Machine Learning to Estimate Biomechanical Forces

Mark Pogson from Quintessa has co-authored a research paper in the journal Medical Engineering & Physics, which uses machine learning to estimate forces exerted on the body while running, based on data from a body-worn sensor.

Biomechanical loads are an important consideration for athletes, but they are difficult to measure outside the laboratory, hence load monitoring tends to focus on physiological metrics such as heart rate, thus giving an incomplete picture of overall training load. Working with collaborators in biomechanics and data science at Liverpool John Moores University, KU Leuven and the University of Birmingham, Mark has developed a machine learning method to estimate biomechanical loads during running activities by using data from an accelerometer attached to the torso, which measures trunk acceleration (TA).

The accelerometer is part of a device which is already widely used by athletes to track their movement using GPS, so the new method provides an easy and cost-effective way to quantify biomechanical loads outside the laboratory by using data which are readily available.

To represent biomechanical load, the method uses the magnitude of ground reaction force (GRF), which is the force exerted on the body when it impacts the ground (Figure 1). GRF can be measured in the laboratory by using a force platform, which allows the relationship with TA, as measured by the accelerometer, to be observed (Figure 2). Although there is a clear association between the two, a mechanical model faces major challenges to predict GRF from the TA signal because the centre of mass of the body continually moves relative to the sensor location.

The method was found to provide excellent estimates of GRF from TA data in many cases, as shown in Figure 3, demonstrating its effectiveness to predict both GRF magnitudes and time dynamics for different athletes and types of impact. Further analysis found that model predictions performed well overall, but some outliers were evident for particularly short and long impacts, which may reflect the distribution of data used for model training.

The method provides promising results which may help athletes to account better for biomechanical loads in their training schedules, thus increasing the effectiveness of training while reducing the risk of injury. With further work, it may be possible to improve GRF estimates by using subject-specific model training and including biophysical data inputs. GRF monitoring is also relevant to healthcare, including for load-dependent conditions such as osteoarthritis, hence similar approaches could be of value in these areas.

A diagram showing the direction of GRF, which opposes the force excerted to the ground by the runner
Figure 1: Ground reaction force (GRF) is the force exerted on the body when it impacts the ground. Being able to measure GRF in sport-specific environments would help to ensure appropriate training intensities for athletes.
Graphs show variation in trunk acceleration (TA) on a scale of 0 to 70 m s-2 and ground force reaction (GF) on a scale of 0 to 3.5 kN with time on a scale of 0 to 0.25 s for a) steady running and b) deceleration. Steady running GRF peaks with a value of 30kN at 0.1 s before decaying, where the Trunck acceleration first peaks at about 0.1s at 30ms-2, then decreases to 0ms-2 and increases again to another local maximum of about 20ms-2 at 0.15s, then decays for the remaining 0.15s.
Figure 2: Trunk acceleration (TA) and corresponding magnitudes for single impacts during steady running and deceleration, as measured in the laboratory. Different types of impact have characteristic GRF and corresponding TA profiles, but the relationship is difficult to model mechanically.
The figure shows three graphs, the first show the predictions of the GRF of steady running agains the observed values. Both lines increase to a local maximum of around 1.5kN between 0 and 0.05s, then decrease and increase again to a similar maximum at around 0.1s, then both steadily decrease until they reach 0. The predicted line reaches 0 at 0.25s where the measured data reached 0 at 0.2s. The next graph shows the data from Acceleration, both the measured and predicted lines show a curve that steadily increases to a maximum of about 1.7kN at 0.15s, then decreases back to 0 at around 0.3s. The third graph shows Deceleration, both lines on this graph show a peak of 4kN between 0 and 0.05s, and then a sharp decay down to around 1.5kN at 0.05s where they remain steady until 0.1s, before decreasing to 0kN at around 0.2s
Figure 3: Examples of GRF predictions from machine learning, compared against test data for different types of impact. Predictions match the different impacts well, with good agreement evident in both the GRF magnitudes and time dynamics.