Feature Selection Optimisation in an Automated Diagnostic Cancellation Task

This paper describes an investigation into feature selection and classification in the automation of a standard target cancellation task for the diagnosis of visuo-spatial neglect. Alongside a conventional assessment based on the number of targets cancelled, a series of time-based dynamic features have been algorithmically defined which can be extracted by capturing the test subject's response on a graphics tablet connected to a computer. We identify the diagnostic capabilities of the individual features and show that dynamic data contains important indicators for neglect detection. Furthermore, employing standard pattern recognition techniques, we establish the optimum feature vector size and classifier for a multi-feature analysis of a test attempt and show that an improvement in diagnostic error rate is achievable over any single individual feature.