Valid and invalid foot contacts on force platforms during gait analysis - A dataset for automated classification
Description
Background The incorporation of force platform data, i.e., ground reaction force (GRF) and center of pressure (COP), in biomechanical gait analysis requires valid foot contacts on the force platforms. Foot contacts are considered valid if the foot has complete and exclusive contact with a force platform while the other foot does not touch this force platform. Compliance with these criteria is usually assessed subjectively by visual inspection by the person conducting the gait analysis. Research question Can the assessment to distinguish between invalid and valid foot contacts on a force platform during gait analysis be automated using a machine learning model? Methods Twenty healthy participants (10 female and 10 male) underwent gait analysis using GRF and COP measurements during the stance phases on one force platform (Kistler, Switzerland). Six typical cases of invalid foot contacts in force platform measurements were simulated, with simple and diffcult valid and invalid foot contacts recorded in each case. Each measurement was classified by two examiners through visual inspection and two video recordings (Qualisys, Sweden) of the lower body. A Support Vector Machine (SVM) was trained to distinguish valid and invalid foot contacts on the force platform based on preprocessed GRF and COP time-series signals. different combinations of GRF and COP data as input to the SVM were evaluated. Results Using a combination of anterior-posterio and medio-lateral COP as input to the SVM achieved the highest accuracy of 96.6% (100% of simple cases and 93.2% of diffcult cases). Significance The development of an automated classification model based on machine learning has the potential to enhance the precision of foot contact assessments on force platforms during gait analysis. This can benefit experimental procedures by improving the quality of data and increasing the usability of (publicly) available datasets through simplified data cleaning. Keywords: ground reaction force, GRF, center of pressure, COP, support vector machines, SVM, pattern recognition Published Dataset The data collected for this study is shared publicly for reproducibility and as a benchmark for similar approaches. The .c3d files contain the raw, unprocessed analog force plate data (Kistler 9287CA, 60x90cm, 400Hz). The pre-processed data is shared as .csv file(s) - either separated by channel or all in one file. The recorded videos cannot be published for data privacy protection reasons. The metadata for each subject and each study is also available in .csv format.
Files
Steps to reproduce
Twenty physically active participants (10 female, 10 male; 27±7 years; 177±9 cm; 72±15 kg) without gait pathology and free of lower extremity injuries were included in the study. The participants walked barefoot along a 10 m-long walkway with an embedded force plate (FP) (Kistler 9287CA, Switzerland; 60x90cm, 400 Hz) under 24 instructed conditions. In these conditions, six cases of invalid foot contact in FP measurements were simulated. These included four cases with partial contacts of a single foot: partial single foot contact at the medial (SCM), lateral (SCL), top (SCT), and bottom (SCB) edge of the FP. Furthermore, two cases included a partial double foot contact of the other foot: partial double foot contact at the bottom (DCB) or top (DCT) edge of the FP. For each case, ”simple” and ”di cult” valid and invalid foot contacts were acquired. Each participant completed at least five measurements for each condition. Only one side (left or right foot) was measured per participant, with the tested side alternating at the beginning of the study. Participants were directed to walk on a straight line towards the FP from different starting positions based on a ground template. The template’s position was calibrated during 10 familiarization walks. Participants were instructed to walk at their preferred (self-selected) speed. The ground truth classification for each measurement was assessed visually by two examiners and controlled with two sagital plane video recordings (Qualisys Miqus Video, Sweden; 85 Hz). The calculated 3D ground reaction forces (GRF) and 2D centers of pressure (COP) were processed using a zero-lag Butterworth low-pass filter (2nd order, cut-off frequency = 15Hz). The trajectories were trimmed to the foot contacts with a vertical force-threshold of 10N for GRFs and 50N for COPs and linearly resampled to 101 values to represent percentage of contact time (0-100%). The amplitude of the GRFs were normalized by the participant’s body weight and the COPs based on the size of the FP. The medio-lateral components were mirrored for the measurements performed with the left foot. The data processing was done using Matlab R2022b and the Biomechanical ToolKit. In total, 1920 measurements were used for the binary validity classification using kernel support vector machines (SVMs), with a balanced number of four measurements randomly selected per condition for each participant. A 5-fold cross-validation was used to validate the performance of the SVMs. For this purpose, 5 folds were formed (two of each sex, left and right foot each). Each fold (384 measurements) was used once to validate the model trained with the remaining folds (1,536 measurements). The hyperparameters (linear or RBF kernel, C and gamma) were optimized for each case based on a 2-fold cross-validation on the training data using grid search. The data analysis was conducted using the Scikit-learn library within the Python 3.11 environment.