| Title: |
Low-dimensional representations of neural time-series data with applications to peripheral nerve decoding |
| Authors: |
Lubba, Carl Henning Thore |
| Contributors: |
Schultz, Simon; Jones, Nicholas; Galvani Bioelectronics |
| Publisher Information: |
Bioengineering, Imperial College London |
| Publication Year: |
2020 |
| Collection: |
Imperial College London: Spiral |
| Description: |
Bioelectronic medicines, implanted devices that influence physiological states by peripheral neuromodulation, have promise as a new way of treating diverse conditions from rheumatism to diabetes. We here explore ways of creating nerve-based feedback for the implanted systems to act in a dynamically adapting closed loop. In a first empirical component, we carried out decoding studies on in vivo recordings of cat and rat bladder afferents. In a low-resolution data-set, we selected informative frequency bands of the neural activity using information theory to then relate to bladder pressure. In a second high-resolution dataset, we analysed the population code for bladder pressure, again using information theory, and proposed an informed decoding approach that promises enhanced robustness and automatic re-calibration by creating a low-dimensional population vector. Coming from a different direction of more general time-series analysis, we embedded a set of peripheral nerve recordings in a space of main firing characteristics by dimensionality reduction in a high-dimensional feature-space and automatically proposed single efficiently implementable estimators for each identified characteristic. For bioelectronic medicines, this feature-based pre-processing method enables an online signal characterisation of low-resolution data where spike sorting is impossible but simple power-measures discard informative structure. Analyses were based on surrogate data from a self-developed and flexibly adaptable computer model that we made publicly available. The wider utility of two feature-based analysis methods developed in this work was demonstrated on a variety of datasets from across science and industry. (1) Our feature-based generation of interpretable low-dimensional embeddings for unknown time-series datasets answers a need for simplifying and harvesting the growing body of sequential data that characterises modern science. (2) We propose an additional, supervised pipeline to tailor feature subsets to collections of ... |
| Document Type: |
doctoral or postdoctoral thesis |
| Language: |
unknown |
| Relation: |
alma; http://hdl.handle.net/10044/1/80541; https://doi.org/10.25560/80541; EP/L016737/1 |
| DOI: |
10.25560/80541 |
| Availability: |
http://hdl.handle.net/10044/1/80541; https://doi.org/10.25560/80541 |
| Rights: |
Creative Commons Attribution Licence |
| Accession Number: |
edsbas.3C403ACE |
| Database: |
BASE |