Author(s): WOOD, L., WATSFORD, M.L., DUFFIELD, R., Institution: UNIVERSITY OF TECHNOLOGY SYDNEY, Country: AUSTRALIA, Abstract-ID: 1804

INTRODUCTION:
Temporal gait outcomes such as contact time (CT), flight time (FT) and temporal asymmetry often directly informs training loads and Return to Play decisions. However, difficulties arise quantifying precise gait event timing outside laboratory contexts without motion capture. Wearable inertial measurement units (IMUs) offer an applied pathway; however, translation to training and competition requires validated, high resolution gait event detection pipelines in controlled settings before broader ecological deployment.
METHODS:
Running-trained athletes (n=14; 7 female) completed repeated constant-speed treadmill running trials (25 s, 10–20 km·h⁻¹). Bilateral ankle IMUs (1080 Hz; 20 mm proximal to the lateral malleoli) were synchronised with high-speed video (240 fps) to generate sample-level contact and flight labels for training and testing. A novel high-resolution 1D convolutional neural network (CNN) classified gait phase from short non-overlapping windows (~2.78 ms) to prioritise precise transition timing, with pre-processing to standardise left-right coordinate conventions and deterministic post-processing to enforce physiological plausibility. Model performance was evaluated using precision, recall, F1 and Youden’s J; metrics were computed per participant × speed to avoid pseudo replication. Agreement of derived CT/FT/asymmetry against video was quantified using Lin’s CCC and Bland–Altman (bias and 95% limits of agreement). Test-retest reliability used ICC(2,1) on athlete means across repeated speed pairs; model repeatability used ICC(2,1) across two independent training runs.
RESULTS:
Across 14 athletes and >9.4 million tested samples, the CNN pipeline demonstrated high discrimination (precision 0.956; recall 0.973; F1 0.965), with consistently strong sensitivity/specificity across speeds (Youden’s J 0.944-0.956). Agreement with video showed small CT/FT bias (0.001 s / -0.002 s) and narrow 95% limits of agreement (CT -0.012-0.015 s; FT -0.020-0.016 s), with proportional bias detected (p<0.001). Reliability of CT/FT improved with speed (CT ICC up to 0.91; FT ICC up to 0.97), while asymmetry reliability was poorest at slower speeds and acceptable at moderate-fast speeds. Model repeatability across independent training runs was near-perfect (CT/FT ICC > 0.99; asymmetry ICC 0.76-0.99), indicating that observed between-trial variability was driven primarily by biological and experimental factors rather than model instability.
CONCLUSION:
A bilateral ankle-IMU + CNN pipeline enables accurate CT and FT estimation across treadmill speeds, supporting scalable athletic performance monitoring. Asymmetry outcomes require cautious interpretation due to error propagation when true asymmetries are small yet demonstrate more robustness when aggregated across sufficient steps. Future work will extend validation to field-based running and optimise pipeline components to enhance ecological robustness and accelerate widescale adoption in applied settings.