The effects of curve registration on linear models of jump performance and classification based on vertical ground reaction forces.

Functional principal components define modes of variation in time series, which represent characteristic movement patterns in biomechanical data. Their usefulness however depends on the prior choices made in data processing. Recent research showed that better curve alignment achieved with registration (dynamic time warping) reduces errors in linear models predicting jump height. However, the efficacy of registration in different preprocessing combinations, including time normalisation, padding and feature extraction, is largely unknown. A more comprehensive analysis is needed, given the potential value of registration to machine learning in biomechanics. We evaluated popular preprocessing methods combined with registration, creating 512 models based on ground reaction force data from 385 countermovement jumps. The models either predicted jump height or classified jumps into those performed with or without arm swing. Our results show that the classification models benefited from registration in various forms, particularly when landmarks were placed at critical points. The best classifier achieved a 5.5 percentage point improvement over the equivalent unregistered model. However, registration was detrimental to the jump height models, although this performance variable may be a special case given its direct relationship with impulse. Our meta-models revealed the relative contributions made by various preprocessing operations, highlighting that registration does not generalise so well to new data. Nonetheless, our analysis shows the potential for registration in further biomechanical applications, particularly in classification, when combined with the other appropriate preprocessing operations.