Development of a novel method for the classification of osteoarthritic and normal knee function
Advances in our understanding of human locomotion can be futile if no practical use is made of them. For the long-term benefit of patients in a clinical setting, scientists and engineers need to forge stronger links with orthopaedic surgeons to make the most use of the recent developments in motion analysis technology. With this requirement as a driving-force, an objective classification tool was developed that uses motion analysis for an application to clinical diagnostics and monitoring, namely knee osteoarthritis (OA) progression and total knee replacement (TKR) recovery. The classification tool is based around the Dempster-Shafer (DS) theory, and as such is built upon the sound foundations of Bayesian statistics. The tool expands on the work of Safranek et al. (1990) and Gerig et al. (2000) who developed and used parts of the classification method in the areas of vision and medical image analysis respectively. Using the data collected during a clinical knee trial, this novel approach enables the objective classification of subjects into an OA or normal group. Each piece of data is transformed into a set of belief values: a level of belief that a subject has OA knee function, a level of belief that a subject has NL knee function and an associated level of uncertainty. The belief values are then represented on a simplex plot, which enables the final classification of a subject, and the level of benefit achieved by TKR surgery to be visualised. The DS method can be used as a fully or partially automated tool. The input variables and control parameters, which are an intrinsic part of the tool, can be chosen by an expert or an optimisation approach. Using a leave-one-out (LOO) approach, the tool was able to classify new subjects with an accuracy of 97.62%. This compares with the 63.89% and 95.24% LOO accuracies of two well-established methods---the Artificial Neural Network and the Linear Discriminant Analysis classifiers respectively. The tool also provides an objective indication of the variables that are the most influential in distinguishing OA and NL knee function. In this case, the variables identified by the tool as important are often cited as clinically relevant variables, which enhances the appeal of the tool to the clinical community and allows for more effective comparison with clinical approaches to diagnosis. Using Simulated Annealing to select the control parameters reduced the LOO accuracy to 95.24%. Automated feature selection using a Genetic Algorithm and Sequential Forward Selection increased the LOO accuracy to 100%. However, further work is required to improve the effect of this process on the overall level of uncertainty in the classification. Initial studies have demonstrated a practical and visual approach that can discriminate between the characteristics of NL and OA knee function with a high level of accuracy. Further development will enable the tool to assist orthopaedic surgeons and therapists in making clinical decisions, and thus promote increased confidence in a patient's medical care.