Recent Advances in Neural Dust,a platform for peripheral and central

nervous system recording

Michel M. MaharbizDJ Seo, Ryan Neely, Konlin Shen

Jose Carmena, Elad Alon, Jan Rabaey

Electrical Engineering and Computer Science

University of California, Berkeley

2016 NSF Nanoscale Science and Engineering Grantees Conference 12 December 2016

the Neural Dust concept

http://arxiv.org/abs/1307.2196

• Idea conceived in March 2013

• Technical rationale published in arXiv in July 2013

• First in vitro demonstration, J Neurosci Meth, Nov 2014

• 1st gen beam-steering system, IEEE EMBC 2015, August 2015

• In vivo PNS data, Neuron, August 2016

[Hochberg Nature 2006][Doerner 2010]

Today’s systems

Bulky, invasive, wired, low-density

Future systems

Low-power, scalable, wireless systems

1.25 mm2, 2.5 µW power consumption

[Biederman JSSC 2013]

[Biederman JSSC 2013] [Yin Neuron 2014]

[Muller JSSC 2015]

How do we get to chronic interfaces?

Extreme MiniaturizationImportant for recording resolution and longevity of sensors

Resolution – observation at the

cellular level

• Neurons are 10 – 100 µm in diameter

• Need spatial measurement resolution

on the scale of 10’s µm[Courtesy of T. Blanche]

• Recorded signal quality reduced by scar tissue

• Small fraction of patient’s lifetime

• Not sufficient for chronic BMI

• Miniaturized probe design can reduce scar

tissue

(slide courtesy of DJ Seo)

Reliability and Longevity

[Seymour 2006]

[Potter 2012]

[Turner 1999]

TX Drive

RX

Two fundamental issues:

• A small form factor (volume) + speed of light fres = 10’s GHz

• Significant tissue loss at such high frequency

• Output power limit due to safety regulations: 10 mW/cm2

• e.g. 1 mm2 interrogator, 100 μm dust node, 2 mm distance received power < 40 pW << 2.5 μW for CMOS

Could we make a vanishing small RF interface to the nervous system?

theorySeo et al., arXiv 2013

Neural Dust system concept

• the interrogator couples ultrasound energy to the motes

• the interrogator can perform both/either spatial and frequency discrimination with sufficient bandwidth/resolution to interrogate each mote

• each mote consists of a piezoelectric transducer, surface electrodes for electrophysiological signal acquisition, and a silicon CMOS die containing electronics for signal amplification/conversion.

• The mote reports recorded signals back to the interrogator by reflecting and modulating the amplitude, frequency, and/or phase of the impinging ultrasound wave.

Neural Dust (ND) SystemMiniaturized ultrasonic platform can be used for CNS & PNS

Sub-Dural Link Model

• TX (interrogator) and RX (neural dust) modeled with KLM

• Match resonant frequency to maximize power transfer

• 2 mm tissue as a lossy transmission line

US Wireless Powering: Ultrasonic power transfer is much more efficient

US EM

0.1 mW/mm2

7.2 mW/mm2

Power Safety Limit

US EM

1 dB/mm

2.5 dB/mm

10 MHz 10 GHz

Tissue Attenuation

• Low acoustic velocity allows operation at a much lower frequency

– e.g. λ = 150 μm @ 10 MHz US vs. λ = 5 mm @ 10 GHz EM

• The acoustic loss is smaller than EM loss

– Safety regulation (0.1 mW/mm2 for EM vs. 7.2 mW/mm2 for US)

• Efficiency of ~7% (or -11.6 dB) at 100 μm• Received power: ~500 μW US vs ~40 pW EM (1 mm2 interrogator)

• Scaling prediction of 3.5 μW at 20 μm node

Significantly more power with US

• Electrode has thermal noise• Electrode |Z| density: Cdl ~ 0.5 pF/μm2, Rs = 18.65 MΩ·μm2

• Voltages are measured differentially• Neural dust: reference electrode on the same footprint

• e.g., measured signal amplitude for d = 100 μm is ~10 μV [Du 2011]

[Du PLoS 2011] d

Randles Model for Electrode

Scaling: Electrode Modeling

Scaling of the mote

• Captured power decreases with mote size

• Extracellular recording is differential, so signal decreases with size• smaller motes need more power to maintain same

SNR

• Fundamental electrode thermal noise

Scaling with an SNR of 10 dB shows operation down to 50 μm

Can exceed FDA safety regulation, but scaling is ultimately limited by electrode thermal noise

How do you build the front end?

Simplified neural front-end with a single FET sensor• Electrical load impedance (FET) varies with vneural

• Instantaneous ultrasonic wave reflectivity changes

• Backscattered wave is modified

First experimental dataSeo et al., J Neurosc Meth, 2014

(a) (b)

Initial validation of power coupling

• Measured power transfer efficiency at various mote sizes matches simulated behavior closely.

• For each mote dimension:• impedance spectroscopy• frequency response of harvested

power on the PZT

reinforces the reliability of the simulation framework.

Initial validation of power coupling

Simulated backscatter sensitivity scaling plot for various impedance levels.

in vivo rodent PNS dataSeo et al., Neuron, in press

dust motes and transceiver

[Seo EMBC 2015]

[Tang ISSCC 2015]

[Seo, Tang TBioCAS 2015]

Noise Floor & Tissue AttenuationSystem is currently sensitive to misalignment & misorientation

• Severe penalty with misalignment & misorientation

• Effect of lateral misalignment follows the beam pattern power fall-off

• Angular misalignment is less severe than lateral misalignment

• Can be addressed by spatial multiplexing and/or more robust reconstruction

algorithms (Future Directions)

EMG preparation

100 µs

• transducer is placed ~1 cm away from mote• 6 pulses, 1.8 MHz every 100 µs• 720 µW total output average power• 8 mW/cm2 average ISPTA (FDA -> 720 mW/cm2)• 0.3 mW/cm2 peak ISPTA (FDA -> 190 W/cm2)• the transceiver consumes 500 µW but heavily

duty-cycled (0.008%) so 1.8 V operation is possible

Graded EMG dataa) voltage recorded at implanted “ground truth”

electrodes vs. electrophysiological electrode

stim intensity

b) voltage reported by dust mote vs.

electrophysiological electrode stim intensity

c) one-to-one comparison of “ground truth” vs.

dust mote electrodes at 100% stimulation

intensity

d) voltage error between signals in c)

e) neural dust EMG peak-to-peak voltage vs.

stim intensity showing classic sigmoidal curve

100 kHz sample rate for “ground truth”

10 kHz sample rate for dust motes

ENG preparation in the sciatic nerve

Adapted from Advanced microsurgery, 2. In vivo

laboratory surgery, The University of Szeged.

ENG datab) compound potential recorded at

implanted “ground truth” electrodes vs.

electro-physiological electrode stim intensity

c) voltage reported by dust mote vs.

electrophysiological electrode stim intensity

d) one-to-one comparison of “ground truth”

vs. dust mote electrodes at 76% stimulation

intensity

e) voltage error between signals in c)

f) neural dust ENG peak-to-peak voltage vs.

stim intensity showing classic sigmoidal

curve

200 kHz sample rate for “ground truth”

10 kHz sample rate for dust motes

the future

Beam-forming for aligning motes, targeting multiple motes and steering

A custom ASIC drives a 7 x 2 PZT transducer array

with 3 cycles of 32V square wave with a specific

programmable time delay to focus the beam at the

800 um neural dust mote placed 50mm away.

The measured acoustic-to-electrical conversion efficiency of the

receive mote in water is 0.12% and the overall system delivers

26.3% of the power from the 1.8V power supply to the transducer

drive output, consumes 0.75uJ per each transmit phase.

Seo et al., EMBC 2015

-600 -400 -200 0 200 400 6000.4

0.5

0.6

0.7

0.8

0.9

1

Input Voltage (mV)

Ba

cksca

tter

(norm

)

Digital front endAmplifying the input signal can fully utilize the dynamic range

• Noise floor of the system limits detecting CNS signals

• Neural signal occupies a small region in the calibration curve

• Potential usable dynamic range is wasted

• Neural front-end, with appropriate amplifying chain, can utilize this by amplifying

the input signal

[Design largely adopted from Biederman

JSSC 2013 & Muller JSSC 2012]

700 µm

400 µ

m

Cap

Rectifier

Amp

DigLDO

Bias

refM

od

US-

US+

IN-

IN+

Piezo

Full-wave

Rectifier

+

Doubler

Reference,

Regulator

AC-DC

ASIC

Modulator

(FET

switch)

10b ADC

Offset cancellation

DAC

LNA

Mixed signal feedback

Digital Comm. ASICNeural front-end and miller modulator are included

Low-power neural front-end with digital backscatter

• Powered by ultrasound and communication with miller modulation

• 400 µm x 700 µm chip taped out in 65nm TSMC LP in November

• Back in January for recording CNS in animal

[w/ Kyoungtae Lee]

what’s coming• ND recording noise floor to < 10 µV• <500 micron dust• ND stimulation• (w/ Prof. Rikky Muller, Ben Johnson)

what’s coming• Backpack for awake behaving

neuromodulation

• Biocompatible materials

• CNS!

• integrating new circuits and sensors• pressure

• temperature

• strain

• O2

• pH

• …

• instrument tissue and organs, not just nerves

• fundamental platform for data in and out

• Three examples:• GI tract microflora species tracking

• chronic deep oxygen measurement

• chronic tissue pressure/stiffness

• fracture healing

• demonstrate chronic viability

3-5 years

Häggström, Mikael. "Medical gallery of Mikael Häggström 2014". Wikiversity

Journal of Medicine 1 (2). DOI:10.15347/wjm/2014.008. ISSN 20018762

recent grad students

faculty collaboratorsElad AlonJose CarmenaJan RabaeyCaroline Ajo-FranklinKristofer PisterHirotaka SatoMurat ArcakAdam ArkinFlip Sabes

undergraduatesRobin HerbertKaylee MannJonathan ChenSvetoslav KolevNimbus GoehausenMyo Myo NyiRyan TsengYoav PeeriEmin BaghoomianChris Berry…

Questions?Thanks!Arda Ozilgen

David PiechBrian KilbergAmy LiaoDJ SeoKonlin ShenMonica LinTom ZadjelTravis MasseyBochao LuStephanie GarciaCamilo Diaz-BotiaAlyssa Zhou

Peter LedochowitschBrian PepinDaniel CohenMichael LorekGabriel LavellaVedavalli Krishnan

• XDCR model using 3-port network, based on KLM model (1970)

• Both electrical and mechanical resonances• Determined by the thickness of the XDCR

• Aspect ratio: Interrogator (10:10:1), neural dust (1:1:1) for density

Piezoelectric XDCR

• 3D loss mechanism: beam spreading

• Dust mote should be placed at interrogator’s Rayleigh distance• Interrogator sized (1 mm2) to match its Rayleigh distance (natural focus) with tissue

transmission distance (d = 2 mm) @ 10 MHz

• Beam steering to enable multi-node interrogation (more later)

Tissue Depth (d = 2 mm)

Rayleigh Distance = 𝑫𝟐

𝟒𝝀

Inte

rro

gato

r

Depth (mm)

Beam spreading and mote position

• Re-radiation along two perpendicular axes due to Poisson’s ratio• For perfect cube, ~66% of the energy kept in the main thickness resonance mode

• Modeled as additional loss

Simulation Result:• @15MHz (1st resonance)

Ex/Etotal = 16.6%Ey/Etotal = 16.6%

Ez/Etotal = 66.8%

• @22.6MHz (2nd resonance)Ex/Etotal = 21.0%Ey/Etotal = 20.2%

Ez/Etotal = 58.8%

Cube: Mode Coupling (Re-Radiation)1 1.5 2 2.5 3 3.5 4

x 107

10-20

10-18

10-16

10-14

10-12

X: 1.5e+07

Y: 5.51e-15

BaTiO3 Cube Thickness Resonance mode (z) in x

X: 2.26e+07

Y: 3.966e-14

1 1.5 2 2.5 3 3.5 4

x 107

10-20

10-18

10-16

10-14

10-12

X: 1.5e+07

Y: 5.5e-15

BaTiO3 Cube Thickness Resonance mode (z) in y

X: 2.26e+07

Y: 3.813e-14

1 1.5 2 2.5 3 3.5 4

x 107

10-20

10-18

10-16

10-14

10-12

X: 1.5e+07

Y: 2.218e-14

BaTiO3 Cube Thickness Resonance mode (z) in z

X: 2.26e+07

Y: 1.109e-13

1 1.5 2 2.5 3 3.5 4

x 107

10-20

10-18

10-16

10-14

10-12

BaTiO3 Cube Thickness Resonance mode (z)

Energy in x

Energy in y

Energy in z

freq (Hz)