From the four sensors in the Sparta Science force plate, data is collected one thousand times per second (1000 Hz) and is ultimately resolved into vertical force and center of pressure (CoP). So for each millisecond, three different pieces of data are generated (i.e. 3 degrees of freedom): resultant vertical force (Z), CoP coordinate in x plane (X), CoP coordinate in y plane (Y). This data is most commonly referred to as time-series data, and in its “raw” form isn’t of much immediate utility, which is why different softwares and systems derive (extract) different metrics (features) from the time-series data that may be of more interest and use to scientists, researchers, and hopefully practitioners and organizations.
What’s with the jargon?
This may all be information you generally are aware of, but you may be seeing a few new terms that are worth explaining. The terms extract and feature can be thought of similarly to derive and metric respectively. This terminology is more commonly used in the machine learning community and thus is a bit more accurate when discussing how we utilize force plate data here at Sparta Science.
Extracting Features: Now and Forever
The majority of software programs paired with force plates derive or calculate (using calculus derivatives based on principles of physics) different force plate metrics to provide users with more easily interpretable information than the time-series data provides. For example, some researchers may have a good understanding of how concentric rate of force development is calculated from a countermovement jump, what typical ranges are for their subjects, and what that metric may represent.
Similar to other softwares, our platform also extracts features from the time-series data utilizing physics-driven principles and provides them after the completion of scan (force plate assessment). However our platform additionally stores all of the time-series data in our secure data lake (database) as part of the Sparta Movement Health Platform ecosystem. This allows us to continuously extract new features from historical data and use them in analysis and modeling. So if there is a new feature (metric, variable, etc) of interest, utilizing our machine learning pipeline we are able to extract this feature from previous data collection and utilize it in analysis, research, and model development.
This is not only a huge benefit for researchers whose constant challenge is recruiting more and more subjects for studies but for organizations that are investing in better leveraging data insights. Only through leveraging a platform is one able to extend the lifespan of data, allowing continuous learning and insights through the use of machine learning.
I just jumped 65 inches!
It is important to understand that all of these physics-derived metrics are calculated, not measured directly. Therefore:
- There is the possibility of error in these calculations
- These metrics simply allow us to better interpret information from the time-series data by viewing the data through a physics-based lens.
Using the time-series data to derive physics-based features accurately can actually be a quite tricky process. The calculations themselves are relatively simple, however, if there are quality issues with the underlying time-series data this error can be magnified making the results of our calculations off. Sometimes, way off.
We will use two different but related features to explain: Jump Height (from impulse) and Jump Height (from flight time). Both of these features are calculating Jump Height utilizing two different methods which give us slightly different but usually close answers. When the quality of underlying time-series data is suspect, however, we can get some pretty wonky results! For example in one data set, we saw an individual with Jump Height (flight time) as 27 inches yet a Jump Height (impulse) as 65 inches! While displaying a 65-inch jump is an obvious error in calculation, the underlying problem is real. At least in this case we cannot assume that the other calculated metrics are correct either… and what about last time? … and next time?
In real-world settings, this can be difficult as athletes, patients, and soldiers can quickly lose trust in the accuracy of or value of technology.
Feature Extraction ≈ Smart Calculation
Because we have years of stored time-series data, our data scientists and developers have been able to iterate and improve on the standardization of our scan protocols to ensure that data is collected in a reliable fashion and features are extracted accurately. In this context it is easy to think of the difference between extraction and calculation in that extraction is simply a “smart calculation.” Not only are we performing the physics-based calculation, but with the depth of data we have we are able to deploy machine learning models directly to our Scan app that make sure to filter time-series data for quality and accurately perform the calculations.