Abstract:
Physical Activity has been associated with fatal diseases such as cancer and heart dis-
ease. Therefore, prediction of physical activity intensity levels is of paramount impor-
tance. This study utilizes raw accelerometer data of 28 healthy individuals of ages 18-79
years to predict physical activity intensity levels. Parametric and non-parametric ap-
proaches were used to develop predictive models from the accelerometer data. As a
parametric approach the proportional odds cumulative logit model was considered both
in the frequentist and Bayesian paradigm. Tree-based methods: bagging, random forest,
boosting and Generalized Unbiased Interaction Detection and Estimation (GUIDE) were
considered as non-parametric methods. All the models were built from four accelerom-
eter placements, i.e., left wrist, right wrist, right hip and right ankle. In terms of the
placement, right ankle has signi cantly highest prediction accuracy and right wrist has
the lowest prediction accuracy. The parametric and non-parametric models have mixed
performance with no particular method outperforming the others. However, considering
overall prediction accuracy, the proportional odds cumulative logit model has the highest
prediction accuracy.
ii
