Innovative solutions for control and sensory feedback restoration in closed-loop upper-limb prostheses

The hand is one of the most crucial body parts that enables interactions with the outside world. Since it is the most distal portion of the upper limbs, it has a highly intricated anatomy that is composed of several bones, joints, ligaments, muscles, and tendons. Additionally, it features an intricated blood vessel network and delicate innervation. Several tasks are made possible by this intricate structure, including the ability to hold items, communicate, and serve as the primary tactile sense organ. Hand loss, or more generally the upper limb loss, can affect the level of autonomy and the capability of performing the activities of daily life (ADL) for amputees, with a significant reduction in patients’ quality of life. Over the years, prosthetic systems have undergone a considerable evolution, both from an aesthetic and a technological/control point of view. Since they most closely mimic the functionality of a normal hand, myoelectric prostheses are considered the gold standard among active upper limb prostheses. This type of prosthesis exploits the information contained in the electromyographic (EMG) signal to actuate the prosthetic joints. Nevertheless, despite these advancements, modern prosthetic systems continue to have several shortcomings that prevent them from completely restoring a human limb functionality. Although it is feasible to grab items of various sizes and shapes, delicate manipulation is still not possible, even using poly-articulated hands. Indeed, current commercial myoelectric prostheses despite the use of advanced control strategies such as those based on pattern recognition, are based on the steady-state EMG signal, which allows robust control, but which can only occur after the time needed to reach steady state. A major advancement in the field of upper-limb prosthetics has been reached with the surgical procedure, namely Targeted Muscle Reinnervation (TMR). The idea is that by reinnervating the residual nerves of the amputated limb to new target muscles, users may be able to control the prosthesis more intuitively. This procedure is also an emerging technique for the treatment and reduction of the phantom limb pain (PLP) and neuroma pain, and for the targeted sensory reinnervation (TSR) of bidirectional neuroprosthetic devices. In fact, a prosthesis that mimics the human limb as closely as possible should not only have robust and intuitive control, but also a reliable sensory feedback system to close the control loop. However, the current limb prostheses do not provide any sensory feedback to the user. Several approaches have been proposed in literature to restore tactile feedback in amputees such as the use of invasive interface with Peripheral Nervous System (PNS). Another solution is the TSR, which allows the clinician to reinnervate the nerves of the afferent system in specific target areas, with the intention of mapping in those sites the same physiological channels that were lost together with the amputated limb. However, such approaches require a surgery and present disadvantages related to the weak long-term stability and postoperative complications. An alternative solution, which does not require a surgical procedure, is the Transcutaneous Electrical Nerve Stimulation (TENS), which was able to evoke somatotopic sensation non-invasively using only electrical stimuli provided to the nerves by means of superficial electrodes. This thesis aims to propose solutions to create a closed-loop prosthesis. These solutions aim to i) improve the performance of current efferent controls by proposing a new time-domain features for classification gestures, with the dual purpose of decreasing the time needed to recognize movement intention and improving stability; and to ii) making chronic use of stimulation possible (invasive or non-invasive depending on the user's characteristics/preferences), by proposing a modular neural stimulator prototype, that can be used to perform both PNS and TENS. Regarding the efferent part, this thesis proposes the use of the muscular activation sequences, calculated in the EMG transient phase, as time-domain features in the classification of hand gestures. Specifically, a features extraction process was developed to encode the training dataset. The classification performances were computed offline on 10 able-bodied persons and on-line on two persons with a transradial (TR) amputation. Then, a comparison among different supervised machine learning techniques is carried out to choose the most suitable classification algorithm for prostheses control. This intends to overcome limitations of the classical features extracted in the transient phase and to propose a valid alternative to traditionally adopted pattern recognition techniques that predominantly rely on the steady state of sEMG signal. For the afferent part, this thesis addresses the specific issue of improving sensory feedback in upper limb prosthesis, by designing a new modular stimulation prototype that can be used both in an implantable scenario (PNS) and in a non-invasive scenario (TENS). To this purpose, firstly an analysis of the existing stimulators was carried out for defining the basic design specifications and considering the possibility of integrating the developed system within an existing prosthetic system. Then, these specifications were used for the design of the modular stimulating system (hardware-firmware) for the restoration of sensory feedback. The system, in both its variants, has been bench tested to validate the waveforms generated. Moreover, for the wearable version, further tests were conducted with able bodied subjects, that reported being able to feel some sensations.