An overarching question the biomechanics community is still trying to answer is: How do we assess an individual against normative data in a meaningful way? For example, a clinician may want to assess whether a patient needs surgery. A sports team may want to determine if an athlete is ready to return to play. In each of these scenarios, the assessor needs to determine what they are comparing the individual against. If this is a normative data set, how has the data been collected? Is there enough data and does it represent enough individuals? How do we summarize the data so that we can make meaningful comparisons? Assessment of Neurological Impairment and Recovery Using Statistical Models of Neurologically Healthy Behavior [1], was a great starting point for discussions about normalized data sets at our recent journal club.
Building a Normal Dataset
Building a normal data set requires gathering enough diverse patients, collecting data, and processing the data, which is most often a several year process. When something takes this much time and effort we want to take careful consideration in how it’s done, so that we don’t waste time. Some of considerations may be:
- When can we stop collecting data? When do we know we have enough participants?
- How diverse should the participants be? Other than the standard, age, sex, height and weight, are there other factors to consider?
- Should one normal dataset contain multiple activities (i.e. multiple speeds of gait) or should they be separated?
- Should we put a focus on building normal datasets with different activities other than gait to get a comprehensive overview of someone’s motion?
- What activities should the participants be doing?
Luckily, markerless motion capture has made these questions easier to tackle, since we can now collect large data sets in a much more efficient way. As Scott et al. stated, it only takes 30-50 individuals in a normal dataset when looking at the mean, and a few hundred when looking at a distribution for 2D movements. 3D movements can require an order of magnitude more participants. Since markerless motion capture allows for thousands of participants to be easily collected for a single dataset, the focus can be on achieving diversity in the participants and performing tasks that will best represent the study question.
Using a Normal Dataset
Once we build a normal dataset, how do we know how to use the data to get the best representation of a ‘healthy performance’? It is common to compare an individual to the mean of the normal data, however we agree with Scott et al. that this is not a sufficient comparison. Only using the mean will eliminate important information and over simplify data comparison. In response to this idea, Scott et al. recommended using a distribution of the normal dataset and then calculating the Mahalanobis distance [2] between the individual and distribution to get a more comprehensive understanding of how the individual compares to the norm. We believe that this process of using measure distances is a step in the right direction, and biomechanists should consider similar methods moving forward.
Special Case Considerations
For certain applications we may not have access to normative data that represents the question we are trying to answer. For example, if we are determining whether or not a top professional pitcher should return to the game after an injury, we could not generate a data set of enough individuals that reflects the players original performance level. In situations where a normative data set can not be generated, is it enough to assess an individual on their own previous performance? What other ways could we assess an individual who is so unique?
Moving Forward…
The discussion on normative data led to more questions than answers. With some questions being ambiguous, we believe that more discussion on best practices in building a normal data set is needed to ensure that we are building and using a dataset that best represents a healthy performance.
[1] Scott SH, Lowrey CR, Brown IE, Dukelow SP. Assessment of Neurological Impairment and Recovery Using Statistical Models of Neurologically Healthy Behavior. Neurorehabil Neural Repair. 2023 Jun;37(6):394-408. doi: 10.1177/15459683221115413.
[2] Maesschalck R, Rimbaud D, Massart D.L.. The Mahalanobis distance. Chemometrics and Intelligent Laboratory Systems. 2000 Jan.;50(1). doi:10.1016/S0169-7439(99)00047-7