Neuromorphic Pattern Generation Circuits for Bioelectronic Medicine
NNeuromorphic Pattern Generation Circuits for BioelectronicMedicine
Elisa Donati , Renate Krause , and Giacomo Indiveri Abstract — Chronic diseases can greatly benefit from bio-electronic medicine approaches. Neuromorphic electroniccircuits present ideal characteristics for the development ofbrain-inspired low-power implantable processing systemsthat can be interfaced with biological systems. Thesecircuits, therefore, represent a promising additional toolin the tool-set of bioelectronic medicine. In this paper,we describe the main features of neuromorphic circuitsthat are ideally suited for continuously monitoring thephysiological parameters of the body and interact withthem in real-time. We propose examples of computationalprimitives that can be used for real-time pattern generationand present a neuromorphic implementation of neuraloscillators for the generation of sequence activation pat-terns. We demonstrate the features of such systems withan implementation of a three-phase network that modelsthe dynamics of the respiratory Central Pattern Generator(CPG) and the heart chambers rhythm, and that could beused to build an adaptive pacemaker.
I. INTRODUCTIONChronic diseases are one of the leading causes ofdeath in the world population. The majority of deathscaused by these diseases are attributed to respiratoryfailure, cardiovascular disease, cancers, and diabetes [1].As the proper treatment and management of chronicconditions is still an open challenge, there is a grow-ing interest in implantable electronic devices that canmonitor physiological parameters and control these ina closed-loop interaction with the body in real-time. Inaddition, the very same technology developed for theseimplantable devices used to support the treatment ofchronic diseases can be used for controlling closed-loopprosthetic devices and brain-machine interfaces.A promising approach toward the development ofthis technology is the one of “neuromorphic engi-neering” [2], [3]. Neuromorphic bio-signal processingsystems built following this approach adhere to de-sign principles that are based on those of biologi-cal nervous systems [4]. Their electronic circuits aretypically designed using mixed-mode analog/digitaltransistors and fabricated using standard VLSI processes,to emulate the physics of real neurons and synapsesin real-time [5], [6]. Similar to the neural processesthey model, neuromorphic systems process informa-tion using energy-efficient asynchronous, event-driven,methods. They are adaptive, fault-tolerant, and can be *This work was supported by the EU-H2020 FET project CResPACE(Grant No. 732170) Institute of Neuroinformatics, University of Zurich and ETHZurich, Zurich, Switzerland [elisa | rekrau | giacomo]@ini.uzh.ch flexibly configured to display complex behaviors bycombining multiple instances of simpler elements. Themost striking difference between neuromorphic systemsand conventional bio-signal information processing sys-tems is in their unconventional (beyond von Neumann)computing architecture: rather than implementing oneor more digital, time-multiplexed, central processingunits physically separated from the main memoryareas, they are characterized by co-localized memoryand computation. Neuromorphic systems comprise bio-physically realistic neuron and synapse circuits thatare at the same time the site of memory storageand of complex non-linear operations which performcollective and distributed computation. This in-memorycomputing feature is the main reason that allowsneuromorphic systems to perform bio-signal processingusing orders of magnitude less power than conventionalelectronic computing systems [7], [8]. Examples of ultra-low power brain-inspired neuromorphic systems ableto process temporal patterns have been recently appliedto electromyography (EMG) signal processing [9], [10]and electrocardiography (ECG) anomaly detection [11],[12].Starting from the faithful emulation of individualneurons [13], neuromorphic processing systems canimplement complex computations by assembling andcombining multiple computational primitives, verymuch like standard computing based on Boolean logicis built from the combination of logic gates [14], [5].In particular, a computational primitive that is usefulfor generating a rich set of complex movements andswitching behaviors is the Central pattern generator(CPG). CPGs are biological neural circuits that producerhythmic outputs to drive stereotyped motor behaviorslike breathing, or chewing, walking, swimming andflying. Thanks to their simple structure and the limitednumbers of parameters, neuromorphic implementa-tions of CPGs have been applied to different fields asrobotics [15], [16] and biomedical applications [17].In this paper we show how a neuromorphic imple-mentation of neural oscillators is able to reproducea sequence of population activation as shown in thebiological respiratory CPG (rCPG) or in the heartchambers activation. An neuromorphic implementationof rCPG can replace its biological counterpart in thevagus nerve stimulation to generate the desired heartrhythm. Otherwise, bypassing the biological rCPG, theneural oscillator can directly generate stimuli for the a r X i v : . [ c s . ET ] F e b time (ms) I s y n ( n A ) Vtau=1.64 VVtau=1.66 VVtau=1.68 V V m e m ( V ) Fig. 1: Neuromorphic processor micro graph with typicalsilicon synapse and silicon neuron response curves. Theneuromorphic processor is the 4-core chip presented in [19].The lower left plot shows the impulse response of a singleneuromorphic analog synapse circuit and the top right plotshows the response of a silicon neuron tuned to exhibit spike-frequency adaptation and bursting behavior. The synapse andneuron circuit details are presented in [5]. heart chambers activation. In both cases, the networkwould use physiological input, such as the concentrationof oxygen or carbon dioxide in the blood to modulatethe heart pacing. II. METHODS
A. Mixed-signal neuromorphic processors
Unlike digital simulators of neural networks, mixed-signal neuromorphic processors use the physics ofsilicon to directly emulate neural and synaptic dy-namics [5]. In this case, the state variables evolvenaturally over time and “time represents itself” [6],bypassing the need to have clocks and extra circuitsto manage the representation of time. Examples ofneuromorphic architectures that follow this mixed-signal approach include the Recurrent On-Line LearningSpiking (ROLLS) neuromorphic processor [18], andthe Dynamic Neuromorphic Asynchronous Processor-scalable (DYNAP-SE) chip [19].In these devices, the silicon neurons circuits imple-ment a model of the adaptive exponential Integrate-and-Fire (IF) neuron[20], and the synapses implement bio-logically realistic first order synaptic dynamics [5]. Thecircuits parameters can be tuned to make them exhibitlinear and non-linear behaviors such as spike-frequencyadaptation, refractory period saturation, regular firing,or bursting (see Fig. 1).
B. Oscillation generation
A neural oscillation is a rhythmic or repetitive patternof activity in the central nervous system. A basicbuilding block for generating sequences with identical
E E
I I
E E
ExcitatoryInhibitory
Central Pattern Generator Neural Oscillator
Fig. 2: Computational primitives to generate oscillations:the central pattern generator CPG and the neural oscillator.The CPG architecture (left) is composed of two excitatorypopulations coupled by mutual inhibition. The neural oscil-lator architecture (right) comprises two excitatory and twoinhibitory populations mutually coupled. It follows Dale’slaw to be more biologically realistic and to provide a higherflexibility in tuning the oscillation properties. cycles such as those in the rCPG or in the heartchambers activation is the “neural oscillator”. In thissection we present a neuromorphic implementation ofa neural oscillator as an extension of a basic CPG (seeFig. 2).
1) Central Pattern Generation architecture:
CPGs arecentral nervous system networks that can generatecoordinated muscle outputs in the absence of continuingpatterned sensory input. Although CPGs are able toproduce autonomously the desired rhythmic patterns,their activation is constantly monitored and modu-lated by high level centers. Together with the constantsensory feedback, modulatory top-down inputs allowhigh adaptability to proprioceptive and environmentalconditions. As a result, the CPG output is typicallya spatio-temporal sequence pattern with phase lagsbetween the temporal sequences that corresponds todifferent rhythms.Traditionally, this concept was applied to simple,innate, rhythmic movements with identical cycles thatrepeat continually (e.g. respiration) or irregularly (e.g.locomotion). More recent studies showed that manynatural movement sequences are not simple rhythms,but include different elements in a complex order wheresome involve learning [21].A key model to understanding rhythm generationis the half-center oscillator network. It consists of twoneurons that have no rhythmogenic ability individu-ally, but produce rhythmic outputs when reciprocallycoupled. In a CPG two pools of neurons are mutuallycoupled with inhibitory connections. Neurons of eachpool make transition between activated and inhibitedphases following an “escape” or a “release” mode. Inthe release mode the inhibited neurons can escapefrom inhibition thanks to the intrinsic properties ofthe membrane and inhibit the other neurons whereasin the escape one the neurons show spike-frequencyadaptation and slowly stop firing, releasing the otherinhibited neurons.
2) Neural oscillator architecture:
An alternative CPGsarchitecture is that of a neural oscillator, which can xcitatoryInhibitory P1 E P2 P3 E I E post-I / Left Atriumaug-E / Right Atrium early-I / Ventriclesor Fig. 3: Three-phase network, to model the rCPG as well as theheart chambers activation. Each population can be modelledusing two different architectures: the CPG unit and the neuraloscillator. produce rhythmic outputs without resorting to adapta-tion features of the neurons. Here we refer to a neuraloscillator consisting of an excitatory and a reciprocallyconnected inhibitory neuron population. The neuronsof the excitatory population are connected to each otherleading to self-excitatory behavior of the excitatorypopulation and receive a constant input which drivesthe neural oscillator.The neural oscillator architecture leads to limit cyclebehavior due to the reciprocal connections betweenthe excitatory and the inhibitory neuron populations.The constant input to the excitatory neuron populationleads to an increase in the activity of the excitatorypopulation. The excitatory population then starts toactivate its inhibitory counterpart which in return willinhibit the excitatory population and reset the network.Similar to the CPGs architecture with adaptation, herethe strength of the inhibition is proportional to theactivity level of the excitatory neurons and thereby leadsto these oscillatory dynamics.Neural oscillators are a useful building block forsystems which generate rhythmic outputs. They can becoupled by adding connections between their excitatorypopulations and another set of connections betweentheir inhibitory populations. This coupling of individualneural oscillators allows it to precisely tune the phaseshift between the individual neural oscillators. Hence, bycoupling multiple neural oscillators among each otherand tuning their phase shifts accordingly, it is possibleto obtain a large variety of complex, periodic outputpatterns.
C. Three-phase network
We model the rCPG as well as the heart chambersactivation as a three-phase network. For the rCPG thethree-phase network directly represents its biologicalcounterpart in the medulla oblongata. The medullarynetwork operates in a three-phase rhythm which consistsof the early inspiration phase (early-I), the post inspira-tion phase (post-I) and the late inspiration phase (aug-
Fig. 4: Silicon neuron measurements. The neuron producesperiodic bursting, lasting approximately 500 ms, with anequally long inter-burst interval.Fig. 5: Raster plot of the three-phase network with CPG.The network consists of 3 excitatory populations (n=4) andoscillates with a frequency of 1 Hz
E) [22]. We also model the heart chambers activation as athree-phase network because it is general practice to onlystimulate a subset of the heart chambers. Our cardiacthree-phase network model does not aim to reproducethe mechanism of the heart but only to reproduce therhythmic output. It generates the stimulation time ofthe two atria separately and ventricles combined. Thisallows us to show the hardware’s ability to generate veryshort (25-30 ms) as well as relatively long (800-900 ms)delays between different activation phases which arethe upper and lower boundaries of delays required forany other heart stimulation pattern.III. RESULTS
A. Central Pattern Generation results
The first step to reproduce the CPG behavior is to setthe correct parameters of neuron and synapse circuitsto generate bursting dynamics. This burst activity canbe reproduced by setting strong adaptation current insilicon neurons and weak synaptic weights and slowtime constants in the synapses, Fig. 4.In the neuromorphic CPG the frequency obtained fora single unit alternate movements ranges from 0 . 𝐻𝑧 toaround 3 . 𝐻𝑧 . Figure 5 shows the raster plot of threesegments, P1, P2, P3, at low frequency, around 0 . 𝐻𝑧 .This frequency allows the emulation of the basic rCPG,as well as, heart chambers activation rhythms. B. Neural Oscillator results
The oscillation frequency of a neural oscillator isdirectly proportional to the constant input current tothe excitatory population. Here we tuned our system of ig. 6: Raster plot of the three-phase network with neuraloscillators. The network consists of 6 neuron populations (3excitatory (n=16) and 3 inhibitory populations (n=4)) andoscillates with a frequency of 1 𝐻𝑧 . three coupled neural oscillators to run at a frequency ofroughly 1 𝐻𝑧 (see Fig. 6) on the DYNAP-SE chip [19].The noise and device mismatch in mixed-signalanalog/digital neuromorphic circuits leads to smallvariations in the population activities of the neuraloscillators over time. This is illustrated in Fig. 7 whichshows the phase diagram of each neural oscillator in thethree-phase network for two adjacent oscillation cycles.All three neural oscillators are similarly affected by thenoise which is shown by the similar deviations betweenthe yellow and black lines for P1, P2 and P3 while theyremain fully synchronized as shown in Fig, 6. Overall,the three-phase network has a standard deviation ofroughly 2 𝑚𝑠 for an oscillation frequency of 1 𝐻𝑧 . Fig. 7: Phase diagram of neural oscillators. The axes indicatethe average activity of the excitatory (x-axis) resp. inhibitory(y-axis) population of the three neural oscillators over twooscillation cycles (Cycle1: black, Cycle2: yellow).
IV. CONCLUSIONSIn this paper we introduced the neuromorphic engi-neering approach applied to bioelectronic medicine andpresented a neuromorphic implementation of a three-phase network that emulates a rCPG and/or a heartchambers activation sequences, as a key technologyfor adaptive pacemakers. The cycle in the three-phasecan be generated by using two different computationalprimitives, the CPG unit or the neural oscillator. Weshow how both primitives can be used to exhibit similarresults, with oscillation frequencies of approximately1 Hz, compatible with the frequencies involved in therCPG and cardiac rhythm.References [1] A. Sav, M. A. King, J. A. Whitty, E. Kendall, S. S. McMillan,F. Kelly, B. Hunter, and A. J. Wheeler, “Burden of treatment forchronic illness: a concept analysis and review of the literature,”
Health Expectations , vol. 18, no. 3, pp. 312–324, 2015.[2] C. Mead, “How we created neuromorphic engineering,”
NatureElectronics , vol. 3, no. 7, pp. 434–435, 2020. [3] G. Indiveri and T. Horiuchi, “Frontiers in neuromorphic engi-neering,”
Frontiers in Neuroscience , vol. 5, no. 118, pp. 1–2, 2011.[4] P. Sterling and S. Laughlin,
Principles of neural design . MIT Press,2015.[5] E. Chicca, F. Stefanini, C. Bartolozzi, and G. Indiveri, “Neuro-morphic electronic circuits for building autonomous cognitivesystems,”
Proceedings of the IEEE , vol. 102, no. 9, pp. 1367–1388,9 2014.[6] G. Indiveri and Y. Sandamirskaya, “The importance of space andtime for signal processing in neuromorphic agents,”
IEEE SignalProcessing Magazine , vol. 36, no. 6, pp. 16–28, 2019.[7] G. Indiveri and S.-C. Liu, “Memory and information processingin neuromorphic systems,”
Proceedings of the IEEE , vol. 103, no. 8,pp. 1379–1397, 2015.[8] M. Rahimiazghadi, C. Lammie, J. K. Eshraghian, M. Payvand,E. Donati, B. Linares-Barranco, and G. Indiveri, “Hardwareimplementation of deep network accelerators towards healthcareand biomedical applications,”
IEEE Transactions on BiomedicalCircuits and Systems , pp. 1–1, 2020.[9] Y. Ma, E. Donati, B. Chen, P. Ren, N. Zheng, and G. Indiveri,“Neuromorphic implementation of a recurrent neural networkfor EMG classification,” in
International Conference on ArtificialIntelligence Circuits and Systems (AICAS),2020 . IEEE, 2020, pp.69–73.[10] E. Donati, M. Payvand, N. Risi, R. Krause, and G. Indiveri,“Discrimination of EMG signals using a neuromorphic imple-mentation of a spiking neural network,”
Biomedical Circuits andSystems, IEEE Transactions on , vol. 13, no. 5, pp. 795–803, 2019.[11] F. Bauer, D. Muir, and G. Indiveri, “Real-time ultra-low powerECG anomaly detection using an event-driven neuromorphicprocessor,”
Biomedical Circuits and Systems, IEEE Transactions on ,vol. 13, no. 6, pp. 1575–1582, December 2019.[12] A. Das, P. Pradhapan, W. Groenendaal, P. Adiraju, R. T. Rajan,F. Catthoor, S. Schaafsma, J. L. Krichmar, N. D. Dutt, and C. V.Hoof, “Unsupervised heart-rate estimation in wearables withliquid states and a probabilistic readout,”
Neural networks , vol. 99,pp. 134–147, 2018.[13] K. Abu-Hassan, J. D. Taylor, P. G. Morris, E. Donati, Z. A.Bortolotto, G. Indiveri, J. F. Paton, and A. Nogaret, “Optimalsolid state neurons,”
Nature communications , vol. 10, no. 1, pp.1–13, 2019.[14] G. Marcus, A. Marblestone, T. Dean et al. , “The atoms of neuralcomputation,”
Science , vol. 346, no. 6209, pp. 551–552, 2014.[15] E. Donati, F. Corradi, C. Stefanini, and G. Indiveri, “A spikingimplementation of the lamprey’s central pattern generator inneuromorphic VLSI,” in
Biomedical Circuits and Systems Conference(BioCAS) . IEEE, Oct 2014, pp. 512–515.[16] D. Gutierrez-Galan, J. Dominguez-Morales, F. Perez-Peña,A. Jimenez-Fernandez, and A. Linares-Barranco, “Neuropod:a real-time neuromorphic spiking cpg applied to robotics,”
Neurocomputing , vol. 381, pp. 10–19, 2020.[17] K. J. Abu-Hassan, J. D. Taylor, J. J. van Bavel, M. A. Vos, andA. Nogaret, “Silicon central pattern generator model of cardiaccontraction behavior,” in . IEEE, 2020,pp. 1–2.[18] N. Qiao, H. Mostafa, F. Corradi, M. Osswald, F. Stefanini, D. Sum-islawska, and G. Indiveri, “A reconfigurable on-line learningspiking neuromorphic processor comprising 256 neurons and128k synapses,”
Frontiers in neuroscience , vol. 9, p. 141, 2015.[19] S. Moradi, N. Qiao, F. Stefanini, and G. Indiveri, “A scalablemulticore architecture with heterogeneous memory structuresfor dynamic neuromorphic asynchronous processors (DYNAPs),”
Biomedical Circuits and Systems, IEEE Transactions on , vol. 12, no. 1,pp. 106–122, Feb. 2018.[20] R. Brette and W. Gerstner, “Adaptive exponential integrate-and-fire model as an effective description of neuronal activity,”
Journalof neurophysiology , vol. 94, no. 5, pp. 3637–3642, 2005.[21] A. Berkowitz, “Expanding our horizons: central pattern gener-ation in the context of complex activity sequences,”
Journal ofExperimental Biology , vol. 222, no. 20, p. jeb192054, 2019.[22] J. C. Smith, A. Abdala, H. Koizumi, I. A. Rybak, and J. F. Paton,“Spatial and functional architecture of the mammalian brain stemrespiratory network: a hierarchy of three oscillatory mechanisms,”