Electing a representative

One of the aspects of gait analysis that I didn’t cover in any particular depth in my book was how to select data from a number of trials that is in someway representative of the patient. I think one of the reasons for this is because I couldn’t easily get my hands on any data that illustrated the issues well. In some ways this speaks for itself – in my experience large inter-trial variability, larger enough to affect how we interpret data, is actually quite rare in clinical practice. If the variability is small then it doesn’t really matter what technique you choose.

representative-trial

My personal preference is to avoid the issue altogether by overplotting data from multiple trials on the same graphs (similar to the graphs on the left above but with data from the other side plotted in a different colour as well). In this way you can take into account both the general pattern and the variability when interpreting the data. Some people object that using this technique it can be difficult to appreciate subtle features in individual traces, but there is a real question as to whether you should be even looking at such features if they are not consistent from trace to trace.

Now I’ve started annotating graphs with symbols to identify specific features in the data, however, I find that the combination of symbols and multiple curves can be a bit too messy and have resorted to looking to a single trace that can be taken of as representative. Although alternatives have been proposed I use the average trace. It leads to a little smoothing of the data – which isn’t perfect – but none of the other techniques are perfect either. It was interesting last week to come across a patient’s data that illustrated beautifully the problems of doing this.

You can see the knee and hip traces here with the individual traces overlaid on the left and the averaged data plotted on the right. The problem is with the left knee. You can see that the patient exhibits two distinct patterns of knee movement in early stance (but remarkable repeatability elsewhere in the gait cycle). She either walks with full knee extension in early stance or with quite marked knee flexion. In the fully extended pattern her knee is more posterior and so there is increased hip extension as well (there is an effect on ankle dorsiflexion as well which I haven’t plotted). The average trace for the knee, however, falls well within normal limits and if you only ever looked at that trace you would never know that there was anything wrong with the knee.

None of the methods that are commonly used to generate a representative trace whether they be through picking one particular trace or providing some sort of averaging (mean or median) will result in something that represents the patient. The fundamental reason these don’t work is that this person’s gait is not characterised by one gait pattern, but by two, which she alternates between. It is not possible to understand her walking on a single curve, you need to look at multiple trials.

So although such issues don’t arise very often we don’t have a good way dealing with them in how we plot out or mark-up our data. What is needed is not a way of selecting single trace but a means of indicating that any single trace is unrepresentative. The broad semi-transparent bands on the right hand graphs are supposed to indicate that the trace cannot be appropriately interpreted without referring to the multiple trial data. I’ve now added them to the list of symbols that we use for marking up our graphs.

(PS The idea for a symbol to indicate underlying variability in the multiple trial data was first suggested to me by Sheila Gibbs in Dundee during an early consultation on the Impairment Focussed Interpretation methodology I outline in my book.)

(PPS if you do feel a need to select a representative trace and want to do this systematically then a robust method is presented by Morgan Sangeux in this recent paper. It is however worth noting that in the example he gives in Figure 1 the trace chosen as most representative overall does not give a good indication of the underlying data for either pelvic tilt or foot rotation despite the rigour of the technique.)

Advertisements

One comment

  1. Hi Richard,
    Thanks for mentioning our method to pick the most representative stride and detect outliers. I agree the trace picked as representative in Figure 1 of our paper does not represent well the distribution of the curves for the pelvic tilt and foot progression graphs. This is however inherent to selecting the most representative stride overall, i.e. no trace would have been satisfying for all graphs simultaneously. This is exemplified in figure 2 of our paper where stride #5 was picked as representative overall, but stride #2 (for hip flexion), #3 (for pelvic obliquity and foot progression) and #6 (for pelvic and hip rotation) were picked based on individual graphs (stride #5 was picked for all the other graphs).
    In practice, we pick the most representative stride overall because the joints in the lower limb are linked in a kinematic chain and, for example, what happens at the knee is correlated to what happens at the hip. In the hip/knee graphs you show in your post, I suspect the two curves differing during loading response at the knee correspond to the same two strides differing at the hip at the same instant. This relationship is easy to identify in this case but may be more difficult in more complex gait patterns, for example in Figure 1 in our paper.
    We do not use the representative stride for clinical decision making but plot all the strides (6 or more) and let our eyes (and brain when it is still connected to them 🙂 identify the variability and where the overall pattern may or may not be trusted. However, we do select a representative stride to analyse results from cohort studies in order to obtain one observation per subject (when the variability of the gait pattern intra-subject is not part of the statistical analysis, which is often the case). Most clinician agree the mean is not an observation, i.e. chances are that the patient would never walk like its average trace. The representative trace on a graph per graph basis is not an observation either, probably even less than the mean, which is why we pick the representative overall.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s