Time-Series autocorrelative structure of cerebrovascular reactivity metrics in severe neural injury: An evaluation of the impact of data resolution
Amanjyot Singh Sainbhi, Nuray Vakitbilir, Alwyn Gomez, Kevin Y. Stein, Logan Froese, Frederick A. Zeiler
Abstract
Goal: Cerebrovascular reactivity (CVR) dysfunction is a contributor to secondary injury in traumatic brain injury (TBI). The issue with applying non-overlapping moving average filters to reduce temporal resolution of high resolution CVR data is that the autocorrelative structure is ignored. It violates the priors of linearity and raises concerns for the level of certainty for any reported models. The goal is to assess if there is a data resolution for cerebral physiology where Box-Jenkin’s time-series statistical structures can be ignored. The CVR indices were derived in varying temporal resolutions from 10-second to 1-day and each signals’ stationarity was assessed. By varying autoregressive order (1–10), integrative order (0–2), and moving average order (0–10), the autoregressive integrative moving average (ARIMA) models were fit to each index in varying temporal resolutions to obtain median optimal ARIMA models. A total of 100 patients were included with 3934.5 minutes of median recording duration. The stationarity analysis showed stationarity in 1st and 2nd order differenced data after temporal reduction. The median optimal ARIMA models for each combination of temporal resolution and CVR indices were found based on Akaike Information Criterion. Autocorrelative function (ACF) and partial ACF plots of residuals confirmed median optimal ARIMA model adequacy. For accurate predictions/trajectory forecasting, the autocorrelative structure needs to be accounted for in CVR data using an autocorrelative order of 8–10 for high frequency data and about 5 for low frequency data. Also, there is the need to understand such ARIMA structures in raw multi-modal cerebral physiology using multi-center high-resolution datasets.