THE IMPACT OF CORRELATIONS IN CONSOLIDATING STANDARDIZED ROBOTIC TASKS INTO A GLOBAL PERFORMANCE SCORE

Abstract

Neurological diseases and disorders that impact the brain contribute a considerable amount to global burden of disease. Informative assessment is needed to guide treatment and identify novel interventions. However, clinically used neurological scales are coarse and commonly have floor and ceiling effects, impacting their suitability for clinical trials. Kinematic-based measures using technologies such as interactive robots possess the fine resolution to detect improvements in motor function and may prove useful to quantify the potential benefits of novel therapeutic interventions. Our lab has developed over a dozen robot based behavioral tasks to quantify sensory, motor, and cognitive function associated with the arm. Detailed spatiotemporal measures are combined to provide a single measure of performance in the task, called a Task Score. The challenge is to combine these Task Scores into a single measure suitable for use in clinical trials. The present thesis explores how Tasks Scores can be combined to create a single measure of overall performance, called the Kinarm Score. While a simple summation of Tasks Scores could be used, the presence of correlations across behaviours may impact the ability to identify impairments. Notably, the use of actual healthy participant data to compute these correlations would require all participants to complete all robot-based tasks whenever a new task is developed. Thus, the addition of one new task requires hundreds of individuals to be assessed. First, simulations were used to quantify how correlations across Tasks Scores would impact a global Kinarm score. We found that with a moderate correlation of 0.6, the number of individuals classified as impaired was 6.86% (± 0.55%), rather than the expected level of 5%. Importantly, each additional task increased this error such that 9 tasks increased the percentage of impaired individuals to 15.5% (± 1.02%), more than 3 times the expected level. Next, techniques were developed to incorporate correlations into a simulated dataset. The inclusion of inter-task correlations improved model performance by identifying only 6.7% of individuals as impaired in contrast to the 10% if correlations were not considered in the simulated dataset (5% is expected).