State Space Methods Using Biologically-Relevant Generative Models to Analyze Neural Signals

Abstract

Neural oscillations have long been recognized for their mechanistic importance in coordinating activity within and between brain circuits. Co-occurring broad-band, non-periodic signals are also ubiquitous in neural data and are thought to reflect the characteristics of population-level neuronal spiking activity. Identifying oscillatory activity distinct from broadband signals is therefore an important, yet surprisingly difficult, problem in neuroscience. Commonly-used bandpass filters produce spurious oscillations when applied to broad-band noise and may be ill-informed by canonical frequency bands. Curve-fitting procedures have been developed to identify peaks in the power spectrum distinct from broadband noise. Unfortunately, these ad hoc methods are prone to overfitting and are difficult to interpret in the absence of generative models to formally represent oscillatory behavior. Similarly, broadband power spectrum log-log slope or “1/f” curve-fitting methods have been developed to identify excitatory-inhibitory balance in the LFP or ECoG, but are not defined in terms of a generative model. Here we present three novel methods that utilize generative models to (1) identify and characterize neural oscillations distinct from broad-band noise (2) apply this oscillatory structure to improve cortical source signal estimates inferred from scalp-level EEG recordings and (3) identify and characterize excitatory and inhibitory neurotransmitter contributions to LFP signals.