Table 1.
Summary of abbreviations.
Fig 1.
Illustration showing the anatomical position of the inertial sensors.
(A) Lower back (LB) sensor attached with double-side tape and held in place with an inelastic adhesive bandage wrapped around the waist; (B) Upper back (UB) sensor attached only with double-sided tape; (C) Left shank (LS) sensor also attached with the same tape and held firmly in place by an adhesive bandage wrapped around the leg. The right shank (RS) sensor (not shown) was attached in the same way.
Table 2.
Data processing parameters with their respective ranges.
Fig 2.
Schematic design of the nested cross validation and optimisation procedure.
(A) The training/validation set is partitioned 10-fold for the outer loop, and then (B) each training outer set is re-partitioned 2-fold for the inner loop. (C) Optimisation works with the observations from the inner loop to determine an ensemble model based on the series of optimal parameters determined by Particle Swarm Optimisation. (D) The ensemble model is then evaluated on the outer training set. (E) The process repeats for each outer fold, adding to the series of optimal parameters used to determine the outer ensemble model. (F) This yields parameter distribution and partial plots. It also produces the final ensemble model that may be applied to the holdout data.
Table 3.
Peak power (W·kg-1) computed from VGRF data.
Table 4.
ANOVA Type I effects for the optimised models’ outer validation RMSE.
Fig 3.
Outer validation RMSE distribution by level for each condition.
Top row: single factors in the GLM, namely (A) Jump Type; (B) Sensor location; (C) Signal representation; (D) Model type. Bottom row: two factors (E) Model type and sensor location for the CMJNA using the resultant signal representation. Horizontal arrows indicate significant differences, where * p < 0.05, ** p < 0.01, *** p < 0.001.
Fig 4.
Data parameter distributions across the intermediate models for the LB-CMJNA data set with ensemble optimal values highlighted.
(A) LR model type; (B) SVM model type; (C) GPR model type. Optimal values are shown by the darker shaded bar for categorical parameters and by a darker vertical line at the peak position for numeric parameters with that optimal value shown.
Fig 5.
Model parameter distributions across the intermediate models for the LB-CMJNA data set with ensemble optimal values highlighted.
(A) LR model type; (B) SVM model type; (C) GPR model type. Optimal values are shown by the darker shaded bar for categorical parameters and by a darker vertical line at the peak position for numeric parameters with that optimal value shown.
Fig 6.
Aggregated surrogate model partial plots for the data parameters from the LB-CMJNA data set showing the predicted AM loss at the global minimum.
(A) LR model type; (B) SVM model type; (C) GPR model type. The central blue line is the central SM estimate. The darker shaded area about this line is the SM fitted noise level. The lighter shaded area covers the standard deviation. Note that for SVM (middle column), the SM range (y-axis) is higher.
Fig 7.
Aggregated surrogate model partial plots for the model parameters from the LB-CMJNA data set showing the predicted AM loss at the global minimum.
(A) LR model type; (B) SVM model type; (C) GPR model type. The central blue line is the central SM estimate. The darker shaded area about this line is the SM fitted noise level. The lighter shaded area covers the standard deviation. Note that for SVM (middle column), the SM range (y-axis) is higher.
Table 5.
Predictive error estimates over progressive optimisations for each model type using the resultant LB sensor for the CMJNA, based on nested cross validation and the independent holdout test.
Table 6.
Ensemble optimal parameters over successive optimisations (1st, 2nd, 3rd, 4th) for each model type using the resultant LB sensor for the CMJNA.