TOWARDS REAL-TIME ASSESSMENT: WEARABLE-BASED ESTIMATION OF 3D KNEE KINETICS IN RUNNING AND THE INFLUENCE OF PREPROCESSING WORKFLOWS
Höschler, Lucas; Halmich, Christina; Schranz, Christoph; Fritz, Julian; Koelewijn, Anne; and Schwameder, Hermann (2024): TOWARDS REAL-TIME ASSESSMENT: WEARABLE-BASED ESTIMATION OF 3D KNEE KINETICS IN RUNNING AND THE INFLUENCE OF PREPROCESSING WORKFLOWS In: ISBS Proceedings Archive: Vol. 42: Iss. 1, Article 72.
This study aimed to estimate knee kinetics in recreational runners during treadmill running based on seven IMUs and pressure insoles using convolutional neural networks (CNN) with two input segmentations. Ground-truth knee moments of 19 runners during sloped and level treadmill running were calculated by conventional lab-based methods. We trained two CNNs on (1) step-segmented and (2) continuously windowed inputs and investigated differences in their joint moment estimations to ground-truth calculations. For both input segmentations, the predictions errors (nRMSE) were below 0.10 and 0.25 for the sagittal and non-sagittal planes, respectively. The continuous inputs led to a slightly decreased accuracy during stance phases (nRMSE<0.33). Using continuous data, we reduced the required amount of pre-processing as well as latency and achieved similar performance as with step-segmented data.
