Microelectronics for Next Generation Neural Interfaces

[EPSRC RESEARCH AREAS]

Brain Machine Interfaces (BMIs) can be defined as a communication pathway that allows information flow between the nervous system and electronics. As such, BMIs introduce the possibility to repairing sensory and motor functions, gathering data on how the neutral system influences the body,and other [ASSISTIVE TECHNOLOGY] applications. A major problem with modern BMIs is that they consist of bulky computers physically attached to patients via wires traversing the skull and the brains protective membranes (meninges: arachnoid, dura, pia). This makes the free movement difficult, and introduces infection risks due to the transmeningeal connection. To be chronically safe and compatible with free movement, BMIs should be wireless.
There are two electrical signals in the brain: extracellular action potentials (or 'spikes') and Local Field Potentials (LFPs). Most BMIs are based on spike detection. However, spikes are incompatible with wireless application, since their high frequency content mandates a high sampling rate, which produces prohibitive amounts of heat in the context of the brain. An additional issue is that spikes are not stable over time, making chronic applications difficult, if not impossible.
Therefore, there has been growing interest in LFP-based BMIs. LFPs are relatively large-amplitude, low-frequency oscillations, and are understood to be the result of the sum of large numbers of action potentials from cells within 0.5-5mm from the electrode sensor. There is growing and plentiful evidence that motor information can be decoded from LFPs. LFPs have been found to be very stable over time, and the low sampling frequency means that the heat dissipation is within acceptable levels for wireless BMI application. As such, LFP recordings seem to be a promising candidate for the basis of chronic, wireless neural probes.
Even when using LFPs, the necessary bandwidth to communicate the data out from the brain remains a limiting issue for wireless BMIs, again due to heat. As such, on-chip strategies for reducing the bandwidth must be investigated. This is what this PhD proposal focuses on, as part of the Next Generation Neural Interfaces team at Imperial College.
Potential strategies for reducing the bandwidth include on-chip [DIGITAL SIGNAL PROCESSING] for bandwidth reduction or using bipolar recordings. On-chip signal processing could help eliminate uninformative or redundant data, reducing the amount of date sent out of the brain while preserving the important information. Strategies for on-chip processing could include channel selection based on some, to be determined, indicator of informative channels, template matching, feature extraction or data compression. The performance of these strategies must be balanced against their power requirements, which in turn is limited by heat dissipation.
Bipolar recordings involve taking a recording between two local electrodes instead of, as is usual in BMIs, between a local electrode and an electrically distant ground. This could help eliminate contributions to the LFP from non-local sources and, therefore, improve the signal content. However, this strategy multiples the number of channels. Hence, some on-chip sorting of useful channels would be required, in conjunction with this strategy.
The steps of this PhD include:
Creating various LFP -> Spike/Behaviour decoders with various machine learning a information theory techniques (e.g.: Principal Component Analysis, Long/Short Term Memory Neural Networks, Unscented Kalman and Wiener Filters, etc. These decoders must be trained and qualified on a wide variety of neural information form different probe geometries and studies for robust results.
Using these decoders to investigate the effectiveness of the different bandwidth reduction strategies across different individuals and circumstances.
Designing the associated ultra-low power [MICROELECTRONICS] for integration with the implantable neural probe.