ARIMA-based forecasting of cerebral physiologic signals in acute traumatic brain injury: a CAnadian high-resolution TBI (CAHR-TBI) cohort study.

backgroundTraumatic brain injury (TBI) remains a major global health issue, with limited progress in reducing morbidity and mortality for TBI patients in need of sedation and intensive care. This has led to increased focus on the mechanisms of secondary brain injury, typically monitored via high-frequency, multi-modal physiologic data reflecting pressure flow and oxygen delivery. However, the complexity and volume of such data pose challenges for clinicians, leading to the use of resolution-reducing techniques, such as moving averages and point sampling. However, data often remains a challenge to utilize clinically for physiologic insult predications and early or pre-emptive interventions. Time series modeling approaches like autoregressive integrated moving average (ARIMA) are valuable in analyzing statistical signal structures, providing insights into temporal dynamics by revealing temporal patterns and forecasting future physiological states.

resultsThis study evaluated the effects of resolution reduction via averaging on point and interval predictions using ARIMA models. Analysis was performed on both raw signals and derived physiologic metrics of cerebral pressure flow, compliance, and oxygen delivery by utilizing the CAnadian High-Resolution TBI (CAHR-TBI) data set. Temporal resolution was reduced by averaging with non-overlapping intervals, ranging from 1-min to 24-h windows. Data from A total of 376 TBI patients requiring intensive care was analyzed across various temporal resolutions. ARIMA models perform best at high temporal resolutions, particularly for derived cerebrovascular reactivity indices, with accuracy decreasing for raw signals at lower resolutions. The choice of data partitioning method affects performance; however, all methods struggle at the lowest resolutions, highlighting ARIMA's limitations for long-term forecasting of cerebral physiologic signals with lower resolution data commonly recorded in patient records.

conclusionsThis study highlights the significant influence of temporal resolution and data partitioning methods on the predictive performance of ARIMA models for cerebral physiological signals. While ARIMA performs well at high temporal resolutions, its accuracy declines for raw physiological signals as resolution decreases. The choice of cross-validation method also impacts forecasting performance. The findings underscore the need for hybrid modeling approaches that integrate ARIMA with machine learning techniques to improve predictive accuracy, particularly for complex cerebral physiological signals.