Abstract for HONS 02/21
A Machine Learning Approach to Modelling and Predicting Biomechanical Measures of Strain Across Bone
James Houghton
Department of Computer Science and Software Engineering
University of Canterbury
Abstract
Accurately assessing bone fracture healing requires years of specialty training assessing radiographs which thereafter can still result in an incorrect diagnosis. Providing an ability to track and predict bone strength over time would allow for physicians to better assess patient recovery time. Our overall problem can be broken down into three sub-problems: determining what activity is occurring at a given time, analysing changes to that activity in subsequent occurrences, and correlating those changes with a biomechanical model. This research project presents activity recognition methods to address sub-problem one and identifies the time series extrinsic regression model to address sub-problem two. A time series forest model is compared with a convolutional neural network for activity recognition, with the time series forest model achieving an average leave one out cross validation accuracy of 85.7% on an interpolated dataset.
