## Brief description of the outcome measures - with reading material

## How the Program Works

The website does the following things to your data before analysing the results:

1) It resamples the data to 100Hz using cubic spline interpolation. this is based on the sampling rate you input, so please make sure it is correct. This should have negligible effect on the results of any adequately sampled data, and is needed to streamline the analysis and reduce the risks of crashing due to overloading of the hardware. For example, running the wavelet analysis on a 5 minute trial sampled at 10000Hz could take days.

2) By default it applies a 6.25Hz low pass filter to the data using a Coiflet-5 wavelet. I prefer these filters over traditional Butterworth/Chebyshev style filters because they do not have the notorious edge effects. If you do choose to use one of the Butterworth filter options the data is mirrored at the start and end of the trace prior to filtering, then the middle section removed. This attenuates any edge effect occurring but may not be perfect.

1) It resamples the data to 100Hz using cubic spline interpolation. this is based on the sampling rate you input, so please make sure it is correct. This should have negligible effect on the results of any adequately sampled data, and is needed to streamline the analysis and reduce the risks of crashing due to overloading of the hardware. For example, running the wavelet analysis on a 5 minute trial sampled at 10000Hz could take days.

2) By default it applies a 6.25Hz low pass filter to the data using a Coiflet-5 wavelet. I prefer these filters over traditional Butterworth/Chebyshev style filters because they do not have the notorious edge effects. If you do choose to use one of the Butterworth filter options the data is mirrored at the start and end of the trace prior to filtering, then the middle section removed. This attenuates any edge effect occurring but may not be perfect.

## Standard Measures

The standard measures included in SeeSway are:

1) Path length (cm). Total COP displacement during the trial. Put simply, if the COP trace was a piece of string it would be the length of it once it was stretched out. To convert this to path velocity, which allows for comparison across studies of different trial duration, simply divide it by the number of seconds the trial went for.

2) Amplitude (cm) on each axis. This is simply the distance between the highest and lowest value on an axis. For example, for the AP axis it is the difference between the most forward and most backward COP position during the trial. Higher amplitudes indicate greater sway range.

3) Standard deviation (SD) of the COP path (cm). This variable quantifies how "bunched" the sway pattern is. If a person oscillates about a central point the standard deviation would likely be lower than someone who shows a gradual shift away from the start position as the trial progresses if the path length is similar.

4) Root-mean-square (RMS) displacement (cm). This variable quantifies the mean absolute COP distance from the starting position.

1) Path length (cm). Total COP displacement during the trial. Put simply, if the COP trace was a piece of string it would be the length of it once it was stretched out. To convert this to path velocity, which allows for comparison across studies of different trial duration, simply divide it by the number of seconds the trial went for.

2) Amplitude (cm) on each axis. This is simply the distance between the highest and lowest value on an axis. For example, for the AP axis it is the difference between the most forward and most backward COP position during the trial. Higher amplitudes indicate greater sway range.

3) Standard deviation (SD) of the COP path (cm). This variable quantifies how "bunched" the sway pattern is. If a person oscillates about a central point the standard deviation would likely be lower than someone who shows a gradual shift away from the start position as the trial progresses if the path length is similar.

4) Root-mean-square (RMS) displacement (cm). This variable quantifies the mean absolute COP distance from the starting position.

## Advanced Measures

The advanced measures can require user input to change the default settings. For the variables reported here it is strongly recommended that you refer to the recommended reading to increase your knowledge of how to best analyse the data and interpret the results.

1) Wavelet decomposition path length (cm). The trace is split into moderate (i.e., fast moving; 3.13 to 6.25Hz), low (1.56 to 3.13Hz), very low (0.78 to 1.56Hz) and ultralow (<0.78Hz) frequency signals using a 9-level Symlet-8 wavelet, with the COP path length in each band reported. This results in easily interpretable data, as higher scores reflect more sway. For more information about how this works and what it means refer to the following papers:

Liang et al. (2014)

Clark et al. (2014)

2) Detrended fluctuation analysis (DFA). This fractal measure examines the long range correlations in the trace to provide a measure of signal complexity.

Practical interpretation of these results is somewhat unclear, however when the DFA < 0.5 or 1 < DFA < 1.5, the signal is anti-persistent (smaller DFA = more anti-persistent). When 0.5 < DFA < 1 or 1.5 < DFA < 2, the signal is persistent (larger DFA = more persistent). Collins et al. (1995) suggest that greater persistence of the COP data indicates a decline in postural stability, while Amoud et al. (2007) stated that greater anti-persistence reflects a more tightly controlled postural system. The default settings of the website are small box sizes between 32 and 100 (DFA alpha 1) and larger box sizes above 100 (DFA alpha 2). A general rule of thumb is to set the minimum box size (N min) based on a multiple of the uppermost frequency expected in the signal (i.e. for 6.25Hz in a 100Hz signal to occur 5 times this is around 32 samples). Set the maximum box size (N max) no higher than 1/10th of the sample length (i.e. for a 30 second trial set this to no higher than 300).

Physionet.org - DFA

The Neurophysiological Biomarker Toolbox - this site has a DFA tutorial which might be helpful.

Linkenkaer-Hansen et al. 2001

Collins et al. (1995)

Amoud et al. (2007)

3)

Ramdani et al. (2009)

Physionet.org - entropy explanation

Physionet.org - sample entropy Matlab code

Borg et al. (2011)

The exact code used is available at:

https://alpha.physionet.org/content/sampen/1.0.0/matlab/1.1-1/sampenc.m

After trying many different methods this one seemed the most robust. It is the “pure” matlab versions compared to their sampen.m file from the physionet website. I also leaned towards using the physionet version because it was so well established in the literature. WE and others have found wide discrepancies in the results of DFA, ranging from scores typically around 0.5 for some studies through to values clustered around 1.5 which is typical from this code. If you are running a different code that would be interesting to see why there are differences. This was first used by me in the early 2010’s, there may be other variants out there now which are more efficient.

1) Wavelet decomposition path length (cm). The trace is split into moderate (i.e., fast moving; 3.13 to 6.25Hz), low (1.56 to 3.13Hz), very low (0.78 to 1.56Hz) and ultralow (<0.78Hz) frequency signals using a 9-level Symlet-8 wavelet, with the COP path length in each band reported. This results in easily interpretable data, as higher scores reflect more sway. For more information about how this works and what it means refer to the following papers:

Liang et al. (2014)

Clark et al. (2014)

2) Detrended fluctuation analysis (DFA). This fractal measure examines the long range correlations in the trace to provide a measure of signal complexity.

Practical interpretation of these results is somewhat unclear, however when the DFA < 0.5 or 1 < DFA < 1.5, the signal is anti-persistent (smaller DFA = more anti-persistent). When 0.5 < DFA < 1 or 1.5 < DFA < 2, the signal is persistent (larger DFA = more persistent). Collins et al. (1995) suggest that greater persistence of the COP data indicates a decline in postural stability, while Amoud et al. (2007) stated that greater anti-persistence reflects a more tightly controlled postural system. The default settings of the website are small box sizes between 32 and 100 (DFA alpha 1) and larger box sizes above 100 (DFA alpha 2). A general rule of thumb is to set the minimum box size (N min) based on a multiple of the uppermost frequency expected in the signal (i.e. for 6.25Hz in a 100Hz signal to occur 5 times this is around 32 samples). Set the maximum box size (N max) no higher than 1/10th of the sample length (i.e. for a 30 second trial set this to no higher than 300).

*For a more detailed description of what this means refer to the following papers and sites:*Physionet.org - DFA

The Neurophysiological Biomarker Toolbox - this site has a DFA tutorial which might be helpful.

Linkenkaer-Hansen et al. 2001

Collins et al. (1995)

Amoud et al. (2007)

3)

__This is a measure of the irregularity of the signal. A higher sample entropy value indicates that the signal is more irregular. Given the known impacts of underlying trends in the force platform data on the sample entropy score (refer to the Ramdani et al. 2009 paper below), this website converts the data to increment values (i.e. changes the units from displacement to instantaneous velocity) prior to analysis. The default settings are pattern length: m=3, tolerance threshold: r=0.30, and time delay: tau=1. This analysis is heavily influenced by the following sources, most notably the Ramdani paper for the settings and the phyionet.org site for the Matlab code:__**Sample entropy.**Ramdani et al. (2009)

Physionet.org - entropy explanation

Physionet.org - sample entropy Matlab code

Borg et al. (2011)

The exact code used is available at:

https://alpha.physionet.org/content/sampen/1.0.0/matlab/1.1-1/sampenc.m

After trying many different methods this one seemed the most robust. It is the “pure” matlab versions compared to their sampen.m file from the physionet website. I also leaned towards using the physionet version because it was so well established in the literature. WE and others have found wide discrepancies in the results of DFA, ranging from scores typically around 0.5 for some studies through to values clustered around 1.5 which is typical from this code. If you are running a different code that would be interesting to see why there are differences. This was first used by me in the early 2010’s, there may be other variants out there now which are more efficient.