Estimating bilateral hand kinetics from upper-limb kinematics in sprint front crawl: a subject-wise validation of an artificial neural network

Hand force contributes substantially to front crawl performance, but measuring it requires hand-worn hardware. We tested whether an artificial neural network driven only by three-dimensional upper-limb kinematics can reconstruct bilateral hand kinetics recorded by a hand-force device, and at which level of the signal (waveform, discrete descriptors, between-swimmer and inter-limb differences) the estimates agree.

Twenty-four male competitive swimmers performed maximal 50-m front crawl without breathing while wearing eo SwimBETTER handsets; 14 upper-body markers were tracked by underwater motion capture at 100 Hz. For each swimmer, the underwater phase (fingertip entry to exit) of each hand was analysed. A laterally mirrored multilayer perceptron using 197 trunk-frame features estimated six targets per hand under subject-wise six-fold cross-validation, with architecture, regularisation and smoothing selected inside each fold. It was compared with four alternative learned models and two baselines (a scalar group-mean null model and a phase-conditioned mean-waveform template), and uncertainty was estimated with participant-cluster bootstraps.

For resultant force, root-mean-square error was 24.5 ± 6.1 N (left) and 23.3 ± 9.9 N (right), per-window correlations were 0.85 and 0.84, and ICC(A,1) was 0.70 and 0.64. The network was comparable to gradient boosting and coordinate-only models but was not more accurate than the template (template RMSE 21.8 and 21.2 N; left-hand difference 2.7 N, 95% CI 0.9-4.5). Between-swimmer correlations of measured and estimated peak force, impulse and mean force ranged from -0.36 to 0.27, and the estimated inter-limb symmetry index did not agree with the measured one (r -0.23 to -0.10; the dominant side was matched in 50%-58% of swimmers). Left-hand peak force was underestimated by 32.6 N, and the absolute error in time to peak averaged 12%-13% of the underwater phase. Modelling both hands in every window, a broader definition that includes each hand's recovery and entry, inflated apparent accuracy (left RMSE 21.6 N, r 0.89).

In a device-on protocol, kinematics-driven estimation recovered the group-typical force-time pattern but added little to a phase template and did not resolve differences between swimmers or between limbs. It cannot yet replace hand-force measurement for individual assessment.