By: Maitreyee Wairagkar
Postdoc Research Associate,
Biomechatronics Lab,
Department of Mechanical Engineering,
Imperial College London
Brain activity is composed of both oscillatory and broadband arrhythmic components; however, neural correlates of movement are commonly characterised by narrowband oscillatory processes such as event-related desynchronisation (ERD). The complex ongoing scale-free dynamics of the arrhythmic broadband electroencephalography (EEG) remain unexplored. In this study, we identified the presence of long-range temporal correlations (LRTC) in the broadband EEG over short timescales using 2 s sliding windows. LRTC is typically considered as an invariant property of narrow-band EEG such as alpha oscillation amplitude fluctuations over long timescales and is assumed to indicate criticality. However, we discovered that the ongoing dynamics of LRTC in broadband EEG undergoes distinct instantaneous changes during voluntary movement over short timescales. The broadband LRTC increased significantly (p < 0.05) during movement. In contrast, the alpha oscillation LRTC computed on longer stitched EEG segments decreased significantly (p < 0.05) during movement, consistently with the literature. This suggests the complementarity of underlying fast and slow neuronal scale-free dynamics during movement generation. Furthermore, eliminating LRTC from broadband EEG by fractional differencing did not affect the alpha ERD and conversely, eliminating alpha ERD by fixing alpha power throughout EEG trial did not affect broadband LRTC indicated that broadband LRTC and oscillatory ERD are also complementary neural correlates of movement. The broadband LRTC can predict movement 0.5 s before its onset with high accuracy of 75.88 ± 6.4% in real-time. This study shows that broadband EEG also undergoes, independently of alpha band, dynamic reorganisation consistent with properties attributed to critical phenomena.
Page 1 – Brain activity is composed of both oscillatory and broadband arrhythmic components; however, neural correlates of movement are commonly characterised by narrowband oscillatory processes such as event-related desynchronisation (ERD). The complex ongoing scale-free dynamics of the arrhythmic broadband electroencephalography (EEG) remain unexplored. In this study, we identified the presence of long-range temporal correlations (LRTC) in the broadband EEG over short timescales using 2 s sliding windows. LRTC is typically considered as an invariant property of narrow-band EEG such as alpha oscillation amplitude fluctuations over long timescales and is assumed to indicate criticality. However, we discovered that the ongoing dynamics of LRTC in broadband EEG undergoes distinct instantaneous changes during voluntary movement over short timescales. We conducted experiments on 14 participants to record EEG while they performed single voluntary finger tap at a random time of their choice in a10s window following an instruction for right tap, left tap or resting state. We performed Detrended Fluctuation Analysis (DFA) on 2s segments of broadband EEG to quantify long-range temporal correlation.
Page 2 – Instantaneous LRTC was obtained from the DFA exponent of a 2s EEG segment shifted by 100ms. We observed that the broadband LRTC increased significantly during right and left finger movement (p<0.05) (red and blue traces). No such increase was observed in the resting state (black trace). The build-up of broadband LRTC started 1s prior to the onset of movement which allows prediction of movement. We did not observe any lateralization in the LRTC during right- and left-finger taps. Channels C3, Cz, and C4 showed maximum changes in the DFA exponent.
Page 3 – We established the complementarity of narrowband ERD and broadband LRTC by examining the effect of LRTC on estimating ERD by removing LRTC and by examining the effect of ERD on estimating LRTC by removing the ERD from EEG. First, we compared the ERD on raw EEG consisting of LRTC (solid line sub-figure A) and ERD in the same EEG after removing the LRTC (dashed line sub-figure A) using fractional differencing. There was no significant difference between ERD when LRTC was present and ERD when the LRTC was removed. We then compared LRTC on raw EEG consisting ofERD (solid line sub-figure B) and LRTC on EEG from which the effect of ERD was removed by fixing the alpha peak throughout the movement trial(dashed line sub-figure B). Again, there was no difference in the estimated LRTCs in both the conditions. This indicated that broadband LRTC and ERDare complementary neural correlates of movement.
Page 4 – Linear Discriminant Analysis classifier was used for classifying right and left finger tapping and resting state trials on a single trial basis. The DFA scaling exponents from channels C3, Cz and C4 were used as features. High accuracy of 75.88 ± 6.4% was obtained for movement detection using single-trial broadband LRTC. We were able to predict movement 0.5 s prior to its onset using LRTC. This opens the possibility of using LRTC in brain-computer interface applications. We demonstrated broadband LRTCs as novel neural correlates of movement intention in EEG. We discovered that the ongoing dynamics of LRTC in broadband EEG undergoes distinct instantaneous changes during voluntary movement over short timescales. This dynamic reorganisation in form of LRTC observed in broadband EEG is consistent with properties attributed to critical phenomena.
Real cool work, thanks! Is the predictive power of LRTC larger than that of oscillatory power? Could it predict movement onset more accurately / earlier? Is the prediction better when both LRTC and oscillatory power are combined? thanks!
Thank you Aline! Yes, we observed that broadband LRTC indeed gave significantly higher classification accuracy of voluntary movement than oscillatory event-related desynchronisation (ERD). LRTC could also predict the movement earlier. We explore precisely these comparisons in our paper https://doi.org/10.3389/fnsys.2019.00066.
In addition, we modelled long and short range correlations the broadband signal using ARFIMA. The classification accuracies of combination of ARFIMA + ERD features together were indeed higher, however, not significantly higher than ARFIMA alone.
thank you, this is super interesting, I have downloaded the paper and will be in touch if I have more questions!