Breaking the Routine: Understanding the Interplay of Control Synergies and Reinforcement Learning to Improve Motor Training and Recovery in Neurologic Conditions
- Author(s): Senesh, Merav Rachel
- Advisor(s): Reinkensmeyer, David J
- et al.
Neurological conditions, such as stroke and Parkinson’s Disease are leading causes of chronic disability mainly due to the motor impairment. The current common solution for reducing motor impairment is to apply intensive movement practice, which is conducted by rehabilitation therapists. To aid patients and therapists in this goal, recent research efforts are focused on designing robotic devices and wearable sensors that help therapists promote optimal forms of practice without directly supervising training. However, it has been hypothesized that recovery may be limited when using these technologies because they encourage practice of ‘incorrect’ movements. Therefore, it is needed to further understand and characterize motor learning of ‘correct’ movements, in order to design beneficial mechanical devices that promote the recovery of patients. This dissertation provides insight into the optimal conditions for machine-assisted movement motor training by first studying the kinematics of movement recovery patterns after stroke to identify what are the ‘correct’ movements needed to progress in recovery, and second, by studying sensor-based feedback for promoting the practice of the desired movements. We first revisited a well-known but controversial model of early movement recovery after stroke – the proportional recovery (PR) model. We utilized a mathematical and behavioral approach to examine the PR model and explain why a portion of stroke survivors do not follow the proposed proportional recovery model. Our results showed that individuals “stuck” in abnormal control synergies (the “flexion” and “extension” arm synergies, defined by certain arm kinematics) recovered less than predicted by the proportional recovery model. Moreover, individuals who shifted from using abnormal control synergies to individuated joint control showed greater responsiveness to robot-assisted movement training. Next, we analyzed clinical assessments of 319 persons’ abilities to perform “out-of-synergy” and “in-synergy” arm movements after chronic stroke using the Upper Extremity Fugl- Meyer (UEFM) scale. Our result showed that for some individuals with moderate impairment, rudimentary dexterity corresponded with reduced ability to move the arm in-synergy, i.e., there is a competition between the two types of movement. This result suggests that at least some individuals with a moderate impairment level may have not yet made the “switch” from practicing the in-synergy movement to more high-level control movements. In such cases, it would seem logical that rehabilitative movement training should focus on promoting those movements. Based on these finding we then analyzed the effectiveness of three different rehabilitation training types on “breaking out” of synergy by introducing a new analysis of the UEFM assessment. We found that diverse (88 different exercises) arm/hand training was better at promoting the transition and was characterized with an early “breakpoint” on the impairment scale. We then asked how can we design an effective sensor-based feedback strategy that promotes a desired switch between movement patterns? We elected to investigate the use of reinforcement feedback, a motor training strategy that, while straightforward to implement with technology and capable of improving learning retention, has been neglected in rehabilitation technology. We did so by first studying the proposed paradigm in the context of shaping dance performance. We chose dancers because they are movement experts, i.e., individuals who work intensively for many years in order to learn a variety of complex movement abilities. Therefore, we hypothesized that they would have high levels of control flexibility, allowing them to move away from their established movement patterns toward novel patterns, and that, by studying this flexibility, we could better understand how to promote such flexibility. Surprisingly, the dancers found the proposed task to be difficult, and we found evidence of learning only in 14% of the motor practice sessions. The successful sessions were characterized by relatively low initial success rate, significantly more wrist than ankle movements in the initial period, and intermediate durations between successful movements. These findings provided guidance for how to make movement training with reinforcement feedback more effective. Specifically, to better shape movements, we next developed an adaptive reinforcement feedback algorithm based on the concept of “hints”, which we implemented by initially presenting easier versions of the goal task, and then rewarding actions that progressively moved closer to the goal task. We tested this approach by comparing it with a conventional and fixed feedback approach with a group of college engineering students who played a novel computer game we developed based on the idea of dancing by moving a computer mouse. The adaptive algorithm caused a significantly higher initial success rate (as designed) but also caused a significantly higher final success rate therefore demonstrating a superior learning process. Moreover, there was a significant decrease in the reported frustration level for some objectives with adaptive reinforcement, although this effect was smaller than expected as frustration was more strongly associated with average success than initial success. Success in finding the target movement could not be predicted by initial success alone for the fixed feedback in this case. Surprisingly we also found that repeated exposure to reinforcement training promoted free movement diversity; a finding with implications related to the beneficial effect of diverse training with stroke patients mentioned before. The results of this dissertation set the ground for development of a novel class of movement training technologies for individuals with neurological conditions based on adaptive reinforcement feedback. For training of the arm after stroke, we envision an implementation using a wearable sensor and an “arm dance” paradigm in which the patient trains by moving the arm to music while receiving adaptive reinforcement feedback about movement individuation, speed, or smoothness. We believe that in this way we can better promote diverse but ‘correct’ movements, compared to existing robotic and wearable sensor training approaches. Another possible application is in the context of physical therapy for individuals with Parkinson’s Disease, where the proposed training paradigm could be used to promote learning to make faster and bigger movements.