reinforcement learning (RL) to recommend intervention timing. Methods: This retrospective cohort study included 263 adult intensive care unit inpatients (194 men, 69 women, age range 18-87), diagnosed with severe TBI requiring trach and/or PEG between 1 January 2016 and 31 December 2023, at a single academic institution. Key outcomes included ICU and hospital length of stay (LOS), complications, time to oral feeding/decannulation, readmission, and mortality. One-step temporal difference (TD) learning and Q-learning were used to predict the expected value of interventions and to recommend optimal timing based on patient states, respectively. Results: Early PEG and trach interventions were associated with significantly shorter ICU and hospital length of stay (LOS) and fewer complications. Delayed PEG placement, however, was associated with a 67% reduction in the odds of mortality (OR: 0.33, Conclusion: Early interventions are associated with improved outcomes; however, delaying PEG or trach placement may be advantageous in select situations to reduce mortality. RL techniques, such as TD and Q-learning, can aid in decision-making regarding interventions.