Design and validation of distributed EEG signal processing methods for a wearable network of galvanically independent recording units.
Epilepsy is one of the most common severely disabling brain conditions, affecting over 46 million people worldwide. It is characterised by recurring, sudden, excessive, synchronous electrical activity in neuronal networks that disrupt ongoing brain activity and causes clinical seizures. Diagnosis and follow-up of epilepsy typically relies on electroencephalography (EEG). Clinical EEG is a non-invasive electrophysiological measurement technique recording the electrical field of the brain cortex from a number of scalp surface electrodes. It is typically acquired for a few dozens of minutes in a controlled condition in the clinic with devices that are ill-suited for long-term EEG-monitoring in daily life.
The development of portable EEG technology has taken advantage of the advent of silicon integrated circuit chips and the subsequent miniaturisation of electronics. However, the practical usability of many current EEG wearable devices is still limited. One major limitation relates to the trade-off between number of electrodes and wearability. More electrodes lead to larger electronics and more wires, which make the platform less user-friendly, and induce more wire-related artefacts. Conversely, the less electrodes, the lower the spatial resolution and the less tasks the system can be used for.
We propose a different platform to offer solutions to some of the current limitations associated with wearable EEG devices. This new platform will consist of a collection of wireless miniature EEG sensor units operating as a sensor network. Each EEG unit should incorporate electrodes, an amplifier, a wireless radio and a processing unit in a single package with a small scalp area footprint. The new topology eliminates the need for wires and will allow for a flexible, discreet, miniature, wearable system with as many EEG sensors as necessary for a particular application or patient. This new topology introduced new challenges which are investigated in this PhD thesis.
Specifically, this thesis presents signal processing methods for the analysis of EEG in a wireless EEG sensor network. It addresses the following main challenges.
- Design algorithms for automatic analysis of epileptiform activity. To do this we developed a novel multi-channel EEG signal processing method for automated epileptiform event detection which is specifically designed to run on a microcontroller with minimal memory and processing power. It is based on a linear multi-channel filter that is precomputed offline in a data-driven fashion based on the spatial-temporal signature of the seizure and peak interference statistics. At run-time, the algorithm requires only standard linear filtering operations, which are cheap and efficient to compute, in particular on microcontrollers with a multiply-accumulate unit. It has been validated on multiple datasets and compared to existing state-of-the-art algorithms.
- Design strategies for optimal EEG sensor selection. For this we propose a channel (or variable) selection algorithm for generalised eigenvalue problems such as those used in our epileptiform event detection algorithm. The method extends and generalises existing work on convex optimisation-based variable selection using semidefinite relaxations toward group-sparse variable selection using the l1,inf-norm. We comprehensively compared our method to state-of-the-art methods for sensor selection for spatio-temporal filter design in a simulated sensor network setting.
- Provide guidelines for the design strategies of miniaturised EEG sensor networks. This was investigated by conducting a study on the limits of miniaturisation of an EEG sensor network by emulating different networks using high-density EEG recordings and analysing interictal spikes in the different simulations.
Results show that the epileptiform event detection algorithm performs on par with state-of-the-art detection algorithms at a much lower computational cost for the detection of absence seizures and interictal epileptiform discharges. The study on channel selection algorithms indicates which algorithm to select in function of computational constraints, number of channels to select, and the topology of the problem. It shows that both the proposed channel selection algorithm and a backward greedy selection method best approximate the optimal solution. The proposed algorithm is also more robust to failure cases. The study on limits of miniaturisation of a network of wireless EEG sensors showed that recording equipment should be specifically designed to measure the small signal power at a short inter-electrode distance by providing an input-referred noise of <300 nV. It also showed that an inter-electrode distance of minimum 5cm in a setup with a minimum of two EEG units is required to obtain near equivalent performance in interictal spike detection to standard EEG.
In summary, this PhD thesis introduces several new signal processing methods for wireless EEG sensor networks for monitoring of people with epilepsy. It contributes to the technological advancement required for the wider adoption of this technology.