Deep State Space Models for Forecasting Intracranial Pressure in Patients with Traumatic Brain Injury

Sammanfattning

Patients hospitalized with traumatic brain injuries (TBIs), such as those resulting from car accidents, face significant risks of secondary complications, including reduced blood flow to the brain. One critical indicator of these complications is intracranial hypertension—elevated pressure within the cranium—which often necessitates urgent medical intervention to prevent severe outcomes such as death or long-term disability. The duration of time a patient spends in this state is strongly correlated with negative outcomes, highlighting the critical need for rapid treatment and stabilization. Numerous predictive models have been developed to forecast episodes of intracranial hypertension and alert healthcare providers. However, these efforts have yielded mixed results, largely due to the challenges of sequence modeling with long-term data. Mamba, a novel sequence modeling architecture, addresses these challenges by combining efficient computational performance with robust long-term memory capabilities, making it a promising candidate for clinical applications. This study evaluates the performance of Mamba using high-resolution patient data, including intracranial pressure (ICP), arterial blood pressure (ABP), and electrocardiogram (ECG) readings. Results indicate that Mamba matches the performance of a simple Long Short-Term Memory (LSTM) model—a class of sequence model—achieving a root mean square error (RMSE) loss of 3.929. Furthermore, MAMBA offers significant advantages in training and inference speed while overcoming memory limitations. Despite these strengths, Mamba’s predictions with a 30-minute forecast horizon did not reliably anticipate secondary insult events far into the future. Comparisons with other studies using similar forecast horizons suggest comparable performance, though differences in methodologies and datasets limit direct conclusions. Nevertheless, Mamba’s computational efficiency and potential for refinement puts it as a promising foundation for future predictive sequence models in clinical care.