Computational Modeling to Predict and Affect Recovery After Stroke

There is currently no direct association between patient movement kinematics and rehabilitation outcomes for stroke survivors. Research has demonstrated minor successes in predicting some clinical measures. Recent advancements in machine learning algorithms and robotic technology enable us to parse through a large amount of patient movement data to find appropriate targets for treatment. This thesis examines the predictability of chronic stroke survivor clinical outcomes following neurological injury using patient movement and demographic data, and whether predictive movement metrics can be influenced to improve recovery. Building on a recent two-week study with stroke survivors, we construct a machine learning model to predict clinical changes, and use the model to determine an appropriate intervention target. We then establish how this target should be altered in a crossover trial with chronic stroke survivors. Lastly, we test a two-week treatment in a chronic stroke survivor case study. Our predictive models determined that movement speed was a proper focus for an intervention. In our crossover trial, we showed that negative viscosity had the strongest effect on movement speed. Treating a patient with negative viscosity for two weeks showed significant improvements in multiple movement metrics, and small gains in clinical outcomes. These results show that designing a rehabilitation intervention based on negative viscosity can lead to improvements in chronic stroke survivors. Our success also establishes the framework we used for this patient population as a potentially generalizable pipeline to identify and test interventions for other populations. Further research could focus on testing treatments for some of the other model predictors, as they could present more potential intervention targets.