Can EEG distinguish between different types of temporal predictions?

When the next economic crisis occurs, will it be just another peak in a very, very slow oscillation? Or will it be triggered by specific circumstances and preceded by warning signs? Or perhaps we will expect a crisis to happen only because it’s been long enough since the last one? And most importantly, will the particular scenario of our prediction make any difference when it comes to the dynamics of an actual crisis and the recovery from it?

In the lab, neural and perceptual temporal predictions can similarly be induced by various experimental factors, including rhythms (periodic streams of stimuli), cues (contingencies between specific events and temporal intervals), and hazards (the contextual probability of an event occurring, given recent history). But are the neural mechanisms of these predictions different? A popular explanation of the first scenario – predictions based on rhythms – is that neural systems can entrain to external rhythms and amplify the processing of stimuli occurring at expected time points. Several measures of entrainment have been used in the past, with inter-trial coherence (ITC) being one of the most popular metrics. However, just like other forms of predictions, rhythmic predictions are also linked to enhanced processing of expected stimuli, as well as several other neural signatures, such as the contingent negative variation (CNV), a slow preparatory potential preceding the expected time point, or alpha-band modulation just before the onset of an expected target.

In this paper, Assaf Breska and Leon Deouell show impressive similarities between rhythm-based and memory-based temporal predictions in terms of their underlying neural signatures based on EEG data. In the rhythm-based paradigm, participants viewed a rhythmic stream of stimuli, followed by a cue and a target – both according the same rhythmic pattern as the preceding stream. In the memory-based paradigm, the rhythmicity of the stream was broken, such that only every second interval had a fixed duration, and the remaining intervals were random. As a result, the interval between the cue and the target could be predicted based on the most frequent preceding interval, but the whole stream would arguably be too jittered to entrain a neural oscillation. Both conditions could be in a faster (dominant interval lasting 700 ms) or slower (1300 ms) regime, and both also contained a subset of trials in which targets were presented at an unexpected (invalid) interval.

The authors have analysed four prominent neural signatures of temporal predictability: two preceding an expected target (the CNV and alpha-band modulation), one around the time of target onset (delta-band phase coherence), and one following target presentation (the latency of a P300 component). Crucially, none of them showed significant differences between the two paradigms. In other words, rhythm-based and memory-based temporal expectations produced strikingly similar neural correlates of target anticipation and processing. However, there was one exception: when a target was expected but omitted, in the rhythmic paradigm the CNV bounced back to baseline immediately after the omission, but in the memory-based paradigm, it took almost 400 ms more for the signal to start returning to baseline. One can interpret this finding, as the Authors do, in at least a couple of ways. On the one hand, rhythm-based predictions are likely more precise, so the CNV can return to baseline as soon as the system “realises” that its expectation was violated. On the other hand, a fast return to baseline might reflect a more automatic nature of rhythmic predictions, as opposed to a more flexible allocation of resources in the memory-based prediction, which might result in a prolonged state of readiness for the omitted (possibly delayed) stimulus.

As one reads through the results section of the paper, these analyses seem to suggest that the neural mechanisms underlying rhythmic and memory-based predictions are largely identical. Regarding the similarity of delta-band phase coherence between the two paradigms, one could even potentially conclude that there is just as much (or little) entrainment in rhythmic as in non-rhythmic temporal expectations. However, this is not the correct conclusion, as noted by the authors. What this paper does show is that the ITC is not a sensitive measure of entrainment. In other words, simply looking at low-frequency phase locking does not allow a differentiation between conditions in which one would expect a different level of low-frequency entrainment.

However, I wonder if – based on their data – the authors could not have focused on this point a bit more, and either show why the metric is not sensitive, or perhaps suggest a better alternative. First of all, I missed a plot showing actual neural entrainment to the streams. Given that the paradigms included both faster (1.42 Hz) and slower (.77 Hz) regimes which were not harmonic, one could quantify differences in entrainment to these specific frequencies between the two regimes. Second, we know that “significant entrainment” might be an artefact of rhythmic evoked potentials, and we also know what neural signatures we might expect from data showing true low-frequency entrainment. In this case, we don’t know if delta-band (here 0.5-3 Hz) phase estimates around target onsets were not contaminated by ERPs evoked by targets. For example, while the authors show that delta-band phase correlates with reaction times, a similar correlation might have been expected between ERP amplitude or latency and behaviour. Again, it would be nice to see whether different conditions show phase concentration (and possibly link with behaviour) in slightly different frequency bands, as suggested by the authors’ oscillatory entrainment model. Finally, I was left wondering whether the difference in CNV resolution time between rhythm-based and memory-based predictions could be picked up by the ITC metric, and if so, whether future research should not indeed concentrate on “resonance” effects (i.e., the persistence of an oscillation after the interruption of external stimulation) as a cleaner metric of rhythmic entrainment.

Nevertheless, the paper convincingly shows that a significant difference in low-frequency inter-trial coherence around stimulus onset between a purely rhythmic and a purely random stream does not constitute strong evidence for neural entrainment by external rhythms. And while the main conclusion here is methodological, the paper does raise a question to what extent different experimental manipulations of temporal predictions rely on qualitatively different neural mechanisms. While recent TMS work does suggest that different networks are involved in rhythm processing and other forms of temporal orienting, most measures – including perceptual sensitivity, fMRI neuroimaging, as well as our standard EEG measures – might not be as sensitive to different types of temporal predictions.

Ryszard Auksztulewicz, Oxford Centre for Human Brain Activity 

Source article: Breska A, Deouell LY (2017) Neural mechanisms of rhythm-based temporal prediction: Delta phase-locking reflects temporal predictability but not rhythmic entrainment. PLOS Biol, February 10, doi: 10.1371/journal.pbio.2001665.