Article
Critique of approximate entropy as complexity measure of intracranial pressure
Search Medline for
Authors
Published: | June 9, 2017 |
---|
Outline
Text
Objective: Approximate entropy (ApEn) was recently introduced to extend the spectrum of cerebrospinal fluid pressure parameters. The algorithm was found to produce low values when intracranial pressure (ICP) was elevated and high values when ICP was at low levels. Applying ApEn to point processes (e.g. neural spike trains) ApEn’s scale features the degree of signal regularity or roughness (ICP dynamics). The scale’s cornerstones outline clear but trivial signal characteristics. The intriguing question is how to interpret ApEn beyond extremes?
Methods: To reveal midrange ApEn scale’s meaning we studied the influence of intrinsic ApEn parameters (embedding and radius) on real-world ICP and simulation data. To mark off the interval ranging from regularity to irregularity in ICP time series, samples were evaluated from a mechanically ventilated patient (highly regular, non-stationary), from nocturnal readings (generally non-stationary) of a patient who underwent long-term ICP measurement for diagnostic reasons, and red/white noise random numbers (mainly irregular). Data were tested by fixing one of the ApEn parameters while the other was varied. A total of 24 parameter combinations were examined in detail.
Results: Variation of parameters influenced the results notably. All parameter combinations resulted in an almost uniform picture of ApEn: starting out from initially high values and converging toward zero with two exceptions. 1) nocturnal ICP showed a concave decay instead of a convex, indicating some stability over parameter variations, and 2) white noise depicted a family of curves covering the complete scale of ApEn. Initial values fitted best for noise data with high ApEn but only for a small numbers of parameter combinations. Paradox ApEn was found for the ventilated patient. A direct assignment of data specifications and corresponding ApEn was impossible.
Conclusion: ApEn did not clearly scale regularity or irregularity. Application of recommended parameter settings from literature and also variations in parameters led to confusing results. ICP data created the impression of a distorting influence of non-stationarity on ApEn. This is fatal for biomedical data is inherently non-stationary. Based on these computations ApEn seems not to be an appropriate candidate for explaining ICP dynamics.