Participants

Ten right handed healthy volunteers (7 males, 3 females, mean age 26.9 years) were enrolled into the study after a neurological screening to exclude potential adverse effects of TMS. Subjects with medical history of seizures, convulsions, loss of consciousness and traumatic brain injury, carriers of intracranial metallic objects and/or of cardiac pace-makers were excluded from the study. The entire experimental procedure was approved by the Local Ethical Committee of the Hospital “L. Sacco” and each volunteer gave written informed consent to participation.

TMS targeting

Structural magnetic resonance images (MRI) were recorded from all subjects at 1 mm³ spatial resolution (1T Philips scanner). Three TMS targets were identified on individual MRIs in the left occipital lobe (Brodmann's area - BA19), the left parietal lobe (BA7) and the left frontal lobe (BA6). Precision and reproducibility of stimulation were achieved by using a Navigated Brain Stimulation (NBS) system (Nexstim Ltd., Helsinki, Finland), that employs a 3D infrared Tracking Position Sensor Unit to map the positions of TMS coil and subject's head within the reference space of individual structural MRI. In addition, the NBS system computes on-line the distribution and intensity (V/m) of the intracranial induced electric field using a locally best-fitting spherical model of the subjects' head and brain and taking into account the exact shape, 3D position and orientation of the TMS coil. Stimulation intensity, expressed as a percentage of the maximal output of the stimulator, was kept between 40–75% for all subjects, corresponding to an electric field between 110–120 V/m on the cortical surface. In each area, the TMS hot spot (i.e. location of the maximum electric field induced by TMS on the cortical surface) was always kept on the convexity of the gyrus, about 1 cm lateral to the midline. These medial stimulation sites were selected because they are easily accessible and far from major head or facial muscles whose unwanted activation may affect EEG recordings. The reproducibility of the stimulation coordinates across sessions was guaranteed by a virtual aiming device that indicated in real-time any deviation from the desired target greater than 3 mm. The TMS stimulator consisted of a Focal Bipulse 8-Coil (mean/outer winding diameter ca. 50/70 mm, biphasic pulse shape, pulse length ca. 280 µs, focal area of the stimulation hot spot 0.68 cm²) driven by a Mobile Stimulator Unit (Eximia TMS Stimulator, Nexstim Ltd., Helsinki, Finland). The coil was always placed tangentially to the scalp, in order to optimize transmission of the magnetic field to the cortical surface. TMS pulses were delivered at an inter-stimulus interval randomly jittered between 700–900 ms (equivalent to ca. 1.1–1.4 Hz).

EEG recording

Continuous hd-EEG was recorded using a 60-channel TMS-compatible amplifier (Nexstim Ltd., Helsinki, Finland). This equipment ensured artefact-free EEG recordings starting from 8 ms after the TMS pulse [1]. Impedance at all electrodes was kept below 5 kΩ. EEG signals were band-pass filtered between 0.1–500 Hz and sampled at 1,450 Hz with 16 bit resolution. Vertical electrooculogram (EOG) was recorded by two extra sensors. A total of about 200 trials were collected for each TMS/hd-EEG session. Contamination of TMS-evoked potentials by auditory responses to the clicks produced by the TMS coil's discharge was prevented by masking noise and by placing a thin layer of foam between coil and scalp. After each session, electrodes' position was digitized using a 3D Infrared Tracking Position Sensor Unit (for more details about the EEG recording procedures see [16], [21], [24]).

General experimental design

The experimental protocol consisted of two main arms, aimed at evaluating the sensitivity and the repeatability, respectively, of TMS-evoked potentials. In order to test for sensitivity, different, randomly ordered, TMS sessions were performed in the same day (day1) varying only one stimulation parameter at a time (either site, or intensity, or angle of the TMS-induced current). TMS-evoked potentials were considered sensitive to the extent that they changed when stimulation parameters were changed. In order to evaluate repeatability, a subset of TMS sessions was repeated later in the same day (day1) as well as one week afterward (day8), without changing any stimulation parameter. TMS-evoked potentials were considered repeatable to the extent that they did not change over time when stimulation parameters were kept constant. In order to quantify sensitivity, single-subject pairwise comparisons were performed between TMS-evoked potentials obtained with different stimulation parameters. We called these comparisons “change comparisons” (C). To quantify repeatability, we performed single-subject pairwise comparisons between TMS-evoked potentials with identical stimulation parameters. We called these comparisons “no change comparisons” (NC). As described below, each pairwise comparison involved applying a non-parametric test based on random permutations between the single-trial TMS-evoked potentials of the two sessions.

Change comparison (sensitivity)

No change comparison (repeatability)

Data Analysis

Statistical analysis.

We implemented a non-parametric permutation-based statistical procedure to perform pairwise comparisons between TMS-evoked potentials, and to synthesize their degree of diversity in a single value (divergence index - DI), corrected for multiple hypothesis testing.

At first, a Wilcoxon rank-sum test was applied to check that the baselines (250 ms pre-stimulus) of the single trials, contributing to the two TMS-evoked potentials to be compared, had the same distribution. In case of a negative result, the most deviated trials were removed and the test was repeated until the baseline distributions of the two groups of trials were statistically equivalent (P>0.05). The percentage of rejected trials was always less than 5%. At this point, we could test the null hypothesis that two sets of TMS-evoked potentials (each one composed by 60 EEG channels by 182 time points, corresponding to 250 ms post-stimulus sampled at 725 Hz) are equivalent. If this is the case, “mixing” together, in any random combination, the single trials collected during the two TMS/hd-EEG sessions should always result in statistically equivalent TMS-evoked potentials. Otherwise, the null hypothesis can be rejected. Thus, for each comparison, 1000 “mixed” TMS-evoked potentials were obtained by randomly mixing and averaging 1000 times the single trials collected in two different sessions (Fig. 1A,B). The set of 1000 values at each post-stimulus time sample represented the instantaneous empirical null probabilistic distribution of the voltage of the TMS-evoked potentials. In order to correct for multiple comparisons in time, we computed a single distribution for the whole time interval as follows: i) all instantaneous distributions were centralized around zero, by shifting them by an amount δ(t) (Fig. 1C); ii) for each centralized distribution, we computed the maximum absolute value (Fig. 1D); iii) the one-tail (1-α)100^th percentile of the distribution of the maximum absolute values was used to estimate a significance threshold G for the whole time window of interest (Fig. 1D); iv) two boundaries were computed as (+G+δ(t)) and (–G+δ(t)). The temporal profile of these boundaries is modulated by δ(t), since G is a fixed threshold. The null hypothesis of equivalence between two TMS-evoked responses at each time sample t was rejected with probability of false positives α corrected for multiple comparisons when at least one of the two original potentials at that time sample lay beyond the significance boundaries (Fig. 1E). Finally, for each comparison the DI was defined as the percentage of significantly different time samples in the first 250 ms post-stimulus in all 60 EEG channels out of the total number of spatial-temporal samples. In this way, the DI was systematically calculated at the sensor level for all pairwise comparisons (n = 92). As a proof of concept, the same statistical procedure was also carried out at the source level on the time series of regional cortical currents in one subject (see Source modeling paragraph below). Cortical meshes were automatically parcellated into subregions (Automated Anatomical Lobules classification) using the masks provided by WFUPickAltas tool (freely available at: http://www.ansir.wfubmc.edu; [24], [26], [27]).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Non-parametric statistical procedure to perform single-subject pairwise comparisons between TMS-evoked potentials.

Single-trial recordings from two different conditions (blue and red lines) were randomly mixed 1000 times (A) and averaged (B). Instantaneous distributions of averaged voltages were computed and centralized around zero by keeping record of the displacement δ(t) (C). The distribution of maximum absolute values of each centralized distribution was computed and used to define a significance threshold G as the (1-α)100^th percentile (D). Significance boundaries (gray dotted lines) were computed as (±G + δ(t)) and used to define the significantly different time samples (red stars) between conditions at a specific channel (E).

https://doi.org/10.1371/journal.pone.0010281.g001

C comparisons resulted in the largest DI values

Generally, comparing the brain responses evoked by TMS pulses delivered over different stimulation sites revealed obvious differences in the space distribution and time course of voltages and currents (Fig. 2, first row). The average DI of all 22 C comparisons between different stimulation sites was 11.45±5.7% (range 3–19.7%). Inspecting single DI values (Fig. 3A) also showed that comparisons between distant areas (i.e. BA6 vs BA19, large cyan dots) was associated with the highest DI values (18.4±1.2%), while pairwise comparisons between nearby sites (i.e. BA6 vs. BA7 and BA7 vs. BA19, small cyan dots) resulted in lower DIs (8.8±4.3%).

Varying stimulation intensity (Fig. 2, second row) resulted in amplitude and latency changes of the main TMS-evoked components, while the general topographical distribution of voltages and currents tended to be preserved. The average DI across the 20 C comparisons between different intensities of stimulation (Fig. 3A, black dots) was 10.88±6% (range 2.31–22.2%).

When changing the angle of the TMS-induced current, the morphology of cortical responses varied in a rather unpredictable way, on a single-case basis (Fig. 2, third row). The average DI value was 3.92±3% (range 0.7–13.8%), with no systematic difference between 0° vs. 45° and 0° vs. 90° pairwise comparisons. In 4 out of 20 angle comparisons, DI values were smaller (Fig. 3A, green dots) than the largest DI obtained comparing experimental sessions with identical stimulation parameters (Fig. 3A, yellow and orange dots).

NC comparisons results in lowest DI values

When TMS was applied with identical stimulation parameters at different times, the morphology and the spatial-temporal dynamics of TMS-evoked potentials were largely preserved (Fig. 2, fourth row). The average DI was 0.28±0.4% (range 0–1.2%) when comparing same-day recordings and 0.43±0.4% (range 0–1.67%) for pairwise comparisons between one-week-apart sessions. The single DI values for all NC comparisons (Fig. 3A, yellow and orange dots) were always below 1.67%

Relative differences in the DI value are preserved across latencies

In order to understand whether the observed differences between TMS-evoked potentials were preserved over the entire post-stimulus period, rather than limited to a specific latency, we computed the DI over three subsequent post-stimulus intervals, namely 0–60 ms, 60–120 ms and 120–250 ms (Fig. 3B). For each type of pairwise comparison, the average DI computed at early latencies (0–60 ms) was significantly higher (P<0.05) as compared to the one computed for late latencies (120–250 ms), although the relative differences among types of comparison were preserved across all time intervals (P<0.01).

ROC Analysis

Results of ROC analysis showed that the optimal DI threshold according to the Younden index was 1.67% and yielded a 95.1% sensitivity and 100% specificity (repeatability), corresponding to an overall accuracy of 96.7%. The efficacy of DI in reliably quantifying the pairwise differences between TMS-evoked potentials was 99.1%, as measured by the area under the curve (Fig. 4).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 4. ROC analysis applied to the DI. ROC curve is depicted as a solid black line, interspersed by blank dots representing the values of sensitivity and specificity (repeatability) associated to single DI values.

The optimal DI value of 1.67% computed according to the Younden index [34] is shown as a black dot. Dashed line represents the ROC curve of a random classifier. Gray-shaded region indicates the area under the curve.

https://doi.org/10.1371/journal.pone.0010281.g004

Sensitivity of TMS-evoked potentials and the differentiation of cortical circuits

Integration and differentiation (or functional specialization) within regions are fundamental organizing principles of thalamocortical networks [35], [36]. While integration refers to the ability of the elements of a system to interact with each other, differentiation may be defined as the system's ability to react in different ways to different perturbations. TMS/hd-EEG, by directly exploring causal interactions (effective connectivity) among cortical areas, may provide a dependable evaluation of thalamocortical integration [16], [37], [38]. On the other hand, to the extent that TMS-evoked potentials are sensitive to changes in perturbation parameters, they may also gauge the degree of differentiation within thalamocortical networks. The present work shows that, at least on a coarse grain, different cortical perturbations result in a degree of response differentiation that is consistently higher compared to random test-retest variability (Fig. 3 and 4).

In the present experiments, changing stimulation parameters almost invariably resulted in higher DIs compared to the no-change conditions (Fig. 3A). Importantly, this finding was not limited to the early latencies, and indeed DI values for the C conditions were significantly larger until 250 ms post-stimulus (Fig. 3B). This evidence demonstrates that the EEG response to TMS is primarily due to direct cortical stimulation and to the ensuing reverberation of activity in a specific network of connected elements. Moreover, if TMS-evoked potentials were heavily contaminated by somatosensory, or auditory event-related potentials, the EEG responses generated when stimulating the head in two sites located a few centimetres away, rotating the stimulation angle, or increasing slightly the intensity of stimulation would have a very similar morphology. Thus, our finding suggests that the collateral stimulation of peripheral nerves by TMS plays a little role in the generation of TMS-evoked potentials.

Changing the site of stimulation resulted in very different responses and in high DI values that were even higher when the responses triggered in areas located far away (area 19 vs. area 6) where compared (Fig. 3A, large cyan dots). This variability in the cortical response reflects specific properties of the stimulated circuits and may be ascribed to local differences in cortical excitability [17], to differences in the frequency tuning of corticothalamic modules [18], [39] and to differences in the pattern of cortico-cortical connectivity [4], [12], [24].

Rotating the coil in the same area produced smaller modifications of TMS-evoked potentials as compared to changes of stimulation site. Indeed, 4 out of the 20 pairwise comparisons between recording sessions with different angles resulted in a DI <1.67%, i.e. the optimal threshold that maximized sensitivity and specificity (repeatability) of TMS-evoked potentials in the ROC analysis. Previous work [40] showed that by stepwise rotating the coil relative to left motor cortex, the largest muscle responses were obtained when the coil was 50° to the parasagittal plane, with the induced current in axis with the main direction of the axons in the motor “hand knob”. In addition, TMS coil orientation has been shown to affect the motor threshold [41], the degree of selectivity when stimulating different peripheral muscles [42] and even cognitive functions [43]. However, while the main orientation of motor cortex axons is quite predictable across subjects, little is known, a priori, about the main orientation of fibers in other brain areas, such as the ones stimulated in this study (BA6, BA7 and BA19). Thus, it is likely that the 4 pairwise comparisons between different stimulation angles resulting in low DI values, may be due to a virtually negligible variation of the induced current direction as compared to cortical axons. In fact, only by integrating TMS/hd-EEG with high resolution structural imaging techniques, such as diffusion tensor imaging (DTI), one may be able to control, in any cortical area, the coil's orientation with respect to the main direction of axons.

Certainly, in order to evaluate the fine grain of cortical differentiation with TMS/hd-EEG, one should design ad hoc experiments where stimulation parameters are varied in a systematic way (i.e. by moving/rotating the coil several times by a constant step). Meanwhile, it would be interesting to test whether the DI resulting from two different perturbations decreases in physiological (sleep, anesthesia) and pathological (coma, epilepsy) conditions, where the capacity for integration and differentiation in thalamocortical circuits is thought to be altered [35], [37], [44], [45].

Repeatability of TMS-evoked potentials and longitudinal changes in cortical circuits

TMS-evoked potentials recorded on the same day, with the same stimulation parameters (Fig. 3, yellow dots) were very stable (mean DI 0.28%), despite the fact that, between the two measurements, several (from 5 to 8) other sessions of repetitive TMS were carried out. It is well known that repetitive TMS pulses delivered at low (<1 Hz) and high (>5 Hz) stimulation frequency can respectively induce a reduction [46] and an increase [22], [47] of brain excitability [48]. Throughout the experiment, we used a stimulation frequency jittering randomly between 1.1–1.4 Hz. Our results indicate that this particular stimulation rate does not induce significant brain reorganization/plasticity processes and may be used to probe repeatedly the excitability of cortical circuits without significant interference.

The mean DI between TMS-evoked responses recorded during identical experimental sessions performed one week apart (0.43%) was slightly higher than mean DI between same-day sessions (0.28%). This minor discrepancy might be ascribed to different factors: i) small co-registration errors, that clearly reduce the reproducibility of navigation; ii) small errors in re-positioning the EEG cap; iii) unavoidable and unpredictable biological variability due to changes in brain excitability that likely takes place on the time-scale of several days. Nevertheless, this comparison clearly demonstrated that TMS-evoked potentials, besides being sensitive to changes, are also very stable over time. Thus, as suggested by the ROC curve in Fig. 4, repeating after one week a given perturbation and observing a DI >1.67% would strongly indicate that, in the mean time, some change occurred in the brain circuits. In principle, identifying a cut-off level above which one can decide whether a change in brain responsiveness occurred, or not, is crucial if one wants to use TMS/hd-EEG to track over time pathological alterations, plastic changes and therapy-induced modifications in cortical circuits. The DI can be automatically computed at the sensor as well as at the source level (Fig. 2) on a large matrix of spatial-temporal data, without having to select particular peaks, or components. By design, the DI tends to better capture changes in amplitude than changes in shape, since two responses with identical shape but with different amplitude will be maximally different. This feature may represent a limitation or an advantage, depending on the case (see below). In the present experiments, increasing TMS stimulation intensity by 10% produced changes that were reliably detected by the DI. This finding suggests that calculating the DI on repeated TMS/hd-EEG sessions may be effective in revealing rather fine modifications in cortical excitability (indexed by the response's amplitude) due to pathological alterations, i.e. stroke, epilepsy and depression [49]–[52] or therapeutic interventions, i.e. electroconvulsive therapy, rTMS, neurorehabilitation or drug administration [51], [52].

Limitations of the study

The DI values computed in the C conditions allowed detecting changes in the stimulation parameters with excellent accuracy: however their absolute values spread over a wide range (Fig. 3A). This dispersion may be due to a differential susceptibility of different cortical sites to changes in stimulation parameters. For example, if delivering TMS pulses on a specific cortical region at I% intensity activates locally most of the axons, increasing the stimulation intensity by 10% in this particular area would not dramatically change the electrical response and would result in a low DI. Similarly, as already discussed above, stimulating at different angles would not make much difference if the fibers in the target area are oriented in all directions (anisotropic arrangement). The contribution of these factors to the morphology of TMS-evoked potentials cannot be easily predicted, even if some insights may be provided by the integration of high-resolution structural neuroimaging (such as DTI). At any rate, a systematic understanding may only be achieved by varying parametrically the perturbation parameters, e.g. stimulating at several locations uniformly distributed on the scalp, delivering TMS pulses at progressively higher intensities from threshold to saturation, rotating gradually the angle of the induced current to span the whole circle. Such an extensive mapping of cortical electrophysiology was clearly beyond the scope of the present work, which was primarily aimed at evaluating, technically, the general level of accuracy of TMS/hd-EEG. Therefore, here, we tested only a limited set of all possible stimulation parameters, a task that still required considerable experimental and methodological efforts, e.g. recording and analyzing 100 TMS/hd-EEG sessions and performing 92 pairwise comparisons between TMS-evoked potentials.

Various classifiers, describing differences between evoked potentials, could have been used to build the ROC curve. In this work, we implemented the DI because it allows to quantify the differences between TMS-evoked potentials starting directly from the entire matrix of spatial-temporal data, without requiring a priori information, and to synthesize them into a single number. However, the ability of the DI in detecting modifications of TMS-evoked potentials due to physiological and/or pathological alterations in cortical circuits should be carefully evaluated. Indeed, the DI is rather conservative (it corrects for multiple comparisons in time) and explores the entire post-stimulus period, at all sensors. Thus, since several channels and latency ranges are often not involved by a TMS-evoked potential, one should generally expect low absolute DI values. In fact, we found that maximum DI values were around 20% (Fig. 3). When testing more specific hypotheses, other methods of classification (e.g., restricting the DI in space and time, template-matching, Mahalanobis distance) could be chosen. Clearly, the aim of the present study was not to develop an optimal classifier, but to use the most general one in order evaluate the accuracy of TMS/hd-EEG with a data-driven approach.