Periodic Reporting for period 3 - DISPATCH Neuro-Sense (Distributed Signal Processing Algorithms for Chronic Neuro-Sensor Networks)
Reporting period: 2022-01-01 to 2023-06-30
The possibility to chronically interface with the brain 24/7 in daily-life activities would revolutionize human-machine interactions and health care. However, such chronic brain computer interfaces (BCIs) will have to satisfy challenging energy and miniaturization constraints. The premise of this project is that these constraints will naturally lead to modular platforms in which multiple networked miniature neuro-sensor modules form a ‘neuro-sensor network’ (NSN).
However, algorithms that extract information from neural signals are designed for traditional neuro-sensor arrays with central access to all data, which makes them unsuited for NSNs where the collected data is distributed over multiple neuro-sensor modules. Centralizing the data collected by an NSN would require unrealistic energy budgets which are not available in highly miniaturized (battery-powered) neuro-sensors.
The central idea of this project is to remove this algorithm bottleneck by designing novel distributed algorithms to let the modules of an NSN jointly process the recorded neural data through in-network data fusion and with a minimal exchange of data. This is viewed as the key enabler to exploit the vast amount of spatial information captured by NSNs, while maintaining a viable energy consumption and bandwidth.
We mainly focus on establishing these ideas within a new non-invasive NSN concept based on electroencephalography (EEG). By combining multiple smart miniature EEG modules into an ‘EEG sensor network’ (EEG-Net), we compensate for the lack of spatial information captured by a single module, without compromising in comfort, flexibility and wearability of such devices. Equipping such EEG-Nets with distributed neural signal processing algorithms will allow to process high-density EEG data at viable energy levels, which is a game changer towards high-performance chronic EEG for, e.g. epilepsy monitoring, neuroprostheses, and BCI applications. In addition to the algorithm design activities, we will also build an actual prototype of an EEG-Net and validate it within the three aforementioned use cases. Furthermore, to demonstrate the general applicability of our novel algorithms, we will validate them on data from other emerging NSN platforms as well, such as modular and/or untethered neural implants.
However, algorithms that extract information from neural signals are designed for traditional neuro-sensor arrays with central access to all data, which makes them unsuited for NSNs where the collected data is distributed over multiple neuro-sensor modules. Centralizing the data collected by an NSN would require unrealistic energy budgets which are not available in highly miniaturized (battery-powered) neuro-sensors.
The central idea of this project is to remove this algorithm bottleneck by designing novel distributed algorithms to let the modules of an NSN jointly process the recorded neural data through in-network data fusion and with a minimal exchange of data. This is viewed as the key enabler to exploit the vast amount of spatial information captured by NSNs, while maintaining a viable energy consumption and bandwidth.
We mainly focus on establishing these ideas within a new non-invasive NSN concept based on electroencephalography (EEG). By combining multiple smart miniature EEG modules into an ‘EEG sensor network’ (EEG-Net), we compensate for the lack of spatial information captured by a single module, without compromising in comfort, flexibility and wearability of such devices. Equipping such EEG-Nets with distributed neural signal processing algorithms will allow to process high-density EEG data at viable energy levels, which is a game changer towards high-performance chronic EEG for, e.g. epilepsy monitoring, neuroprostheses, and BCI applications. In addition to the algorithm design activities, we will also build an actual prototype of an EEG-Net and validate it within the three aforementioned use cases. Furthermore, to demonstrate the general applicability of our novel algorithms, we will validate them on data from other emerging NSN platforms as well, such as modular and/or untethered neural implants.
In the first 2,5 years of the project, we have performed the following work:
1) We have designed various novel distributed algorithms for various core tasks in the analysis of neural signals such as (1) artefact removal, (2) neural signal enhancement, extraction & decoding, and (3) spatial feature extraction and classification. These distributed algorithms are key enablers towards the exploitation of spatial information across the modules of a neuro-sensor network (NSN), while facilitating a viable bandwidth usage and energy consumption. All the developed algorithms are fully scalable in terms of network size and bandwidth requirements. In addition, the insights obtained during the algorithm design have been leveraged towards the development of a generic and more fundamental algorithm design framework, which contains several of these algorithms as special cases. This framework will be used in future distributed algorithm design activities within the project (and beyond). A particularly interesting feature of the framework is that, if an existing centralized algorithm satisfies some relatively mild properties, it automatically generates a distributed version of that algorithm that is applicable in an NSN with any possible topology.
2) We have collected and analyzed a high-density EEG data set to inform and validate the algorithm design, and to gain insight in the properties of the signals collected by an EEG-Net. The latter is necessary to establish EEG-Nets as a new chronic EEG recording platform, as the modular design will impact the signal content itself (due to short electrode distances and galvanic separation of the modules). To this end, we have emulated the sensors of an EEG-net by re-referencing the electrodes to their nearest neighbors, in order to analyze miniaturization effects and the lack of a common reference across the different sensors. Furthermore, in order to select the optimal locations to place these (emulated) sensors, we have designed automatic electrode selection algorithms for various neural signal processing tasks.
3) We have developed a prototype of a wireless miniature EEG sensor node in which an EEG amplifier, bluetooth LE radio, processing unit, and coin cell battery are all integrated in a small package of less than 1.5 square centimetres. In parallel, we are developing a dry electrode with a low-impedance for long-term use, with the goal of integrating these in future designs of the EEG sensor module.
1) We have designed various novel distributed algorithms for various core tasks in the analysis of neural signals such as (1) artefact removal, (2) neural signal enhancement, extraction & decoding, and (3) spatial feature extraction and classification. These distributed algorithms are key enablers towards the exploitation of spatial information across the modules of a neuro-sensor network (NSN), while facilitating a viable bandwidth usage and energy consumption. All the developed algorithms are fully scalable in terms of network size and bandwidth requirements. In addition, the insights obtained during the algorithm design have been leveraged towards the development of a generic and more fundamental algorithm design framework, which contains several of these algorithms as special cases. This framework will be used in future distributed algorithm design activities within the project (and beyond). A particularly interesting feature of the framework is that, if an existing centralized algorithm satisfies some relatively mild properties, it automatically generates a distributed version of that algorithm that is applicable in an NSN with any possible topology.
2) We have collected and analyzed a high-density EEG data set to inform and validate the algorithm design, and to gain insight in the properties of the signals collected by an EEG-Net. The latter is necessary to establish EEG-Nets as a new chronic EEG recording platform, as the modular design will impact the signal content itself (due to short electrode distances and galvanic separation of the modules). To this end, we have emulated the sensors of an EEG-net by re-referencing the electrodes to their nearest neighbors, in order to analyze miniaturization effects and the lack of a common reference across the different sensors. Furthermore, in order to select the optimal locations to place these (emulated) sensors, we have designed automatic electrode selection algorithms for various neural signal processing tasks.
3) We have developed a prototype of a wireless miniature EEG sensor node in which an EEG amplifier, bluetooth LE radio, processing unit, and coin cell battery are all integrated in a small package of less than 1.5 square centimetres. In parallel, we are developing a dry electrode with a low-impedance for long-term use, with the goal of integrating these in future designs of the EEG sensor module.
We have proposed a crucial novel algorithm design methodology, where data-driven distributed neural signal processing algorithms are able to fully exploit the dynamically changing neural correlation structure across all modules of a neuro-sensor networks. The major gain is that a data centralization can be bypassed, which would result in an expensive and non-scalable energy consumption. The proposed algorithms obtain a provable equivalent performance as the centralized counterpart neural signal processing algorithms, i.e. as if all modules would have access to all data, but -paradoxically- without sharing all the neural data. Instead, each module broadcasts only a minimal amount of fused data to its neighbors. Such a design goes beyond common ‘work-horse’ distributed signal processing tools in the state of the art, in the sense that we do not address traditional parameter estimation problems. Instead, our algorithms are able to solve challenging spatial filtering or source separation problems, in which fused signals (rather than parameters) are shared through the network.
Furthermore, based on our analysis of emulated EEG-Net signals, we were able to demonstrate that an EEG-Net with galvanically separated miniaturized EEG sensors can achieve the same neural decoding accuracy as with standard EEG with an equal amount of electrodes if (a) optimal sensor locations are selected in both cases, and (b) if the distance between the electrodes within a single EEG-Net module is at least 3cm.
Finally, our new wireless miniature EEG sensor prototype is the smallest to date, and has some novel functionality tailored to the EEG-Net concept that is not available in existing devices, such as a synchronization module based on human-body communication that allows to synchronize a multitude of these nodes with each other.
In the next phase of the project, we will leverage our algorithm design framework to design a larger toolbox of algorithms for neural decoding and feature extraction. We will further develop the framework itself towards a larger class of problems, and a faster convergence. Furthermore, we will test these algorithms in several use cases and validate them in terms of performance and energy consumption on actual EEG-Net hardware prototypes. The latter will be based on our EEG sensor node prototype, in combination with a new design of dry EEG electrodes. We will also validate our algorithms on data from other emerging neuro-sensor network types.
Furthermore, based on our analysis of emulated EEG-Net signals, we were able to demonstrate that an EEG-Net with galvanically separated miniaturized EEG sensors can achieve the same neural decoding accuracy as with standard EEG with an equal amount of electrodes if (a) optimal sensor locations are selected in both cases, and (b) if the distance between the electrodes within a single EEG-Net module is at least 3cm.
Finally, our new wireless miniature EEG sensor prototype is the smallest to date, and has some novel functionality tailored to the EEG-Net concept that is not available in existing devices, such as a synchronization module based on human-body communication that allows to synchronize a multitude of these nodes with each other.
In the next phase of the project, we will leverage our algorithm design framework to design a larger toolbox of algorithms for neural decoding and feature extraction. We will further develop the framework itself towards a larger class of problems, and a faster convergence. Furthermore, we will test these algorithms in several use cases and validate them in terms of performance and energy consumption on actual EEG-Net hardware prototypes. The latter will be based on our EEG sensor node prototype, in combination with a new design of dry EEG electrodes. We will also validate our algorithms on data from other emerging neuro-sensor network types.