A functional Hidden Markov Model to incorporate dynamics into Bayesian optimal stopping problems: Helping physicians manage traumatic brain injuries

Abstract

This paper describes a decision support system for neurosurgeons to determine the release schedule for polytrauma patients with concurrent traumatic brain injuries (TBIs). We present a novel functional Hidden Markov Model (fHMM) designed to mitigate the computational burden caused by ever-growing historical data. Our fHMM is nested into a Bayesian optimal stopping problem, effectively representing emergency physicians' crucial decisions during the discharge process. Our model is unique in its ability to adapt to different physician decision-making styles while continuously monitoring patient progress. We validate our approach using simulated and real data from a major hospital in Tampa Bay. The analysis reveals that dynamic prognostic measures can help balance the competing needs of rapid patient release and neurologic stability. Additionally, our results show variability in physicians' expectations regarding neurological recovery. Counterfactual analyses suggest that implementing this decision model could potentially save up to $27,500 per TBI patient under certain conditions.