UC Berkeley

Wearable biosensing devices are becoming more and more ubiquitous, and new medical devices are being used to relieve previously untreatable diseases. Smart biomedical interfaces such as these depend on some form of device intelligence, enabling local processing of recorded biosignals to perform tasks such as hand gesture recognition or detection of neurological disease states. The devices and the algorithms they employ, however, must be robust to a variety of interferers and confounding factors that often prevent their practical use.

This thesis descries methods to improve robustness in implantables and wearables through a variety of means, demonstrating them in two different applications. The first part presents the Wireless Artifact-free Neuromodulation Device, or WAND, which enables closed-loop neuromodulation by cancelling self-interference artifacts that normally prevent simultaneous neural recording and stimulation. The second part then describes the FlexEMG system for hand gesture recognition using electromyography, which employs a hyperdimensional computing algorithm capable of incrementally learning hand gestures in a variety of situational contexts such as changing limb positions. These systems have been deployed in realistic animal and human subject experiments that better approximate real-world use to demonstrate the hardware and algorithmic improvements towards robustness.