Time-resolved parameterization of aperiodic and periodic brain activity

Abstract

Macroscopic neural dynamics comprise both aperiodic and periodic signal components. Recent advances in parameterizing neural power spectra offer practical tools for evaluating these features separately. Although neural signals vary dynamically and express non-stationarity in relation to ongoing behaviour and perception, current methods yield static spectral decompositions. Here, we introduce Spectral Parameterization Resolved in Time (SPRiNT) as a novel method for decomposing complex neural dynamics into periodic and aperiodic spectral elements in a time-resolved manner. First, we demonstrate, with naturalistic synthetic data, SPRiNT’s capacity to reliably recover time-varying spectral features. We emphasize SPRiNT’s specific strengths compared to other time-frequency parameterization approaches based on wavelets. Second, we use SPRiNT to illustrate how aperiodic spectral features fluctuate across time in empirical resting-state EEG data (n=178) and relate the observed changes in aperiodic parameters over time to participants’ demographics and behaviour. Lastly, we use SPRiNT to demonstrate how aperiodic dynamics relate to movement behaviour in intracranial recordings in rodents. We foresee SPRiNT responding to growing neuroscientific interests in the parameterization of time-varying neural power spectra and advancing the quantitation of complex neural dynamics at the natural time scales of behaviour.

Human sensorimotor beta event characteristics and aperiodic signal are highly heritable

“…The 1/f signal is thought to stem from passive dendrite filtering properties (Halnes et al, 2016) but it is also modulated in an activity-dependent way (Pettersen et al, 2014). It has been shown to be affected by processes such as brain maturation (McSweeney et al, 2021; Hill et al, 2022; Tröndle et al, 2022) and aging (Voytek et al, 2015; Wilson et al, 2022) as well as in neurological (Semenova et al, 2021) and psychiatric diseases (Ostlund et al, 2021). Furthermore, 1/f reflects the attentional state (Waschke et al, 2021) and may contribute to integration of signals over longer periods of time (Maniscalco et al, 2018).…”

“…The rhythmic and aperiodic components are important to disentangle as they are probably generated by different mechanisms. Aperiodic signal is believed to represent excitation-inhibition balance (Gao et al, 2017) and is modulated, e.g ., by brain maturation (McSweeney et al, 2021; Hill et al, 2022; Tröndle et al, 2022), aging (Voytek et al, 2015; Wilson et al, 2022) and several neurological and psychiatric conditions (Molina et al, 2020; Ostlund et al, 2021; Semenova et al, 2021).…”

Singular superlet transform achieves markedly improved time-frequency super-resolution for separating complex neural signals

“…The problem of over-merging or over-disregarding neuronal clusters can also manifests itself as significant discrepancies between the number of theoretically available neurons and actually recorded neurons [7, 25] or as a misrepresentation of neural population dynamics. SST can alleviate these problems from the early stages of spike sorting pipeline and thereby both increase the number of neurons detected and the quality of the neuronal clusters created, increasing the reliability of the findings within the field.…”

“…However, as we will show, for the higher frequencies in a wide frequency dynamic range, SLT experiences a similar problem as CWT: the high frequency resolution gets smeared when combined with high temporal resolution observations. But since SLT also has a drastically increased computational load compared to CWT, this effect alone renders SLT impractical for investigations of large data sets [25] especially when the signals involve a wide dynamic frequency range, such as large scale extracellular neural recording data [26]. Here, we build on the principles of the SLT [24], but rather than relying on calculating multiple discrete observations across resolutions to generate their mean, we instead derive a singular mathematical expression of a wavelet, called the singular superlet, to generate a similar fundamental effect.…”

“…However, as we will show, for the higher frequencies in a wide frequency dynamic range, SLT experiences a similar problem as CWT: the high frequency resolution gets smeared when combined with high temporal resolution observations. But since SLT also has a drastically increased computational load compared to CWT, this effect alone renders SLT impractical for investigations of large data sets [25] especially when the signals involve a wide dynamic frequency range, such as large scale extracellular neural recording data [26].…”