Causal inference in neuronal time-series using adaptive decomposition.

Background: The assessment of directed functional connectivity from neuronal data is increasingly common in neuroscience by applying measures based in the Granger causality (GC) framework. Although initially these consisted in simple analyses based on directionality strengths, current methods aim to discriminate causal effects both in time and frequency domain.
New Method: We study the effect of adaptive data analysis on the GC framework by combining empirical mode decomposition (EMD) and causal analysis of neuronal signals. EMD decomposes data into simple amplitude and phase modulated oscillatory modes, the intrinsic mode functions (IMFs), from which it is possible to compute their instantaneous frequencies (IFs). Hence, we propose a method where causality is estimated between IMFs with comparable IFs, in a static or time-varying procedure, and then attributed to the frequencies corresponding to the IF of the driving IMF for improved frequency localization.
Results: We apply a thorough simulation framework involving all possible combinations of EMD algorithms with causality metrics and realistically simulated datasets. Results show that synchrosqueezing wavelet transform and noise-assisted multivariate EMD, paired with generalized partial directed coherence or with Geweke's GC, provide the highest sensitivity and specificity results.
Comparison With Existing Methods: Compared to standard causal analysis, the output of selected representative instances of this methodology result in the fulfillment of performance criteria in a well-known benchmark with real animal epicranial recordings and improved frequency resolution for simulated neural data.
Conclusions: This study presents empirical evidence that adaptive data analysis is a fruitful addition to the existing causal framework.
(Copyright © 2015 Elsevier B.V. All rights reserved.)