Classification of Error Types in Physiotherapy Exercises

Abstract

Systems of healthcare and the field of physical therapy in particular are likely to come under increased stress in the decades to come as populations age faster than their workforces can handle. Much of healthcare is dependent on skilled human effort. Intelligent systems could be developed to support the efforts of healthcare specialists and enable patients and users to continue effective treatment with less human supervision, helping healthcare systems to adapt to the change in demographics and their associated rising costs. This thesis aims to work towards that goal by investigating methods of detecting mistakes made during physiotherapy exercises, which would be a strictly necessary component in any intelligent system that is able to give feedback to a user about her performance. Special consideration is given to the requirements of systems which must function in a home-environment without specialist supervision. For the purposes of this thesis a dataset is built by recording a set of exercises with the help from volunteer test subjects. The dataset is then used to develop a model for classifying exercises and mistakes made during the exercises based on discrete-time quaternion sequences, after which the model is tested using several validation schemes. The model manages to successfully predict exercises across test subjects and error types with high rates of accuracy given the validation schemes used. However, when classifying error types across subjects, the accuracy proves unsatisfactory. Error types are classified with high rates of accuracy when using template exercises for a given test subject. These results highlight important limitations of the proposed model.