TSS

Abstract: This research models signals and noise for extracellular neural recording. Although recorded data approximately follow Gaussian distribution, there are slight deviations that are critical for signal detection: a statistical examination of neural data in Hilbert space shows that noise forms an exponential term while signals form a polynomial term. These two terms can be used to estimate a spiking probability map which tells the probability of spike presence in any time window. Both synthesized data and animal data are used for the algorithm verification and detection performance comparison against other popular detectors. Experiment results suggest the predicted the spiking probability map is consistent with the benchmark and robust to different preparations.

The proposed algorithm has been implemented in a .13um CMOS process to substantially reduce data volume, potentially enabling a low power, wireless operation. The chip takes spike raw data as input and output three data streams simultaneously: field potentials down sampled at 1Hhz (1Mbps), band-pass filtered data sampled at 40KHz (20Mbps), and a spike probability map that indicates the presence of neural spikes (10Kbps). The chip has been tested in both benchtop experiments and in-vivo experiments: during the training phase (2.5sec), one channel consumes 85uW. After training, the power consumption is 40uW, 10uW, and 1uW for output spiking probability, bandpass filtered data, and field potentials respectively.

Research team members: Tong Wu (VLSI), Yin Zhou (detection algorithm), Jian Xu (ADC), Mohammadreza Keshtkaran (similarity measure algorithm)

Abstract: The objective of the project is to develop a neuromorphic attentional and sense-making system that processes and integrates the vast amount of stream data and interact quickly with the complexity of the real-world as humans do. Tightly integrating tools and theories from multiple areas from neuroscience to computational modeling, to neuromorphic hardware, the proposed system targets to set state-of-the-art performance specifications including intelligence, reliability, versatility, and reaction speed while overcoming some of the limitations of human operators such as fatigueness, illusions, and memory bottlenecks. We envision the system to have powerful applications in surveillance, unmanned systems, battlefield awareness, interactive data acquisition and analysis, and so on.

Our group's contribution to this project is the development of an innovative hardware system (mechanics and VLSI) that mimics biological visual sensing, with high resolution and accuracy, to support active sensing and analysis in closed-loop.

Research team members: Xiuxiang Huang (system and algorithm), Baitong Wang (mechanical system), Lei Cai (finite element analysis), Tony Wu (VLSI), Azam Khalili (micromotion analysis)

Abstract: In 2002 Maass has reported an elegant computational model based on recurrent circuits of integrated-and-fire neuron without requiring specifying circuit connections and tasks. The model is biologically plausible and mathematically proven to be able to simulate any Turing Machine. In several subsequent investigations, the model has been demonstrated to perform well in controlled computational and cognitive tasks. Due to multiple faced engineering challenges to realize a large-scale neural network, there are very few studies attempt to devise the model into real circuits/system, which we believe is of fundamental importance and can lead to new types of computational devices that mimic the brain. In this joint investigation with local neuroscientists, we propose to build a new computation technology coined “neurocomputor”, where neurons perform computation and storage like a computer, while stimulator and recorder are the I/O interfaces for information sensing and communication. Research and development of the hardware will be supported under this project, where the deliverables include 1.) an optical stimulator array to induce large-scale neuromodulation patterns, through which information is written into a genetically modified neural network (input interface); 2.) a spike train processor to interpret the evoked neural responses, though which the computational results are collected (output interface). In a relatively short term, the proposed hardware system will be integrated into neuroscience experiments in Dr. Antonius Van Dongen's lab at Duke-NUS for studying biological memory. In a longer term, the proposed technologies will serve as a generic tool for neural modulation and processing, which supports our collaborations in both neuroscience and medicine.

Research team members: Baitong Wang (LED array, mechanical system), Lei Cai (finite element analysis), Xiuxiang Huang (electronics, protocol), Amir Rastegarnia (neural network)

Abstract: There is a growing demand of chronic, wireless neurosenor interface with on-the-fly processing capabilities. Such neurosensor interface is designed with low power, low noise operation, thus meeting the urgent clinical needs of providing long-term, neurological health monitoring for patients who suffer from conditions such as epilepsy, Alzheimer’s disease, and sleep apnea. In this research, we propose to develop an ASIC to record and process sub-scalp EEG data for seizure detection and treatment. Major functional blocks consist of recorder, filter, ADC, signal processor, power management, and wireless data links. The faced challenges include the requirement of low-noise, long-term recording, low-power operation, wide system dynamic range, on-chip signal processing, wireless power and data management, and system integration and miniaturization. To resolve the challenges on noise, dynamic range, and long-term reliability, we propose a novel EEG recorder based on our newly invented frequency-shaping recording architecture. To allow less constrained environment for patients, we propose a customized wirelessly power and data links to deliver 1mW power and sufficient data bandwidth to support up to 16 recording channels. To provide the neurosenor interface re-configurability and on-the-fly processing capability, we propose to develop signal processing hardware. To reduce implant heat dissipation which causes discomforts, we plan to adopt a closed-loop data-link that enables the earpiece to adaptively control implanted circuitry (e.g., adaptively power on and off individual channels). Once the system is developed, we envision it has powerful applications in neurological health monitoring that could potentially generate appreciable economic returns and also serves as an enabling technology for a new type of brain machine interface that contributes to a better understanding of the brain.

Abstract:For invasive BCI experiments, the idea of neural assemblies has always been closely associated with the occurrence of spike patterns in convergently − divergently connected networks, where recording a sufficient number of neurons is a prerequisite to establish casual connectivities. Over the past 50 years, progresses in neural recording instruments have allowed the number of recorded neurons to double every 7.4 years, a mimicked “Moore’s Law”. Current BCI recording systems allow simultaneous data acquisition from about 100 channels, equivalently up to 200 neurons in total, which is small in number compared with a natural system (human brain ~100 billion, octopus ~300 million, bee ~1 million, aplysia ~20,000). To continue with the exponential increase in recording density, it faces multifold challenges from several aspects. First, given the electrode material, an increased recording density needs a proportionally increased amplifier input impedance to avoid degeneration in signal quality, especially in long-term experiments. This is because the electrode impedance tends to increase due to the elicited adverse tissue response, which results in electrical isolation from target neurons. A second challenge is that a full-spectrum recording is desired, from very low frequency field potentials to individual spikes at several kHz. It is further complicated by requiring removing electrode offset, which is achieved at a cost of compromised system specifications. Third, strong artifacts and interferences can cause electronics saturation, increasing the requirement on system dynamic range. A piece of electronics robust to artifacts is in great demand.

As a principled contribution of this research, we present a study focusing on neural recording system architecture. From carefully designed experiments and analyses, we reason that the proposed architecture can achieve the following targets.

Abstract: In this Project, we plan to develop a memristor based neurostimulator technology that meets the urgent need of medical device market for minimally invasive stimulation. The research in phase I will cover the feasibility study of memstimulator technology and its initial hardware prototyping.  This part is expected to be completed within two (2) years, contributing to three quarters (3/4) of the project deliverables. As a second mission of this research, we propose to research on synergic typed network storage, where data is stored in network connectivity. This part is in conjunction with the Artificial Cognitive Memory Program at the Data Storage Institute. It is expected to be completed within two (2) years, contributing to one quarter (1/4) of the project deliverables.

Abstract: Recent advancements in neural recording systems enable neuroscientists and clinicians to observe neurons communicate with one another by way of electrical activity, which is known as action potentials or simply as spikes. The direct applications for these multi-channel neural recording and processing capable systems are the enabling technologies for neuroprosthetic devices — devices those can be controlled by thoughts. As reported in literature, neural signals recorded from monkey’s motor cortex were analyzed to build a relationship between neural activities and intended limbs movements then used to control a cursor on a computer screen or a robotic arm. The positive achievements of the emerging technologies in brain-machine interface yearn for feedback mechanisms enabling the brain perceive information through prosthetic sensors. Building a realistic bionic arm is an example research that incorporates multi-channel neural recording and stimulation technologies for actuating a robotic arm and perceiving senses from the prosthetic sensors respectively.To surmount the challenging requirements of an implantable neuroprosthetic device that is low power, small footprint, high-performance signal processing, and limited wireless data rate, great efforts are aimed at developing hardware efficient algorithms and architectures. On-chip neural signal processing can reduce the wireless data transmission. For example, it potentially provides real-time computing solutions to complex spike detection and sorting problem and enables a closed-loop neuroprosthetic framework. In this work, we intend to develop neural signal processing algorithms and their corresponding microchip implementations, capable of processing neural signals on-the-fly over a large number of channels. Such DSP chips would be proof-of-concept prototypes that demonstrate the feasibility of on-chip neural signal processing, enabling new technologies for building intelligent, closed-loop, and robust implantable biomimetic systems.

Abstract: This research proposes to develop a closed-loop, minimally charged functional electrical stimulator as a platform technology. The targeted applications relate to neuromodulation and neurological disorder treatment. The technology components include neural stimulation, neural recording, signal processing, and wireless data management, which are to be implemented in integrated circuits. Compared with the current state-of-the-art, the proposed research will have a number of distinguished innovations/features to better cater to unmet clinical needs, which are summarized below.
Unmet clinical needs: functional electrical stimulation has been used as a treatment option for many neurological diseases and functional losses. Despite major success in improving patient life quality and treatment of intractable diseases, there could be substantial side effects and risks during and after surgeries that discourage the usage of an implant. To benefit patients and expand the market size, it is important to target unmet clinical needs and make improvements on safety, reliability, size, power consumption, endurance, and wireless management. In the following, we outline a few unmet clinical needs and the proposed engineering solutions.
Electrical stimulation may introduce undesirable, long-term side effects: for example, in deep brain stimulation, the patients may experience changes in personality, sleep apnea, cognitive dysfunction, compulsive gambling, depression, apathy, and inability to speak and learn. Recent studies suggest that they are caused by stimulation of unwanted neural circuits. These circuits further modulate their next stage neurons and may produce sustained long-term effects. The challenge here is inherent as both targeted and unwanted neurons are close to each other. Possible solutions are to reduce electrode size, stimulus strength, or distance to the targeted neurons, which usually associate with compromised long-term treatment outcomes and increased surgical challenges.
It is difficult to evaluate treatment outcomes, complicating optimization of stimulation protocols: although in many applications, the treatment may produce immediate effects (e.g., DBS, VGS, etc). there are also situations where treatment outcome cannot be easily accessed. In addition, stimulation may contribute to a variety of long-term effects. These long-term effects along with treatment outcomes are implicitly correlated with stimulation protocols, varying case by case and trial to trial. As a result, a patient and case specific protocol optimization is very challenging to perform. A possible solution is to record and analyze stimulation evoked neuronal responses and further correlate them with clinical experiences for protocol optimization.
Stimulation charge induces tissue damage and electrode interface erosion: it is generally more of a concern for neural prosthesis where the clinical usefulness depends on the ability to chronically provide reliable and safe stimulation over a large number of microelectrodes that tend to have a low-charge density threshold. For neural prosthesis, there is an important but less noticed trade-off between reliability and safety: different stimulation channels have different thresholds because the targeted neurons usually have diverse ion channel populations, geometry size, and distance to electrodes. Also, the threshold may change over time. As a result, the administered stimulation charge tend to be large, causing increased tissue damage and corrosion of electrode surface. Added that the stimulator supply voltages have to be elevated to meet a worst case scenario, which causes extra heat dissipation and tissue damage. An example is the epiretinal implant, the averaged threshold is below 5V while the supply voltage is designed to be ±14V resulting in a 180% increase in heat dissipation.
Wireless power and data management are required: both transcutaneous power and data telemetry methods have been studied for decades and need to be integrated into the full system in the next wave of translational research and product development. Current research efforts on power telemetry are to reduce the device size, increase power harvesting efficiency, and develop new materials to increase coupling, which translate to smaller surgical cut, less weight, less heat, and less electromagnetic stress on tissues. Studies on data telemetry on the other hand constantly derive ideas from the RF community. With recent technology support, the goal is to transmit and receive data efficiently and reliably, allowing remote diagnosis and monitoring of outpatients thus improving quality of life and patient safety.
In response to the unmet clinical needs listed, we propose to develop a novel technology platform. Compared with the state-of-the-art research papers and commercially available products, this work will have the following important features:
1. Develop a stimulator that allows delivering the "just needed" amount of charge for each stimulation channel. This feature is to achieve improved spatial resolution (less side effects), minimal charge injection (less tissue damage and electrode surface erosion), and "just needed" supply voltage (less heat dissipation). To do so the stimulator has to be reconfigurable: automatic configuration of both compliance voltage and stimulus pulse width for each individual pixel. It also requires new stimulator circuits along with non-stop data acquisition, signal processing, and power management on-chip.
2. Develop a novel recording circuit to acquire broadband signals from 1Hz to 10kHz, with sufficient dynamic range to record field potentials, neural spikes, artifacts, and interference. Special efforts will be made to dynamic range extension for recording before-, during-, and after- stimulation. The involved circuit techniques including frequency shaping modulation [24], high precision ADC, digital waveform reconstruction. Compared with its conventional counterpart, the proposed frequency shaping recorder can also provide 20 times improved input impedance, and thus more tolerance to glial cells encapsulation and electrode interface erosion.
3. Design and integrate a set of biomedical signal processing algorithms on-chip, including newly invented EC/PC spike detection, incremental/EM-PCA, unsupervised classification, adaptive noise reduction, noise-shaping, artifact detection and removal, and power line interference cancellation. These algorithm modules will contribute to evaluate treatment outcome, e.g., characterizing the evoked neuronal responses by stimulation thus optimizing treatment protocol and stimulation charge injection. An example is shown in Section IV, where spiking probability map has been derived in real-time with 80uW power consumption.
4. Design corresponding power and data management systems. The power management includes two components: 1.) transcutaneous power link to wirelessly deliver power and 2.) an implant side power management system to adaptively provide the required supply voltages with high efficiency (less heat) and small size. The challenge in data management is mainly a wireless transceiver that can reliably and efficiently transmit data to external devices (enabling remote diagnosis) and receive commands from external devices (enabling remote operation).