Revealing the Physiological Origin of Event-Related Potentials using Electrocorticography in Humans

The scientific and clinical value of event-related potentials (ERPs) depends on understanding the contributions to them of three possible mechanisms: (1) additivity of time-locked voltage changes; (2) phase resetting of ongoing oscillations; (3) asymmetrical oscillatory activity. Their relative contributions are currently uncertain. This study uses analysis of human electrocorticographic activity to quantify the origins of movement-related potentials (MRPs) and auditory evoked potentials (AEPs). The results show that MRPs are generated primarily by endogenous additivity (88%). In contrast, P1 and N1 components of AEPs are generated almost entirely by exogenous phase reset (93%). Oscillatory asymmetry contributes very little. By clarifying ERP mechanisms, these results enable creation of ERP models; and they enhance the value of ERPs for understanding the genesis of normal and abnormal auditory or sensorimotor behaviors.


29
The brain produces electrical responses to sensory, cognitive, and motor events. These responses, 30 called event-related potentials (ERPs), can be detected by averaging the electrical activity recorded 31 from electrodes placed on the scalp (electroencephalography (EEG)) (Makeig et al., 2004). ERPs 32 have been used for decades to study different aspects of information processing in the brain (such 33 as attention (Coull, 1998) or cognitive workload (Isreal et al., 1980)) or to diagnose specific neuro-34 logical disorders (such as deficits in the auditory or visual system (Coats, 1978; Parisi et al., 2001)). 35 ERPs are characterized by the type of event that causes them, e.g., auditory stimulation (result- 36 ing in auditory evoked potentials (AEPs)) or movements (resulting in movement-related potentials 37 (MRPs)), and by the polarity (positive/negative) and latency (e.g., 100 ms) of the dominant peak in 38 the response elicited by the event (Picton et al., 1974). Prior    . 44 Together, these three mechanisms could account for much of the interactions between stimu-45 lus, neural oscillations, and ERPs. For the additivity mechanism, the stimulus may induce a direct 46 response. For the phase reset mechanism, the stimulus may induce a phase reset in an ongoing 47 oscillation. For the asymmetry mechanism, the stimulus may induce a variation in the biased oscil-48 lation (i.e., an oscillation with asymmetrically distributed peak/trough amplitudes). The differential 49 effects of these three mechanisms produce an ERP. Consequently, while the additive mechanism 50 directly affects the ERP, phase reset and asymmetry mechanisms indirectly affect the ERP through Electrode locations for all subjects (dots). Locations that exhibited task-related activity during auditory stimulation or motor movements are highlighted in red or green, respectively. B. The green trace shows the ERP produced by a button press, averaged across all task-related locations and subjects. C. The red trace shows the ERP produced by auditory stimulation, averaged across all task-related locations and subjects. These average ECoG ERPs are similar to ERPs reported in this or previous studies that used scalp-recorded EEG (Figure 2-Figure Supplement 5).      terpretation of the results of thousands of studies using scalp-recorded ERPs, and the generation 85 of more generalized models of brain function that such interpretation could inform.

86
In our study, we recorded signals from the surface of the brain to determine the physiological 87 origin of auditory and movement-related potentials. We recorded electrocorticographic (ECoG) 88 activity from eight human subjects while they executed a simple reaction-time task. In this task, 89 the subjects responded to a salient auditory stimulus by pressing a push-button with the thumb of the overall signal accounted for by each generating mechanism. Prior to this assessment, we 105 verified that our data fulfilled the three preconditions for investigating phase reset (Figure 3). For 106 this, we first determined that the frequency band of the ongoing oscillation across all eight subjects Locations exhibiting evoked potentials resulting from auditory stimulation (red dots) or a button press (green dots) in subject S3. B. Time course of ECoG activity during auditory stimulation (left) and button presses (right) for locations A1 and M1, and their across-trial average. Single-trial ECoG responses at location A1 are phase-locked at stimulus onset, and demonstrate the same N1, P1, and P2 components as seen in the across-trial average. In contrast, single-trial ECoG responses at location M1 are not phase-locked at movement onset, and thus no evoked potentials are exhibited in the average across all trials. Instead, a slow cortical potential arises from the average across all trials. C. Average AEPs (red traces on the left) and MRPs (green traces on the right) for locations A1-3 and M1-2 and their average in subject S3. All auditory locations exhibit a clear N1, P1, and P2 components, and all motor locations exhibit a prominent slow cortical potential. D. Time courses of ERPs from locations A1 and M1 in subject S3 in two different frequency bands (<3 Hz and 3-40 Hz). The characteristic components of the AEP are captured by the 3-40 Hz band. In contrast, the slow negative potential in the MRP can only be seen in the <3 Hz band. E. ECoG power in the <3 Hz and 3-40 Hz bands for baseline (-400 to 0 ms) and ERP (0 to 400 ms) periods (top and bottom, respectively), calculated across all task-related locations and all subjects. Baseline activity is mostly comprised of 3-40 Hz band power (p<0.001, paired t-test     (Figure 3-Figure Supplement 1). To remove the additive effect, we removed frequency components outside of the ongoing oscillation (see C). For this, we subtracted the frequency components outside the frequency band of the ongoing oscillation (see B) from the evoked potentials (see A). To remove the effect of phase reset, we recomposed the evoked potentials with constant phase velocity. For this, we decomposed each trial into 1 Hz-wide frequency bands between 1 and 200 Hz (see E). Next, we adjusted the signal's ongoing oscillation to have constant phase velocity (see F). Finally, we combined the time series across all frequency bands into our recomposed signal (see G). To remove the effect of asymmetric amplitude, we subtracted the asymmetric bias from each evoked potential within the frequency band of the ongoing oscillation. For this, we first detected the peak and troughs within the ongoing oscillation (see I). Next, we determined the relationship between the amplitude at these peaks and troughs in the ongoing oscillation, and the voltage at the same time points in the original signal (see J). This analysis yielded two linear relationships (i.e., voltage-to-voltage functions), one for the peak (see the blue line in K), and one for the trough (see the purple line in K). The average between these two relationships represents the asymmetry between peak and trough amplitude as a function of the amplitude of the ongoing oscillation (see black line in K). We used this function to translate the envelope of ongoing oscillation (see L) into the asymmetric bias (see M). Finally, we subtracted the time-varying asymmetric amplitude from each trial's original signal (see N).

112
After verifying that all three preconditions for our phase reset analysis were fulfilled, we deter-113 mined the contribution of each of the three generating mechanisms by removing their influence 114 from the individual trials. For the additivity mechanism, we removed frequency components out-115 side of the ongoing oscillation. For the phase reset mechanism, we recomposed each trial's signal 116 within the frequency band of the ongoing oscillation with constant phase velocity (Freeman, 2004). 117 For the asymmetric amplitude mechanism, we subtracted the asymmetric bias from each trial's sig-  The results of our spectral composition analysis shows that MRPs are comprised mostly of low 124 frequency components (<3 Hz), while P1 and N1 of AEPs are comprised mainly of 3-40 Hz compo-125 nents ( Figure 3E). Further, we observed that higher pre-stimulus oscillation power yields a bigger 126 N1 peak amplitude in the AEP ( Figure 3F), but doesn't influence the MRP amplitudes ( Figure 3G). 127 The analysis of the specific contribution of each mechanism in the rise of the AEPs and MRPs shows 128 that the additivity mechanism explains 61% and 88% of the energy in the AEPs and MRPs, respec-129 tively ( Figure 5B). The phase reset mechanism explains 41% and 12% of the energy in the AEPs and  Finally, we found that the combination of additivity and phase reset mechanisms explains al-136 most the entire amount of energy in the ERPs (97% and 99% of the energy in the AEPs and MRPs) 137 ( Figure 5). Our results show that the individual components of the AEPs (i.e., P1 and N1 compo-138 nents) are mostly generated by phase reset (93%). In contrast, the P2 component is generated by 139 both additivity (57%) and phase reset (36%) ( Figure 5B). 140

141
In this study, we investigated the specific contributions of additivity, phase reset, and asymmetric  our main analysis, this shows that MRPs are generated by additivity and not by phase reset (Fig-168 ure 5). 169 In contrast to MRPs, our results show that AEPs meet all preconditions for phase reset. This  (Figure 5-Figure Supplement 1). 259 To confirm that ERPs obtained from ECoG represented valid AEPs and MRPs, we performed 260 a control experiment with 7 subjects in which we recorded EEG and eye-movement data. The 261 satisfactory comparison between AEPs and MRPs obtained from ECoG and those obtained from 262 EEG (Figure 2-Figure Supplement 4), as well as those described in the literature (Figure 2-Figure   263 Supplement 5), allowed us to reject this potential confound. Finally, because pre-stimulus saccades 264 are known to induce phase reset into cortical signals (Ito et al., 2011), we verified that the subjects 265 maintained eye-gaze throughout the pre-stimulus period. We detected saccades from 1 s before 266 to 1 s after stimulus onset. However, our analysis showed that only 13 of 2370 trials over 7 subjects 267 (EEG data) exhibited saccades before stimulus onset, making this an unlikely confound.  Eight human subjects (S1-S8, 4 males, 4 females, average age = 41±14) participated in this study at 293 the Albany Medical Center in Albany, New York. The subjects were mentally and physically capable 294 of participating in our study (average IQ = 96±18, range 75-120, Wechsler 1997). All subjects were 295 patients with intractable epilepsy who underwent temporary placement of subdural electrode ar-296 rays to localize seizure foci prior to surgical resection.

297
The implanted electrode grids were approved for human use (Ad-Tech Medical Corp., Racine, 298 WI; and PMT Corp., Chanhassen, MN). The platinum-iridium electrodes were 4 mm in diameter (2.3 299 mm exposed), spaced 10 mm center-to-center, and embedded in silicone. The electrode grids were 300 implanted in the left hemisphere for seven subjects (S1, S3, S6, and S7) and the right hemisphere  (Coon et al., 2016). 306 Patients with ECoG coverage extending from auditory to motor cortex are a relatively rare oc-307 currence. Thus, most previous ECoG studies were limited to subjects with either motor or au-  (Sharbrough, 1991). 316 All subjects provided informed consent for participating in the study, which was approved by  Data collection 320 We recorded ECoG signals from the subjects at their bedside using the general purpose Brain-

321
Computer Interface (BCI2000) software (Schalk et al., 2004), interfaced with eight 16-channel g.USBamp 322 biosignal acquisition devices, or one 256-channel g.HIamp biosignal acquisition device (g.tec., Graz, 323 Austria) to amplify, digitize (sampling rate 1,200 Hz) and store the signals. To ensure safe clinical 324 monitoring during the experimental tasks, a connector split the cables connected to the patients 325 into a subset connected to the clinical monitoring system and a subset connected to the amplifiers. 326 We recorded EEG signals and eye-movement coordinates from the subjects in our control group 327 using the same g.USBamp setup and a Tobii T60 eye-tracking monitor (Tobii Tech., Stockholm, 328 Sweden) that was positioned at eye level 55-60 cm in front of the subject and was calibrated for 329 each subject at the start of each experimental session.

331
The subjects performed an auditory reaction task in which they responded with a button press to 332 a salient 1 kHz tone. For this, the subjects used their thumb contralateral to their ECoG implant. In 333 total, the subjects performed between 134 and 580 trials. Throughout each trial, the subjects were 334 first required to fixate gaze onto the screen in front of them. Next, a visual cue indicated the start 335 of the trial, which was followed by a random 1-3 s pre-stimulus interval and subsequently, the au-336 ditory stimulus. The stimulus was terminated by the subject's button press, or after a 2 s time out, 337 after which the subject received feedback about his/her reaction time. This feedback motivated 338 the subjects to respond as fast as possible to the stimulus. To prevent false starts, we penalized 339 subjects with a warning tone if they responded too fast (i.e., less than 100 ms after stimulus onset). 340 We excluded false-start trials from our analysis. In this study, we were interested in the auditory 341 and motor response to this task. This required defining the onset of these two responses. For the 342 auditory response, we defined this as the onset of the auditory stimulus (as measured by the volt-343 age between the sound port on the PC and the loudspeaker). For the motor response, we defined 344 the onset as the time when the push-button was pressed. To ensure the temporal accuracy of 345 these two onset markers, we sampled them simultaneously with the ECoG signals using dedicated 346 inputs in our biosignal acquisition system. We defined baseline and task periods for the auditory 347 and motor response. Specifically, we used the 0.5-s period prior to the stimulus onset as the base-348 line for the auditory response, and the 1-s to 0.5-s period prior to the button press as the baseline 349 for the motor response. Similarly, we used the 1-s period after stimulus onset as the task period 350 for the auditory response, and the period from 0.5-s before to 0.5-s after the button press as the 351 task period for the motor task.

352
Data pre-processing 353 As our amplifiers acquired raw unfiltered ECoG signals, we first removed any offset from our sig-354 nals using 2nd order Butterworth highpass filter at 0.05 Hz. Next, we removed any common noise, 355 using a common median reference filter. For the creation of the common-mode reference, we ex-356 cluded signals that exhibited an excessive 60 Hz line noise level (i.e., ten times the median absolute 357 deviation). To improve the signal-to-noise ratio of our recordings and to reduce the computational 358 complexity of our subsequent analysis, we downsampled our signals from 1200 to 400 Hz using 359 MATLABs "resample" function, which uses a polyphase antialiasing filter to resample the signal at 360 the uniform sample rate.

361
Electrode selection 362 To select the appropriate electrodes, we needed to determine which electrodes exhibited an early 363 auditory or motor-related response. We accomplished this in three steps. In the first step, we 364 selected only those cortical locations that exhibited a task-related response in the high gamma 365 band. For this purpose, we performed a statistical comparison between baseline and task peri-366 ods across all trials. Specifically, we calculated the Spearman's correlation coefficient between the 367 power of baseline/task periods, and a corresponding label (i.e., -1 for baseline and +1 for task pe- from the permutation test. Finally, we identified the cortical locations exhibiting a task-related re-375 sponse in the high gamma band that had a p-value smaller than 0.001 (Bonferroni-corrected for 376 the number of cortical locations). In the second step, for each subject, we restricted our selection 377 to the single auditory and single motor-related cortical locations that exhibited the earliest onset.

378
To perform this selection, we first calculated the Spearman's correlation coefficient between the 379 power of baseline/task periods for individual time points of the task period (i.e., auditory: 500 ms 380 after stimulus onset for auditory; motor: 250 ms before button press). Next, we determined the  Figure 4E). Next, for frequency bands above 3 Hz, we   (Mazaheri and Jensen, 2008). To remove this effect, we estimated and 407 removed the time-varying asymmetric amplitude from each trial's original signal (see Figure 4H-N). 408 As the asymmetry between peak and trough amplitude is a function of the peak and trough ampli-409 tudes themselves (Mazaheri and Jensen, 2008), we first needed to determine this function for each 410 cortical location. For this purpose, we first detected the peak and troughs in the ongoing oscilla-411 tion (i.e., the 3-40 Hz filtered signal). Next, we identified the relationship between the amplitude 412 at these peaks and troughs in the ongoing oscillation and the voltage at the same time points in 413 the original signal. This analysis yielded two linear relationships (i.e., voltage-to-voltage functions), 414 one for the peak, and one for the trough. The average between these two relationships represents

Phase-reset Mechanism
Phase-resetting of ongoing oscillations at the same time point In the additivity mechanism, in each trial, components exhibiting the same frequency characteristic as the ongoing oscillation are added to the ongoing oscillations. The average across trials reveals an ERP. The main characteristic of additivity is the independence between ongoing oscillation and evoked potential. Thus, ongoing oscillations are considered background noise that averages out to zero across trials, and only external stimuli directly affect ERPs. C. In the phase-reset mechanism, for each trial, the phase of the ongoing oscillation is reset at a certain time point after stimulus onset. The average across trials reveals an ERP. In contrast to additivity, the generation of an ERP is dependent on the phase of the ongoing oscillation. Thus, external stimuli can affect ERPs only indirectly through inducing a phase-reset in the ongoing oscillation. While the additivity and phase-reset mechanisms can both explain the generation of ERPs, their physiological pathways may be fundamentally different. In the additivity mechanism, in each trial, an ERP is added to the ongoing oscillation. The average across trials reveals the ERP, while individual trials appear to exhibit asymmetry. C. In the asymmetric amplitude mechanism, for each trial, the ongoing oscillation exhibits asymmetry during or after the amplitude changes. The average across trials reveals the ERP. In contrast to additivity, the generation of an ERP is dependent on the amplitude envelope of the ongoing oscillation. Thus, external stimuli can affect ERPs only indirectly through affecting the amplitude and asymmetry in the ongoing oscillation. While the additivity and asymmetric amplitude mechanisms can both explain the generation of ERPs, their physiological pathways may be fundamentally different.