Abstract
Therapeutic management during the acute phase of traumatic brain injury (TBI) relies on continuous multimodal cerebral physiologic monitoring to detect and prevent secondary injury. These high-resolution data streams come from various invasive/non-invasive sensor technologies and challenge clinicians, as they are difficult to integrate into management algorithms and prognostic models. Data reduction techniques, like moving average filters, simplify data but may fail to address statistical autocorrelation and could introduce new properties, affecting model utility and interpretation. This study uses the CAnadian High-Resolution TBI (CAHR-TBI) dataset to examine the impact of temporal resolution changes (1 min to 24 h) on autoregressive integrated moving average (ARIMA) modeling for raw and derived cerebral physiologic signals. Stationarity tests indicated that the majority of the signals required first-order differencing to address persistent trends. A grid search identified optimal ARIMA parameters (p,d,q) for each signal and resolution. Subgroup analyses revealed population-specific differences in temporal structure, and small-scale forecasting using optimal parameters confirmed model adequacy. Variations in optimal structures across signals and patients highlight the importance of tailoring ARIMA models for precise interpretation and performance. Findings show that both raw and derived indices exhibit intrinsic ARIMA components regardless of resolution. Ignoring these features risks compromising the significance of models developed from such data. This underscores the need for careful resolution considerations in temporal modeling for TBI care.