3.2 MEG Sensor Data
3.2.1 Event-related Fields
As can be seen in Figure 3-7, a strong peak in the ERF responses over occipitotemporal sensors was observed at 90 ms (+/-20 ms), referring to the prominent M100, followed by a second at 140 ms (+/-20 ms) expanding to temporal sensors which we refer to as the prominent M170. Few global peaks can be discriminated best in the Figure 3-8 at: 220 ms, 300 ms, 340 ms and 390 ms, mostly with a width of 30 to 40 ms.
A permutation t-test revealed a difference in ERF between the first and the third sequential positions at 157 – 272 ms (p=0.034) for left occipitotemporal sensors. For these sensors and period, a lower amplitude was recorded for the third sequential position than for the first sequential position. For right occipitotemporal sensors, the reverse pattern was observed. The time course indicated a peak at 220 ms over more occipital sensors and a peak 20 ms later (at 240 ms) over frontotemporal sensors. The M250 is ill-defined in time and is usually detected in a broader time window between 200 and 300 ms. However, the M250 is determined by a subsequent global peak after the M100. In the present time course, the M250 component could be the earlier measured and less pronounced peak at 220ms or the later, broader peak at 290ms. A compelling
Figure 3-7: Series of topographical representations of event-related fields (ERF) from 0 to 300 ms for the first sequential position. Each topography represents an average of non-overlapping consecutive 10ms of ERF signal time-locked to stimulus. Left and right temporal sensors recorded two strong peaks of magnetic field strength, one at 80 to 90 ms and another one at 140 to 150ms. The sign reverses for each side from the one to the other peak. Occipital sensors recorded the two peaks of magnetic field strength at 100 to 110 ms and 170 to 180 ms. Frontotemporal sensors recorded a peak at 230 to 240 ms. The numbers below each topography indicate the starting point and the end of the time in seconds for the represented time average.
reason to assume that M250 might correspond to the earlier peak at 220ms is that the components M100 and M170 are also detected earlier and the peak at 290 ms could refer to one of the adjacent stages proposed by Schweinberger (2016). However, the
Figure 3-8: Topography and time-locked signal of significant difference between the first and the third sequential positions. The permutation t-test across all sensors from 1 to 500 ms revealed a significant difference between the first and the third sequential positions in occipitotemporal sensors expanding to frontotemporal sensors from 150 to 270ms. A: The left and right topographies show group averages for the first (left) and the third (right) sequential positions. These sequential positions are indicated by icons below the topographies. These two topographies show the earlier effect which revealed peak activation in the occipital sensors with a significant increase on the left side. B: The two topographies show the later effect and revealed spatially four global peaks. The peak in frontotemporal sensors and occipitotemporal sensors showed a significant difference between the time course of the first and third sequential positions. C: The ERF time course of the left sensor selection (indicated by the icon top right): The ERF-signal was significantly lower for left occipitotemporal sensors of the third than of the first sequential positions with peaks at 220ms and 265ms. Further peaks without a significant difference were at 300ms, 340ms and 390ms. D: Here the corresponding right sensors to C are shown. E: The ERF time course of left frontotemporal sensors revealed one global peak at 230ms and a significant decrease from the first to the third sequential positions. F: Here the corresponding right sensors to F are shown.
Figure 3-9: Accuracies of SVM classification analysis on ERF-data. A SVM pattern classification was performed on time- and sensor average ERF for each participant separately. A: The participant average accuracies across 41 participants for the left ERF components are displayed. All reached significant results (binomial test), however, the M250 component showed a higher average accuracy than the others. B: The average accuracies across 41 participants for the right ERF components are displayed. All reached significant results (binomial test), but here the M250 component showed a slightly lower accuracy than the others. C: The confusion matrices of a single participant’s result (240 to 270 ms) visualized that trials of the first sequential position are better determined by a decision boundary than trials of the third sequential position. D: The confusion matrices of the corresponding (to C) permuted data set revealed accuracies around chance level (50%) as expected. The error-lines at each bar represent the standard deviation.
components could be refined in parsing the time window of interest in an M200 and an M250 component.
A classifier (SVM) analysis was applied to the two different time windows which the t-test revealed and additionally two more later components which were determined by the ERF time course. This resulted in five time-windows. The classification performance is illustrated by the median accuracy across the participants for each component for the left and right sensor selection in Figure 3-9-A|B. The significance of a component across the participants was determined by a binomial test. For all components, half or more of the participants revealed significant accuracies above chance level (component 200-240: 22 of 41 participants; component 240-270: 27 of 41; component 280-380: 21 of 41; component 380-500: 22 of 41 participants). The best accuracies were observed for the component 240-270 on the left occipitotemporal sensors (median accuracy=65.28%, SD=6,39). The lowest accuracies were observed for the same component but the right
occipitotemporal sensors selection (median accuracy=56.81%, SD=5.96). On the right occipitotemporal sensors selection, the second-best accuracy was observed for the component 200-240 (median accuracy = 60%, SD = 6.64).
Figure 3-9-C shows the result of a pattern classification for one participant by means of a confusion matrix (participant 31, component 170-200). The average confusion matrix (fivefold cross validation) indicates that the first class, which was the first sequential position, could be correctly predicted by 89% (32 correct and 4 false trials), whereas the second class, which was the third sequential position, could be predicted by 86% (25 correct and 11 false trials). This resulted in an overall performance of 79%.
The significance of a classification performance was determined by locating the accuracy of a classification in a generated null-distribution of accuracies. This distribution was created by applying 100 classifications on permuted data sets. The accuracies of a permuted data set were expected to reach chance level only. The confusion matrix of these accuracies confirmed this (Figure 3-9-D).