| dc.description.abstract | Accelerometers are widely used in industry, engineering, and consumer electronics; however, they have been increasingly utilized in biology and healthcare research. In this thesis, we explore and develop the use of several machine and statistical learning models for the prediction of activity in accelerometry data. Accelerometry datasets are often collected by mailing accelerometers to participants, who then wear the accelerometers for a period of time to collect data on their activity. The purpose of this study was to develop models to classify a given day in an accelerometry dataset as either activity from the delivery process or actual human wear. These models can then be used to automate the cleaning of accelerometer datasets adulterated with delivery data that needs to be removed before the analysis. In Chapter 1, this thesis provides an introduction to several popular statistical and machine learning models: random forests, logistic regression, and neural networks. These models are applied to a benchmark accelerometer dataset in Chapter 2, and then used to develop our mail delivery classification models on an accelerometry dataset in Chapter 3. We found that a hybrid convolutional recurrent neural network performs best in the classification task but simpler models such as logistic regression and random forest also have excellent performance. In practice, the user can weigh the large computational cost and greater performance of a convolutional recurrent neural network against the much faster but slightly less powerful random forest or logistic regression in choosing which model they prefer. | |
