Combining wearable sensor signals, machine learning and biomechanics to estimate tibial bone force and damage during running

Abstract

There are tremendous opportunities to advance science, clinical care, sports performance, and societal health if we are able to develop tools for monitoring musculoskeletal loading (e.g., forces on bones or muscles) outside the lab. While wearable sensors enable non-invasive monitoring of human movement in applied situations, current commercial wearables do not estimate tissue-level loading on structures inside the body. Here we explore the feasibility of using wearable sensors to estimate tibial bone force during running. First, we used lab-based data and musculoskeletal modeling to estimate tibial force for ten participants running across a range of speeds and slopes. Next, we converted lab-based data to signals feasibly measured with wearables (inertial measurement units on the foot and shank, and a pressure-insole) and used these data to develop two multi-sensor algo rithms for estimating peak tibial force: one physics-based and one machine learning. Additionally, to reflect current running wearables that utilize foot impact metrics to infer musculoskeletal loading or injury risk, we estimated tibial force using the ground reaction force vertical average loading rate (VALR). Using VALR to estimate peak tibial force resulted in a mean absolute percent error of 9.9%, which was no more accurate than a theoretical step counter that assumed the same peak force for every running step. Our physics-based algorithm reduced error to 5.2%, and our machine learning algorithm reduced error to 2.6%. Further, to gain insights into how force estimation accuracy relates to overuse injury risk, we computed bone damage expected due to peak force. We found that modest errors in tibial force translated into large errors in bone damage estimates. For example, a 9.9% error in tibial force using VALR translated into 104% error in bone damage estimates. Encouragingly, the physics-based and machine learning algorithms reduced damage errors to 41% and 18%, respectively. This study highlights the exciting potential to combine wearables, musculoskeletal biomechanics and machine learning to develop more accurate tools for monitoring musculoskeletal loading in applied situations.

Effect of pressure insole sampling frequency on peak force accuracy during running

“…For each step, the active force peak was computed separately for each sampling rate. We focused on active peak force from the pressure insoles because we previously found it to be the most important input signal for a wearable sensor system that was trained to estimate peak tibial bone force (Matijevich et al, 2020).…”

“…Pressure sensing insoles enable us to measure forces under the feet during activities such as running, which can provide valuable insight into human movement and clinical care (Chuckpaiwong et al, 2008; Dixon, 2008; Hullfish and Baxter, 2020; Mann et al, 2016). For instance, we have previously shown that peak force under the foot, which can be estimated from a pressure insole, is an important input for biomechanical algorithms trained to estimate forces on and understand injury risk to musculoskeletal structures inside the body (Matijevich et al, 2020). Pressure insoles can also be used in real-world data collections outside the laboratory, to collect more data (e.g., more strides), in more realistic and representative environments, and over longer periods than can be achieved in laboratory studies.…”

“…Presently, there is little empirical data in the scientific literature to help elucidate the degree to which lower sampling frequencies degrade estimates of key biomechanical metrics extracted from pressure insoles. We encountered this barrier ourselves when trying to assess what range of sampling frequencies would be acceptable for monitoring running in the real-world, given our desire to capture the active force peak under the foot each step as a key input to musculoskeletal force estimation algorithms we previously developed (Matijevich et al, 2020). To overcome this barrier, the purpose of this study was to characterise the effect of decreasing sampling frequency on the accuracy of peak force estimates from pressure insoles during running.…”

Rethinking running biomechanics: a critical review of ground reaction forces, tibial bone loading, and the role of wearable sensors

“…As presented in Table 5, three studies employed data-driven approaches to predict GRF metrics using acceleration data (Komaris et al, 2019;Derie et al, 2020;Tan et al, 2020), and one study used this approach to predict tibial loading force using IMU signals (Komaris et al, 2019). Additionally, three studies were conducted on treadmills (Komaris et al, 2019;Matijevich et al, 2020;Tan et al, 2020), and one was conducted overground (Derie et al, 2020). One study utilized IMU sensors (Tan et al, 2020), one used tri-axial accelerometers (Komaris et al, 2019), and two used virtual accelerometers (Derie et al, 2020;Matijevich et al, 2020), where the acceleration data were derived from kinematic measurements.…”

“…Notably, the prediction of GRF metrics from inertial sensors using deep learning algorithms has shown high accuracy, as evidenced in studies (Ngoh et al, 2018;Johnson W. R. et al, 2020;Tan et al, 2020). Similarly, projections of inner tibial bone load have been successfully explored through machine learning (Matijevich et al, 2020). Understanding the role of external TA in both external Frontiers in Bioengineering and Biotechnology frontiersin.org impact loading and internal tibial bone loading, therefore, becomes crucial (Matijevich et al, 2019).…”