Detection of <12 µV

RMS

Extracellular Action Potential and Local

Field Potential by Optimum Design of a Single Pixel

Electrolyte-Oxide-MOSFET Interface in CMOS 28 nm

David Tomasella

a

, Elia Vallicelli

b

, Andrea Baschirotto

c

and Marcello De Matteis

d

Department of Physics, University of Milano Bicocca, Piazza della Scienza 3, Milano, Italy

Keywords: Biological Neural Networks, Biosensors, Neural Engineering, Analog Integrated Circuits, Low-Noise

Amplifier.

Abstract: Microelectrode-Arrays (MEAs) allow monitoring thousands of neurons/mm

2

by sensing: extracellular Action

Potentials and (in-vivo) Local Field Potentials. MEAs arrange several recording sites (or pixels) in a spatial

grid, planarly and capacitively coupled with in-vitro cell cultures and/or integrated in electrocorticography

grids. This paper focuses on Electrolyte-Oxide MOS Field-Effect-Transistors (EOMOSFET) MEAs for cell-

level recording and presents a complete model of the neuron-electronics junction that reduces to a single

electrical scheme all the biological (the neuron) and physical layers (the electrolyte, the Diffuse/Helmoltz

capacitances, the oxide and the MOS transistor) composing the interface. This allows to predict the noise

power coming from biological environment (electrolyte bath) and to optimize all electrical parameters with

the main aim to minimize the final sensing Noise Figure and thus enhance the acquisition Signal-to-Noise-

Ratio. Frequency domain simulations from the proposed model demonstrates that there is an optimum design

point for all parameters involved in the building EOMOSFET pixel that allows to perform >9 dB Signal-to-

Noise-Ratio at <12 µV

RMS

extracellular neuro-potentials power at the electrode node. This will finally enable

high-resolution recording of ultra-weak neuro-potentials signals flowing by the electrolyte cleft that have not

been never explored adopting planar capacitive coupling interfaces.

1 INTRODUCTION

There is a rather large lack of information between

neuron membrane electrical activity of single

neurons, and physiological or whole brain

behavioural events. To fill this gap, we need to

understand the activity of individual neurons and how

it contributes to neural circuits functioning. Such

ambitious perspective cannot be achieved by

macroscale neural recording techniques

(electroencephalogram, magnetic resonance, etc.),

nor by patch clamps (monitoring single cell unit). One

of the best options is to adopt Microelectrode-Arrays

(MEAs, (Obien, 2015; Thomas, 1972; Pine, 1980))

that allow monitoring thousands of neurons/mm

2

by

sensing: extracellular Action Potential (EAP in 300

a

https://orcid.org/0000-0001-8413-5751

b

https://orcid.org/0000-0003-0905-151X

c

https://orcid.org/0000-0002-8844-5754

d

https://orcid.org/0000-0003-1061-1262

Hz – 5 kHz bandwidth) and (in-vivo) Local Field

Potential (LFP up to 300 Hz bandwidth). They are

used as planar probes in neuron cells cultures forming

a cell-electrode capacitive coupling. Implanted

MEAs are typically needle-shaped probes that deeply

penetrate the cortex for tissues recording/stimulating

and for increasing proximity and signals detection

rate.

Both in-culture and implantable MEAs can be

integrated in commercial CMOS silicon substrates

with an additional post-processing step consisting on

covering CMOS metal aluminium electrodes by

noble metal films (Pt/Au (Gross, 1982)) or dedicated

oxide layers (TiO2 (Cianci, 2012)). Active MEAs

embed both analog signal processing channels (by

neural amplifier, low-pass filter for antialiasing and

A-to-D conversion) and digital circuits synthesizing

66

Tomasella, D., Vallicelli, E., Baschirotto, A. and De Matteis, M.

Detection of <12 VRMS Extracellular Action Potential and Local Field Potential by Optimum Design of a Single Pixel Electrolyte-Oxide-MOSFET Interface in CMOS 28 nm.

DOI: 10.5220/0010346300660076

In Proceedings of the 14th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2021) - Volume 1: BIODEVICES, pages 66-76

ISBN: 978-989-758-490-9

Copyright

c

2021 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved

Figure 1: Microelectrode Arrays State-of-the-Art (Noise Power vs. Electrodes/Pixels Density).

Figure 2: Microelectrode Arrays State-of-the-Art (Noise Power vs. CMOS Minimum Channel Length).

advanced Digital Spikes Detection ((Vallicelli, 2018;

Shadid, 2009) DSD) algorithms. DSDs exploit the

large array spatial resolution for separating relevant

extracellular events from background noise by spatial

correlation post-processing algorithms. They perform

complex digital algorithm that require a certain

computing power leading to a non-negligible

dynamic power consumption and preventing

integration in the same MEA silicon area.

Despite outstanding advancements in neural probes

development, there are still many phenomena that

state-of-the-art MEA technology cannot observe at

high spatial resolution:

at membrane level, subthreshold events, such as

synaptic potentials can influence cell rest status

without producing an action potential;

propagation of the action potentials (AP) in

axons, the backpropagation of AP in the

dendrites and the generation of dendritic spikes,

and weak extracellular synaptic field potential;

traces of membrane oscillations in the

extracellular space, never observed by planar

probes since they require single-cell patch

recording techniques.

Simply speaking, these events and phenomena cannot

be observed because noise power floor in cell-

electrolyte-electrode-electronics junction, is to date

higher than the signal power of those extracellular

neuro-potentials carrying information coming from

subthreshold or ultra-weak events.

Figure 1 shows some of most relevant MEAs state-

of-the-art (DeBusschere, 2001; Huys, 2012; Frey,

2010; Maccione, 2013; Eversmann, 2003; Wang,

2019; Lopez, 2018; Dragas, 2017) in terms of noise

power (in µV

RMS

measured at the electrode node) vs.

Pixel Density (electrodes count per mm

2

). Both axes

are in log scale. There is a clear trend of noise power

increasing with larger pixels density. Such trend is

justified by the smaller electrode area (at higher pixel

density) that increases flicker noise power coming

from both the TiO

2

/Metal film charge traps and the

MOS transistor (MOST) placed just below the

electrode.

In other words, to decrease the noise power and to

detect sub-threshold events, it is necessary to increase

the area of the single electrode but this would again

lead to a reduction of the pixel density, degrading the

neural recording spatial resolution.

Morevoer, this approach has already reached the

point of maximum efficiency because ref. (Huys,

2012; Dragas, 2017) in CMOS 0.18 µm and ref.

(Wang, 2019; Lopez, 2018) in CMOS 0.13 µm

(rectangular box in both Figure 1 and Figure 2) show

practically the same noise power (i.e. approximately

22 µV

RMS

) at the electrode equivalent node.

Detection of <12 VRMS Extracellular Action Potential and Local Field Potential by Optimum Design of a Single Pixel

Electrolyte-Oxide-MOSFET Interface in CMOS 28 nm

67

Figure 3: Drain-Source Current (I

DS

), Transconductance

(g

m

) and Efficiency (g

m

/I

DS

) in CMOS 0.18 µm and 28 nm

for Standard-Process MOST.

Furthermore, at such noise power levels,

increasing electrode area (and thus reducing pixel

density) does not involve a significant 1/f (flicker)

noise power reduction because MOS transistors

thermal noise (in first approximation independent on

electrode area and dependent on MOST dc current)

becomes more dominant.

More specifically in-band thermal noise power

spectral density is inversely proportional to the MOS

transistor transconductance (g

m

). Such g

m

is

proportional to the current consumption, that (in bio-

signals processing analog stages) operate the MOST

with few µA current (Harrison, 2003). Thus the

efficiency of the MOST (defined as g

m

/I

DS

ratio

(Sansen, 2007) or in other words the amount of g

m

that can be synthesized by a MOST operating at a

certain drain-source current I

DS

) becomes a key

parameter for both noise and power minimization in

next generation MEAs.

Unfortunately, all most performant MEAs (Huys,

2012; Wang, 2019; Lopez, 2018; Dragas, 2017) are

implemented in very old CMOS technologies (0.13

µm processes nodes and beyond for flicker noise

power reduction), where MOST efficiency is very

much lower than nm-range CMOS nodes.

Figure 2 shows state-of-the-art MEAs noise

power vs. CMOS process generation in terms of

minimum MOST channel length.

Maximum (sub-threshold voltage) MOST

efficiency in 0.13 µm (or 0.18 µm CMOS node) is

about 20 V

-1

, whereas 28 nm CMOS improves g

m

/I

DS

up to 26 V

-1

, resulting in approximately +3dB

electrode noise attenuation at same current.

These last considerations are validated in Figure

3, where MOST Drain-Source Current (I

DS

),

Transconductance (g

m

) and Efficiency (g

m

/I

DS

) are

plotted vs. gate-source voltage in CMOS 0.13 µm and

28 nm for Standard-Process MOSTs.

Moreover, older CMOS processes increases

system complexity (Baschirotto, 2009; De Matteis,

2006) and harness, because Digital Spikes Detection

stages cannot be on-chip integrated due to its very

high dynamic power in low scaled-down CMOS

nodes. Effectively such power will be incompatible

with a stable spatial temperature distribution for cell

integrity in neuron cultures and with the limited

power budget required by the portability of

implantable devices.

In this context, this paper proposes a dedicated

design of EOMOSFET MEAs (for the case of a single

recording site/pixel and that can be easily extended to

spatial matrix MEAs) that efficiently sets the basis for

definitely overcome the above issues by:

adopting the CMOS 28 nm technology node and

then including in the model the technology node

parameters like dielectric constant, gate

capacitance per unit area, sub-threshold slope

factor coefficient, etc. This enables two key

improvements:

o to take advantage of the higher g

m

(and

lower thermal noise power spectral

density) for a given current of analog

CMOS 28 nm SP MOST against 0.13 µm

and beyond;

o to use on-chip DSD stages without

exceeding in extra-power (by ultra-low

digital dynamic power for Standard-

Process MOST);

rejecting the flicker noise by maintaining the

electrode area equal to 100 µm

2

(approximately

three times lower than neuron area) enabling

about 1 k neuron recording for 1 mm

2

active

MEA area.

This paper is organized as follows. Section 2

describes the cross-section layers of the CMOS 28 nm

EOMOSFET pixel and all relevant signal/noise

transfer functions as a function of the interface

electrical parameters. Section 3 presents the

BIODEVICES 2021 - 14th International Conference on Biomedical Electronics and Devices

68

Figure 4: EOMOSFET Single Pixel Cross-Section.

simulation results of the model and the selected

design point that allows to perform >9 dB SNR with

< 12 µV

RMS

neuro-potentials signals power. At the

end of the paper conclusions will be drawn.

2 EOMOSFET CROSS-SECTION

AND ELECTRICAL

CHARACTERISTICS

EOMOSFET biosensors are spatially organized in a

matrix of pixels. The proposed model is based on the

the single pixel electrical scheme shown in Figure 4.

The neuron population and the silicon die separated

by an electrolyte bath (NaCL at 0.1 Moles

concentration). The TiO

2

post-processing layer

(Cianci, 2012) isolates the silicon circuits from the

biological environment. The extracellular ionic

currents flow by the electrolyte equivalent resistance

(R

E

) inducing a small voltage variation. Thus, the

voltage across R

E

is coupled with the TiO

2

capacitance (C

TiO2

) by the C

D

-C

H

series, where C

D

and

C

H

are the double-layer region capacitances (Diffuse

and Helmotz layers capacitances, respectively

(Massobrio, 2016)). Just beneath the TiO

2

there is an

on-chip metal electrode, whose area (A

ELE

) is in this

case 100 µm

2

(here M

SENSE

area is 50% lower to

maintain the imaging resolution). This oxide layer has

approximately 6 nm thickness and builds a specific

series capacitance (C

TiO2

).

Notice that this scheme is referred to the worst

case scenario of scarce neuron-chip adhesion

(Massobrio, 2016), where neuro-potentials signals

are very weak and are spread across the electrolyte

bath. Thus the voltage source (v

in

) models both EAP

and LFP signals.

Thus, the equivalent capacitance (C

DHT

) between

the cells and the MOS transistor (MOST) M

SENSE

gate

is given by eq. (1):

𝐶

/

/

/

(1)

C

D

and C

H

capacitances depend on the charge

concentration at the electrolyte-oxide interface.

More specifically such Helmoltz and diffuse

layers capacitances can be calculated using the metal

electrode area (A

ELE

), the water permittivity

(ε

w

=78.4ꞏε

0

where ε

0

is the vacuum permittivity equal

to 8.85 pF/m) and two physical lengths L

D

(Deybe

Detection of <12 VRMS Extracellular Action Potential and Local Field Potential by Optimum Design of a Single Pixel

Electrolyte-Oxide-MOSFET Interface in CMOS 28 nm

69

Figure 5: Signal Frequency Response.

Table 1: Neuron-Silicon Electrical Model Parameters.

Parameter Explanation Value

A

ELE

Metal Electrode Area 100 µm

2

R

E

Electrolyte Bulk Resistance 125 kΩ

C

D

Stern Ca

p

acitance 35.9

p

F

C

H

Helmotz Ca

p

acitance 17.3

p

F

C

T

TiO2 Ca

p

acitance 3.3

p

F

M

SENSE

W

S

/L

S

M

SENSE

Aspect Ratio

50 µm / 1

µ

m

g

ms

M

SENSE

Transconductance 50

µ

A/V

r

ds

M

SENSE

Out

p

ut Resistance 187 kΩ

τ

LP

Low-Pass Filter Dominant

Time-Constant

100 µs

(

10 kHz

)

R

F

Pseudo-Resistor Value 5 GΩ

length, equal to 1 nm) and x

2

(Stern length, equal to 2

nm) depending on the electrolyte-oxide interface

(Massobrio, 2016).

𝐶

∙𝐴

𝑎𝑛𝑑 𝐶

∙𝐴

(2)

The TiO

2

isolation layer capacitance has 45ꞏε

0

and 6

nm permittivity and thickness, respectively (Cianci,

2012):

𝐶

∙𝐴

(3)

Table 1 resumes the main values of the neuron-

electronics junction electrical model.

Hence, the M

SENSE

MOST is the core of a

dedicated neural Low-Noise-Amplifier (LNA) that

drives an ideal low-pass filter whose main aim is to

limit the signal bandwidth at 10 kHz. M

SENSE

small-

signal current is proportional (by its transconductance

(g

ms

)) to the induced extracellular AP and LFP (v

in

).

The M

SENSE

MOST is here biased by a feedback

pseudo-resistor (M

F

) to synthesize a very high

resistance (R

F

in the order of few GΩ) that sets the

low frequency pole (where LNA bandwidth starts).

The electrolyte bulk plays a key role for both

signal and noise transfer function. Its equivalent

resistance R

E

depends on the electrolyte bulk

conductibility k

’

and on the metal electrode area A

ELE

(Deen, 2006) as follows:

𝑅

∙

𝜋/𝐴

(4)

Assuming the same number of carriers (N

C

=N

P

=N

N

)

for both positive (N

P

, cations) and negative (N

N

,

anions) charges, then the bulk electrolyte

conductivity is related with the moles concentration

(N

MOL

) and the water density (ρ=1000 Kg/m

3

) as

expressed in eq. 5 (Park, 2016):

𝑘

𝑞∙

𝜇

∙𝑁

𝜇

∙𝑁

𝑞∙

𝜇

𝜇

∙𝑁

𝑞∙

𝜇

𝜇

∙𝜌∙𝑁

(5)

where µ

P

-µ

N

are the mobility coefficients for cations-

anions, respectively. Thus, the bulk electrolyte

conductivity is equal to 12.6 mA/(Vꞏm) and the bulk

resistance R

E

is then 125 kΩ at 100 mM.

2.1 EAP and LFP Signal Transfer

Function

Previous considerations definitively fix the numerical

values of all parameters involved in neuron-

electronics junction as a function of the physical size

and characteristics of the several layers composing

the EOMOSFET pixel. It is thus possible to calculate

the small signal transfer function for all relevant

signal and noise contributions. This will provide a

frequency domain map of the achievable Signal-to-

Noise-Ratio and, more importantly, a clear limit in

terms of maximum allowable noise power for analog

stages with the main aim to avoid significant SNR

degradation.

Input signal v

in

includes both EAP and LFP neuro-

potentials. The transfer function between input signal

BIODEVICES 2021 - 14th International Conference on Biomedical Electronics and Devices

70

Figure 6: Noise Sources Frequency Response.

(v

in

) and Low-Pass Filter (LPF) output node (v

out,lpf

/v

in

in Laplace domain) is given by eq. 6:

,

𝑠

∙

∙

∙

∙

∙

∙

∙

∙

∙

∙

(6)

where g

ms

and r

ds

are M

SENSE

MOST small-signal

parameters, and τ

LP

is the low-pass filter dominant

time constant (fixing the LPF -3dB passband at 10

kHz). Figure 5 shows the corresponding eq. 6

frequency response and includes the analog signal

processing (Neural Amplifier and Low-Pass Filter)

frequency response. Thanks to the very high

resistance synthesized by pseudo-resistors (5 GΩ),

the dominant pole time constant (C

HDT

ꞏR

F

/(g

ms

ꞏr

ds

))

has a frequency of 136 Hz and more importantly the

gain is higher than 0 dB (i.e.) from 13 Hz (channel

starts to amplify the input signal) enabling recording

of ultra-weak slow-oscillation and/or low-frequency

neuro-potentials.

2.2 Noise Transfer Function

Figure 6 shows the EOMOSFET pixel equivalent

circuit with the most relevant noise sources (coming

from electrolyte resistance R

E

(v

n,Re

), TiO

2

film

(v

n,TiO2

), feedback pseudo-resistor R

F

(v

n,RF

) and

sensing MOST M

SENSE

(v

n,MSENSE

)). The analog front-

end is composed by the cascade of a neutral Low-

Noise-Amplifier (LNA, for neural signal read-out and

amplification) driving a Low-Pass Filter (LPF) at 10

kHz -3dB bandwidth for signal selection and out-of-

band noise rejection. The amount of noise at the

output of the analog signal processing chain (v

out,lpf

,

i.e. the LPF output node) depends on the specific

noise power of each noise source and the frequency

response associated to each noise source.

2.2.1 R

E

Thermal Noise

The electrolyte resistance (R

E

) generates thermal

noise (v

n,RE

) whose in-band power spectral density is

given by eq. 7:

〈

,

〉

4∙𝑘∙𝑇∙𝑅

(7)

k and T are Boltzman constant and temperature (300

K) and R

E

is 125 kΩ as reported in eq.4 and Table 1.

The v

out

/v

n,RE

transfer function (in Laplace domain)

gives the total amount of voltage signal at the output

of the filter as a function of C

HDT

capacitance (C

H

, C

D,

and C

TiO2

serie), the feedback pseudo-resistor (R

F

)

and M

SENSE

MOST main small-signal parameters

(transconductance (g

ms

) and output resistance (r

ds

)):

,

,

𝑠

∙

∙

∙

∙

∙

∙

∙

∙

∙

∙

(8)

Figure 7 illustrates the corresponding frequency

response and demonstrates that R

E

thermal noise has

the same signal transfer function. This is even

demonstrated by Figure 8 where the noise power

spectral density (PSD) for every EOMOSFET noise

source with the total noise PSD is plotted.

Effectively R

E

noise PSD (green line) perfectly

overlap the v

in

signal bandwidth and thus it is one of

the most relevant ineliminable noise sources in

EOMOSFET pixels and more generally in planar

capacitively-coupled neural probes/MEAs.

2.2.2 TiO

2

Flicker Noise

TiO

2

film (used to separate or couple the biological

environment from silicon chip) mainly generates

Detection of <12 VRMS Extracellular Action Potential and Local Field Potential by Optimum Design of a Single Pixel

Electrolyte-Oxide-MOSFET Interface in CMOS 28 nm

71

Figure 7: Noise Sources Frequency Response.

Figure 8: Output Noise PSD.

flicker noise. Its power source is inversely

proportional to the frequency f, as follows:

〈

,

〉

,

∙

(9)

where k

pf,TiO2

is the specific flicker constant and C

TiO2

is the equivalent TiO

2

capacitance. v

n,TiO2

has the

same transfer function as v

n,RE

(eq. 8). In this model

TiO

2

has been preferred to other neuron-silicon

junction coupling options, because it has lower flicker

noise power comparing with Pt/Au noble metal films

that can exhibit an 1/f

2

noise, increasing the low

frequency noise power at the output of the neutral

amplifier.

Nonetheless the intrinsic neuron-electronics

junction ac-coupling allows some filtering of TiO

2

flicker at low frequency. Afterwards, starting from

136 Hz, v

n,TiO2

frequency response has 20 dB gain but,

at the same time, TiO

2

1/f noise behaviour has already

reduced its noise power. This generate a TiO

2

noise

PSD maximum at 132 Hz and, after this frequency,

the circuit attenuates the TiO

2

flicker noise power.

2.2.3 M

SENSE

Noise

The main noise contributions of the neural amplifier

come from M

SENSE

and R

F

(feedback pseudo-resistor).

M

SENSE

is a MOST biased in subthreshold region

synthesizing 50 µA/V transconductance g

ms

. This

value allows to minimize M

SENSE

thermal noise at few

µA current consumption and thus making feasible the

integration of the proposed EOMOSFET setup in

thousands of pixels resolution spatial grid/matrix

without excess of power. The equivalent model for

M

SENSE

noise source is:

〈

,

〉

∙

∙

∙4∙𝑘∙𝑇∙

(10)

where k

F

is the CMOS 28 nm flicker constant and W

S

and L

S

are M

SENSE

width and length. The transfer

function associated to the v

n,MSENSE

noise source is:

,

,

𝑠

≅

∙

∙

∙

∙

∙

∙

∙

∙

∙

∙

(11)

BIODEVICES 2021 - 14th International Conference on Biomedical Electronics and Devices

72

At low frequency (<< 13 Hz) the C

HDT

capacitor

behaves like an open circuit. M

SENSE

is then diode-

connected and thus its noise source transfers to the

output by unitary gain. At very high frequency (>> 13

Hz) C

HDT

features very low impedance (ideally a

short circuit), then the feedback is broken and

v

n,MSENSE

is directly applied between M

SENSE

gate and

ground. Thus output voltage noise is 20 dB higher.

One of the objectives of this model is to minimize

such noise contribution by acting on both M

SENSE

area

(WꞏL for flicker noise power reduction, here set at 50

µm

2

) and transconductance (for thermal noise power

reduction at 50 µA/V resulting in 14.8 nV/√Hz in-

band noise PSD).

2.2.4 R

F

Pseudo-resistor Noise

To enhance the pass-band gain of the EOMOSFET

pixel and to enable observation of ultra-weal ultra-

low frequency neuro-potential signals, the feedback

resistance must be in the GΩ order. Integrated

resistors in silicon technologies are not a feasible

option for such a large resistance value, hence this

model adopts a MOST in off region (pseudo-resistor)

where gate-source nodes are shorted, preventing any

conductive channel between drain and source and

exploiting the parasitic diodes formed by source-bulk

and drain-bulk junctions. Such diodes are

automatically biased in reverse region synthesising a

very high resistance (5 GΩ in this case). Starting from

these considerations, the R

F

pseudo-resistor noise

source has two main noise sources: shot (from diodes)

and flicker as follows:

〈

,

〉

∙

∙

(12)

〈

,

〉

2∙𝑞∙𝐼

(13)

where W

F

and L

F

are width and length of the pseudo-

resistor MOST, I

LEAK

is the leakage current mainly

coming from M

SENSE

gate oxide and approximately

equal to few pA. Notice that since R

F

is connected in

feedback, its noise is relevant at very low frequency

(< 136 Hz). So, this effect will be important in

13 Hz – 136 Hz bandwidth. The transfer functions for

v

n,RF

and i

n,RF

noise sources are respectively:

,

,

𝑠

≅

∙

∙

∙

∙

∙

∙

∙

∙

∙

∙

(14)

,

,

𝑠

≅𝑅

∙

∙

∙

∙

∙

∙

∙

∙

∙

∙

∙

(15)

Table 2: Noise Summary.

Bandwidth Noise Source Value [

µ

V

RMS

]

< 300 Hz

R

E

5.5

TiO

2

15.7

R

F

10.4

M

SENSE

1.73

Total Noise 19

300 Hz – 5 kHz

R

E

30

TiO

2

27.2

R

F

1

M

SENSE

9.5

Total Noise 42

< 5 kHz

R

E

30.6

TiO

2

31.4

R

F

10.4

M

SENSE

9.66

Total Noise 46.08

3 NOISE SUMMARY AND

SIGNAL-TO-NOISE-RATIO

The small-signal electrical and analytical model

based on Figure 6 scheme allow to easily evaluate the

noise and signal behaviour across the EOMOSFET

pixel.

More in details, by reducing the M

SENSE

MOST

thermal/flicker noise power and taking into account

both shot and flicker noise sources coming from MF

pseudo-resistor, the presented setup enables very low

noise performances (1.9 µV

RMS

and 4.2 µV

RMS

) at the

electrode (or at the M

SENSE

gate node).

Table 2 presents the noise summary of this

specific setup where, in both LFP and EAP

bandwidths, dominant noise contributions come from

electrolyte bath (thermal) and TiO

2

(flicker) coupling.

In other words by a dedicated design of the

interface in terms of both biological and circuital

electrical parameters, it is possible to set MOSTs

(M

SENSE

and M

F

) dc current and aspect ratio with the

main aim to minimize read-out noise power and in

first approximation remaining with the only interface

noise due to the electrolyte-electrode junction.

Effectively in LFP bandwidth, total noise

(measured at the low-pass filter output node) is 19

µV

RMS

(with 5.5 µV

RMS

R

E

noise power and 15.7

µV

RMS

TiO

2

noise power, respectively).

In EAP bandwidth, total output noise is 42 µV

RMS

when R

E

and TiO

2

have 30 µV

RMS

and 27.2 µV

RMS

noise power, respectively.

About SNR performances, Table 3 reports a final

resume of the achieved SNR in both EAP and LFP

bandwidths, also including Noise Figure

performances.

Electrolyte-Oxide-MOSFET Interface in CMOS 28 nm

73

Table 3: Noise Summary, SNR and Noise Figure.

Parameter Explanation Value

< 300 Hz Output Noise LFP Bandwidth Output Noise 19 µV

RMS

< 300 Hz Electrolyte Noise LFP Bandwidth TiO

2

+R

E

Output Noise 16.63 µV

RMS

< 300 Hz Electrode SNR at v

in

=10 µV

RMS

LFP Bandwidth SNR at the Electrode

(considering only TiO

2

and R

E

Noise)

15.5 dB

< 300 Hz LPF SNR

LFP Bandwidth Output SNR

(at the LPF Output Node)

14.42dB

< 300 Hz NF LFP Noise Figure 1.15 dB

300 Hz - 5 kHz Output Noise EAP Bandwidth Output Noise 42 µV

RMS

300 Hz - 5 kHz Electrolyte Noise EAP Bandwidth TiO

2

+R

E

Output Noise 40.5 µV

RMS

300 Hz - 5 kHz Electrode SNR at v

in

=10 µV

RMS

EAP Bandwidth SNR at the Electrode

(considering only TiO

2

and R

E

Noise)

7.8 dB

300 Hz - 5 kHz LPF SNR

EAP Bandwidth Output SNR

(at the LPF Output Node)

7.5 dB

300 Hz - 5 kHz NF EAP Noise Figure 0.31 dB

Figure 9: SNR vs. Frequency for LFP and EAP.

Assuming 10 µV

RMS

LFP signal at electrolyte bath

level (state-of-the-art (DeBusschere, 2001; Huys,

2012; Frey, 2010; Maccione, 2013; Eversmann,

2003; Wang, 2019; Lopez, 2018; Dragas, 2017)

operates with >22 µV

RMS

detection threshold), the

SNR at the electrode (without considering the LNA

additional noise) is 15.5 dB and after amplification

14.42, resulting in very low noise figure of 1.15 dB.

More interestingly, since most of DSD detects neuro-

potential spikes with SNR≥9 dB (Shahid, 2009), then

the proposed EOMOSFET setup enables the

detection of < 10 µV

RMS

slow ultra-weak events.

At 10 µV

RMS

EAP, Noise Figure is 0.31 dB,

meaning that the system introduces a very negligible

egradation of the SNR.

Finally, Figure 9 shows the SNR vs. frequency

when LFP/EAP input signal power equals the noise

power in the corresponding bandwidths (0 dB SNR)

and when such signal power is 9 dB higher than the

noise.

Effectively, the proposed analog read-out has a

certain frequency-dependent channel response

(mainly due to the R

F

low frequency bandwidth

limitations). Thus, some neuro-potential signals can

experience different gain values and, at the same

noise power, this results in different SNR.

BIODEVICES 2021 - 14th International Conference on Biomedical Electronics and Devices

74

Maximum 7.5dB SNR is achieved at 201 Hz with

5.5 µV

RMS

LFP signal power. For EAP neuro-

potentials, SNR reaches 9 dB with 11.8 µV

RMS

input

power, against 22 µV

RMS

state-of-the-art in Figure 1

and Figure 2.

Hence, this demonstrates that the presented setup

can be adopted for thousands of pixels resolution

MEAs with the key advantadge of improving the

noise performances and thus decreasing the minimum

detectable signals power.

4 CONCLUSIONS

In this paper a complete electrical model of a single-

pixel Electrolyte-Oxide MOS Field-Effect-

Transistors neural interface has been presented. The

model includes all biological and electrical

parameters building the interface. Thanks to specific

noise and signal simulation results, the proposed

setup allows optimum design and sizing of all MOS

transistors embedded in the analog signal processing,

minimizing noise power, and enabling ultra-weak

slow oscillation detection. More specifically the

proposed optimum design features 9 dB SNR at 11.8

µV

RMS

extra-cellular Action Potentials power and 7.8

dB SNR for 5.5 µV

RMS

Local Field Potentials, at the

electrode node.

ACKNOWLEDGEMENTS

This work has been supported by Brain28 PRIN

Project founded by the Italian Ministry of the

University, Education and Research.

REFERENCES

Obien, M. E. J., Deligkaris, K., Bullmann, T., Bakkum, D.

J., & Frey, U. (2015). Revealing neuronal function

through microelectrode array recordings. Frontiers in

neuroscience, 8, 423.

Thomas Jr, C. A., Springer, P. A., Loeb, G. E., Berwald-

Netter, Y., & Okun, L. M. (1972). A miniature

microelectrode array to monitor the bioelectric activity

of cultured cells. Experimental cell research, 74(1), 61-

66.

Pine, J. (1980). Recording action potentials from cultured

neurons with extracellular microcircuit electrodes.

Journal of neuroscience methods, 2(1), 19-31.

Gross, G. W., Williams, A. N., & Lucas, J. H. (1982).

Recording of spontaneous activity with photoetched

microelectrode surfaces from mouse spinal neurons in

culture. Journal of neuroscience methods, 5(1-2), 13-

22.

Cianci, E., Lattanzio, S., Seguini, G., Vassanelli, S., &

Fanciulli, M. (2012). Atomic layer deposited TiO2 for

implantable brain-chip interfacing devices. Thin solid

films, 520(14), 4745-4748.

Vallicelli, E. A., Reato, M., Maschietto, M., Vassanelli, S.,

Guarrera, D., Rocchi, F., ... & De Matteis, M. (2018).

Neural Spike Digital Detector on FPGA. Electronics,

7(12), 392.

Shahid, S., Walker, J., & Smith, L. S. (2009). A new spike

detection algorithm for extracellular neural recordings.

IEEE Transactions on Biomedical Engineering, 57(4),

853-866.

DeBusschere, B. D., & Kovacs, G. T. (2001). Portable cell-

based biosensor system using integrated CMOS cell-

cartridges. Biosensors and Bioelectronics, 16(7-8),

543-556.

Huys, R., Braeken, D., Jans, D., Stassen, A., Collaert, N.,

Wouters, J., ... & Verstreken, K. (2012). Single-cell

recording and stimulation with a 16k micro-nail

electrode array integrated on a 0.18 μm CMOS chip.

Lab on a Chip, 12(7), 1274-1280.

Frey, U., Sedivy, J., Heer, F., Pedron, R., Ballini, M.,

Mueller, J., ... & Kirstein, K. U. (2010). Switch-matrix-

based high-density microelectrode array in CMOS

technology. IEEE Journal of Solid-State Circuits, 45(2),

467-482.

Maccione, A., Simi, A., Nieus, T., Gandolfo, M., Imfeld,

K., Ferrea, E., ... & Berdondini, L. (2013, June).

Sensing and actuating electrophysiological activity on

brain tissue and neuronal cultures with a high-density

CMOS-MEA. In 2013 Transducers & Eurosensors

XXVII: The 17th International Conference on Solid-

State Sensors, Actuators and Microsystems

(Transducers & Eurosensors XXVII) (pp. 752-755).

IEEE.

Eversmann, B., Jenkner, M., Hofmann, F., Paulus, C.,

Brederlow, R., Holzapfl, B., ... & Gabl, R. (2003). A

128× 128 CMOS biosensor array for extracellular

recording of neural activity. IEEE Journal of Solid-

State Circuits, 38(12), 2306-2317.

Park, C. H., & Chung, I. Y. (2016). Modeling of Electrolyte

Thermal Noise in Electrolyte-Oxide-Semiconductor

Field-Effect Transistors. Journal of Semiconductor

Technology and Science, 16(1), 107.

Wang, S., Garakoui, S. K., Chun, H., Salinas, D. G., Sijbers,

W., Putzeys, J., ... & Lopez, C. M. (2019). A Compact

Quad-Shank CMOS Neural Probe with 5,120

Addressable Recording Sites and 384 Fully Differential

Parallel Channels. IEEE transactions on biomedical

circuits and systems, 13(6), 1625-1634.

Lopez, C. M., Chun, H. S., Wang, S., Berti, L., Putzeys, J.,

Van Den Bulcke, C., ... & Van Helleputte, N. (2018). A

multimodal CMOS MEA for high-throughput

intracellular action potential measurements and

impedance spectroscopy in drug-screening

applications. IEEE Journal of Solid-State Circuits,

53(11), 3076-3086.

Electrolyte-Oxide-MOSFET Interface in CMOS 28 nm

75

Dragas, J., Viswam, V., Shadmani, A., Chen, Y., Bounik,

R., Stettler, A., ... & Hierlemann, A. (2017). In

vitromulti-functional microelectrode array featuring 59

760 electrodes, 2048 electrophysiology channels,

stimulation, impedance measurement, and

neurotransmitter detection channels. IEEE journal of

solid-state circuits, 52(6), 1576-1590.

Harrison, R. R., & Charles, C. (2003). A low-power low-

noise CMOS amplifier for neural recording

applications. IEEE Journal of solid-state circuits, 38(6),

958-965.

Sansen, W. M. (2007). Analog design essentials (Vol. 859).

Springer Science & Business Media.

Baschirotto, A., Delizia, P., D’Amico, S., Chironi, V.,

Cocciolo, G., & De Matteis, M. (2009). Low power

analog design in scaled technologies.

De Matteis, M., D'Amico, S., & Baschirotto, A. (2006,

June). Power-minimization design procedure for Rauch

biquadratic cells. In 2006 Ph. D. Research in

Microelectronics and Electronics (pp. 141-144). IEEE.

Massobrio, P., Massobrio, G., & Martinoia, S. (2016).

Interfacing cultured neurons to microtransducers

arrays: a review of the neuro-electronic junction

models. Frontiers in neuroscience, 10, 282.

Deen, M. J., Shinwari, M. W., Ranuárez, J. C., & Landheer,

D. (2006). Noise considerations in field-effect

biosensors. Journal of applied physics, 100(7), 074703.

BIODEVICES 2021 - 14th International Conference on Biomedical Electronics and Devices

76