Load carriage is a routine part of military training and combat exercises, and as recreational pursuits such as hiking and backpacking increase in popularity, load carriage factors into the athletic activities of the general population as well. Various musculoskeletal injuries are associated with load carriage. The ability to quantify the body’s reaction to load, as demonstrated in this small-scale study employing machine learning models, may help in preventing these injuries.
One of the fundamental physical competencies for members of the armed forces is carriage of externally borne loads, typically represented by body armour and/or rucksacks filled with equipment as seen in image 1.
Image 1: Rucksack
Representative external frame rucksack or backpack
While routine, this practice is associated with injuries including stress fractures, back pain and damage to the hip, knee, ankle and foot joints [1,2] due to pathological methods of biomechanical compensation for increased weight. Loads carried by soldiers are variable and depend on the necessities of the mission, but may be up to 68 kg (150 lbs), approaching 90% of the bodyweight of the average male soldier at 77kg (169 lbs); the recommendation for external carriage load is 33% of bodyweight [3].
The ubiquity of manual materials handling in many industries [4] and recent increase in popularity of leisure activities such as backpacking and hiking [5] puts the civilian public at risk for the same types of injuries.
The situational need for increased carriage load must be balanced with the increased injury risk, and if possible, mitigated by alternate carriage strategy [6]. Analysis of how individuals respond to different external carriage loads may aid in preventing these injuries by allowing intervention when loads or carriage strategies are not optimal for health.
The Plantiga platform is able to measure ground reaction force (g-force, or total gs) at the foot for each foot strike over the course of walking, running and jumping activities as well as more idiosyncratic movement types. This is accomplished through the use of a sensor-embedded insole paired with a web app displaying data visualized through the use of traditionally programmed algorithms.
The form factor of this system allows its use in monitoring the effects of load carriage in real world environments, such as training marches or treks in the wilderness, in contrast to lab based studies which are often performed using non-mobile instrumented treadmills or force plates. This reduces the potentially confounding effects of measurement technologies which require users to alter their movement strategies in order to achieve successful measurements while at the same time allowing factors such as terrain and environment to be taken into account when measuring movement.
The use of machine learning models is advantageous in analyzing datasets that combine complex movements with complex external factors. In this paper, we discuss the results of two small-scale experiments employing machine learning models to test the Plantiga system’s sensitivity to movement changes influenced by load carriage: load prediction of a static weight following a train/test cycle and load prediction of a variable weight using principal component analysis and a linear regression model.
In this test a total of fourteen subjects performed two test sessions each. The test sessions consisted of one unweighted and one weighted walk. The subjects all self reported as physically active, defined as engaging in a minimum of 3 sessions per week of rigorous physical exercise (participants: Males: n=11, Females: n=3, Age Range = 18-34).
In the first test session, participants performed an unweighted 5 km (3.1 mile) walk. In the second session, participants performed a 5 km walk wearing an external frame backpack with hip and chest belts weighted with 22.6 kg (50 lbs), approximating a military fighting load. Testing sessions were performed at least seven days apart and were performed on the same flat gravel track surface.
To create the supervised linear regression machine learning model, two datasets were excluded from the train/test cycle for use as a validation set. The first validation dataset included the loaded condition and the second included the unloaded condition. The datasets were used to validate that the model could accurately detect the external carriage load for participants on which the model had not been trained (i.e. in unseen data).
The remaining dataset were split into individual steps. Seventy five percent of the steps were randomly selected to compose the training set to train the model. Twenty five percent were selected to compose the test dataset to prevent overfitting the model. The model made predictions after every 500 steps.