Brain signals exhibit a wealth of dynamic patterns that are thought to reflect ongoing neural computations. Multiscale sample entropy (MSE) intends to describe the temporal irregularity of such patterns at multiple time scales of brain function. However, the notion of time scales may often be unintuitive. In particular, traditional implementations of MSE are sensitive to slow fluctuations at fine time scales, and fast dynamics at coarse time scales. This conceptual divergence is often overlooked and may lead to difficulties in establishing the unique contribution of MSE to effects of interest over more established spectral power. Using simulations and empirical data, we highlight these issues and provide evidence for their relevance for valid practical inferences. We further highlight that standard MSE and traditional spectral power are highly collinear in our example. Finally, our analyses indicate that spectral filtering can be used to estimate temporal signal irregularity at matching and intuitive time scales. To guide future studies, we make multiple recommendations based on our observations. We believe that following these suggestions may advance our understanding of the unique contributions of neural signal irregularity to neural and cognitive function across the lifespan.
Neural times series exhibit a wealth of dynamic patterns that are thought to reflect ongoing neural computations. While some of these patterns consist of stereotypical deflections [e.g., periodic neural rhythms; 1, 2], the framework of nonlinear dynamics and complex systems also emphasizes the importance of temporal irregularity (or variability) for healthy, efficient, and flexible neural function [3–6]. Specifically, functional network dynamics may reflect the non-linear interaction of local and global population activity, for which intermediate levels of network noise theoretically afford high network capacity and dynamic range [7–10]. In parallel with such conceptual advances, multiscale entropy (MSE) [11, 12], an information-theoretic index that estimates sample entropy [13] at multiple time scales (Fig 1A), has become a promising tool to quantify the irregularity of neural time series across different brain states, the lifespan, and in relation to health and disease [14–22]. However, we argue that outstanding methodological issues regarding the mapping of neural-to-MSE time scales reduce the current interpretability of MSE results, and–if not properly accounted for–limit MSE’s utility for investigating substantive neurocomputational questions of interest.
In general, sample entropy quantifies the irregularity of temporal patterns in a given signal (for an example of its calculation, see Fig 1B). Whereas signals with a repetitive structure (like stationary signals or rhythmic fluctuations) are estimated as having low entropy, less predictable (or random) signals are ascribed high entropy. As an extension of this principle, MSE aims to describe temporal irregularity at different time scales–varying from fine (also referred to as ‘short’) to coarse (or ‘long’). In conventional Fourier analysis of time series data, time scales are quantified in terms of lower and higher frequencies present in the signal. This has been shown to be a principled time scale descriptor that relates at least in part to structural properties of the generating neural circuits [2, 23–26]. Given this meaningful definition of fast and slow events, it is a common assumption–including in guides to MSE’s interpretation in neural applications [27]–that fine-to-coarse scales characterize the irregularity of high-to-low frequency dynamics, respectively. However, here we highlight one methodological and one conceptual issue regarding the computation of MSE that challenge such a direct scale-to-frequency mapping. First, we show that the traditional definition of temporal patterns may lead to an influence of high frequencies on coarse entropy time scales (Issue 1). Second, we highlight that the signal content at fine time scales renders entropy estimates sensitive to a conjunction of scale-free and narrowband signals, including slow fluctuations (Issue 2).
Due to its assessment of temporal patterns rather than sinusoidal oscillatory dynamics, MSE has been motivated as a complementary measure to spectral variance/power that is sensitive to multi-scale, potentially non-linear, signal characteristics, such as phase shifts or cross-frequency coupling. [Note that we use the terms power and variance interchangeably, as a time domain signal’s broadband variance is proportional to the integral of its power spectral density, while narrowband variance in the time domain is identical to narrowband power in the spectral domain.] However, the overlap between these measures is often unclear in application because the mapping between spectral power and scale-wise entropy is ambiguous. Such ambiguity affects both the ability to compare individuals at any scale, and the ability to compare entropy levels across scales within person. We argue that a clarification of these issues is thus necessary for valid inferences of time scale-specific ‘neural irregularity’ in a growing number of neuroscientific MSE applications.
A principle assumption of sample entropy is that “the degree of irregularity of a complex signal […] cannot be entirely captured by the SD [i.e., standard deviation]” [28; i.e., square root of variance]. To ensure this, sample entropy is typically assessed relative to the standard deviation of the broadband signal to intuitively normalize the estimation of irregularity for overall distributional width [13, 14, see also 28]. In particular, the similarity bound– defined by a constant r, by which the signal SD is multiplied–reflects the tolerance for labeling time points as being similar or different, and thus, determines how liberal the algorithm is towards detecting ‘matching patterns’ (Fig 2A and 2B). While wider bounds decrease entropy estimates, narrower bounds increase them [13, 29, 30] (S2 Fig). Crucially, the similarity bound is often not equally liberal across time scales, resulting in an entropy estimation bias. Specifically, to characterize temporal irregularity at coarser time scales, signals are typically successively low-pass filtered [or ‘coarse-grained’; 31] (Fig 2C), whereas the similarity bound typically (in its ‘Original’ implementation) is set only once–namely relative to the SD of the original unfiltered signal. Due to the progressive filtering, coarse-graining successively reduces the signal’s SD, yet a single global (i.e., scale-invariant) similarity bound remains based on the cumulative variance of all estimable frequencies (Fig 2D and 2E). As a result, the similarity bound becomes increasingly liberal towards pattern similarity at coarser scales, thereby reducing entropy estimates. This is most clearly illustrated by the observation that white noise signals, which should be characterized as equally random at each time scale, exhibit decreasing entropy values towards coarser scales when global similarity bound s are used [27, 29, 32]. This issue has been recognized previously [29], and provided a rationale for recomputing the similarity bound for each time scale [29, 33–35]. But despite the benefits of this refinement that was already proposed fifteen years ago, our review of the literature revealed that the use of global bounds remains dominant in over 90% of neuroscientific MSE applications (see S1 Text) and in previous validation work [27]. Crucially, the consequences of this bias for practical inference remain unclear. We therefore argue that a comprehensive assessment of the resulting bias is needed to highlight this issue, both to clarify previous results and to guide future studies.
While matched similarity bounds account for total signal variation at any specific time scale, sample entropy remains related to the variance structure (i.e., the power spectrum) of the signal as one indicator of its temporal irregularity [4]. Most neural signals exhibit a scale-free power distribution [36–38], for which the exponent x indicates the prevalence of low-to-high-frequency components in the signal. This ratio is also referred to as the power spectral density (PSD) slope. Smaller exponents (indicating shallower PSD slopes) characterize signals with relatively strong high-frequency contributions (i.e., reduced temporal autocorrelations, and less predictability) compared to larger exponents that indicate steeper slopes. This conceptual link between PSD slopes (or high-to-low frequency power ratios that may have strong broadband slope contributions [39]) and sample entropy has been empirically observed across subjects, wakefulness and task states [14, 17, 40]. However, the sensitivity of fine-scale entropy to PSD slopes–a multi-scale characteristic–highlights that the contribution of slow-to-fast signal content to fine-scale entropy is unclear. This ambiguity arises from the algorithm that derives scale-wise signals. In particular, ‘Original’ MSE implementations use low-pass filters to derive signals at coarser time scales, which increasingly constrains entropy estimates to slower fluctuations. As such, each scale defines an upper bound for the range of included frequencies (see methods). However, the opposite is not true, resulting in a lack of high-frequency specificity. Hence, finer time scales characterize the entire broadband signal (see Fig 3A) which represents a non-specific mixture of low and high-frequency elements across scale-free and rhythmic signal contributions [41, 42]. Crucially, the contribution of these elements to neural broadband signals is not equal. Rather, the variance of signals is dominated by the amplitude of low frequencies, which may thus disproportionally impact the assessment of pattern irregularity [35]. As a result, broadband signal characterization challenges the assumption that fine-scale entropy mainly describes ‘fast’ events. More generally, this highlights large uncertainty regarding the frequencies that are represented at any particular time scale.
The projection of narrowband rhythms into simulated noise signals provides a well-controlled situation in which to study the mapping of neural irregularity to MSE, due to their clearly defined time scale (i.e., period = inverse of frequency) and regularity (added rhythmic variance = more regular signal = decreased entropy). Moreover, rhythmic structure remains a dominant target signal in neuroscience [1, 36, 43] for which entropy, as a complementary descriptor, should provide an anti-correlated reflection. However, previous simulations on the mapping of rhythms onto MSE time scales have produced puzzling results that have received little attention in the literature so far; while a linear mapping between rhythmic frequency and entropy time scales has been observed, added rhythmic regularity has been shown to increase entropy above baseline in previous work [4, 22, 44]. This notably contrasts with the intuition that added signal regularity should reduce observed entropy. Thus, additional simulations are necessary to assess the intuitive notion that rhythmicity should be anticorrelated with entropy, and to investigate whether this phenomenon indeed occurs at specific time scales, as previously assumed [4, 22, 44]. In particular, we probed the feasibility of using high-pass and band-pass filters (relative to standard low-pass options) to control the MSE time scales at which rhythmicity would be reflected (Fig 3B).
In summary, Issue 1 suggests a coarse-scale bias introduced by global similarity bounds, and Issue 2 highlights a mixture of narrow- and broadband contributions to fine scales. In worst-case scenarios, a conjunction of these issues may lead to a reflection of fast dynamics in coarse entropy and a reflection of slow dynamics in fine entropy, thus paradoxically inverting the intuitive time scale interpretation. These issues have not been jointly assessed, however, and there is little evidence of whether and how these methodological issues may impact practical inferences motivated by neurobiological questions of interest. We focus on two example scenarios in the current study.
One principal application of multiscale entropy is in the domain of lifespan covariations between neural dynamics and structural brain network ontogeny [for a review see 45]. Within this line of inquiry, it has been proposed that structural brain alterations across the lifespan manifest as entropy differences at distinct time scales [16, 18, 40, 46]. Specifically, it has been suggested that coarse-scale entropy decreases and fine-scale entropy rises with increasing adult age as a reflection of senescent shifts from global to increasingly local information processing [16, 18]. Crucially, this mirrors observations based on spectral power, where age-related decreases in the magnitude of low-frequencies [47, 48] are accompanied by increases in high-frequency activity, conceptualized also as a flattening of power spectral density (PSD) slopes [16, 18, 40, 49]. These results seemingly converge towards a joint decrease of low-frequency power and coarse-scale entropy in older adults (and an increase for both regarding fast dynamics). However, this correspondence is surprising upon closer inspection given the presumed anticorrelation between the magnitude of signal regularity (as indicated by heightened spectral power) and entropy. In light of concerns regarding the interpretation of entropy time scales (see above), we assessed cross-sectional age effects on both MSE and spectral power as a test case for potential mismatches in scale-dependent inferences.
Identifying the time scale contributors to MSE is further relevant due to the assumed functional separability of narrow- and broadband brain dynamics. Whereas narrowband rhythms have been closely associated with synchronous population spiking at the service of temporal information coordination [50], scale-free broadband dynamics may provide a complementary index of the level of neocortical activation and aggregate spiking activity in humans [38, 51–53]. In particular, shallower PSD slopes have been proposed as a signature of enhanced cortical excitability (or ‘neural noise’) [54]. Such excitability in turn may regulate the available range of network dynamics as reflected in information entropy [10]. Notably, interactions between narrow- and broadband activity are neurobiologically expected. In particular, as the magnitude of narrowband alpha synchronization increases, population output is thought to decrease [55]. However, the methodological conflation of narrow- and broadband contributions to entropy (see “Issue 2” above) may complicate principled investigations regarding their neurobiological coupling in practice. As a corollary goal in the present work, we therefore investigate whether a principled separation of narrow- and broadband contributions to entropy is tractable.
Here, we aimed to address two issues of frequency-to-scale mapping and their relevance for empirical applications. First, we simulated variations in rhythmic power and frequency to probe the relationship between rhythmicity and MSE time scales. Primarily, our goal was to assess how global similarity bounds (Issue 1) and the scale-wise spectral content of the analyzed signal (Issue 2) influence the time scales at which added rhythmicity is observed. Then, we attempted to replicate reported cross-sectional age differences in human electroencephalography (EEG) signals recorded during rest. We assessed whether younger adults would show increased coarse scale and decreased fine-scale entropy compared to older adults, and we probed the extent to which such scale-specific results depend on mismatched spectral power via the issues above. As corollary goals, we assessed the potential of band-pass and band-stop approaches for deriving more intuitive insights regarding the time scales of signal irregularity. First, we probed the potential of ‘frequency-specific’ estimates of signal irregularity via band-pass filters, and assessed age differences therein. Second, we assessed the relation between alpha rhythms and broadband signal irregularity, after accounting for their methodological coupling. We refer to traditional settings that use global bounds and low-pass filtering as ‘Original’ throughout the remainder of the manuscript (see methods for details).
Our first aim was to probe how scale-specific events, namely rhythms of a given frequency, modulate MSE time scales. For this purpose, we simulated 10 Hz (alpha) rhythms of varying power on top of pink noise and calculated the MSE of those signals. First, we probed the influence of global similarity bounds (as used in ‘Original’ implementations) on the time scale mapping (Issue 1). Crucially, as a result of using a global similarity bound for all time scales, strong rhythmic power decreased MSE estimates across a range of time scales, including time scales at which added 10 Hz rhythmicity did not contribute to the scale-wise signal (Fig 4A, upper panel). As highlighted in Issue 1, this can be explained by a general increase in the liberality of bounds (Fig 4A, lower panel) that introduced a bias on coarse-scale entropy below 10 Hz. In contrast, when scale-dependent similarity bounds were used with low-pass filters (Fig 4B and 4C), strong rhythmicity systematically affected entropy only at finer time scales than the simulated frequency (i.e., to the left of the vertical line in Fig 4C, albeit in a diffuse manner, which we will examine next).
Second, we assessed the influence of the scale-wise filters (and hence, the spectral signal content) on frequency-to-scale mapping (see Issue 2, Fig 3B). In particular, we expected that low-pass filters (A-C) would lead to entropy decreases at finer time scales than the simulated frequency, whereas high-pass filters would lead to a rhythm representation at coarser time scales (Fig 3B). In line with these expectations, low-pass filters constrained the influence of narrowband rhythms to finer time scales (Fig 4C). As in previous work [33], Butterworth filters (Fig 4C) improved the removal of 10 Hz rhythms at coarser time scales and produced less aliasing compared with ‘Original’ point-averaging (see methods, Fig 4A and 4B), with otherwise comparable results. Hence, low-pass filters rendered multiscale entropy sensitive to variance from low frequencies, suggesting that slow events (e.g. event-related potentials) are reflected in a diffuse manner across time scales. In contrast, high-pass filters constrained rhythm-induced entropy decreases to coarser time scales that included 10 Hz signal content, hence leading to estimates of high frequency entropy that were independent of low frequency power (Fig 4D). Finally, when band-pass filters were used (Fig 4E), rhythmicity decreased sample entropy at the target scales (despite producing edge artifacts surrounding the time scale of rhythmicity). In sum, these analyses highlight that rhythmic power increases will diffusely and non-specifically modulate MSE time scales as a function of the coarse-graining filter choice, unless a narrowband filter is applied.
Such diffuse reflection of rhythms across MSE time scales is at odds with previous simulations suggesting a rather constrained, linear mapping between the frequency of simulated rhythms and entropy time scales [4, 22, 44]. Furthermore, those studies indicated entropy increases with added rhythmicity, in contrast with the marked (and expected) decreases in entropy observed here. Crucially, increased entropy relative to baseline runs counter to the idea that the addition of a stereotypic pattern should decrease rather than increase pattern irregularity. To assess whether these seemingly divergent results can be reconciled, we repeated our simulation for different frequencies. We focused on a comparatively low level of rhythmicity (amplitude level = 2; SNR ~ 1.3 (see methods); S3 Fig displays exemplary time series), for which Fig 4A–4C suggested transient entropy increases above baseline. Similar to previous reports, we observed a positive association between simulated frequencies and peak entropy time scales (Fig 5) across implementations, such that rhythms of a given frequency increased entropy at slightly finer time scales (see increases in entropy above baseline to the left of the dotted vertical lines in Fig 5A–5C). However, as shown in Fig 4A–4C, such increases were counteracted when rhythmic strength increased, while global similarity bound s (Fig 5A) liberally biased, and thus decreased, entropy at coarser time scales (i.e., to the right of the dotted lines in Fig 5A) independent of rhythmic strength. While the mechanistic origin of entropy increases remains unclear, previous conclusions may thus have overemphasized the scale-specificity of rhythmic influences.
In sum, our simulations highlight that the choice of similarity bound and the signal’s spectral content grossly affect one’s ability to interpret MSE time scales. Our frequency-resolved simulations suggest that a previously argued direct frequency-to-scale mapping is not tenable when typical estimation procedures are used. Supplementing these narrowband contributions to MSE, we report results from simulations of varying spectral slopes in S2 Text and S7 Fig.
Our simulations suggest profound influences of the choice of similarity bound (Issue 1) and spectral content (Issue 2) on scale-dependent MSE estimates. However, whether these issues affect inferences in empirical data remains unclear. Entropy differences across the lifespan are an important application [6], where ‘Original’ MSE implementations suggest that older adults exhibit higher entropy at finer time scales and lower entropy at coarser time scales compared to younger adults [for a review see 45]. Importantly, a shallowing of PSD slopes with age has also been reported, as represented by higher power at high frequencies and lower power at low frequencies [40, 49]. The raised issues of a potential (1) reflection of high frequency power on coarse scales and (2) diffuse reflection of slow spectral content thus question whether traditional MSE group differences reflect veridical differences in signal irregularity at matching time scales. Given those two issues, we specifically hypothesized that:
To assess these hypotheses, we first attempted to replicate previously reported scale-wise age differences in MSE and spectral power during eyes open rest. ‘Original’ settings replicated scale-dependent entropy age differences (Fig 6A1). Specifically, compared with younger adults, older adults exhibited lower entropy at coarse scales, and higher entropy at fine scales (Fig 6A1). Mirroring these results in spectral power, older adults had lower parieto-occipital alpha power and increased frontal high frequency power (Fig 6A2) compared to younger adults. This was globally associated with a shift from steeper to shallower PSD slopes with increasing age (Fig 6D). At face value, this suggests joint shifts of both power and entropy, in the same direction and at matching time scales. Crucially, however, the spatial topography of entropy differences inverted the time scale of power differences (Fig 6B and C; cf., upper and lower topographies), such that frontal high frequency power topographies resembled coarse entropy topographies (Fig 6B), while parieto-occipital age differences in slow frequency power resembled fine-scale entropy differences (Fig 6C). This rather suggests scale-mismatched associations between entropy and power.
Next, we assessed the impact of scale-wise similarity bounds and different scale-wise filters on the indication of MSE age differences (Fig 7).
Briefly, we observed three main results that deserve highlighting:
In the following sections, we assess these results more closely.
Scale-dependent entropy effects in the face of global similarity bounds (as observed in the ‘Original’ implementation; Fig 7A) may intuitively suggest scale-specific variations in signal irregularity in the absence of variance differences. However, global similarity bounds increasingly diverge from the scale-wise signal variance towards coarser scales (Issue 1; Fig 8A). This introduces a liberal bias that systematically varies as a function of the removed variance, thereby rendering coarse MSE scales sensitive to differences in higher frequency power (i.e., Issue 1), as observed in the case of aging (Fig 8A and 8B).
To assess whether global bounds introduced an association between high frequency power and coarse scale entropy in the case of aging, we probed changes in similarity bounds and MSE between the use of global and scale-varying bounds. As expected, we observed a strong anti-correlation between inter-individual changes in similarity bounds and MSE (Fig 8C). That is, the more similarity bounds were re-adjusted to match the scale-wise variance, the more entropy estimates increased. Crucially, this difference was more pronounced for older adults (paired t-test; r: p = 5e-6; MSE: p = 3e-4). Due to their increased high frequency power, coarse-graining decreased older adults’ scale-wise variance more so than younger adults’ variance. Thus, global similarity bounds presented a more liberal threshold at coarser scales for older adults than for younger adults, in turn producing lower MSE estimates. In line with this assumed link between high frequency power and coarse scale entropy as a function of global bounds, individual high frequency power at frontal channels was anticorrelated with coarse-scale entropy estimates when a global similarity bound was applied (Fig 8D), but was dramatically weaker when the similarity bound was recomputed for each scale (YA: r = -0.15; p = .302; OA: r = .20, p = .146). This is in line with our observation that coarse-scale age differences (Fig 7A) were not found when scale-wise bounds were used (Fig 7B).
Taken together, these results indicate that increased high frequency power with age can account for entropy decreases at coarse time scales, whereas the pattern irregularity of slow dynamics per se was not modulated by age.
A common observation in the MSE literature is that MSE is highly sensitive to task and behavioral differences at fine time scales, which are assumed to reflect fast dynamics. This is surprising given that high-frequency activity remains challenging to measure [56]. Moreover, previous studies suggest that fine-scale entropy reflects power spectral density (PSD) slopes [e.g., 14, 40]. Given that ‘Original’ MSE implementations contain both high- and low-frequency components due to the assessment of broadband signals, we probed whether fine-scale associations with PSD slopes depend on the presence of slow fluctuations and whether age-related slope variations can account for fine-scale entropy age differences (Hypothesis B).
As expected, individual fine-scale entropy was strongly and positively related to PSD slopes (Fig 9A) in both younger and older adults. Notably, after high-pass filtering the signal, the positive relation of fine-scale entropy to PSD slopes disappeared in both age groups (Fig 9B, dotted lines), and turned negative in older adults (see S6 Fig for scatter plots), while age differences in fine-scale entropy disappeared (Fig 7D). Relations between entropy and PSD slopes–and age differences–re-emerged once low-frequency content was included in the entropy estimation (Fig 9B, dashed and dotted lines), indicating that the presence of slow fluctuations was necessary for PSD slope relations. To assess whether varying PSD slopes accounted for fine-scale age differences in ‘Original’ MSE, we computed partial correlations between the measures. No significant prediction of age group status by fine-scale entropy was observed when controlling for the high collinearity with PSD slopes (r = -.04, p = .69), whereas PSD slopes significantly predicted age group status when controlling for fine-scale entropy (r = .37, p = 2e-4).
Finally, spectral slopes were anticorrelated with coarse-scale entropy when global similarity bounds were used (Fig 9B, solid lines), but not when criteria were scale-wise re-estimated (Fig 9B, dashed and dotted lines). This again suggests a presence of the scale-wise bias noted in Issue 1 (i.e., scale-wise bound divergence); subjects with shallower slopes (more high frequency power) had increasingly liberally-biased thresholds at coarser scales, resulting in overly low entropy estimates.
In sum, age differences in fine-scale entropy were conditional on the presence of both low- and high-frequency dynamics and reflected differences in PSD slopes; while the pattern irregularity of fast dynamics per se was not modulated by age.
The previous analyses highlighted how the spectral content of the signal can give rise to MSE time scale mismatches. However, our simulations also suggest a far more accurate mapping between entropy and power when scale-wise bandpass filters are used (Fig 4E). Concurrently, application of the band-pass implementation indicates a partial decoupling between entropy and variance (as reflected in the similarity bound ) age differences (Fig 7E). Specifically, older adults exhibited higher parieto-occipital entropy at alpha time scales (˜8–12 Hz) and lower central entropy at beta time scales (˜12–20 Hz) than younger adults (Fig 7; Fig 10A and 10B). Whereas alpha-band entropy was moderately and inversely correlated with alpha power (Fig 10C) and the age difference was inversely reflected in the similarity bound in a topographically similar fashion (Fig 10E), the same was not observed for entropy in the beta range for both age groups (Fig 10D and 10F). Promisingly, this indicates evidence for what many who employ MSE measures in cognitive neuroscience presume–that power and entropy can be decoupled, providing complementary signatures of neural dynamics.
This divergence of entropy and power in the beta band is particularly interesting as beta events have been observed to exhibit a more transient waveform shape [57, 58], while occupying a lower total duration during rest than alpha rhythms [42]. Indeed, it should be the rate of stereotypic spectral events that reduces pattern irregularity rather than the overall power within a frequency band. To better test this assumption in our data, we applied single-trial rhythm detection to extract the individual rate of alpha (8–12 Hz) and beta (14–20 Hz) events. As predicted, alpha events had a more sustained appearance compared with beta events as shown in Fig 10G and 10H (events were time-locked to the trough of individual events; see methods). Importantly, both alpha and beta event rate were inversely and moderately correlated with entropy estimates (Fig 10I and 10J) at matching time scales in the band-pass version. Correlations were also numerically higher than between power and entropy (Fig 10C and 10D), suggesting that entropy captured the non-stationary character of the rhythmic episodes that are not captured by sustained power estimates. The relationships remained stable after controlling for individual event rate and entropy in the age effect cluster of the other frequency band (partial correlations: alpha for younger adults: r = -.52, p = 2e-4; alpha for older adults: r = -.71, p = 8e-9; beta for younger adults r = -.49, p = 6e-4; beta for older adults: r = -.56, p = 2e-5), indicating separable associations between event rate and entropy between the two frequency bands. This is important, as our simulations suggest increased entropy estimates around narrow-band filtered rhythmicity (see Fig 4E). Furthermore, a permutation test indicated age differences in beta rate that were opposite in sign to the entropy age difference (see Fig 10L). In particular, older adults had a higher number of central beta events during the resting state compared with younger adults, thus rendering their beta-band dynamics more stereotypic. In sum, these results suggest that narrowband MSE estimates approximate the irregularity of non-stationary spectral events at matching time scales.
Finally, the neurobiological relation between narrowband rhythms and broadband signal characteristics (spectral slopes in particular; Fig 9) is a substantive question of considerable interest [59–61]. Rhythmic alpha events have been theorized to phasically modulate cortical excitability, with higher amplitudes of alpha events thought to reflect an overall reduction in population activity due to reduced excitability [55]. Such activation levels in turn have been related to scale-free broadband characteristics in human electrophysiological data [38, 51–54], which strongly contribute to fine-scale entropy estimates (Fig 9; S7 Fig). It is thus conceivable that alpha rhythms transiently reduce broadband irregularity. In line with this notion, negative associations between alpha power and fine-scale entropy have been observed [40, 62]. However, sample entropy’s joint sensitivity to broad- and narrowband dynamics (“Issue 2”) (see Fig 4) makes it ambiguous whether such associations truly reflect shifts in broadband features. We confirm this ambiguity in simulations (Fig 11A; sample entropy calculated for 250 ms signals consisting of varying slope coefficients in the presence or absence of alpha rhythms), where we observe that increased rhythmic regularity during alpha events concurrently decreases sample entropy, even when no change has occurred in the aperiodic signal component (Fig 11A: red panels). Controlling the spectral signal content via band-stop filters (here: 8–15 Hz) removes such circular entropy decreases due to increased narrowband regularity in the alpha band, while accurately indicating entropy changes due to changes in spectral slopes (Fig 11: green panels).
We used fine-scale sample entropy’s sensitivity to aperiodic slopes determined above (Fig 9; S7 Fig) to probe the relationship between broadband irregularity and rhythmic alpha events with high temporal precision in empirical data. To test transient modulations of irregularity during alpha rhythms, we leveraged the temporal on- and offsets of individual alpha segments (8–15 Hz; > 3 cycles) during eyes-open rest as uniquely identified by rhythm detection (see Fig 11B; see S8 Fig for exemplary traces). We created 250 ms segments surrounding the on- and offsets of alpha activity, followed by the calculation of sample entropy. To investigate potential differences as a function of magnitude, we median-split high- and low-amplitude alpha events. For both splits, we observed that sample entropy decreased upon alpha onset, whereas it recovered to high levels following alpha offset (Fig 11C1 and 11D1; red panels). However, due to the aforementioned circularity, the observation of transient entropy decreases during alpha periods offers little unambiguous insight beyond the successful identification of rhythmic event on- and offsets by the eBOSC algorithm. Importantly, transient entropy decreases during high-amplitude alpha events were also observed after removal of the alpha band (Fig 11C1; green panel), indicating that narrowband amplitude increases in the alpha-band were not sufficient to explain the observed entropy differences. This provides evidence that spontaneous, large-amplitude alpha rhythms during eyes open rest transiently decrease broadband signal irregularity, supporting their suggested role in the modulation of cortical excitability. We did not observe an interaction between alpha status and age for any of the contrasts (all p > .05), suggesting that decreased irregularity during transient alpha events is a preserved characteristic of cortical alpha rhythms across the adult lifespan. To further investigate a broadband effect, we calculated spectral slopes (using an auto-sandwiching approach, see methods). This analysis revealed a transient steepening of slopes during alpha events, in line with a broadband shift towards decreased excitability (Fig 11C2). In contrast to high-amplitude events, entropy decreases were not indicated for low-amplitude events after accounting for circularity bias (Fig 11D1, green panel). Similarly, no shift in aperiodic slopes was observed (Fig 11D2). This suggests that the originally indicated entropy decreases during low-amplitude events do not represent broadband shifts. This analysis highlights sample entropy’s potential to indicate fluctuations in signal irregularity with high temporal precision. Notably, the analysis reinforces the need for a targeted modulation of spectral content to avoid circular inferences, and reduce the ambiguity of results. Our findings suggest an alternative use case for dedicated bandpass filters that retains high sensitivity to broadband effects of interest. Specifically, the mechanistically informed use of band-stop filters here affords analyses into the modulators of signal irregularity and thereby can reveal non-trivial neurocomputational/-biological insights.
MSE aims to characterize the temporal irregularity of (neural) time series at multiple temporal scales. In the present study, we have highlighted two primary issues that may render the interpretation of time scales unintuitive in traditional applications: (Issue 1) biases from global similarity bounds, and; (Issue 2) the characterization of broadband, low-frequency dominated signals (see Fig 12A for a schematic summary). In the following, we discuss these effects and how they can impact traditional inferences regarding signal irregularity, in particular with regard to empirical age differences. Then, we discuss age effects in narrowband signal irregularity at interpretable temporal scales. Finally, we recommend procedures to improve scale-specific MSE inferences.
The ability to estimate entropy at coarser time scales provides the main motivation for a multi-scale implementation. Towards coarser scales, entropy is generally thought to represent the irregularity of increasingly slow dynamics. However, MSE’s traditionally global similarity bounds systematically bias coarse scale entropy estimates. Given that scale-wise variance decreases across scales, the liberality of global similarity bounds increases, causing entropy to decrease despite no ostensible shift in pattern irregularity. This bias is independent of the values of the global similarity bound–which did not differ across groups here–but rather depends on the removed variance at the time scale of interest. This issue has led to puzzling results in past work. For example, several papers using ‘original’ MSE have shown that in white noise signals (which by definition should be equally irregular at all time scales due to their randomness), entropy unintuitively decreases towards coarser scales, whereas pink noise signals undergo less entropy reduction across initial scales due to the removal of less high-frequency content [29] (S7 Fig). Strikingly, such puzzling effects have been used to validate the most common implementation of MSE [e.g., 27, 32] rather than to indicate the presence of a systematic bias in estimation. This appears motivated by the assumption that “changes of the variance due to the coarse-graining procedure are related to the temporal structure of the original time series, and should be accounted for by the entropy measure” [12]. We rather consider the similarity bound divergence a clear bias that invalidates the intuitive interpretation of time scales in MSE applications, and highlight that more intuitive broad-scale offsets are indicated when bound biases are removed (see S2 Text for elaboration on this issue).
Importantly, we highlight that this bias affects practical inferences. In the current resting-state EEG data, an age-related increase in high frequency power manifested unintuitively as a decrease in coarse-scale entropy via systematic group differences in the divergence of similarity bounds. Note that we presume that this age difference arises from a relative bias. As such, variations in high-frequency power suffice, even at low levels in 1/f scenarios, to systematically impact coarse-scale estimates and to specifically explain variance in a third variable of interest (e.g., age; see Fig 12B). Given that global similarity bounds remain prevalent in applications (see S1 Text), we hope that our practical example motivates the adoption of scale-varying parameters. Overall, we perceive little justification for the use of scale-invariant parameters in MSE estimation in future work. Indeed, as most previous work included biased, global bounds, reported coarse-scale effects may dominantly reflect false positives, while the sensitivity to true coarse-scale effects may have suffered, hence jointly increasing false negatives. Hence, results obtained with global bounds are ambiguous and hard to interpret. A critical task for future work (potentially including the re-analysis of existing data) will thus be to establish specific coarse-scale effects that provide empirical evidence for the practical utility of a multi-scale entropy computation. Recent advances for the robust estimation of coarse-scale entropy from sparse neuroimaging data [34, 63, 64] may be required to better estimate coarse-scale effects in in vivo data.
In parallel to the assumption of dominantly slow signal contributions to coarser scales, fine-scale entropy is often interpreted as a signature of “fast” temporal irregularity. However, it is typically estimated from broadband signals. As such, slow trends [35], neural rhythms at characteristic time scales [65] (Fig 4) and scale-free ‘background’ or ‘noise’ activity with a power-law form [38, 50, 53] (Fig 9; S7 Fig) jointly contribute to fine-scale entropy estimates. By linking fine-scale entropy to broadband PSD slopes, we replicated previous observations of increasing sample entropy with shallower slopes [14, 17, 29, 40, 46, 66] and shorter temporal autocorrelations [4, 27, 67]. However, we qualify this association by highlighting that the joint presence of slow and fast dynamics in the signal is necessary to produce such effects, hence verifying a broadband origin. At a mechanistic level, differences in spectral slopes and fine-scale entropy may jointly index variations in cortical excitability. Cortical neurons constantly receive a barrage of synaptic inputs. Variations in the excitatory and inhibitory summary statistics of these inputs robustly alter the conductance state of membrane potentials [for a review see 68], thereby producing variations in the irregularity of spike output and the appearance of global EEG signals [for a review see 69]. Whereas excitability is reduced during synchronized states characterized by strong low-frequency fluctuations, “desynchronized” cortical states feature enhanced sensitivity to external stimuli [70–72]. From a functional perspective, cortical information capacity, approximated via the entropy of cortical activity, may non-linearly vary alongside such excitation/inhibition (E/I) ratio, with highest information capacity afforded at intermediate levels of systemic excitability [10]. From a technical perspective, spectral (PSD) slopes have been proposed as a functional index of such an E/I ratio [49, 54, 73–75]. However, frequency-dependent filtering of current flow in the extracellular medium [76] or at the dendrite [77] may also contribute to the observed inter-individual differences in spectral slopes.
More generally, the association between broadband signal entropy and spectral slopes coheres with the notion that shallower slopes have a more ‘noisy’ or irregular appearance in the time domain. Thus, spectral slopes and temporal predictability are–at least in part–different perspectives on the same signal characteristic. Practically however, the correspondence between fine-scale entropy and 1/f slopes should nonetheless be tested, given that these scales are also sensitive to other signals characteristics, such as narrowband rhythmicity (Fig 4). Such necessity for narrowband control is highlighted by our analysis of transient fine-scale entropy changes during non-stationary alpha events (Fig 11). Only the removal of narrowband rhythmic regularity afforded non-circular insights. Specifically, we observed that broadband entropy transiently reduces following the onset and prior to the offset of parieto-occipital alpha rhythms, alongside a steepening of spectral slopes. This result is in line with alpha rhythms reflecting synchronized states with reduced cortical excitability [55, 59, 60, 78–81], but extends prior applications by characterizing non-stationary events at the single-trial level with high temporal precision, rather than temporal averages. Notably, our results contradict a prior observation that increased spontaneous alpha amplitudes at rest relate to a shallowing of low-frequency slopes, both in time and space [61]. Whether differences in frequency range, temporal specificity, or the stability of slope estimates contribute to this difference is an interesting question for future research that sample entropy may help to resolve. Notably, the fine-scale sensitivity of this effect highlights that single-scale broadband (sample) entropy–in the absence of multiscale implementations–is per se sensitive to broadband effects of interest, benefitting applications with limited available data and time [e.g., closed-loop setups: 62].
For entropy to be a practical and non-redundant measure in cognitive neuroscience, both its convergent and discriminant validity to known signal characteristics should be established. Multiple features can influence the temporal irregularity of neural time series. These include traditional ‘linear’ PSD features, (e.g., temporal autocorrelation, rhythmicity, etc.) as well as ‘non-linear’ features (e.g., phase resets, cross-frequency coupling, etc.). It is therefore worth noting that associations between spectral power characteristics and entropy estimates are partly anticipated (Fig 12B). For example, as noted before, entropy should reduce with increased rhythmic irregularity, and increase with shallowing of PSD slopes (and hence, shortening of temporal autocorrelations). However, the use of MSE is often motivated by its partial sensitivity to non-linear properties of brain dynamics [27, 46] that cannot be captured by traditional PSD analyses [e.g., 82, 83, 84]. In extreme cases, the absence of linear contributions may be erroneously inferred from the use of variance-based similarity bounds. Contrary to such orthogonality assumptions, our analyses highlight that differences in spectral variance (as captured by the similarity bound, which is typically neglected as a measure of interest when estimating MSE) can account for a large proportion of reported MSE effects [see also appendix in 27]. As such, non-linear characteristics per se may often do little to drive MSE estimates (see also results from a surrogate analysis in S3 Text, S9 Fig). This is in line with dominant linear power contributions to non-linear measures [85]. Conversely, the specificity to valid and unique non-linear effects increases after methodologically accounting for linear contributions.
Although the highlighted issues broadly apply to applications in which MSE is a measure of interest (e.g., assessment of clinical outcomes [e.g., 22]; prediction of cognitive performance [e.g., 46]), our results are especially relevant for MSE differences across the lifespan. Previous applications indicated that older adults exhibit lower coarse-scale entropy and higher fine-scale entropy compared with younger adults [16, 18, 27, 86]. While we conceptually replicate these results with the standard MSE implementation, our analyses question the validity of previous interpretations. In particular, our results suggest that age-related increases in coarse-scale entropy do not reflect valid differences in the irregularity of slow dynamics, but rather reflect differential high frequency power biases [see also 19]. Moreover, our analyses ascribe age differences in fine-scale irregularity to a flattening of PSD slopes, as observed from child- to adulthood [46] and towards old age [16, 18, 40, 49]. Such shallowing of scale-free slopes suggests relative shifts from distributed to local processing, and coheres with the notion of increased “neural noise” due to increases in the local excitation/inhibition ratio [54].
Across development, altered time scales of neural computations (as indicated by broadband changes in autocorrelations) [87] may reflect changes in intra- and inter-cortical connectivity [88], arising from reductions in grey matter density [89, 90], the integrity of associative white matter tracts [91], and changes in local receptor distributions and neuromodulation [92–96]. Dynamic interactions between such morphological changes may jointly shape control over local excitability and ‘neural noise’ across the lifespan [97]. Two alternative functional consequences of developmental noise increases have been proposed. On the one hand, intermediate levels of noise may provide beneficial stochastic resonance effects [9, 98–100], in line with relations between higher entropy and behavioral benefits in child- and adulthood [46], as well as in older adults [86]. In contrast, overwhelming amounts of local noise can produce adverse consequences [49, 101], supported by the observation that shallower slopes with advanced adult age relate to impaired working memory performance [49]. While further work including longitudinal assessments and behavioral probes will be necessary to disentangle the functional relevance of developmental changes, we argue that a principled separation of narrow- and broadband changes [102] will help to guide the search for neurobiological mechanisms driving entropy effects.
Taken together, our results suggest that entropy age differences dominantly arise from linear power differences, and appear at counterintuitive time scales. We confirmed the dominant contribution of age group differences in power characteristics using a surrogate analysis (see S3 Text, S9 Fig). Our surrogate analysis replicates a previous surrogate analysis that attributed age group differences mainly to linear auto-correlative properties [see appendix in 27, see also 85]. As we exclusively focused on univariate entropy, it remains an interesting question for future work whether our results are applicable to age-related decreases in ‘distributed’ entropy that capture the mutual information between distinct sensors [16].
Complementing traditional broadband applications, our use of narrowband MSE suggested age-related entropy increases in the posterior-occipital alpha band and decreases in central beta entropy that inversely tracked the regularity of alpha and beta events, respectively. Posterior-occipital decreases in alpha power and frequency with age are fundamental findings in many age-comparative studies [103]. While age-related increases in beta power are not observed as consistently [see e.g., 103 for a review], age-related increases in their prevalence have been observed during eyes open rest [104]. In addition, beta power increases over contralateral motor cortex during rest may reflect greater GABAergic inhibition in healthy aging [105]. While our results are not hemisphere-specific, they may similarly reflect increased inhibition in older adults, potentially reflected in an increased number of stereotypical beta events [58]. However, further work is required to establish the functional interpretation of narrowband age differences, as well as technical impacts of filter bandwidth, and individual center frequencies on narrowband results, especially given age differences in rhythmic peak frequencies [103]. Nevertheless, these results highlight that scale-specific narrowband filtering can provide novel, frequency-specific, insights into event/signal irregularity.
Notably, a narrowband approach may warrant different use cases than broadband entropy. In particular, the sensitivity to multi-scale information, such as cross-frequency interactions and waveform shape, is a defining characteristic of (and motivation for using) entropy as opposed to spectral analysis. However, this sensitivity trades off with specificity when a narrowband approach is chosen, which by definition enforces a more rhythmic appearance than the raw signal may convey [106]. Nonetheless, frequency-specific phenomena such as variations in the amplitude or the presence of rhythmic events are complementary signatures of irregularity in their own right. For example, long-range temporal correlations (LRTCs) of narrowband amplitudes provide an alternative window on the irregularity of temporal dynamics [107–109]. As such, targeted filter applications–either chosen a priori or as a follow-up to broadband entropy effects–may prove useful to delineate spectrally specific effects at directly interpretable neural time scales. Hence, we do not regard narrowband MSE as a replacement for the traditional low-pass implementation of MSE, but rather as a parallel tool for the exploration and probing of broadband effects. Moreover, sensitivity to broad-scale phenomena remains high when band-stop filters are used (e.g., Fig 11), highlighting the general feasibility of applying narrowband filters to derive broadband insights beyond the band-stop range.
The issues raised here suggest that additional steps need to be taken to achieve valid scale-wise estimates of MSE, and to support the perceived complementary nature of MSE relative to more typical measures (such as spectral power, etc.). We are optimistic that the following recommendations (Fig 12C), which have already been partially proposed [33–35, 63, 110], improve the utility of MSE as a principled tool for the estimation of complex brain dynamics.
In combination, such controls may go a long way toward establishing unique, complementary, and valid contributions of MSE in future work.
Many inferences regarding multiscale entropy in cognitive/clinical neuroscience rely on the assumption that estimates uniquely relate to pattern irregularity at specific temporal scales. Here we show that both assumptions may be invalid depending on the consideration of signal normalization and spectral content. Using simulations and empirical examples, we showed how spectral power differences can introduce entropy effects that are inversely mapped in time scale (i.e., differences in the high frequency power may be reflected in coarse entropy and vice versa; see Fig 12A). As these results suggest fundamental challenges to traditional MSE analysis procedures and inferences, we highlight the need to test for unique entropy effects (Fig 12B) and recommend best practices and sanity checks (Fig 12C) to increase confidence in the complementary value of pattern irregularity for cognitive/clinical neuroscience. While the warranted claim has been made that “it would be unreasonable simply to reduce sample entropy to autocorrelation, spectral power, non-stationarity or any of their combinations” [4], this should not mean that we cannot test whether one or more of these contributors may sufficiently explain MSE effects of interest. We thus propose that MSE effects may be taken as a starting point to explore the linear and nonlinear features of brain signals [e.g., 126]. We believe that empirical identification of the unique predictive utility of MSE will advance the quest for reliable mechanistic indicators of flexible brain function across the lifespan, and in relation to cognition, health, and disease.
To assess the influence of rhythmicity on entropy estimates, we simulated varying amplitudes (0 to 7 arbitrary units in steps of 0.5) of 10 Hz (alpha) rhythms on a fixed 1/f background. This range varies from the absence to the clear presence of rhythmicity (see S3 Fig for an example). The background consisted of -filtered Gaussian white noise (mean = 0; std = 1) with x = 1 that was generated using the function f_alpha_gaussian [127]. The background was additionally band-pass filtered between .5 and 70 Hz using 4th order Butterworth filters. Eight second segments (250 Hz sampling rate) were simulated for 100 artificial, background-varying trials, and phase-locked 10 Hz sinusoids were superimposed. To analyze the reflection of rhythmic frequency on time scales and to replicate a previously observed linear frequency-to-timescale mapping between the spectral and entropy domains [4, 22, 44], we repeated our simulations with sinusoids of different frequencies (5 Hz, 10 Hz, 20 Hz, 40 Hz, 80 Hz), that covered the entire eight second-long segments. For a specified amplitude level, the magnitude of frequency-specific power increases (or narrowband signal-to-noise ratio) increased alongside simulated frequencies due to the decreasing frequency power of pink noise, while the ratio of rhythmic-to-global signal variance (or global signal-to-noise ratio (SNR)) remained constant across simulated frequencies. We used the following definition: SNRglobal = , where RMSnoise is the root mean square of the pink noise time series and RMSsignal characterizes the pink noise signal with added rhythmicity.
To investigate the influence of similarity bounds and filter ranges in empirical data, we used resting-state EEG data collected in the context of a larger assessment prior to task performance and immediately following electrode preparation. Following exclusion of three subjects due to recording errors, the final sample contained 47 younger (mean age = 25.8 years, SD = 4.6, range 18 to 35 years; 25 women) and 52 older adults (mean age = 68.7 years, SD = 4.2, range 59 to 78 years; 28 women) recruited from the participant database of the Max Planck Institute for Human Development, Berlin, Germany (MPIB). Participants were right-handed, as assessed with a modified version of the Edinburgh Handedness Inventory [128], and had normal or corrected-to-normal vision. Participants reported to be in good health with no known history of neurological or psychiatric incidences, and were paid for their participation (10 € per hour). All older adults had Mini Mental State Examination (MMSE) [129, 130] scores above 25. All participants gave written informed consent according to the institutional guidelines of the Deutsche Gesellschaft für Psychologie (DGPS) ethics board, which approved the study.
Participants were seated at a distance of 80 cm in front of a 60 Hz LCD monitor in an acoustically and electrically shielded chamber. Following electrode placement, participants were instructed to rest for 3 minutes with their eyes open and closed, respectively. During the eyes open interval, subjects were instructed to fixate on a centrally presented fixation cross. An auditory beep indicated to the subjects when to close their eyes. Only data from the eyes open resting state were analyzed here. EEG was continuously recorded from 64 active (Ag/AgCl) electrodes using BrainAmp amplifiers (Brain Products GmbH, Gilching, Germany). Sixty scalp electrodes were arranged within an elastic cap (EASYCAP GmbH, Herrsching, Germany) according to the 10% system [131], with the ground placed at AFz. To monitor eye movements, two electrodes were placed on the outer canthi (horizontal EOG) and one electrode below the left eye (vertical EOG). During recording, all electrodes were referenced to the right mastoid electrode, while the left mastoid electrode was recorded as an additional channel. Online, signals were digitized at a sampling rate of 1 kHz.
Preprocessing and analysis of EEG data were conducted with the FieldTrip toolbox [132] and using custom-written MATLAB (The MathWorks Inc., Natick, MA, USA) code. Offline, EEG data were filtered using a 4th order Butterworth filter with a pass-band of 0.2 to 125 Hz. Subsequently, data were downsampled to 500 Hz and all channels were re-referenced to mathematically averaged mastoids. Blink, movement and heart-beat artifacts were identified using Independent Component Analysis [ICA; 133] and removed from the signal. Artifact-contaminated channels (determined across epochs) were automatically detected using (a) the FASTER algorithm [134], and by (b) detecting outliers exceeding three standard deviations of the kurtosis of the distribution of power values in each epoch within low (0.2–2 Hz) or high (30–100 Hz) frequency bands, respectively. Rejected channels were interpolated using spherical splines [135]. Subsequently, noisy epochs were likewise excluded based on FASTER and on recursive outlier detection. Finally, recordings were segmented to participant cues to open their eyes, and were epoched into non-overlapping 3 second pseudo-trials. To enhance spatial specificity, scalp current density estimates were derived via 4th order spherical splines [135] using a standard 10–05 channel layout (conductivity: 0.33 S/m; regularization: 1^-05; 14th degree polynomials).
MSE characterizes signal irregularity at multiple time scales by estimating sample entropy (SampEn) at each time scale of interest. A schematic of the estimation pipeline is shown in S1 Fig. The mMSE code is provided at https://github.com/LNDG/mMSE. A tutorial for computing mMSE has been published on the FieldTrip website (http://www.fieldtriptoolbox.org/example/entropy_analysis/).
The estimation of SampEn involves counting how often patterns of m successive data points reoccur in time (pm) and assessing how many of those patterns remain similar when the next sample m+1 is added to the sequence (pm+1). Given that amplitude values are rarely exactly equal in physiological time series, a similarity bound defines which individual data points are considered similar. This step discretizes the data and allows to compare data patterns rather than exact data values. The similarity bound is defined as a proportion r of the time series standard deviation (SD; i.e., square root of signal variance) to normalize the estimation of sample entropy for total signal variation. That is, for any data point k, all data points within k ± r × SD are by definition equal to k, which forms the basis for assessing sequence patterns. SampEn is finally given as the natural log of pm (r)/pm+1 (r). Consequently, high SampEn values indicate low temporal regularity as many patterns of length m are not repeated at length m+1. In our applications, m was set to 2 and r was set to .5, in line with prior recommendations [13] and EEG applications [27, 46, 136].
To extend sample entropy to multiple time scales, MSE ‘coarse-grains’ the original time series for multiple scale factors τ (here 1 to 42, where 1 refers to the original signal). The ‘Original’ MSE method [11, 12] averages time points within non-overlapping time bins (i.e., ‘point averaging’). Such point averaging is equivalent to a low-pass finite-impulse response (FIR) filter, which can introduce aliasing however [33, 137] and constrains the specificity towards increasingly slow signals, while not allowing specificity to fast dynamics or any particular frequency range of interest. To implement control over the scale-wise filter direction and to reduce aliasing, we applied either low- [31, 33, 137], high-, or band-pass filters at each scale factor. The low-pass cut-off was defined as and was implemented using a 6th order Butterworth filter, with nyquist defined as half the sampling rate of the signal. Similarly, the high-pass cut-off was defined as , implemented via 6th order Butterworth filters. Note that these cut-offs describe the upper and lower frequency bounds at each time scale, respectively. Finally, band-pass filters were applied to obtain narrowband estimates by sequentially applying Chebyshev Type I low- and high-pass filters (4th order with passband ripple of 1dB; chosen to achieve a fast filter roll-off), thus ensuring that each scale captured frequency-specific information. The passband was defined as BP = LP +- 0.05*LP. To avoid pronounced passband ripple for broad passbands, 10th order Butterworth filters replaced the Chebyshev filters at scales where the passband was larger than 0.5*Nyquist. At scale 1, only a high-pass 10th order Butterworth filter was applied as the sampling rate of the signal set the upper (Nyquist) frequency bound. These settings were chosen to optimize the pass-through of signals within the pass-band and the attenuation of signals outside the pass-band. Two-pass filtering using MATLAB’s filtfilt function was applied to achieve zero-phase delay. S4 Fig shows the spectral attenuation properties [138] of the filters. To avoid edge artefacts, input signals were symmetrically mean-padded with half the pseudo-trial duration (i.e., 1500 ms). After filtering, we implemented a point-skipping procedure to down-sample scale-wise signals (see S1 Fig). Since point-skipping allows for increasing starting point permutations k for increasing scale factors τ, we counted patterns separately for each starting point k, summed the counts of pattern matches and non-matches across them, and computed sample entropy based on the summed counts as described above: . This implementation is equivalent to “refined composite MSE” [110] and can improve the stability of entropy results for short or noisy signals [31, 110]. Note that no point skipping was performed in the ‘high-pass’ implementation to avoid low-pass filtering. As a result, the signals at increasing scale factors remained at the original sampling rate. To alleviate computational cost, scale factors were sampled in step sizes of 3 for empirical data (only for the ‘high-pass’ implementation) and later spline-interpolated. An adapted version of MSE calculations was used for all settings [64], in which scale-wise entropy was estimated across discontinuous data segments. The estimation of scale-wise entropy across trials allows for reliable estimation of coarse-scale entropy without requiring long, continuous signals, while quickly converging with estimates from continuous segments [64].
Following scale-specific filtering, all implementations re-calculated sample entropy for the scale-specific signal. Crucially, in ‘Original’ applications [11, 12], the similarity bound is calculated only once from the original broadband signal. As a result of filtering, the scale-wise signal SD decreases relative to the global, scale-invariant similarity bound [29]. To overcome this limitation, we recomputed the similarity bound for each scale factor, thereby normalizing MSE with respect to changes in overall time series variation at each scale (.5 x SD of scale-wise signal).
As the interpretation of estimates at each scale is bound to the scale-wise spectral content, our Figures indicate spectral bounds of the scale-wise signals alongside the scale factor as follows: for the low- and band-pass implementation, we indicate the low-pass frequency as calculated above as the highest resolvable (i.e., Nyquist) frequency in the scale-specific signal. Likewise, for the high-pass implementation, we indicate the high-pass limit as the lowest resolvable frequency in the scale-specific signal. In the main text, we refer to higher scale factors as ‘coarser’ scales’ and lower scale factors as ‘finer’ scales, in line with the common use in the literature. Note that the sampling rate of the simulated data was 250 Hz, whereas the empirical data had a sampling rate of 500 Hz.
Power spectral density estimates were computed by means of a Fast Fourier Transform (FFT) over 3 second pseudo-trials for 41 logarithmically spaced frequencies between 2 and 64 Hz (employing a Hanning-taper; segments zero-padded to 10 seconds) and subsequently averaged. Spectral power was log10-transformed to render power values more normally distributed across subjects. Power spectral density (PSD) slopes were derived by linearly regressing power values on log10 -transformed frequencies (i.e., log-log fit). The spectral range from 7–13 Hz was excluded from the background fit to exclude a bias by the narrowband alpha peak [40, 49].
Spectral power, even in the narrowband case, is unspecific to the occurrence of systematic rhythmic events as it also characterizes periods of absent rhythmicity [e.g., 139]. Specifically detecting rhythmic episodes in the ongoing signal alleviates this problem, as periods of absent rhythmicity are excluded. To investigate the potential relation between the occurrence of stereotypic spectral events and narrowband entropy, we detected single-trial spectral events using the extended BOSC method [42, 140, 141] and probed their relation to individual entropy estimates. In short, this method identifies stereotypic ‘rhythmic’ events at the single-trial level, with the assumption that such events have significantly higher power than the 1/f background and occur for a minimum number of cycles at a particular frequency. This effectively dissociates narrowband spectral peaks from the arrhythmic background spectrum. Here, we used a one cycle threshold during detection, while defining the power threshold as the 95th percentile above the individual background power. A 5-cycle wavelet was used to provide the time-frequency transformations for 49 logarithmically-spaced center frequencies between 1 and 64 Hz. Rhythmic episodes were detected as described in [42]. Following the detection of spectral events, the rate of spectral episodes longer than 3 cycles was computed by counting the number of episodes with a mean frequency that fell in a moving window of 3 adjacent center frequencies. This produced a channel-by-frequency representation of spectral event rates, which were the basis for subsequent significance testing. Event rates and statistical results were averaged within frequency bins from 8–12 Hz (alpha) and 14–20 Hz (beta) to assess relations to narrowband entropy and for the visualization of topographies. To visualize the stereotypic depiction of single-trial alpha and beta events, the original time series were time-locked to the trough of individual spectral episodes and averaged across events [c.f., 57]. More specifically, the trough was chosen to be the local minimum during the spectral episode that was closest to the maximum power of the wavelet-transformed signal. To better estimate the local minimum, the signal was low-pass filtered at 25 Hz for alpha and bandpass-filtered between 10 and 25 Hz for beta using a 6th order Butterworth filter. A post-hoc duration threshold of one cycle was used for the visualization of beta events, whereas a three-cycle criterion was used to visualize alpha events. Alpha and beta events were visualized at channels POz and Cz, respectively.
The relation of narrowband alpha events to broadband irregularity represents an empirical question of interest (see Introduction). We examined the relation between these signatures, while controlling for the circular, intrinsic relation between alpha-based regularity and entropy. To highlight the issue of circularity, we first simulated expected links between the two signals by creating 250 ms of data, consisting of (a) aperiodic slopes of , (b) aperiodic slopes of , as well as equivalent versions with superimposed alpha rhythms of unit amplitude (c, d). We probed the practical potential of a 8–15 Hz band-stop filter (6th order Butterworth) to remove the influence of alpha on broadband entropy. Entropy was calculated for the first MSE scale, reflecting broadband sample entropy. Next, in empirical data, we leveraged the temporal on- and offsets of individual alpha segments (8–15 Hz; > 3 cycles) as identified via rhythm detection and segmented the original data to include 250 ms preceding and following event on- and offsets (see S8 Fig for empirical examples). For each subject, all events across posterior-occipital channels at which event number was highest (see Fig 11B) were included in this analysis. At each channel we performed a median split of events according to their amplitude (high/low). We created versions with and without application of 8–15 Hz bandstop filters (S8 Fig), followed by the calculation of sample entropy. We assessed the impact of transient alpha events on irregularity via paired t-tests between alpha on vs. off contrasts, both at event on- and the offset, and individually for low and high amplitude events. As post-hoc tests, we assessed potential interactions between alpha presence and age via linear mixed effect models (random subject intercept). To probe the presence of a broadband effect, we assessed the spectral slopes for the same segments. To improve spectral resolution, we”auto-sandwiched” each 250 ms segment by appending it in x- & y-inverted forms at the original segment’s on- and offset. This effectively increased segment duration to 750 ms, while retaining autocorrelative properties. We then calculated an FFT of each segment (2–90 Hz; 45 2x steps; Hanning taper; 4 Hz smoothing box; zero-padded to 10 s). Linear slopes were fit in log-log space, after excluding the 5–20 Hz range to remove the influence of the rhythmic alpha peak. Individual entropy estimates were averaged across alpha on- and offsets to remove measurement noise, and were statistically compared between alpha on & off periods via paired t-tests.
Spectral power and entropy were compared across age groups within condition by means of independent samples t-tests; cluster-based permutation tests [142] were performed to control for multiple comparisons. Initially, a clustering algorithm formed clusters based on significant t-tests of individual data points (p < .05, two-sided; cluster entry threshold) with the spatial constraint of a cluster covering a minimum of three neighboring channels. Then, the significance of the observed cluster-level statistic, based on the summed t-values within the cluster, was assessed by comparison to the distribution of all permutation-based cluster-level statistics. The final cluster p-value that we report in figures was assessed as the proportion of 1000 Monte Carlo iterations in which the cluster-level statistic was exceeded. Cluster significance was indicated by p-values below .025 (two-sided cluster significance threshold). Effect sizes for MSE age differences with different filter settings were computed on the basis of the cluster results in the ‘Original’ version. This was also the case for analyses of partial correlations. Raw MSE values were extracted from channels with indicated age differences at the initial three scales 1–3 (>65 Hz) for fine MSE and scales 39–41 (<6.5 Hz) for coarse MSE. R2 was calculated based on the t-values of an unpaired t-test: [143]. The measure describes the variance in the age difference explained by the measure of interest, with the square root being identical to Pearson’s correlation coefficient between continuous individual values and binary age group. Effect sizes were compared using the r-to-z-transform and a successive comparison of the z-value difference against zero: [144]. Unmasked t-values are presented in support of the assessment of raw statistics in our data [145].
We thank our research assistants and participants for their contributions to the present work.
17 Feb 2020
Dear Mr. Kosciessa,
Thank you very much for submitting your manuscript "Standard multiscale entropy reflects spectral power at mismatched temporal scales: What’s signal irregularity got to do with it?" for consideration at PLOS Computational Biology.
As with all papers reviewed by the journal, your manuscript was reviewed by members of the editorial board and by several independent reviewers. The reviews (below this email), show a general appreciation of the methodology and attention to an important problem, as well as suggesting some possible issues and improvements.
Another important aspect is the relevance to the scope of the journal, which raised some concerns both with the reviewers and in editorial discussions. It would then be crucial to show an actual relevance in terms of better understanding of the biological mechanisms behind the signal properties. This might be difficult to show, and in that case, upon a resubmission with convincing answers to the other points, we would offer you immediate publication in PLOS One.
We cannot make any decision about publication until we have seen the revised manuscript and your response to the reviewers' comments. Your revised manuscript is also likely to be sent to reviewers for further evaluation.
When you are ready to resubmit, please upload the following:
[1] A letter containing a detailed list of your responses to the review comments and a description of the changes you have made in the manuscript. Please note while forming your response, if your article is accepted, you may have the opportunity to make the peer review history publicly available. The record will include editor decision letters (with reviews) and your responses to reviewer comments. If eligible, we will contact you to opt in or out.
[2] Two versions of the revised manuscript: one with either highlights or tracked changes denoting where the text has been changed; the other a clean version (uploaded as the manuscript file).
Important additional instructions are given below your reviewer comments.
Please prepare and submit your revised manuscript within 60 days. If you anticipate any delay, please let us know the expected resubmission date by replying to this email. Please note that revised manuscripts received after the 60-day due date may require evaluation and peer review similar to newly submitted manuscripts.
Thank you again for your submission. We hope that our editorial process has been constructive so far, and we welcome your feedback at any time. Please don't hesitate to contact us if you have any questions or comments.
Sincerely,
Daniele Marinazzo
Deputy Editor
PLOS Computational Biology
Daniele Marinazzo
Deputy Editor
PLOS Computational Biology
***********************
Reviewer's Responses to Questions
Comments to the Authors:
Please note here if the review is uploaded as an attachment.
Reviewer #1: This technical paper investigates thoroughly, in simulations and with representative EEG data, two issue that affect the interpretability of multiscale entropy (MSE), a widely used tool for the quantification of the complexity of physiological time series performed at multiple time scales: the impact of one specific parameter (the “similarity bound”) on the assessment of complexity at long (coarse) time scales, and the impact of dominant low frequency oscillations on the assessment of complexity at short (fine) time scales.
The two issues are actually well known in the literature, and (as also partially acknowledged by the authors) corrective methodological approaches have been proposed to successfully deal with them (see specific comments below). This aspect should be better emphasized in the paper. Nevertheless, in spite of the limited methodological novelty and neurophysiological appeal of the results, the messages conveyed by the paper are important, because it is true that MSE is often misused or at least used not appropriately, and that inferences based on this measure must be made with full awareness of its meaning and limitations (which is often not the case in neuroscience and physiology applications). Therefore, I recommend acceptance of this paper after the content is simplified, in order to focus on the specific message, some interpretations are better placed in the context of the existing methodological literature, and the general claims are smoothed a bit, in order to better acknowledge that existing methodological solutions already make MSE a useful tool - provided that it is used consciously.
Specific:
- The effect of the signal variance (here reflected in the term “similarity bound”) on the original formulation of MSE is long known (e.g., from Valencia et al. [R1]). Here it is confirmed in simulations, and its effects are highlighted in the practical application studying age differences in resting state EEG. The proposed solution to compute sample entropy based on the variance of the signal after the change of scale (here denoted as “use of local similarity bound”) is also well established (it is for example used in linear parametric estimators of MSE (e.g., Faes et al. [R2])). The recommendation to abandon the original MSE formulation (line 678) is valid, but should be substantiated more clearly in terms of existing modifications of the MSE algorithm (like the refined MSE).
- Also the diffuse reflection of slow oscillations on entropy estimated even at fine time scales (i.e., a contribution of low frequencies to broadband signals) is expected, since at fine time scale the signal contains both low and high frequencies which can both contribute to the irregularity. In particular, the dominance of very low frequency components (where ‘very low’ relates to the sampling frequency and the length of the time series) introduces trends in the analyzed signal which can be assimilated to nonstationary behaviors; this precludes a proper evaluation of Sample Entropy, because stationarity is a formal prerequisite to the evaluation of entropy measures. This issue is thoroughly dealt with in recent works (e.g., Xiong et al. [R3]). Solutions have been proposed, either based on simple detrend of the time series via highpass filters in order to allow MSE to focus on short-term dynamics without the biasing effect of slow trends, or based on incorporating the long-range correlations into parametric models for the estimation of MSE (Faes et al. [R4]) in order to describe both long- and short-term dynamics in MSE analysis. These aspects should be better emphasized in the paper, in order both to better place the work in the context of the existing literature and to substantiate the recommendation of using spectral filters to deal with the second issue highlighted in this work.
- The paper correctly questions the interpretation often given to MSE that it reflects the complexity of the analyzed signal observed at specific time scales. In doing this, however, it should be stated more explicitly that the mismatch between specific temporal scales and MSE values is a defining feature of this measure, which distinguishes it from simple spectral analysis. The use of lowpass filters encompassing several temporal scales is indeed important in order to allow the measure to capture complex (possibly nonlinear) behaviors like the superposition of rhythms and at different frequency and their weaker or stronger interaction. Therefore, it should emerge more clearly from the paper that, while it is incorrect to interpret MSE as a scale-specific measure, the “scale” focused by this measure is actually an upper limit of the range of scales analyzed (i.e., at “scale 1”, all signal frequencies are considered, and at scale tau all frequencies lower than fs/(2+tau) are considered – where fs is the sampling frequency).
Related to the comment above, two issues should be also considered:
(i) not only it is wrong to assume that MSE reflects a specific time scale, but also it makes little sense to look scale-specific MSE values through narrowband filters; the authors correctly show that narrowband filters reconcile MSE with the spectral content (e.g. in Fig. 4E), but they should acknowledge that in that case MSE itself becomes of little use, since it merely reflects the presence of specific oscillations and thus it adds little or no information to much simpler spectral analysis (the evaluation of nonlinear effects and cross-frequency interactions is precluded by the narrowband filter).
(ii) the message that MSE is of little use because it reflects spectral content at mismatched temporal scales (given first of all in the title of the article) is misleading and should be tuned down. While it is true that a the balance among spectral content at different frequencies has an effect on the MSE value, such a balance is one of the factors that determines signal complexity (e.g., white noise is highly complex, an oscillation embedded in noise is less complex, the presence of another oscillation at different frequency increases complexity, the complexity increases with the bandwidth of stochastic oscillations, and so on)
- The paper is very well written, but it is likely too long compared to the message that is given. An effort to focus on the main message would be helpful to better convey it.
[R1] Valencia, J. F., et al. (2009). Refined multiscale entropy: Application to 24-h holter recordings of heart period variability in healthy and aortic stenosis subjects. IEEE Transactions on Biomedical Engineering, 56(9), 2202-2213.
[R2] Faes, L., et al. (2017). Efficient computation of multiscale entropy over short biomedical time series based on linear state-space models. Complexity, 1768264.
[R3] Xiong, W., et al. (2017). Entropy measures, entropy estimators, and their performance in quantifying complex dynamics: Effects of artifacts, nonstationarity, and long-range correlations. Physical Review E, 95(6), 062114.
[R4] Faes, L., et al. (2019). Multiscale information storage of linear long-range correlated stochastic processes. Physical Review E, 99(3), 032115.
Reviewer #2: Overview
This manuscript, by Kosciessa and colleagues, investigates methodological properties of applying multiscale entropy (MSE) to neural field data, such as EEG. The research question is broadly, whether MSE actually measures signal irregularity across scales, in the way that it is typically interpreted to do. More specifically, this investigation centres on two key methodological points about how typical approaches compute and analyze the 'scales' that are foundation of the 'multi-scale' approach, specifically:
whether using a global similarity bound biases measures across scales, since the coarse-graining (or analogous) procedure reduces the variance as one restricts the data to different scales, and so a globally computed similarity bound effects each scale differently and thus conflates signal regularity and signal variance; and
whether the coarse graining procedure is specific to presumed scales, or whether, for example, scales that include, but are not limited to, high frequency activity capture entropy measures that are specific to high frequency activity, or whether the low frequency activity also impacts the measures and thus conflates the typical interpretations of measures at particular scales.
The authors demonstrate, in a series of simulations, that these two methodological issues are indeed problems, such that global similarity bounds and broadband data lead to biases that violate how the resulting measures are typically interpreted. In analyses of EEG data, they further demonstrate how this is the case, showing how patterns of MSE relate to different similarity bounds and coarse graining approaches. These analyses also demonstrate how MSE measures relate to power. Through this work, they propose and investigate solutions to these problems: notably to recompute similarity bounds specific to the variance at each scale, and to use filters to isolate data more specifically to desired scales.
Overall, I commend the authors on a rigorous and methodologically minded paper, that demonstrates methodological insight into a relatively common method in the field, demonstrated through what I find to be a compelling and well chosen empirical analysis. I was very impressed by the methodological rigour of the work, the open nature of the code and data, and the clear demonstration of how these methodological points work, and what they mean. This seems to me to be important work for the field, and impactful work in terms of thinking of how to interpret many studies that have employed these methods.
Overall, I did not find any major issues with the overall framing, logic or main results of this paper. I also find it to be well written. My review is therefore focused on a what I consider to be some minor revisions that include some clarifications regarding their proposed adaptations MSE, a possible small extension to the simulations, and some possible tweaks to the figures.
Comments:
1) Bandpass Filtering
My main conceptual question / comment is about the use, interpretation and recommendation for the bandpass filtered version of the entropy. I like the idea of this approach, and the justification of how it addresses the issues raised, and is a design that actually specifies scales precisely. However, it still seems slightly unclear what it captures, and if (or to what extent) this is a) the same as the typical conceptualization of MSE and b) distinct from other measures (such as power and burst analysis).
In particular, the demonstration that alpha and beta differ in bandpass MSE between groups is very interesting. However, given that the alpha does relate to power, and the beta does relate to the burst analysis, I found it unclear if the there is a conclusion as to whether this MSE approach adds information over and above a combined power & burst analysis of rhythmic activity. The implication in the results appears to be that the difference in beta is proposed to be related to different bursting statistics of rhythmic activity between groups. If this is the appropriate interpretation, it seems to me that a) it is not entirely clear to me that tracking burst properties (such as probably and duration of bursts) of rhythmic activity is quite the same idea as MSE has been proposed to relate to - of signal irregularity (seemingly, more broadly). This could perhaps just be a conceptual issue, as burst dynamics are clearly an aspect of signal (ir-)regularity, but my impression is that entropy measures are more typically thought of as representing the 'eracticness' of time series, and not necessarily the propensity for bursts of rhythmicity. It is also somewhat unclear to me if this MSE analysis offers something more than could be gleaned from a power and burst analysis of these bands.
Specifically, I would appreciate a comment in the manuscript that addresses 1) whether this MSE measure adds something more than would be gleaned from a power and burst analysis and 2) whether, if one does try this, the results should be interpreted somewhat differently than how 'original MSE' has typically been interpreted and discussed.
Some other questions came to mind about this narrowband MSE measure:
It seems very likely the estimate of MSE calculated from bandpassed data would be dependent on the width of the bandpass filters applied. Is this the case? And if so, what does this relation mean, in practice, for computing this measure, and how should one choose bandwidths?
Narrow bandpass filters (and the bandpass filters applied here seem quite narrow, at the lower frequencies) enforce sinuisoidality onto the filtered data. Narrowband filtering then removes what may be considered interesting waveform shape properties that one might have considered could be an interesting driver of entropy differences. This seems potentially related to the notion that this form of MSE may not capture some desired notions of signal irregularity. The question is then: how does the narrowband filtering effect waveform shape in a way that might affect entropy measures?
Should one consider potential differences in centre frequency if one is comparing between groups? The two groups in this analysis differ in age in a manner that would be expected to demonstrate a difference in peak alpha frequency. Given this, could or should one consider aligning compared bands to individual centre frequencies? If not, to what extent might difference in entropy simply relate to differences in centre frequency?
To be clear, I think that a full investigation of all the properties of the bandpass filter version of MSE could quickly become large, and that is out of scope of this particular manuscript. I don't consider that these 'other questions' must all be addressed or answered in this manuscript. They are suggested to motivate that it is perhaps worth adding a quick discussion point noting that there are open questions (and that there may be some limitations) to the narrowband MSE approach, and that further work is needed on this.
2) Simulating Variable 1/f
If there is one asymmetry between what is done in the simulations and what is presented in the real data, I would say it is that a key point made in the EEG data is how the difference in 1/f backgrounds of the data is a key factor in affecting MSE measures, and yet this point is not demonstrated in the simulated data. All of the simulations use a single value for x in 1/f^x. While I think the inference and interpretation of the 1/f effects are clear, if it is relatively straight-forward to compute and add, a set of simulations across different 1/f components of the simulated data (perhaps just 2 different values for x, representing each group) might be a nice addition. This would clearly show, in the simulated data, that this effect can be isolated to changes in the 1/f structure of the data.
3) Values for the PSD Slopes
A minor comment, for clarification, is regarding the actual values for the PSD slopes that are plotted, for example in Figure 6 D1 and in Figure S6. The actual values are very small, and not typical of values for PSD slopes, which are typically closer to values of -1 (equivalent to 1/f^1, or pink noise). A 'back of the envelope' estimated calculation of the slope of the spectrum plotted in Figure 6 A2 is also more consistent with this value or around -1. Have the potted numbers been transformed in some way? I found no mention in the methods that would explain this, and so some clarification of the values plotted would be useful. Perhaps related, it is unclear why the words 'minus 0' appear in the title of the panel D1.
4) Figure Tweaks
Finally, I have is actually a collection of notes and small suggestions for the figures. There are a lot of figures which nicely present a lot of data and findings, but I feel like some small updates would help make the figures a little more accessible to the new reader.
Figure 1
In 1B, the 'similarity bound' is indicated as '(r x SD)', where SD presumably standards for standard deviation. However, in the figure legend, similarity bounds are defined as 'r * 0.5 of signal variance'. It would be clearer if one form of the definition was used, which based on the other figures, I presume should be the one using standard deviation. It would be useful to define 'SD' here.
Figure 2
It's a small aesthetic nitpick: but the y-axis labels for A & B are not aligned
Figure 4
I think it's a mistake that the legend refers to a 'blue-to-red' line gradient?
Fig 6:
I initially found it slightly unclear on how to read and interpret the contrasts and topographies. If the paper is to be organized with the methods at the end (as current), I think it would be useful to note in the figure legend a brief note on what the p-values are that are plotted (these are presumably the results of the significance of the cluster-level statistic?)
The legend suggests asterisks are used to indicate significance, but this does not appear to be used in the plot?
I initially found the labelling to be unclear and wasn't sure what the 'B1', 'C1', etc labels were in panels 'A1' and 'A2'. Assuming these refer to the other sub-panels (and seem to motivate the labelling) briefly noting what these refer to in the legend could be useful.
Is A1 the 'Original' measure of entropy? This should indicated in the title and/or legend.
Figure 7:
I'm unclear on why the x-scale for Row 1 appears to be different from how it was plotted in Figure 6A1. Is what is plotted in Figure 6A1 not a copy of one of the plots in Figure 7 Row 1? It would be easier to highlight this if the scale were the same. Perhaps there is a reason they are different, but this could be noted, as otherwise it initially make the plotted measures between figure seem more different than they actually are.
Could you label the rows? In particular, I found it somewhat unclear as to what comparisons rows 2 and 4 were exactly, and especially row 4 (I believe it is the difference of similarity bounds, between groups, across channels?). This could be indicated more clearly in the figure legend.
It would be useful to you indicate (probably in the legend) in which direction the comparison was made, so that the direction of the t-values can be interpreted with respect to groups.
Figure 8:
The title in C says that entropy difference is <8 Hz, but the figure legend reports it's < 6 Hz.
Figure 9:
The labels for the lines in B are not clearly differentiated. The long dash and full solid labels look the same in the way they are plotted in the legend.
Figure 10:
Are all the topographies on the same color bar? The color bar is not consistently plotted, and if one is to be used across all topographies, this could be noted in the legend.
For G & H, are these collapsed across age groups? This would be useful to clarify.
Figure 11:
A stylistic suggestion is that I don't think you need the 'higher frequencies' arrow twice (at both the top and the bottom).
I initially found it somewhat unclear as to how to interpret the dashed coloured arrows. I'm not entirely sure if there is a change to be made, though perhaps they could be briefly mentioned in the figure legend, for clarification?
Reviewer #3: This is a comprehensive and thought-provoking assessment of the relation between spectral power (SP) and multiscale entropy (MSE), both of which have been used to characterize changes in brain dynamics across a wide range of conditions (mainly from EEG data). One of the perpetual challenges with MSE is the attempt to relate the scales for downsampling to something akin to frequency. This has some support in that there does appear to be some similarity in scale differences in MSE and frequency differences in SP. For example, the authors show age-related differences in MSE (higher fine-scale, lower coarse-scale) map to SP changes with higher power in faster frequencies and lower at slower frequencies - similar to what we (McIntosh) and others have published. While it is tempting to link these two metrics, I think most agree that it is not so straightforward (figure 11 sort of suggests this).
The first point that is emphasized is that the "original" MSE algorithm does not change the similarity criteria with successive downsampling of the signal. This appears to introduce a bias at coarse scales. This is a concern and definitely needs to be considered for MSE applications. The authors quite convincingly show that this can affect interpretations of group differences (Fig 7).
The second point, if I understand correctly, is that a direct mapping between scale and frequency is difficult because of cross-spectral dependencies. This isn't too surprising and in my read of the extant literature, I think most would agree. The Courtiol paper, for example, does make this point, and most of the work we did notes the similarities in MSE and SP, but also that they are sensitive to different aspects of brain signals. For example, nonlinearities (e.g., cross-spectral dependencies) do affect MSE, but by definition do not affect SP. Indeed, the surrogate analysis the authors mention (which was done in the McIntosh et al, 2008 paper), shows this quite nicely. Also some other work (Bruce, Bruce, Vennelaganti, 2009) suggested that better predictor of sample entropy was found to be related to the power ratio of higher to lower frequencies:(alpha + beta)/(delta + theta). Moreover, the direct link between distance (local vs distributed), SP and MSE is also not straightforward. For example, while the age-effects reported in the present paper and by others do replicate nicely, there does appear to be a age-related change in local entropy (higher in ageing) and mutual information (lower in ageing), which maps to coherence estimates (McIntosh et al, Cereb Cortex 2014). These "connectivity" effects, however, are not constrained to a particular scale or frequency, which emphasizes the inherent nonlinearity in these effects. The Courtiol et al work also notes this sensitivity to nonlinearities.
Thus, I am left wondering if the second message is as conclusive as the authors seem to suggest. It's complicated, no doubt, as the Venn diagram in Figure 11 suggests. Which takes me to the main point. I get the impression that the authors advocate SP as the 'gold standard' on which to map the MSE effects, but I am not sure this is a valid. The simulation used to support this changes power at the alpha frequency in an 1/f signal, which makes a very strong assumption on the underlying biological processes. Thus, insofar as the brain signals actually do this -- add power to a restricted frequency band -- the simulation is useful. But, if there are changes in the brain signal that span frequencies (e.g., cross-spectral dependencies), then the spectral filters won't really help and may in fact obscure things. I suggest the authors actually do the surrogate analyses they propose as suggestion (d) on page 32 to really disentangle this. Otherwise, I am concerned that someone who reads this will come away thinking that measuring SP is really all we need to do, which I don't think is the message the authors want to convey.
Reviewer #4: This is a well-done paper. The authors should be congratulated for their thorough and didactic exposition of an intricate problem. However, the study left me unconvinced that, at least in its current state, it has a place in a leading computational neuroscience/biology journal. I will list my concerns in more detail below, but will have to leave it to the editors to weigh my comments against the broader context of the scope of their journal.
The paper starts out with multi-scale entropy (MSE) being a key measure, “increasingly”, "often", "commonly" applied in a specific way, which the authors go on to explain in admirable detail (i.e., applying a broad-band variance scaling). The authors then go on to show how this creates a severe band-specific spectral power confound. Using cross-sectional data from young and older brains they go on to argue that conclusions on entropy/variability changes over the life-span could as well be confounded by power. They end on a set of cook-book-style recommendations. I have not much to quibble with in the way they set out this argument.
The overriding impression, however, was that of a data analysis tutorial or a methods paper. The implications for our understanding of neural functioning and/or neuro-cognitive computation did not become clear.
From the outset, I had been expecting a closer link to the neurobiology of (multi-scale) entropy, and I do think that such a link is necessary to lift the current paper, and its potential impact, beyond the ranks of a methodological comment.
Not least, the mixture of simulations and data re-analysis left me furthermore unclear what I am to take from this study. For a hardcore signal analysis/digital signal processing audience, some of the treatment might be too shallow, while for the applied neuroscientist the particular example of age-related changes in spectral slope might be too specific to fully take away what the consequences for future MSE applications might be.
Should a revision be invited, I would suggest to focus the paper more on what we learn about age-related change in the spectral composition / E:I imbalance in the aging brain and would suggest to tone down the overly specific, cook-book-like recommendations on how to analyse MSE instead.
Minor: SD and variance are used a bit too interchangeably, especially aroudn Fig. 1/caption thereof.
**********
Have all data underlying the figures and results presented in the manuscript been provided?
Large-scale datasets should be made available via a public repository as described in the PLOS Computational Biologydata availability policy, and numerical data that underlies graphs or summary statistics should be provided in spreadsheet form as supporting information.
Reviewer #1: Yes
Reviewer #2: Yes
Reviewer #3: Yes
Reviewer #4: Yes
**********
PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.
If you choose “no”, your identity will remain anonymous but your review may still be made public.
Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.
Reviewer #1: Yes: Luca Faes
Reviewer #2: Yes: Thomas Donoghue
Reviewer #3: Yes: Anthony Randal McIntosh
Reviewer #4: No
Figure Files:
While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email us at figures@plos.org.
Data Requirements:
Please note that, as a condition of publication, PLOS' data policy requires that you make available all data used to draw the conclusions outlined in your manuscript. Data must be deposited in an appropriate repository, included within the body of the manuscript, or uploaded as supporting information. This includes all numerical values that were used to generate graphs, histograms etc.. For an example in PLOS Biology see here: http://www.plosbiology.org/article/info%3Adoi%2F10.1371%2Fjournal.pbio.1001908#s5.
Reproducibility:
To enhance the reproducibility of your results, PLOS recommends that you deposit laboratory protocols in protocols.io, where a protocol can be assigned its own identifier (DOI) such that it can be cited independently in the future. For instructions, please see http://journals.plos.org/compbiol/s/submission-guidelines#loc-materials-and-methods
18 Apr 2020
Dear Mr. Kosciessa,
We are pleased to inform you that your manuscript 'Standard multiscale entropy reflects neural dynamics at mismatched temporal scales: What’s signal irregularity got to do with it?' has been provisionally accepted for publication in PLOS Computational Biology.
I just have one request, which will not influence the decision, but that I would like you to address: in figure 11 please consider removing bar plots and error bars and substituting them with scatter plots, or at least boxplots with quantiles,in order to have an idea on the distribution of data and of the presence of possible outliers, which ultimately will justify the choice of the statistical test used.
Also, p values come from a uniform distribution, and discretizing it in levels marked by asterisks, even if it's a common practice, is in my opinion misleading. Just report p values with effect sizes. Significance is a subjective measure anyway.
Before your manuscript can be formally accepted you will need to complete some formatting changes, which you will receive in a follow up email. A member of our team will be in touch with a set of requests.
Please note that your manuscript will not be scheduled for publication until you have made the required changes, so a swift response is appreciated.
IMPORTANT: The editorial review process is now complete. PLOS will only permit corrections to spelling, formatting or significant scientific errors from this point onwards. Requests for major changes, or any which affect the scientific understanding of your work, will cause delays to the publication date of your manuscript.
Should you, your institution's press office or the journal office choose to press release your paper, you will automatically be opted out of early publication. We ask that you notify us now if you or your institution is planning to press release the article. All press must be co-ordinated with PLOS.
Thank you again for supporting Open Access publishing; we are looking forward to publishing your work in PLOS Computational Biology.
Best regards,
Daniele Marinazzo
Deputy Editor
PLOS Computational Biology
***********************************************************
Reviewer's Responses to Questions
Comments to the Authors:
Please note here if the review is uploaded as an attachment.
Reviewer #1: The Authors have replied exhaustively and satisfactorily to all comments raised in my first revision. I suggest acceptance of the paper, which represents an important contribution that will be hopefully taken into consideration by the high number of researchers employing Multiscale Entropy and related tools for the analysis of neurobiological signals.
Reviewer #3: The authors have done a tremendous job of address my concerns (and those of the other reviewers in my opinion). I commend them on good work
**********
Have all data underlying the figures and results presented in the manuscript been provided?
Large-scale datasets should be made available via a public repository as described in the PLOS Computational Biologydata availability policy, and numerical data that underlies graphs or summary statistics should be provided in spreadsheet form as supporting information.
Reviewer #1: Yes
Reviewer #3: Yes
**********
PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.
If you choose “no”, your identity will remain anonymous but your review may still be made public.
Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.
Reviewer #1: Yes: Luca Faes
Reviewer #3: Yes: AR McIntosh
28 Apr 2020
PCOMPBIOL-D-20-00041R1
Standard multiscale entropy reflects neural dynamics at mismatched temporal scales: What’s signal irregularity got to do with it?
Dear Dr Kosciessa,
I am pleased to inform you that your manuscript has been formally accepted for publication in PLOS Computational Biology. Your manuscript is now with our production department and you will be notified of the publication date in due course.
The corresponding author will soon be receiving a typeset proof for review, to ensure errors have not been introduced during production. Please review the PDF proof of your manuscript carefully, as this is the last chance to correct any errors. Please note that major changes, or those which affect the scientific understanding of the work, will likely cause delays to the publication date of your manuscript.
Soon after your final files are uploaded, unless you have opted out, the early version of your manuscript will be published online. The date of the early version will be your article's publication date. The final article will be published to the same URL, and all versions of the paper will be accessible to readers.
Thank you again for supporting PLOS Computational Biology and open-access publishing. We are looking forward to publishing your work!
With kind regards,
Laura Mallard
PLOS Computational Biology | Carlyle House, Carlyle Road, Cambridge CB4 3DN | United Kingdom ploscompbiol@plos.org | Phone +44 (0) 1223-442824 | ploscompbiol.org | @PLOSCompBiol
Tweets
Standard multiscale entropy reflects neural dynamics at mismatched temporal scales: What’s signal irregularity got to do with it?
Twitter
Facebook
WhatsApp
LinkedIn
