RESUMEN
The purpose of the present study was to identify the performance determinant factors predicting 15-m backstroke-to-breaststroke turning performance using and comparing linear and tree-based machine-learning models. The temporal, kinematic, kinetic and hydrodynamic variables were collected from 18 age-group swimmers (12.08 ± 0.17 yrs) using 23 Qualisys cameras, two tri-axial underwater force plates and inverse dynamics approach. The best models were obtained: (i) with Lasso linear model of the leave-one-out cross-validation in open turn (MSE = 0.011; R2 = 0.825) and in the somersault turn (MSE = 0.016; R2 = 0.734); (ii) the Ridge of the leave-one-out cross-validation (MSE = 0.016; R2 = 0.763) for the bucket turn; and (iii) the AdaBoost tree-based model of the leave-one-out cross-validation for the crossover turn (MSE = 0.016; R2 = 0.644). Model's selected features revealed that optimum turning performance was very similarly determined for the different techniques, with balanced contributions between turn-in and turn-out variables. As a result, the relevant feature's contribution of each backstroke-to-breaststroke turning technique are specific; developing approaching speed in conjunction with proper gliding posture and pull-out strategy will result in improved turning performance, and may influence differently the development of specific training intervention programmes.
Asunto(s)
Rendimiento Atlético , Humanos , Niño , Natación , Fenómenos Biomecánicos , Modelos Lineales , HidrodinámicaRESUMEN
A new concept of human-machine interface to control hand prostheses based on displacements of multiple magnets implanted in the limb residual muscles, the myokinetic control interface, has been recently proposed. In previous works, magnets localization has been achieved following an optimization procedure to find an approximate solution to an analytical model. To simplify and speed up the localization problem, here we employ machine learning models, namely linear and radial basis functions artificial neural networks, which can translate measured magnetic information to desired commands for active prosthetic devices. They were developed offline and then implemented on field-programmable gate arrays using customized floating-point operators. We optimized computational precision, execution time, hardware, and energy consumption, as they are essential features in the context of wearable devices. When used to track a single magnet in a mockup of the human forearm, the proposed data-driven strategy achieved a tracking accuracy of 720 µm 95% of the time and latency of 12.07 µs. The proposed system architecture is expected to be more power-efficient compared to previous solutions. The outcomes of this work encourage further research on improving the devised methods to deal with multiple magnets simultaneously.
Asunto(s)
Magnetismo , Redes Neurales de la Computación , Mano , Humanos , Fenómenos Magnéticos , ImanesRESUMEN
We aimed to compare multilayer perceptron (MLP) neural networks, radial basis function neural networks (RBF) and linear models (LM) accuracy to predict the centre of mass (CM) horizontal speed at low-moderate, heavy and severe swimming intensities using physiological and biomechanical dataset. Ten trained male swimmers completed a 7 × 200 m front crawl protocol (0.05 m.s-1 increments and 30 s intervals) to assess expiratory gases and blood lactate concentrations. Two surface and four underwater cameras recorded independent images subsequently processed focusing a three-dimensional reconstruction of two upper limb cycles at 25 and 175 m laps. Eight physiological and 13 biomechanical variables were inputted to predict CM horizontal speed. MLP, RBF and LM were implemented with the Levenberg-Marquardt algorithm (feed forward with a six-neuron hidden layer), orthogonal least squares algorithm and decomposition of matrices. MLP revealed higher prediction error than LM at low-moderate intensity (2.43 ± 1.44 and 1.67 ± 0.60%), MLP and RBF depicted lower mean absolute percentage errors than LM at heavy intensity (2.45 ± 1.61, 1.82 ± 0.92 and 3.72 ± 1.67%) and RBF neural networks registered lower errors than MLP and LM at severe intensity (2.78 ± 0.96, 3.89 ± 1.78 and 4.47 ± 2.36%). Artificial neural networks are suitable for speed model-fit at heavy and severe swimming intensities when considering physiological and biomechanical background.