Artificially Intelligent Closed-Loop Neurostimulators: Trade-Offs between Local and Remote Computing

The current state of the art in electronic neuromodulation is closed-loop devices that employ machine learning inference to perform automated diagnostics and treatment of neurological disorders. This paper discusses general trade-offs between several choices of how such inference is implemented - locally, remotely or in a hybrid fashion - mainly based on the energy cost considerations. Several topologies for the loop are explored, including computing within the implant, as well as computing performed on a wearable, handheld or stationary processor. Any computing external to the body necessitates a wireless communication link. Common wireless communication requirements are thus also included in the consideration. As two opposite case studies, we summarize two of our recent design approaches where one performs machine learning inference in an implant and the other - on a wearable computer, enabling artificially intelligent closed-loop neurostimulation in the central and peripheral nervous systems, respectively.