NET-TEN a silicon neuromorphic network for low-latency detection of seizures in local eld potentials Margherita Ronchini Yasser Rezaeiyan Milad Zamani

2025-05-02 0 0 2.39MB 14 页 10玖币

侵权投诉

NET-TEN: a silicon neuromorphic network for low-latency

detection of seizures in local ﬁeld potentials

Margherita Ronchini, Yasser Rezaeiyan, Milad Zamani,

Gabriella Panuccio, Farshad Moradi

October 20, 2022

Abstract

Therapeutic intervention in neurological disorders still relies heavily on pharmacological so-

lutions, while the treatment of patients with drug resistance remains an open challenge. This

is particularly true for patients with epilepsy, 30% of whom are refractory to medications. Im-

plantable devices for chronic recording and electrical modulation of brain activity have proved a

viable alternative in such cases. To operate, the device should detect the relevant electrographic

biomarkers from Local Field Potentials (LFPs) and determine the right time for stimulation. To

enable timely interventions, the ideal device should attain biomarker detection with low latency

while operating under low power consumption to prolong the battery life. Neuromorphic networks

have progressively gained reputation as low-latency low-power computing systems, which makes

them a promising candidate as processing core of next-generation implantable neural interfaces.

Here we introduce a fully-analog neuromorphic device implemented in CMOS technology for ana-

lyzing LFP signals in an in vitro model of acute ictogenesis. We show that the system can detect

ictal and interictal events with ms-latency and with high precision, consuming on average 3.50 nW

during the task. Our work paves the way to a new generation of brain implantable devices for

personalized closed-loop stimulation for epilepsy treatment.

1 Introduction

Pharmaceutical treatment currently represents the prevailing therapy in epilepsy. Yet, about one-

third of the patients fail to respond to anti-epileptic drugs [1,2]. Resorting to other strategies is

therefore necessary in these cases to prevent or suppress seizures. Deep Brain Stimulation (DBS) holds

great promise for treating medically intractable epilepsy [3,4,5,6,7]. In this regard, a distinction

must be made between open-loop and closed-loop DBS. In the former approach, the stimulation is

either continuous or cyclic; the electrical stimulus is repeated periodically following a pre-programmed

schedule, regardless of the dynamic state of the targeted brain circuitry. Conversely, closed-loop

devices operate adaptively, delivering a stimulation pattern upon the detection of speciﬁc electrographic

biomarkers. Both methods can eﬀectively reduce the duration and/or the frequency of seizure [3,4].

However, empirical evidence points to a higher eﬃcacy and fewer adverse eﬀects of the closed-loop

over the open-loop paradigm [8,9], as the stimulation is informed by the ongoing brain activity [10].

A closed-loop system requires the integration of three essential components: (1) a recording inter-

face to amplify and ﬁlter the signal, (2) a processing unit to analyze the recorded signals and extract

informative features, and (3) a stimulation back-end to deliver an electrical feedback. The processor

plays a pivotal role, since success in anticipating and averting seizures depends on its responsiveness and

accuracy. Highly accurate tracking of seizures can be achieved oﬄine using software-based algorithms

[11,12], but their computational complexity forces them to run oﬄine on high-performance computers

[13]. Local real-time analysis of data is therefore necessary to reduce wireless communication power

overhead [14], though on-chip processing tightens the already stringent requirements in terms of area

and power consumption [15]. In this respect, plenty of wearable and implantable classiﬁcation systems

have been proposed [14,13,16,17,15]. Unfortunately, most devices still lack an intelligent control

algorithm able to work around the large inter-patient variability [18]. To address all these challenges,

neuromorphic processing cores promise to become an integral part of next-generation neural implants.

Neuromorphic circuits operate on the same principles as biological information-processing systems,

drawing inspiration from the physical phenomena that govern the electrical behavior of neurons and

arXiv:2210.10565v1 [cs.HC] 19 Oct 2022

synapses to implement computational primitives [19]. Thanks to the spike-based representation of

information, the parallelism of multiple processing elements and their colocalization with memory

units, neuromorphic systems achieve low-latency and low-power performance, for which they have

emerged as a worthy opponent for von Neumann architectures in computing systems [20]. These

features also make them an appealing candidate for implantable neural interfaces, as their structure is

inherently suited to simulating spiking neural networks, which are a powerful tool to solve problems of

spatiotemporal pattern recognition [21,22,23,24]. As such, neuromorphic systems have already found

successful application for processing a broad range of biological electrical signals [25,26,27,28,29,30],

including high-frequency oscillations as pathological biomarkers of epilepsy [31,32,33] and seizure

detection [34].

Here, we describe NET-TEN, a fully-analog subthreshold neuromorphic network implemented in

a standard 180 nm Complementary Metal-Oxide-Semiconductor (CMOS) process for the detection of

ictal and interictal events from rodent brain slices coupled to Multi-Electrode Array (MEA).

2 Results

2.1 NET-TEN implementation

Figure 1: Example of a detection task completed by NET-TEN. a) From top to bottom: A 100 seconds

long sample of prerecorded Local Field Potential (LFP) data. One single ictal event emerges from the

baseline; UP and DW spikes generated from the conversion of the LFP data by the Step-Forward

Encoding (SFE) algorithm and used as input to NET-TEN; Graphical representation of the density

of UP and DW spikes fed to NET-TEN. The height of each bar corresponds to the number of spikes

accumulated within one second; Label as manually assigned (0 = baseline, 1 = interictal, 2 = ictal);

Voltage signal of one of the output neurons of NET-TEN as measured by the oscilloscope. b) Zoom-in

of the LFP signal with the corresponding UP and DW spikes generated by the SFE algorithm. Positive

slopes in the signal lead to the emission of UP spikes; negative slopes result in DW spikes.

Figure 1shows the results of a detection task performed by NET-TEN. Prerecorded Local Field

Potential (LFP) data of 100-second duration were converted into spikes in software using the Step-

Forward Encoding (SFE) algorithm proposed by Kasabov et al. [35] and delivered to the fabricated

network through an arbitrary waveform generator. The encoding process gives rise to two spike trains:

UP spikes, associated with positive-going signal deﬂections, and DW spikes, associated with negative-

going signal deﬂections. A second waveform generator synchronized with the ﬁrst one was employed

to carry the label signal and be able to display it on the oscilloscope together with NET-TEN neuronal

spiking. Firing activity at the output signaled the detection of a pathological pattern in the LFP

recordings.

Figure 2: a) Rendering of the prototyping Printed Circuit Board (PCB) comprehensive of potentiome-

ters to adjust the circuit biases. The 48-pin Quad Flat No-Lead (QFN48) footprint outlined in yellow

hosts the packaged NET-TEN chip. b) Photograph of the fabricated NET-TEN circuit die. c) Die

containing the stand-alone components: neuron block and synapse.

Figure 2provides a summary of the designed hardware components. The neuromorphic system

was implemented in a 180nm CMOS technology node and its layout covered an area of 1.08 mm2.

The circuit voltage biases were adjusted by tuning the relative potentiometers on the Printed Circuit

Board (PCB, depicted in Figure 2a), while at the same time the corresponding ﬁring activity generated

at the output was compared with the label. In this way, it was possible to determine heuristically in

which direction to steer each bias, in order to improve the classiﬁcation accuracy. The fabricated chip is

portrayed in Figure 2b. The voltage supply was set at 250 mV. The total static power consumption was

0.68 pW. The average power consumed during a detection task calculated across samples was 3.50 nW.

To evaluate the electrical behavior of the various functional modules that make up the neuromorphic

network, these were fabricated as stand-alone components on a separate chip (Figure 2c).

Figure 3illustrates the schematic of the three principal building blocks that constitute NET-TEN,

namely the neuron, the Excitatory Post-Synaptic Current (EPSC) and the Spike-Timing Dependent

Plasticity (STDP) circuits. As can be seen in Figure 3a, NET-TEN network has a feed-forward

architecture composed by three sparsely-connected layers of ten neurons each. Every neuron of the

input layer project to two neurons of the hidden layer, whose neurons also have a fan-in of two. Hidden

and output layers are connected together in a one-to-one fashion, meaning every neuron of the hidden

layer forms a synapse with one and only one neuron of the output.

The neuron, depicted in Figure 3b, was initially proposed in [37] and further described in [34]. The

circuit replicates biologically plausible dynamics, following the Izhikevich model. After performing the

Monte Carlo analysis, the transistor sizes were modiﬁed with respect to the original implementation

and the capacitance values were set to Cv= 76.86 fF and Cu= 614.88 fF, to alleviate the eﬀect of

process variations. One single neuron cell occupied an area of 1476.84 µm2. In the schematic, the two

state variables of the Izhikevich model, i.e. the transmembrane voltage Vmem and the slow recovery

variable U, are represented by the voltage across the integrating capacitors Cvand Cu, respectively.

Isyn is the synaptic current generated on the basis of the previous network layer and, when injected

into the neuron, it perturbs its equilibrium. M1−M3 current mirror serves as a positive feedback to

accelerate the rise of Vmem and bring it closer to the switching threshold of M9−M10 inverter. This

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

NET-TEN:asiliconneuromorphicnetworkforlow-latencydetectionofseizuresinlocaleldpotentialsMargheritaRonchini,YasserRezaeiyan,MiladZamani,GabriellaPanuccio,FarshadMoradiOctober20,2022AbstractTherapeuticinterventioninneurologicaldisordersstillreliesheavilyonpharmacologicalso-lutions,whilethetreatmentof...

展开>> 收起<<

NET-TEN a silicon neuromorphic network for low-latency detection of seizures in local eld potentials Margherita Ronchini Yasser Rezaeiyan Milad Zamani.pdf

共14页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

NET-TEN a silicon neuromorphic network for low-latency detection of seizures in local eld potentials Margherita Ronchini Yasser Rezaeiyan Milad Zamani

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: