Prediction and classification of foot strike during running using the LoadsolTM insole pressure sensors: An ecologically-valid follow-up study
Moore, S R; Kranzinger, C; Taudes, ; Stöggl, T; Kröll, J; Strutzenberger, G; Schwameder, H (2020): Prediction and classification of foot strike during running using the LoadsolTM insole pressure sensors: An ecologically-valid follow-up study In: Book of Abstracts of the 25th Annual Congress of the European College of Sport Science
The foot strike pattern performed during running is an important variable for runners, performance practitioners, and industry specialists. Foot strike estimation models generated within laboratory constraints (i.e. short runway distance or speed restrictions), may not be applicable for lab-to-field research applications. Thus, the purpose of the current study was to determine the accuracy and precision of foot strike angle prediction and pattern classification models (generated from running foot falls performed within laboratory constraints) when applied to ecologically valid running conditions.
Two Random Forest machine learning models (1) were trained to i) predict the foot strike angle and ii) classify the foot strike pattern of laboratory-constrained running (LAB) foot falls using independent variables from LoadsolTM insoles (30 participants; 70% training set = 2442 steps, validation set = 1047 steps). A sample of ecologically valid foot falls (ECO) were collected from a new set of participants (n = 19; steps = 2202) and applied to the models. Foot strike pattern classes (fore foot – FF, mid foot – MF, rear foot – RF) were consistent with predefined ranges (2). Prediction model accuracy metrics (via root mean square error – RMSE and mean absolute error – MAE) (3) and Bland-Altman bias and precision (4) were calculated for the LAB validation set and the ECO sample separately. Classification model accuracy, recall, and precision were calculated from confusion matrices (3) for the LAB validation set and the ECO sample.
The ECO data set resulted in lower prediction accuracy and precision than the LAB validation set (RMSE: ECO=8.55, LAB=3.65; MAE: ECO=6.86, LAB=2.69; Bland-Altman bias: ECO=-6.14, LAB=-0.11; Bland-Altman maximum precision: ECO=23.33°, LAB=14.30°). The classification model had higher model performance with the LAB set than the ECO set (overall accuracy: ECO=65.8%, LAB=94.1%). Of the three foot strike patterns, the MF condition classified with the least recall and precision (ECO recall: FF=22.9%, MF=60.3%, RF=95.3%, ECO precision: FF=94.5%, MF=59.4%, RF=68.3%). This was consistently lower than the LAB recall and precision (LAB recall: FF=96.3%, MF=76.7%, RF=96.4%, LAB precision: FF=95.9%, MF=74.8%, RF=97.0%).
The machine learning models generated for the prediction and classification of foot strike perform better in the LAB environment, which was consistent with the data used to train the models. However, the evidently good classification recall of RF ECO running suggests that similar models may be generated from ECO and LAB RF foot falls. Importantly, when applying future prediction and classification models, the modelling environment should reflect the environment in which the models will be applied.
1. Breiman, Mach Learn, 2001
2. Altman & Davis, Gait Posture, 2012
3. Galdi & Tagliaferri, Ency Bioinformatics Comput Biol, 2019
4. Bland & Altman, Int J Nurs Stud, 2010