Stationarity in neuromonitoring data

Abstract

Purpose: Signals reflecting the metabolic and circulatory status of an injured central nervous system are normally corrupted systematically. The patient is part of a therapeutic control-loop and the signals acquired are rather determined by the quality of control (stationarity of signals) than by the underlying pathological process.

Methods: To verify the control-loop hypothesis, neuromonitoring data from 12 randomly selected severely head injured patients (initial GCS ≤ 8, 7 men, 5 women) were analysed for circulatory (blood pressure, intracranial pressure [ICP], cerebral perfusion pressure [CPP]) and metabolic (arterial blood gases, jugular bulb oxygenation [SjvO(2)], brain tissue oxygen partial pressure [ptiO(2)]) variables (n = 10). A total of 120 time series of generally not equidistant sample intervals were assessed for stationarity by Wallis & Moore's runs test.

Results: Non-stationarity could only be proven in 23 time series, i.e. the control-loop hypothesis was violated. Trends were mainly found in CPP (n = 5) and ICP (n = 4). The remaining cases spread out on all but one (temperature) signal. Nine patients showed at least one time series with a trend. One patient had clear trends in five out of ten variables that focused on SjvO(2), ptiO(2), ICP and CPP.

Conclusions: Absence of stationarity in about 20% of time series is credited to an effective therapeutic control-loop. For analytical purposes, however, the benefit seems to be overestimated. Consequently, neuromonitoring should be considered the analysis of short-term disturbances that are intentionally compensated for by a short response time. Information content is thus reduced even if the number of sensor devices increases.