What to Visualize

This current section discusses the choices behind the visualizations of Sections 5.2 and 5.3. Firstly, typical signal pair examples are shown. Secondly, the channel-pairs used to visualize the waveform results are selected for both movements, smile and eyebrow lift. Moreover, the same channel-pair selection is completed for the delay results.

Figure 5.1 illustrates the different behavior types observed between the con-tralateral channel-pairs. In the figure, the green background is used for the healthy side, and red for the paralyzed side. Thus, the first row represents the healthy side of three different patients, and the second row shows the paralyzed

sides for the same patients. Smile movement and the channel number four are chosen to work as an example. In each plot, all ten repetitions are shown; each with their own color. Concentration should now be paid on the overall trends, not to the details. To put it briefly, the channel-pair behavior may be divided into three groups: same direction, opposite direction, and shape difference. To begin with the first group; the movement waves are pointing to the same direction. This is the case with the patient number 15 in the figure. There the difference between the healthy and paralyzed sides is in the magnitude and details of the shape. The patient number five’s plots are shown in the middle, and illustrate another type of channel-pair behavior; the waves are to opposite directions. Also, the magnitude may vary and slight differences in the shape be present. Finally, in the rightmost plot, patient number 12 serves as an example of the case where the wave shapes differ significantly. On the paralyzed side, the signal is fluctuating very close to zero and the shape is on most repetitions far away from the healthy side’s shape.

The patient data fits into different levels of these three groups.

Figure 5.1 Different channel-pair behavior types illustrated. The green background signals healthy side, and the red one the paralyzed side. The smile movement and channel 4 have been chosen to provide an example.

The patient data is preprocessed, chopped into movements, and then plotted for data explorationpurposes. The data is visualized before analysis to gain overview and understanding of it; are the analysis approaches described during previous BML feedback loops suitable now, or should adjustments be made - in the pivot point it was decided that this cycle is the patient data PoC. Table 5.1 summarizes the completed data analysis steps of the patient data PoC. In short, the robust parameters are analyzed. In more detail, the cross-correlation is computed and its coefficients used

to assess the direction of the vector of excursion (Section 5.2), and the delay to investigate the temporal difference (Section 5.3). The spatial difference or the speed of the vector of excursion are not evaluated due to the presence of significantly differing channel-pair signals.

Table 5.1 A summary for patient data analysis phase’s data analysis methods.

Quantity Parameter Completed

Vector of excursion 1. Direction; cross-correlation

Yes 2. Speed; min and max derivative

No

Appendix H contains similar plots than Figure 5.1 for each patient. Section H.1 shows the smile movement signals, and Section H.2 the eyebrow lift data. These plots are added to provide insight in data exploration sense and to offer additional information for evaluating the results. Because to the best of knowledge this is the first time facial palsy patients’ facial movements are measured capacitively.

To show the movement signals in the appendices for each patient and for both movements, and to visualize the computed results on the interpatient study, there is a need to limit the amount of data. In other words, an informed choice is needed about which channels to visualize. In the preliminary part this same topic was visited (see Subsection 4.5.2) and two channel pairs chosen for both movements.

Here, in addition to the preliminary study’s decision, Figures 5.2-5.5 are taken into account.

Figures 5.2 and 5.3 show the channelwise boxplots over all patients of the best cross-correlation coefficients of smile and eyebrow lift (EBL), respectively. Each of the 17 patients repeated the movement types 10 times. From those ten repetitions, cross-correlation between the contralateral sides is computed. The highest value of the cross-correlation function for each channel-pair is chosen. Thus, each patient has ten sets of best cross-correlation coefficients for the channelpairs. An arithmetic average is computed for each patient over the ten sets producing an average value for each channel-pair. Therefore, each patient has a channelwise average value computed from the highest cross-correlation coefficients. These average values are then boxplotted over the patient set thus producing Figures 5.2 and 5.3.

Figure 5.2 Best smile cross-correlation coefficient averages per channel over all patients.

Figure 5.3 Best EBL cross-correlation coefficient averages per channel over all patients.

Figures 5.4 and 5.5 are constructed in similar manner than the two previous figures.

The difference is, that Figures 5.4 and 5.5 show the boxplots of the highest cross-correlation coefficients’ delays. In other words, the delays corresponding the highest cross-correlation coefficient are picked, averaged, and plotted over all patients. Figure 5.4 has the delays for smile movement, and Figure 5.5 for eyebrow lift.

Figure 5.4 Best smile delay averages per channel over all patients.

Figure 5.5 Best EBL delay averages per channel over all patients.

The four previous figures are shown here a) to provide an overview to cross-correlation coefficient and delay results over all patients, and b) to form a basis for brief discussion which channels to visualize in the coming sections. Figure 5.2 shows that cross-correlation coefficients’ medians are above 0.7 and a bit below 1.0 for smile. The interquantile ranges (IQR) are above 0.6 and a bit below 1.0. The whiskers are limited to maximum and minumum of 1.5 IQR. No outliers are present.

The medians and IQRs are highest on channel-pairs 4-6; these are the most central midextension channels. This is the area where the smile is expected to be visible in the data. Based on this figure, also the most central sensor from the midextension, channel four, is to be visualized.

Figure 5.3 shows that the cross-correlation coefficient medians are above 0.65 and below 0.8 for eyebrow lift. The IQRs are above 0.5 and below 0.95. Again, no outliers are present. The medians and IQRs are highest on channel-pairs 1 and 2; these are the most central upper extension channels. This is the area where the eyebrow lift is expected to be visible in the data. Based on this figure, the same channels chosen for preliminary study’s visualization are used.

The both boxplots for delays have medians fluctuating around zero. In other words, the best cross-correlation coefficients for both movements are gained close to zero delay. For smile, the IQRs for the channels 4-6 are the smallest. For other channels the midranges are between -2 and 1 seconds. For eyebrow lift, the IQRs for the channels 1-2 are the smallest, too. The other channels have midspreads between -2 and 2 seconds. There are outliers present for both movements, in other words the 1.5 IQR is exceeded several times. The whiskers extent from almost -4 to 2 seconds for smile, and from -6 to 6 seconds for eyebrow lift when observing all the channels.

If limiting to the channels 4-6 for smile and 1-2 for eyebrow lift, the whiskers are approximately -1 to 1 seconds and -4 to 2 seconds, respectively.

The argument can be made that using the existing results in Figures 5.2-5.5 to choose which channels to further analyze is having circular reasoning, or cherry-picking for the least. However, the choice is made acknowledging the problem, and with the relevance in mind; the symmetry of eyebrow lift should be measured close to the eyebrows as that is where the muscles lifting them and the movement is. The same is true for the smile; due to the muscles used the movement is concentrated on the prototype’s midextension area. The most relevant locations for the movements are being focused on, recognizing the problems and limitations of the approach.

To conclude, the channels 4-6 are used for analysing smile and channels 1-2 for computing results for eyebrow lift in the following sections. The measured data is compared to an existing system, the Sunnybrook method, next.