|
Abstract:
|
Since the 1980's accelerometers have been used to predict activity intensity in physiology research. The accuracy of conventional accelerometer analysis methods is suboptimal, but newer methods that use raw data from the accelerometer for prediction have been developed. Ensemble methods, parametric linear and nonlinear models, and neural networks, among others, are types of models that have been used for this purpose. As responses are correlated sequentially, time-series methods can be implemented for prediction. As prior responses are not available at testing and in practice, we consider lag autocorrelation and recursive prediction, using prior predictions in models trained on prior observations. This is done with ensemble methods, particularly with methods that take into account the ordinality of responses (sedentary, light, moderate and vigorous). The time series nature and the ordinality of the responses are crucial aspects that have only been touched on in the literature and we show better predictive accuracy than existing methods by incorporating these considerations in our predictive models.
|
