ML and optimisation techniques for closed-loop neuromodulation of multiple nerve models

: (Alternative Format Thesis)

Abstract

Neural interfaces that stimulate the peripheral nervous system are already being used to treat a range of conditions and injuries, such as epilepsy. However, these interfaces typically operate in an open-loop manner, meaning stimulation parameters can only be changed manually by a physician. Off-target modulation and the nerve/electrode interface changing over time might also reduce stimulation efficacy and produce unwanted side-effects. There is therefore significant research interest in developing closed-loop neural interfaces which incorporate sensing and decoding as well as electrical stimulation to automatically inform stimulation parameters. Electrical recordings have been identified as containing important information about nerve activity that could be used to close the loop in neural interfaces, but in the case of less-invasive peripheral nerve interfaces, these recordings are often multi-dimensional, complex to process, and contain time-varying noise.

Bearing this in mind, this thesis focuses on techniques for processing and decoding recordings from peripheral nerve interfaces with multiple recording sites along the nerve length, particularly using multi-electrode cuffs. On the data collection side, an ex-vivo system was developed and compared with in-vivo experiments, with the overall aim of maximising nerve data collection for data-driven approaches and minimising animal numbers used for experiments. On the processing and decoding side, ML-based techniques were reviewed and implemented for adaptive denoising of peripheral nerve data and efficient fibre area detection from histology imaging. The distribution of conduction velocities in a whole nerve was identified as a valuable set of features to investigate, since detailed information about specific fibre types can be obtained. The N-CAP technique was therefore devised to estimate these distributions in minimally-invasive way and thoroughly validated with simulated and experimental data. In summary, this thesis makes a number of contributions to key challenges in developing closed-loop interfaces that leverage ML and optimisation techniques.

Date of Award	26 Mar 2025
Original language	English
Awarding Institution	University of Bath
Supervisor	Benjamin Metcalfe (Supervisor), Michael Proulx (Supervisor), Christof Lutteroth (Supervisor) & Paulo Rocha (Supervisor)