Design Methodology for Energy-constrained AI Edge Inference in Implantable Medical Devices.

This paper provides design guidelines for enabling edge-computing machine learning blocks in implantable medical devices. Energy consumption is a critical design factor for such implants, as the dissipated heat can damage the surrounding tissue and battery size is limited. For wirelessly powered devices, the power budget is also often limited by safety guidelines for specific absorption rates imposed on the radiated electromagnetic field. The paper examines several case studies of closed-loop neural implantable systems and analyzes the design choices for placing inference computation blocks along the implantwearable-handheld-cloud chain. The case studies cover examples of our recent work for both the central nervous system (CNS) and peripheral nervous systems (PNS). Depending on the data and the algorithm's complexity, the examples use either in-implant or remote inference, or a hybrid approach that combines both. For each example, the paper shows trade-offs between local and remote computation along the signal path.