Peripheral Nerve Recording and Stimulation
Direct communication with the peripheral nervous system has been investigated through the use of implanted intraneural electrodes and nerve cuff approaches. Studies in amputee subjects using longitudinal intrafascicular electrodes (LIFEs) (Rossini et al. 2010), transverse intrafascicular multichannel electrodes (TIMEs) (Raspopovic et al. 2014), the Utah Slant Array (which penetrates the nerve fiber bundles), and the flat interface nerve electrode (FINE) have all shown the ability to use sensory input feedback with simultaneous motor control in amputee subjects. Research is ongoing to demonstrate long-term viability of the implants, but some have been in place as long as 2 years. These investigations are exciting advances that will likely lead to significant changes in the approach to sensory motor restoration in the future through the ability to directly tap into neural control and feedback signals.
Advanced Control Paradigms Pattern Recognition
Pattern recognition, an approach to classifying a user’s intended motions based on learned patterns of myoelectric activity, has seen recent commercial availability. However, there is significant ongoing work under way to improve the interpretation of myoelectric patterns from users (Castellini et al. 2014; Micera, Carpaneto, and Raspopovic 2010). Recent work includes approaches to the simultaneous pattern recognition control of multiple functions and to increase the robustness of pattern recognition to the rigors of daily life—for example, decreasing the sensitivity of classifiers to the position of the residual limb, to other simultaneous bodily activities, or to ongoing fatigue (Hargrove, Lock, and Simon 2013; Hargrove, Simon et al. 2013; Scheme and Englehart 2011; Scheme, Hudgins, and Englehart 2013). Another active area of ongoing research is supervised adaptation (by way of prespecified or intermittent retraining of a pattern recognition system) and unsupervised adaptation (automatic retraining or updating of a pattern recognition system, without the need for specific training periods) to allow a device to modify its operation to new users or new situations (Sensinger, Lock, and Kuiken 2009; Tommassi 2013). Continual, real-time adaptation of pattern recognition is considered to be a major area of clinical interest (Scheme and Englehart 2011), as are ways to better structure the training of pattern recognition systems. For the interested reader, Castellini et al. (2014) provided a comprehensive review of ways in which pattern recognition is being enhanced to better leverage sEMG for more precise and user-friendly pattern recognition.
Machine Learning, Intelligent Systems, and Shared Control Pattern recognition represents one form of autonomy and machine learning on the part of a robotic prosthesis (Oskoei and Hu 2007). The prosthetic control system is observing complex patterns from the user and making moment-by-moment decisions regarding which of the many functions on the device the user will control. In this case, the control system’s choices are based on learned predictions of a user’s motor intent. Pattern recognition and other forms of autonomy have been demonstrated to be desirable for the users of robotic prostheses (Castellini et al. 2014). One simple example of autonomy now deployed in commercial systems is slip detection for grasping, such that a system will hold on to an object even when a user is not attending to his or her grasp. These examples suggest how even modest intelligence on the part of an assistive technology can help support the user of that technology. More advanced examples include research into ways that a prosthetic hand could automatically preshape its grip to accommodate specific objects in an environment (as reviewed by Castellini et al. 2014) or how a robotic system may in fact build up knowledge about the user in the form of predictions and control policies to better inform the simultaneous control of multiple movements or functions (Edwards et al. 2016; Pilarski, Dawson et al. 2013; Pilarski, Dick, and Sutton 2013; Pulliam, Lambrecht, and Kirsch 2011).
Taken as a whole, there is a growing body of evidence to suggest that intelligence and agency on the part of a robotic prosthesis will extend the potential abilities of a prosthetic user (e.g., the idea of a prosthesis-human partnership developing shared communicative capital through ongoing interactions, as proposed by Pilarski, Sutton, and Mathewson 2015). More specifically, increased and more general machine intelligence in the control systems of robotic prostheses is expected to greatly increase the robustness, adaptability, and situational awareness of current systems to better meet the needs of users (Castellini et al. 2014). This hypothesis remains to be rigorously proved or disproved.