Prediction of ground reaction forces in running from wearable instrumentation and algorithmic models

Measurement of ground reaction force (GRF) in running provides a direct indication of the loads to which the body is subjected, at each foot-ground contact, and can provide an objective explanation for performance outcomes. Traditionally, the collection of three orthogonal component GRF data in running requires an athlete to complete a series of return loops along a laboratory based runway, within which a force platform is embedded in order to collect data from a discrete footfall. The major disadvantages associated with this GRF data collection methodology includes, the inability to assess multiple consecutive foot contacts and the fact that measurements are typically confined to the laboratory. The objective of this research was to investigate the potential for wearable instrumentation to be employed, in conjunction with artificial neural network (ANN) and multiple linear regression (MLR) models, for the estimation of GRF in middle distance running. A custom, wearable data acquisition system was developed to acquire in-shoe force (ISF) and centre of mass acceleration (CMA) data simultaneously. Matched data sets from wearable instrumentation (source data) and force plate (target data) records were collected from elite middle distance runners under controlled laboratory conditions for the purposes of ANN and MLR model development (MD) and model validation (MV). Using a range of source data groupings, including ISF and CMA in isolation as well as in combination, it was found that ISF data, employed separately, provided the highest ANN and MLR model prediction accuracy for all three components of GRF. In general terms, an intra-subject, running speed interpolation based prediction scheme along with the MLR model was found to provide the highest prediction accuracy for the vertical (Correlation Coefficient [CC]: 0.997-1.000, Mean Absolute Error [MAE]: 14.494-46.658N) and medio-lateral (CC: 0.875-0.974, MAE: 10.680-39.890N) components of GRF. Alternatively, under the same prediction scheme, the ANN model provided the most accurate predictions of the anterior-posterior (CC: 0.974-0.992, MAE: 14.910-31.118N) component of GRF. The prediction accuracy of each component of GRF was found to be governed by the inherent signal variability, in which case the vertical and anterior-posterior components were more reliable and subsequently predicted significantly more accurately than the medio-lateral component of GRF. In order to achieve accurate GRF predictions, the wearable instrumentation and algorithmic models need to have sufficient sensitivity to capture this inherent GRF signal variability. Findings from this research provide a proof of concept for the prediction of GRF from wearable instrumentation in middle distance running. The emerging capability for obtaining continuous GRF records from wearable instrumentation has the potential to permit unprecedented quantification and analysis of training stress and competition demands.