Joint and tendon loading during running: towards estimating musculoskeletal loads from IMUs with self-organising maps

Abstract-Text (inkl. Referenzen und Bildunterschriften)

Rationale

Running-related injuries are frequently researched. However, there is little evidence that the currently accepted risk factors have predictive value. The need for prospective and longitudinal data collection is hampered by the lack of information about the loading of joints and muscles when data is gathered in the field. While the combination of inertial measurement units (IMUs) and supervised learning tends to overcome this problem, a new network needs to be trained for each of the desired outcome variables.

It was hypothesised that an unsupervised learning algorithm is capable of detecting running patterns from wearable sensor data and that these running patterns can be used to accurately estimate joint loading variables as well as ground reaction forces (GRFs) and tendon forces.

Methods

To test this, a publicly available data set of 28 recreational runners, running on a treadmill at three different speeds was used to (1) generate artificial IMU data from the existing retroreflective marker trajectories, (2) train a Kohonen"s self-organising map (SOM) with the artificial IMU data and assign 3D GRFs, hip, knee and ankle joint moments and reaction forces as well as the patellar tendon and Achilles tendon forces. Subsets of the artificial IMUs were used to identify the minimum number of IMUs necessary for achieving predictions at an accuracy of above 85%.

Results

The assigned variables were estimated with good to excellent accuracy (R²>0.86) except for the mediolateral GRF and the hip joint reaction force (R2=0.39 and 0.57, respectively). Normalised root mean squared errors were at an acceptable level (<16%) again with the exception mentioned before (27% and 26%). To achieve these results, only the data of four artificial IMUs on the shanks and the feet were required.

Conclusion

The proposed approach used a novel approach by utilising unsupervised learning to allow a flexible yet reliable estimation of joint loading variables from two IMU sensors.