Abstract
Accurate estimation of electrophysiological properties is crucial for understanding and finding treatments for numerous neurological disorders. Changes in long-range temporal correlations (LRTCs) were demonstrated to be associated with several conditions, for instance, epilepsy. Currently, LRTCs are primarily measured using either spectral methods in broad-band signals [1] or Detrended Fluctuation Analysis (DFA) in the envelope of oscillatory components [2]. However, theoretical work suggests that the Hurst exponent, a measure of LRTC, can be estimated using multiple methods [3]. Other fields, like Geology, have established frameworks for comparing such methods [4]. With electrophysiological data, such as EEG, in mind, we compare spectral methods, DFA, auto-correlation, wavelet methods, and rescaled range analysis. We built a comprehensive framework for comparing these methods using artificially generated data. Fig. 1 (A-F) demonstrates how each method extracts the Hurst exponent from the signal. Fig. 1G shows that the performance is diverse and that spectral methods, DFA, and wavelet methods perform best, which agrees with the literature [4]. To emulate measured data, we generate a typical electrophysiological signal with appropriate broad-band and amplitude envelope scaling. For analyzing broad-band signals with oscillations, the specparam method [1] works best and is recommended due to its' explicit handling of oscillations. Nonetheless, we suggest that an adjusted version of DFA can serve as a valuable control method. For investigating LRTCs in the amplitude envelope of these oscillations, the most popular method DFA [2] is among the most accurate methods. Moreover, this method can be supplemented or combined with spectral or wavelet methods to enhance measurement robustness. Overall, our findings validate that the most commonly used methods are appropriate for the respective components of the signal. We raise awareness about the importance of choosing the correct method and suggest possibilities for more robust estimation, such as combining methods. We plan to develop comprehensive usage guidelines and test our approaches on public datasets.
