Gait signature changes with walking speed are similar among able-bodied young adults despite persistent individual-specific differences

Abstract

Understanding individuals' distinct movement patterns is crucial for health, rehabilitation, and sports. Recently, we developed a machine learning-based framework to show that "gait signatures" describing the neuromechanical dynamics governing able-bodied and post-stroke gait kinematics remain individual-specific across speeds. However, we only evaluated gait signatures within a limited speed range and number of participants, using only sagittal plane (i.e., 2D) joint angles. Here we characterized changes in gait signatures across a wide range of speeds, from very slow (0.3 m/s) to exceptionally fast (above the walk-to-run transition speed) in 17 able-bodied young adults. We further assessed whether 3D kinematic and/or kinetic (ground reaction forces, joint moments, and powers) data would improve the discrimination of gait signatures. Our study showed that gait signatures remained individual-specific across walking speeds: Notably, 3D kinematic signatures achieved exceptional accuracy (99.8%, confidence interval (CI) 99.1-100%) in classifying individuals, surpassing both 2D kinematics and 3D kinetics. Moreover, participants exhibited consistent, predictable linear changes in their gait signatures across the entire speed range. These changes were associated with participants' preferred walking speeds, balance ability, cadence, and step length. These findings support gait signatures as a tool to characterize individual differences in gait and predict speed-induced changes in gait dynamics.

Fig. 1

Visualization of representative (a) kinematic and (b) kinetic treadmill walking data. Representative continuous right angle flexion/extension (a-i) and right vertical ground reaction force (b-i) for one participant across three speed conditions: extreme slow, self-selected, and extreme fast (walk to run transition). Representative phase-averaged left knee flexion/extension (a-ii, top), left ankle flexion/extension (a-ii, bottom), normalized left ankle moment flexion/extension (b-ii, top), and normalized left anterior–posterior ground reaction force (b-ii, bottom) data for three participants colored by individual (color) and speed (gradient).

Fig. 2

Pipeline of the gait signatures framework and outcomes. (a) 3D motion capture data from 17 able-bodied young adults walking on a treadmill across 9 speeds each was conducted. (b) Continuous timeseries kinematics and kinetics were extracted from all trials. (c) A sequence-to-sequence RNN was trained using subsets of the data recorded in (b), and individual-specific gait signatures were extracted for all individuals’ trials. (d) Principal component analysis was applied to reduce the dimensionality of the high dimensional latent states, each trial was phase-averaged, and the first 3 dominant PCs visualized as 3D loops. (e) 3D projections of low-dimensional gait signatures using multidimensional scaling reveal individual-specific gait signatures among 3 representative able-bodied young adults.

Fig. 3

Variance explained in the original signals by each of the principal components of the extracted dynamics. (a) To determine the variance explained in the original data by the dynamics: (1) the RNN model was trained, (2) the trained model weights (internal activations) were extracted and the dynamics corresponding to each PC was isolated. (3) For each PC, the original signals were reconstructed across participants. (4) The coefficient of determination was calculated between the measured input signals and the reconstructed signals for each PC. These values were used to construct eigenvalue plots for each signal type. (b) Eigenvalue plots of the cumulative variance explained by increasing number of principal components of the gait signature for each of the four signal types.

Fig. 4

Schematic outlining the comparison of individual discriminatory power between the gait signatures generated using four different signals (2D kinematic, 3D kinematic, 3D kinetic and a combination of all signals). 3D motion capture of treadmill walking was obtained from 17 able-bodied young adults, encompassing 9 different speed conditions. The four data types were created, each with varying number of features. RNN models were trained for each data type and respective gait signatures were generated. The classification accuracy of individuals across different speeds was assessed using support vector machine (SVM) classifiers, and the classification accuracies between the four gait signature types were compared.

Fig. 5

Kinematic gait signatures are individual-specific. i) 3D looped trajectories of the first 3 principal components (PCs 1–3) of (a) 2D kinematic and (b) 3D kinematic gait signatures are individual-specific across speeds. Individual’s trials (same color) are grouped closely together with similar shapes. ii) 3D looped representations of the second set of principal components (PCs 4–6) reveal greater individual specific clustering in (a) 2D kinematic signatures, and stronger differentiation observed in (b) 3D kinematic signatures. iii) Individual classification accuracy is lower using (a) 2D kinematics vs. (b) 3D kinematics across varied number of speed trials in the classification model training set. 3D kinematic signatures exhibited robust classification of individuals, achieving a mean accuracy of 99.8% (CI 99.1 to 100%), with a minimum of four speed trials in the training set, surpassing the accuracy of 2D kinematics, which achieved 96% (CI 88.2–100%). Please note that significant differences exist between 2 and 3D kinematic signature accuracies regardless of the number of speed trials in the training set. However, for clarity, we specifically emphasize the statistical comparison in the illustration using only four speeds in the training set as this is where 3D kinematic signatures attain near perfect (100%) classification accuracy. iv) A representative confusion matrix from a single classification model run shows that several individuals were misclassified when four speed trials per individual were included in the training set. v) The intra-individual trial distances in MDS space for 3D kinematic signatures are smaller than the inter-individual distances, further showcasing the individual-specificity across all speed trials of an individual.

Fig. 6

Gait signatures hold information about walking speed. i) 3D looped trajectories of principal components 1–3 colored by gait speed show that similar speed trials are shaped similarly in (a) 2D kinematic and (b) 3D kinematic signatures. Slower speed signatures (blue) are concentrated in the center of all signatures and faster speed signatures (red) on the outskirts, fanning outward and upward in PC3. ii) 3D MDS visualizations of all signatures colored by speed illustrates that slower speeds (blue) across individuals are in the top left region of the map for (a) 2D kinematic signatures and in the right half region of the gait map for (b) 3D kinematic signatures. iii) 3D MDS visualizations of all signatures colored by individual fit with linear mixed effects models show that individuals gait signatures change similarly and linearly with change in gait speed.

Fig. 7

Individual’s gait signatures change linearly with speed. Simple linear regression of individuals’ 3D MDS coordinates vs. speed show similar (a) R² values and (b) slopes across individuals. (c) Hierarchical bootstrapping of linear mixed effects shows that the linear relationship of MDS coordinates with speed is robust across variability in model input data, the number of speed trials selected per individual, and the randomness of the selected speed trials used in model.

Fig. 8

Balance ability may be associated with the extent to which individuals modulate their gait signatures with changes in speed. (a) A moderately positive linear relationship exists between balance ability and the change (Euclidean distance) between SS and extreme slow speed gait signatures- individuals with better balance modulate their gait signatures more when reducing speed from SS to extreme slow walking speeds. (b) A moderately positive linear relationship exists between self-selected walking speed and balance ability- individuals with better balance prefer walking at faster self-selected walking speeds. (c) A strong positive linear relationship exists between change in gait signatures between SS and extreme slow speed and self-selected walking speed. (d) 3D MDS representation of all individuals’ gait signatures colored by their narrowing balance beam score reveal no clustering of signatures in this space.

Fig. 9

A strong positive linear relationship exists between the Euclidean distance between self-selected and extreme slow speed signatures and changes in (a) right leg cadence and (b) left leg step length.