Table 1.
Participant demographics and testing conditions.
Fig 1.
Tracking markers used for each dataset.
The shank was tracked using clusters at different positions and different marker configurations. CALC: calcaneus, CALC2: lateral calcaneus, MTB5: 5th metatarsal base, MPT5: 5th metatarsal head, MTB2: 2nd metatarsal base, MTP2: 2nd metatarsal head, MTP1: 1st metatarsal head.
Fig 2.
Graphical representation of an LSTM unit at timepoint t, where c and h represent the cell state and hidden state respectively, x represents the input features at the current timepoint, and f, i o are the forget, input and output gates controlling the information passed on to the cell state.
Fig 3.
FootNet takes the distal tibia anteroposterior velocity, ankle plantar/dorsiflexion angle and the foot COM anteroposterior and vertical velocities as input and produces a sequence of probabilities of non-contact (0) and contact (1). Note that the input features are first standardised (z scores), 0-padded at the start of the sequence to a standard length of 200 data-points and concatenated into a 1x200x4 array where 1 is the number of cycles, 200 is the number of timepoints and 4 is the number of input features. Zero padding was added to batch the training cases and 200 was selected as a sufficient number of timepoints to accommodate a stride cycle at any of the tested speeds. The padding is ignored during training, hence the 0.5 values at the beginning of the output sequence.
Fig 4.
From each dataset, 30% of the participants were extracted, concatenated and left aside as test set for model testing (22 participants, 8712 cycles). The remaining 70% of the participants were divided in five groups and concatenated in five folds with a representative number of participants from each dataset (10 to 14 participants each, 4999 ± 842 cycles in each fold). These five folds were used for cross-validation, whereby five models (i.e. sets of parameters) were developed using four folds for training and the remaining fold for validation, with a different validation fold for each model. The best performing model was selected, retrained using the five folds altogether and then tested on the test set.
Fig 5.
Algorithm performance for foot-strike, toe-off and contact time.
Top row: Bland-Altman plots. The top and bottom border of the grey patch represent the 95% limits of agreement. Regression lines are included to aid visual interpretation. Correlations between instant within the gait cycle and error were trivial for foot-strike (r = -0.07, p < 0.001) and toe-off (r = 0.04, p < 0.001) hence, line coefficients and coefficient of determination are not reported. Average contact time length and error in contact time (r = 0.1, p < 0.001) line coefficients and coefficient of determination are reported in the Results section. Second row: error histograms. The frequency of errors contained within ±1 motion capture frame of the ground truth event is highlighted.
Fig 6.
Scatter plots for the relation between running speed (top row), foot angle at contact (middle row) and incline (bottom row) and errors in the detection of foot-strike (left), toe-off (middle) and contact time (right) respectively. Note that only data on flat running surfaces was included for the running speed and foot angle at contact scatter plots and correlation analyses. Regression lines are included in every plot to aid visual interpretation, but line parameters and coefficient of determination are only reported in the Results section for nontrivial correlations.
Fig 7.
Top row: mean (± standard deviation) hip, knee and ankle sagittal plane angles at 2.5 (cyan), 4 (blue) and 5 (black) m/s during the gait cycle beginning from highest foot COM position to highest foot COM position. Average foot-strike and toe-off are indicated with vertical dashed lines to aid interpretation. More positive angles refer to dorsiflexion in the ankle plot. Using 4 m/s as example, the violin plots show the error (°) distribution in hip, knee and ankle sagittal plane angles at foot-strike (2nd row) and toe-off (4th row) as a function of anticipated (negative time values on the x axis) or delayed (positive time values) step event detection, where the correct value of the variable is taken from the event time provided by the “gold standard” method (force plate). The horizontal black line within each violin represents the mean. Error classification in acceptable (light blue), reasonable (dark blue) and unacceptable (red) are displayed for foot-strike (third row) and toe-off (fifth row) respectively.