MICROPOWER NEURAL AMPLIFIER WITH ADAPTIVE INPUT-REFERRED NOISE

Abstract

A micropower neural amplifier with adaptive power biasing for use in multi-electrode arrays is provided. The micropower neural amplifier includes a low noise gain stage. The low noise gain stage is implemented using an amplifier and pseudoresistor elements.

Description

BACKGROUND OF THE INVENTION

The invention relates to the field of neural recording amplifier, and in particular to an ultra-low-power neural recording amplifier.

Large-scale chronic multi-electrode neural-recording systems are being built to enable us to understand how the brain works. With the help of such systems, a number of experiments have shown that it is possible to predict intended limb movements by simultaneously recording from many neurons, and interpreting their cortical activities. For example, brain-machine interfaces are being built to help a paralyzed patient move a computer cursor by thoughts alone. Portable, chronic use of such interfaces may eventually play an important role in treatment of paralyzed patients, and enable large-scale monitoring of the brain in experimental neuroscience.

One of the most important parts in the development of brain-machine interfaces is the neural signal amplifier. Neural signals from extracellular recording are very weak (typically between 10 μV and 500 μV). As a result, amplification is needed before such signals can be processed further. Next generation multi-electrode recording systems will be entirely implanted within the skull and incorporate a large number of neural amplifiers (on the order of 100-1000, one for every electrode). For such applications, ultra-low-power operation is very important to minimize heat dissipation in the brain, preserve long-battery life, and maximize the time between recharges. To get clean neural signal recordings, it is important that the input-referred noise of the amplifier be kept low. Practically, the input-referred noise of the amplifier should be kept below the background noise of the recording site (5 μV-10 μV). However, designers must address the tradeoff between low-noise and low-power designs of the amplifier. For an ideal thermal noise-limited amplifier with a constant bandwidth and supply voltage, the power of the amplifier scales as 1/υ²_n where υ_n is the input-referred noise of the amplifier. This relationship shows the steep power cost of achieving low-noise performance in an amplifier.

Many designs of neural amplifiers have been reported in the literature. Most amplifiers consume power near 100 μW to achieve less than 10 μV V_rms input-referred noise for bandwidths of 5-10 kHz. The designs consume power near 100 μW to achieve about 8-9 μV_rms input-referred noise with approximately 10 kHz of bandwidth. The design in [8] achieves an input-referred noise of 2.2 μV V_rms with 7.2 kHz of bandwidth while consuming 80 μW of power. If such amplifiers are to be used in a multi-electrode array, with a power near 100 μW per amplifier for most designs, the power required for the neural amplifiers can become the limiting factor for the whole multi-electrode system.

The best prior design presents many useful techniques for designing a neural amplifier: The use of an MOS-bipolar pseudoresistor element as a high-resistance element and on-chip AC-coupling capacitors enable the amplifier to reject large DC offsets at electrode-tissue interfaces while being able to pass the neural signals of interest. Since high-resistance elements can be implemented in a small area on chip, large off-chip components are not needed. The amplifier in uses a standard wide-output swing operational transconductance amplifier (OTA) with capacitive feedback to realize a gain of approximately 40 dB, and presents design techniques that minimize the input-referred noise of the amplifier by operating some devices of the OTA in strong inversion to minimize their noise contributions. Even though the latter design achieves a power-noise tradeoff near theoretical limit of its particular OTA topology, the topology used is not power-efficient since a large portion of the current in it is wasted in its current mirrors. The power efficiency of an amplifier can be greatly improved if a new OTA topology that makes efficient use of the supply current can be created.

SUMMARY OF THE INVENTION

According to one aspect of the invention, there is provided a micropower neural amplifier with adaptive power biasing for use in multi-electrode arrays. The micropower neural amplifier includes a low noise gain stage, the low noise gain stage implemented using an amplifier and pseudoresistor elements. A band pass filter stage shapes a pass band of the amplifier.

According to another aspect of the invention, there is provided a micropower neural amplifier with adaptive power biasing for use in multi-electrode arrays. The micropower neural amplifier includes a low noise gain stage. The low noise gain stage is implemented using an amplifier and pseudoresistor elements. The MOS-bipolar pseudoresistor element's noise being in the low frequencies, thereby allowing the low frequency noise to be filtered prior to pass band and does not appear in a frequency of interest. A band pass filter stage shapes a pass band of the amplifier.

According to another aspect of the invention, there is provided a micropower neural amplifier with adaptive power biasing for use in multi-electrode arrays. The micropower neural amplifier includes a low noise gain stage. The low noise gain stage is implemented using an amplifier and pseudoresistor elements. The amplifier is implemented using a modified version of a standard folded-cascode topology. The modification involves biasing an operational transconductance amplifier such that only currents in transistors of a folded branch are only a fraction of the current in the input differential pair transistors. A band pass filter stage shapes a pass band of the amplifier.

According to another aspect of the invention, there is provided a method of forming a micropower neural amplifier with adaptive power biasing for use in multi-electrode arrays. The method includes implementing a low noise gain stage using an amplifier and pseudoresistor elements. The pseudoresistor element's noise being in the low frequencies, thereby allowing the low frequency noise to be filtered prior to pass band and does not appear in a frequency of interest. Also, the method includes implementing a band pass filter stage shapes a pass band of the amplifier.

According to another aspect of the invention, there is provided a method of performing operations of a micropower neural amplifier with adaptive power biasing for use in multi-transconductance amplifier and pseudoresistor elements in a low noise gain stage. The pseudoresistor element's noise is in the low frequencies, thereby allowing the low frequency noise to be filtered prior to pass band and does not appear in a frequency of interest. Also, the method includes shaping a pass band of the amplifier in a band pass filter stage.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a schematic diagram of the inventive neural amplifier;

FIG. 2 is a block diagram illustrating the neural amplifier including the input noise source;

FIG. 3 is a schematic diagram illustrating a low-noise OTA used in accordance with the invention;

FIG. 4 is a circuit schematic for analyzing current scaling in the source-degenerated current mirrors;

FIG. 5 is a circuit schematic illustrating a standard folded-cascode OTA used in accordance with the invention;

FIGS. 6A-6B are circuit schematics for obtaining admittance formulas in accordance with the invention;

FIGS. 7A-7B are circuit schematics for analyzing mismatches in source-degenerated current mirrors used in accordance with the invention;

FIG. 8 is a die micrograph of the inventive neural amplifier;

FIGS. 9A-9B are graphs illustrating the transfer function of the neural amplifier configured for recording neural spikes;

FIG. 10 is a graph illustrating measured and simulated input-referred noise spectra of the neural amplifier configured for recording neural spikes;

FIG. 11 is a graph illustrating noise efficiency factor (NEF) of neural amplifiers as a function of the total supply current;

FIGS. 12A-12B are graphs illustrating CMRR and PSRR measurements of the neural amplifier configured for recording action potentials;

FIGS. 13A-13C illustrate a neural recording from a zebra finch's brain;

FIG. 14 is a graph illustrating the transfer function of the amplifier configured for recording Local Field Potentials (LFPs); and

FIG. 15 is a graph illustrating the measured and simulated input-referred noise spectra for the amplifier configured for recording LFP.

DETAILED DESCRIPTION OF THE INVENTION

The invention involves an ultra-low-power neural recording amplifier. The amplifier appears to be the lowest power and most energy-efficient neural recording amplifier reported to date. Low-noise design techniques are used that help the neural amplifier achieve input-referred noise that is near the theoretical limit of any amplifier using a differential pair as an input stage.

The overall schematic of a neural amplifier 2 is shown in FIG. 1. The neural amplifier 2 includes a bandpass filter stage following a gain stage 6 to shape the passband of the amplifier 2. The low-frequency high-pass cutoff of the gain stage 6 is created by the MOS-bipolar pseudoresistor element formed by M_b1-M_b2 and the capacitance C_f. The capacitive feedback formed by C_f and C_in sets the midband gain of the amplifier 2 to approximately 40.8 dB. The high-pass cutoff and the low-pass cutoff frequencies of the amplifier 2 can be adjusted via V_tune and the bias current of the g_m-OTA in a bandpass-filter stage 4 respectively. With the addition of the bandpass-filter stage 4, the amplifier 2 can be configured to record either LFPs (<1 Hz to 100 Hz) or neural spikes (100 Hz to >1 kHz). The bandpass-filter stage 4 includes C and C_L to filter low frequency noise.

For low-bandwidth LFP recording, the bias current of the gain stage OTA 8 can be lowered to conserve power. It is worth mentioning that the high-pass cutoff frequency of the gain stage 6 should be kept as low as possible. Weak-inversion MOS transistors (M_b3-M_b4) are placed in parallel with C_f to create a high-pass filter with a cutoff frequency at a few hundred Hz introduces low-frequency noise that rolls off as 1/f² in power units due to the noise from the transistor being low-pass filtered by C_f. This low-frequency noise appears at the front-end and gets amplified by the gain of the amplifier thereby degrading the minimum detectable signal. In the design, however, the MOS-bipolar pseudoresistor element's noise is at very low frequencies since the MOS-bipolar pseudoresistor element has much higher impedance than a weak-inversion MOS transistor. Therefore, low-frequency noise due to this element is filtered out well before the passband and does not appear in the frequency band of interest.

The operation of the inventive amplifier 2 can easily be understood by the block diagram 14 of FIG. 2. The capacitance C_p,in is included to model parasitic gate capacitances at input terminals of the gain-stage OTA 8. The input referred-noise of the OTA 8 is modeled as a υ²_n term added to the system at the input of the gain-stage OTA 8. The gain-stage OTA 8 is used as a high-gain amplifier and is modeled by G_m and R₀ blocks where G_m and R₀ represents the transconductance and the output resistance of the gain-stage OTA 8 respectively. In the bandpass-filter stage, R_p is the resistance of the series PMOS transistors operating in the triode regime. The value of R_p is set by V_tune. The combination of C and R_p realizes the highpass cutoff frequency for the amplifier. From the small-signal block diagram in FIG. 2, assuming that G_mR₀ is much higher than 1, one can express the transfer function of the neural amplifier as

$\begin{array}{cc}H\left(s\right)=\frac{{\upsilon }_{\mathrm{out}}\left(s\right)}{{\upsilon }_{\mathrm{in}}\left(s\right)}=-\frac{C_{\mathrm{in}}}{C_f}·\frac{{\mathrm{sR}}_pC}{1+{\mathrm{sR}}_pC}·\frac{1}{1+{\mathrm{sg}}_mC_L}.& \left(1\right)\end{array}$

The midband gain of the amplifier is A_v=−C_in/C_f. The highpass cutoff frequency is at f_HP=1/(2πR_pC) whereas the lowpass cutoff frequency is at f_LP=g_m/(2πC_L). One can relate the input-referred noise υ²_n of the gain-stage OTA 8 to the input-referred noise υ²_n,amp of the overall amplifier as

$\begin{array}{cc}\stackrel{\_}{{\upsilon }_{n,\mathrm{amp}}^2}={\left(\frac{C_{in}+C_f+C_{p,in}}{C_{in}}\right)}^2·\stackrel{\_}{{\upsilon }_n^2}.& \mathrm{Eq}.2\end{array}$

The input-referred noise of the bandpass filter stage is insignificant and is not included in the block diagram since the gain of 40 dB of the gain stage alleviates the bandpass-filter stage's input-referred noise requirement. As a result, the power consumption of the bandpass filter stage 4 is much smaller than that of the gain stage 6.

Thus, to achieve low-noise performance, it is important to design the gain-stage OTA to have low input-referred noise.

The schematic of the low-noise OTA 16 is shown in FIG. 3. It is a modified version of a standard folded-cascode topology. The OTA 16 in FIG. 3 is biased such that the currents of the transistors in the folded branch M₇-M₁₂ are only a small fraction of the current in the input differential pair transistors M₁ and M₂. In the invention, the channel current in M₇-M₁₂ is scaled to approximately 1/16^th of the current in M₁ and M₂. The much lower current in M₇-M₁₂ makes the noise contributed by them negligible compared to that from M₁ and M₂. As a result, one can simultaneously lower the total current and the total input-referred noise of the OTA 16.

To ensure that such severe current scaling is achieved, one can carefully set the bias currents of M₅ and M₆ through the use of the bias circuit formed by M_b2, M_c2, and M_c3. The current sources M_b1, M_b2 are cascoded to improve their output impedances and thereby ensure accurate current scaling. They operate in strong inversion to reduce the effect of threshold voltage variations. The source-degenerated current mirrors formed by M_c3, M₅ and M₆ and resistors R₁ and R₂ set the currents in M₅ and M₆ such that the currents in M₇ and M₈ (the difference between the current in M₃ and M₅ and between the current in M₄ and M₆) are a small fraction of the currents in M₁ and M₂.

In order to save power in the bias circuit, the current scaling ratio between M_b1 and M_b2 is 16:1 (2I_B/32) as shown in FIG. 3. To set the currents in the folded-branch transistors to be (I_B/32), which is 1/16^th of the currents in differential-pair transistors, one can set the current in M₅ and M₆ to be (17I_B/32). Such current ratioing is achieved by making R₃ to be 17R₁/2=17R₂/2, and constructing M_c3 as a parallel combination of two unit transistors while M₅ and M₆ are each constructed from 17 unit transistors in parallel. To clarify this scaling further, the current mirror formed by M_c3, R₃, and M₅, R₁ in FIG. 3 is transformed into an equivalent circuit comprised of many source-degenerated unit transistors 18 as shown in FIG. 4. All source-degenerated unit transistors are identical and have the same gate voltage. For any gate voltage there is only one source voltage at which a unit resistor's current equals a unit transistor's current. Thus, the nominal channel currents in all unit transistors are identical and the total current in M₅ is 17/2 times the current in M_c3 as desired.

For the amplifier to have low input-referred noise, the transconductance G_m of the OTA needs to be maximal for a given current level. For the standard folded-cascode OTA 20 shown in FIG. 5, the impedance looking into the sources of M₅ and M₆ is much smaller than the impedance looking into the drains of M₁-M₄. As a result, the standard folded-cascode OTA 20 achieves an overall transconductance G_m near g_m, the g_m of M₁. However, if one lower s the current in M₅-M₁₀ to be a small fraction of the current in M₁ and M₂, the impedance looking into the sources of M₅ and M₆ can be a significant portion of the impedance looking into the drains of M₁-M₄ such that incremental currents do not almost all go through the sources of M₅ and M₆ in the current divider formed between the sources of M₅ and the drains of M₁ and M₃. Therefore, G_m is significantly less than g_m1.

In the standard folded-cascode topology 20 shown in FIG. 5, the current sources formed by M₃ and M₄ contribute a significant amount of noise due to their large channel currents. In the design, one can replace the current-source transistors M₃ and M₄ in FIG. 5 with source-degenerated current sources formed by M₅ and M₆ and degeneration resistors R₁ and R₂ as shown in FIG. 3. With an appropriate choice of degeneration resistance, the noise contributions from the source-degenerated current sources are mainly from the resistors and can be made much smaller than the noise contributions from MOS transistors operating at the same current level. Another benefit of using source-degenerated current sources is that the noise from resistors is mainly thermal noise while NMOS current sources contribute a large amount of 1/f noise unless they are made with very large area. As a result, the 1/f noise in the neural amplifier is mainly from the input differential pair. Therefore, the input-differential pair is made with large-area PMOS transistors, which have lower 1/f noise than similarly-sized NMOS transistors in most CMOS processes.

To achieve low input-referred noise, it is important that the transconductance of the OTA 20 be maximized for a given total current. The maximum transconductance of the standard folded-cascode OTA 20 that can be achieved is the transconductance of one of the transistors in the input-differential pair, say g_m1. As a result, it is advantageous to operate M₁ and M₂ in the subthreshold regime where a transistor's g_m is maximized for a given current level. Therefore, M₁ and M₂ need to have large W/L ratios. The lengths of M₁ and M₂ then need to be small such that their widths stay relatively small and the input capacitance of the amplifier is not too large.

In order to make sure that all the incremental current caused by the differential input goes through the sources of M₇ and M₈, the input differential-pair transistors is cascaded with M₃ and M₄ to increase their output impedances. The source-degenerated current sources formed by M₅ and R₁ and by M₆ and R₂ are designed to have large output impedances as well. The output impedances of the cascoded input-differential pair and the source-degenerated current sources need to be much larger than the impedance looking into the sources of M₇ and M₈ such that Gm is near g_m1.

Before the operation of the OTA 16 in FIG. 3 is analyzed, a brief review of two useful admittance formulas is needed. FIG. 6A shows the circuit 22 and small signal circuit 24 illustrating the admittance looking into the source of an MOS transistor when its drain is connected to an impedance to incremental ground. FIG. 6B shows the circuit 26 and small signal circuit 28 illustrating the admittance looking into the drain of a cascode transistor. Using a nodal analysis, one obtains the two admittances to be

$\begin{array}{cc}{G}_s=\frac{i_d}{{\upsilon }_s}=\frac{g_{s1}+1/r_{o1}}{1+Z_L/r_{o1}},& \mathrm{Eq}.3\\ G_d=\frac{i_d}{{\upsilon }_d}=\frac{1}{r_{o1}}·\left(\frac{1}{1+g_{s1}Z_L+Z_L/r_{o1}}\right)& \mathrm{Eq}.4\end{array}$

Let G_s3 be the admittance looking into the sources of M₃ and M₄, G_d5 be the admittance looking into the drains of M₅ and M₆, and G_s7 be the admittance looking into the sources of M₇ and M₈ of the OTA 16 in FIG. 3. One can express the transconductance G_m of the OTA 16 as

$\begin{array}{cc}{G}_m=g_{m1}·\left(\frac{G_{s7}}{G_{s7}+G_{d5}}\right)\left(\frac{G_{s3\mathrm{ro}1}}{1+G_{s3\mathrm{ro}1}}\right).& \mathrm{Eq}.5\end{array}$

One can express G_s3, G_s7 and G_d5 by using Eqs. 3 and 4 as

$\begin{array}{cc}{G}_{s3}=\frac{g_{s3}+1/r_{o3}}{1+1/\left(r_{o3}\left(G_{s7}+G_{d5}\right)\right)}\approx \frac{g_{s3}}{1+1/\left(r_{o3}\left(G_{s7}+G_{d5}\right)\right)},& \mathrm{Eq}.6\\ G_{s7}=\frac{g_{s7}+1/r_{o7}}{1+\left(1/g_{m11}\right)/r_{o7}}\approx \left(\frac{g_{m11}r_{o7}}{1+g_{m11}r_{o7}}\right)·g_{s7},\mathrm{and}& \mathrm{Eq}.7\\ G_{d5}=\frac{1}{r_{o5}}\frac{1}{1+R_1/r_{o5}+g_{s5}R_1}& \mathrm{Eq}.8\end{array}$

where g_si and r_oi are the incremental source admittance of M_i with its drain at incremental ground, and the output resistance of M_i respectively. The expressions from Eqs. 6-8 present the design constraints for sizing and biasing each device to achieve G_m close to g_m1. The size, the channel current and the simulated intrinsic gain (g_sr_o) of each transistor in the OTA 16 can be easily be computed. From Eq. 5, in order to make G_m close to g_m1, the ratios G_s7/(G_s7+G_d5) and G_s3ro1/(1+G_s3ro1) should be made as close to 1 as possible. The ratio G_s7/(G_s7+G_d5) represents the incremental current gain from the drain of M₃ and M₄ to the output. The incremental current gain from the input differential pair transistors to the drain of the cascode transistors M₃ and M₄ is G_s3ro1/(1+G_s3ro1).

In order to maximize the ratio G_s7/(G_s7+G_d5), one can try to make G_d5<<G_s7. Since M₁₁ and M₇ have the same channel current, g_m11≈g_m7. Therefore, g_m11r_o7≈g_m7r_o7>>1 and one can have G_s7≈g_s7. In order to make G_d5<<G_s7, one needs to minimize G_d5. From Eq. 8, one can minimize G_d5 by making r_o5 large and also making g_s5R₁>>1. Therefore, one can make M₅ and M₆ with large W/L ratios and with long channel lengths to achieve large g_s5 and r_o5 respectively. Then one can choose R₁ such that g_s5R₁>>1.

In order to maximize the ratio G_s3ro1/(1+G_s3ro1), one needs to make G_s3ro1>>1. From Eq. 6, G_s3 is approximately g_s3 if G_s7ro3 is much greater than 1. Since G_s7≈g_s7, one can have G_s7ro3≈g_s7ro3. Since the current in M₇ is about 1/16 of the current in M₃ and both transistors are operating in subthreshold, g_s7≈g_s3/16. From simulation, one can achieve g_s3ro3 of 119 which results in a g_s7ro3 of 7.43. The expression in Eq. 6 is thus reduced to G_s3≈0.88g_s3. Note that M₁ and M₃ have the same currents and the same channel lengths. Thus M₁ and M₃ should have r_o1=r_o3. As a result, G_s3ro1≈G_s3ro3≈(0·88983)r_o3=104. Therefore, the ratio G_s3ro1/(1+G_s3ro1) is close to 1. As a result, one is able to achieve G_m close to g_m1 even with sixteen-fold current scaling between the input differential-pair transistors and the folded-branch transistors.

The noise in cascode transistors typically contributes little to the overall noise in an OTA 16 because these transistors self shunt their own current noise sources: A cascode transistor's current noise is attenuated by a factor of 1/(1+g_sR)² where g_s is its incremental source transconductance and R is the effective source-degeneration resistance respectively. Therefore, the only noise sources that are significant in FIG. 3 are due to non-cascode transistors, i.e., the differential-pair input transistors M₁ and M₂, the resistors R₁ and R₂, and the current-mirror transistors M₁₁ and M₁₂. One can now perform an OTA noise analysis.

The admittances looking into the sources of M₃, M₅, and M₇ are approximately g_s3, g_s5, and g_s7 respectively. Then the current transfer function from each significant current noise source in the OTA to an incrementally grounded output can be calculated to be

$\begin{array}{cc}\begin{array}{c}\frac{\stackrel{\_}{i_{n,\mathrm{out}}^2}}{\stackrel{\_}{i_{n,M1}^2}}={\left(\frac{G_{s3\mathrm{ro}1}}{1+G_{s3\mathrm{ro}1}}·\frac{G_{s7}}{G_{s7}+G_{d5}}\right)}^2\\ \approx {\left(\frac{g_{s3\mathrm{ro}1}}{1+g_{s3\mathrm{ro}1}}·\frac{G_{s7}}{g_{s7}+G_{d5}}\right)}^2,\end{array}& \mathrm{Eq}.9\\ \begin{array}{c}\frac{\stackrel{\_}{i_{n,\mathrm{out}}^2}}{\stackrel{\_}{i_{n,R1}^2}}={\left(\frac{G_{s5}R_1}{1+G_{s5}R_1}·\frac{G_{s7}}{G_{s7}+G_{d3}}\right)}^2\\ \approx {\left(\frac{g_{s5}R_1}{1+g_{s5}R_1}·\frac{g_{s7}}{g_{s7}+G_{d3}}\right)}^2,\end{array}\mathrm{and}& \mathrm{Eq}.10\\ \frac{\stackrel{\_}{i_{n,\mathrm{out}}^2}}{\stackrel{\_}{i_{n,M11}^2}}=1& \mathrm{Eq}.11\end{array}$

Since the inventive circuit is biased such that g_s3ro1>>1, g_s5R₁>>1 and g_s7>>G_d5, G_d3 as explained herein, the expressions from Eqs. 9-11 are reduced to 1. For the following discussion, the MOSFET's current noise is modeled as

i²_n=4γkTg_m  Eq. 12

where k is Boltzmann's constant, T is the absolute temperature, g_m is the transconductance of the MOSFET, and γ=⅔ for above-threshold operation and γ=1/(2κ) for subthreshold operation. From this noise model, one can calculate the input-referred noise of the OTA 16 as the total output current noise divided by its transconductance g²_m1 to be

$\begin{array}{cc}\stackrel{\_}{{\upsilon }_n^2}=\frac{1}{g_m^2}\left(\frac{4{\mathrm{kTg}}_{m1}}{\kappa }+\frac{8\mathrm{kT}}{R_1}+\frac{16}{3}{\mathrm{kTg}}_{m11}\right)& \mathrm{Eq}.13\end{array}$

where M₁ and M₂ operate in weak inversion and M₁₁ and M₁₂ operate in strong inversion. Let IC be the inversion coefficient of the transistor which is defined as the ratio of its channel current I_D to the moderate inversion characteristic current I_S where I_S is given by

$\begin{array}{cc}{I}_S=\frac{2\mu C_{\mathrm{ox}}U_T^2}{\kappa }·\frac{W}{L}& \mathrm{Eq}.14\end{array}$

where U_T is the thermal voltage and is equal to kT/q, where q is the electron charge. Using the EKV model, one can estimate the g_m of each transistor to be

$\begin{array}{cc}{g}_m=\frac{\kappa I_D}{U_T}·\frac{2}{1+\sqrt{1+4·\mathrm{IC}}}.& \mathrm{Eq}.15\end{array}$

One can rewrite Eq. 13 as

$\begin{array}{cc}\stackrel{\_}{{\upsilon }_n^2}=\frac{1}{g_{m1}}·\frac{4\mathrm{kT}}{\kappa }\left(1+\frac{2U_T}{I_1R_1}+\frac{4}{3}\kappa \alpha \frac{I_{11}}{I_1}\right)& \mathrm{Eq}.16\end{array}$

where α=2/1+√{square root over (1+4·IC₁₁)}, which is less than 1, and IC₁₁ is the inversion coefficient of M₁₁ and M₁₂. Eq. 16 suggests that in order to minimize the input-referred noise of the OTA, I₁R₁ should be large compared to 2U_T. Furthermore, the current ratio I₁/I₁₁ should be large compared to 4/3κα. For the implementation, the ratio I₁/I₁₁ is 16. For a total supply current of 2.7 μA and 5.3 kHz bandwidth, I₁ and I₁₁ are approximately 1.18 μA and 68 nA respectively. For R1=240 kΩ, the second and the third terms in Eq. 16 are 1.8×10⁻¹ and 5.4×10⁻² respectively, assuming a temperature of T=300 K, κ=0.7 and α=1. Equivalently, Eq. 13 is reduced to

$\begin{array}{cc}\stackrel{\_}{{\upsilon }_n^2}=\frac{2\mathrm{kT}}{\kappa g_{m1}}\times 2.47& \mathrm{Eq}.17\end{array}$

Eq. 17 can be interpreted as 2.47 times the input-referred noise of a MOS transistor operating in weak inversion with a transconductance of g_m1. This means that the OTA effectively has only 2.47 subthreshold devices that contribute noise. This value is close to the theoretical limit of 2 noise sources in any OTA that uses two subthreshold MOS differential-pair transistors as an input stage. Effectively, the design has minimized almost all other sources of noise except for that of M₁ and M₂.

The key techniques for achieving good power-noise tradeoff in the amplifier are the uses of source-degenerated current mirrors and the severe current scaling ratio between the input-differential pair transistors and the folded-branch transistors. The severe current scaling scheme can work only if the current errors due to mirroring are well controlled: The amplifier would not work if the error due to current scaling is too large such that none of the current flows in M₇-M₁₂ in the OTA of FIG. 3. Thus, one can address and investigate this concern to ensure the correct operation of the amplifier.

Now consider the current matching between two unit transistors in FIG. 4 due to variations in the threshold voltage and variations in the source-degeneration resistance. These variations are modeled as errors in the parameters of each of the unit transistors of FIG. 4. Let the nominal current in one of the unit transistors of M_c3 be I_D and consider the deviation in current ΔI_D in one of the unit transistors of M₅ from its nominal value due to deviations in the threshold voltage ΔV_T and deviations in the source-degeneration resistor ΔR as shown in circuit diagram 30 of FIG. 7A. To model the threshold-voltage mismatch, one can use the body-referenced current equation in saturation for an MOS transistor operating in weak inversion. Let the nominal current in each unit transistor be described by

I _D =I _s e ^{κ(V GS −V T )/U T} ·e ^{((1−κ)V BS /U T )}.  Eq. 18

where I_s is a constant scaling current which is the same for all unit transistors. Let V be the nominal DC voltage drop across R such that I_D=V/R. One can define

$\begin{array}{cc}{g}_T=\frac{\partial I_D}{\partial V_T}=-\frac{\kappa }{U_T}·I_D=-g_m,& \mathrm{Eq}.19\\ g_R=\frac{\partial I_D}{\partial R}=\frac{\partial }{\partial R}\left(\frac{V}{R}\right)=-\frac{1}{R}·\frac{V}{R}=-\frac{I}{R}·I_D\mathrm{and}& \mathrm{Eq}.20\\ g_{\mathrm{mb}}=\frac{\partial I_D}{\partial V_{\mathrm{BS}}}=\frac{1-\kappa }{U_T}·I_D.& \mathrm{Eq}.21\end{array}$

By assuming that ΔV_T and ΔR are small, one can use a small-signal circuit model 32 as shown in FIG. 7B to calculate the variation in nominal current ΔI_D when ΔV_T and ΔR are considered as inputs to the system. With some analysis, the variation in the channel current due to variations in V_T and R is obtained to be

ΔI _D =g _T ·ΔV _T−(g_m+g_mb+1/r_o)·(ΔI_D−g_R·ΔR)·R.  Eq. 22

Combining Eq. 22 with the results from Eqs. 19 and 20 and using the relationship g₅=g_m+g_mb, one can obtain the fractional change in channel current as a function of the fractional change in V_T and R to be

$\begin{array}{cc}\frac{\Delta I_D}{I_D}=-\frac{1}{1+g_sR+R/r_o}·\frac{\Delta V}{I_D/g_m}-\frac{g_sR+R/r_o}{1+g_sR+R/r_o}·\frac{\Delta R}{R}.& \mathrm{Eq}.23\end{array}$

Since M_c3, M₅ and M₆ are biased in weak-inversion regime, their I_D/g_m is approximately 40 mV at room temperature. As seen from (23), the mismatch in threshold voltage as a fraction of 40 mV is attenuated by a factor of 1+g_sR+R/r_o and is negligible if g_sR>>1. In the design, one can have g_sR≈12, thus, the fractional mismatch in threshold voltage is attenuated by more than a factor of 10 and does not play a significant role in current mirror mismatch. In contrast, the fractional mismatch in channel current scales almost 1:1 to the fractional mismatch in R. However, the matching of passive components in most CMOS processes is much better controlled than the matching of transistors' threshold voltages. In the design, therefore, one can try to achieve good resistor matching with careful layout.

To compare the power-noise tradeoff among amplifiers, one can adopt the noise efficiency factor (NEF) proposed in and widely used to compare neural-amplifier designs:

$\begin{array}{cc}NEF=V_{\mathrm{ni},\mathrm{rms}}\sqrt{\frac{2I_{\mathrm{tot}}}{\pi ·U_T·4\mathrm{kt}·BW}}& \mathrm{Eq}.24\end{array}$

where V_ni,rm5 is the total input-referred noise, I_tot is the total supply current, and BW is the −3 dB bandwidth of the amplifier respectively. The theoretical limit of the NEF of an OTA that uses a differential pair as an input stage is when the two differential-pair transistors are the only noise sources in the circuit. The input-referred noise of the OTA is then V2 ni=2×2kT/(κg_m)=4kT/(κg_m) where g_m is the transconductance of a single differential-pair transistor. For minimum input-referred noise, the transistors should run in subthreshold, such that g_m=κI_D/U_T. Assuming a first-order roll-off of the frequency response, the input-referred noise of the ideal OTA is expressed as

$V_{\mathrm{ni},\mathrm{rms}}=\sqrt{\frac{4\mathrm{kt}·U_T}{{\kappa }^2I_D}·\frac{\pi }{2}·BW}$

Combining Eqs. 24 and 25 and setting I_tot=2I_D, one can obtain the theoretical limit for NEF of any OTA that uses a subthreshold MOS differential pair to be

$NEF=\frac{\sqrt{2}}{\kappa }\approx 2.02$

assuming a typical value of κ=0.7. The experimental NEF is near this value, and the theoretical NEF was computed to be 2.47.

The inventive amplifier was fabricated in a 0.5 μm CMOS process through the AMI foundry. It was designed to give a gain of approximately 110 (40.8 dB) by setting the value of C_in to 14 pF and C_f to 120 fF. The OTA in the bandpass filter stage is a wide common-mode range OTA to reduce signal distortion in the case of large input amplitudes. The amplifier occupies a chip area of 0.16 mm². A chip micrograph of the amplifier 34 is shown in FIG. 8.

Four chips were tested on the lab bench and they exhibited very similar performance characteristics, indicating that the severe current-scaling scheme worked robustly. The measured transfer function of one of the neural amplifiers is shown in FIGS. 9A-9B. The amplifier consumes 2.7 μA including the current from the bias circuit (M_b2, M_c2 and M_c3) from a 2.8 V supply. The current I_bias shown in FIG. 3 is not included, since it can be shared by many amplifiers in the array. The −3 dB cutoff frequencies are adjusted to be at 45 Hz and 5.32 kHz. The amplifier is configured as an inverting amplifier, thus the phase is approximately −180° near the midband frequency.

FIG. 10 shows the measured input-referred noise spectrum together with a circuit simulation of the noise spectrum with a similar noise model to the theoretical calculations (the smooth curve). There is good agreement between the measured and simulated curves. The measured thermal noise level is 31 nV/Hz. Integrating under the area of the measured curve from 10 Hz to 98 kHz yields a total input-referred noise of 3.06 μV_rms, while the simulated result is 3.1 μV_rms. With a high-pass cutoff frequency at 45 Hz, 1/f noise is filtered out and is not noticeable in the passband.

The NEF of the amplifier is calculated from the achieved experimental measurements to be 2.67. This value is close to 2.02 which is the theoretical NEF limit and also near the expected theoretical calculation of 2.47. The good power-noise tradeoff of the amplifier is a result of minimizing the effective number of transistors that contribute noise. Moreover, almost all the power is consumed by the input-differential pair. Therefore, little power is wasted in less critical parts of the amplifier. FIG. 11 compares NEF of previously reported neural amplifiers as a function of the total supply current. The amplifier exhibits the best NEF and lowest power consumption reported to date.

The measured CMRR and PSRR are shown in FIGS. 12A-12B respectively. CMRR is calculated as the ratio of the differential-mode gain to the common-mode gain. PSRR is calculated as the ratio of the differential-mode gain to the gain from power supply to the output. The measured CMRR and PSRR exceed 66 dB and 75 dB (over the range of 45 Hz to 5.32 kHz) respectively.

It has been verified the neural amplifier works in a real recording environment by using it to record action potentials in the RA region of a zebra finch's brain, as shown in FIG. 13A. Data were taken with a Carbostar electrode that had an impedance of approximately 800 kΩ. A long extracellular trace and a short extracellular trace recorded from the amplifier normalized by the gain are shown in FIGS. 13B-13C. They were found to be identical to that recorded by a commercial neural amplifier.

Local Field Potentials (LFPs) instead of action potentials are often used in some brain-machine interfaces, for example, those used in paralysis prosthetics. Therefore, one can also measure the performance characteristics of the amplifier configured with lower bandwidth (and power) for such applications. Since the LFP contains energy in the frequency range of 1 Hz to 100 Hz, one can simply lower the −3 dB lowpass cutoff frequency of the amplifier by lowering the supply current of the OTA in the bandpass filter stage. The highpass cutoff frequency can also be lowered to be below 1 Hz by adjusting V_tune. If one can just change the bandwidth in this manner, the input-referred noise of the amplifier becomes excessively low. From a hand-analysis, if one can adjust the bandwidth of the amplifier to be 0.5 Hz-300 Hz while maintaining the same supply current for the gain-stage OTA, the input-referred noise of the amplifier is less than 1 μVrms. Such low input-referred noise is unnecessary and is wasteful of power. From, the input-referred noise power is inversely proportional to g_m1, therefore inversely proportional to I₁. Thus, more power can be saved by lowering the current in the gain-stage OTA as well.

The amplifier can be adjusted to have a highpass cutoff frequency of 392 mHz and a lowpass cutoff frequency of 295 Hz for LFP-suitable configuration. The total current of the amplifier was measured to be 743 nA, corresponding to a power consumption of 2.08 μW from a 2.8 V supply and 1.66 μVrms total input-referred noise integrated from 0.2 Hz to 1 kHz. The measured transfer function for the amplifier configured for recording LFP is shown in FIGS. 14A-14B. The measured input-referred noise spectrum and expected noise curve from simulation are shown in FIG. 15. The measured NEF for LFP recording is then 3.21, still better than any other amplifier to date. Note that the NEF is worse than that of the amplifier configured to record neural spikes. This degradation in NEF is due to the fact that the thermal noise from the resistors R₁ and R₂ becomes more significant once the current in the input differential pair is low. Moreover, 1/f noise becomes significant as well since the highpass cutoff has been decreased to 395 mHz.

The invention provides a micropower low-noise neural recording amplifier. Many low-noise design techniques were employed to enable the amplifier to achieve an input-referred noise near the theoretical limit of two devices of an input differential pair. The inventive amplifier appears to be the lowest power and most energy-efficient neural amplifier reported to date. It can be configured to record either action potentials or local field potentials. The inventive amplifier can thus be useful in brain-machine interfaces for paralysis prosthetics, visual prosthetics, or experimental neuroscience systems for chronic monitoring.

Although the present invention has been shown and described with respect to several preferred embodiments thereof, various changes, omissions and additions to the form and detail thereof, may be made therein, without departing from the spirit and scope of the invention.

Claims

1. A low-power amplifier with adaptive power biasing for use in multi-electrode arrays.

2. The low-power amplifier of claim 1, wherein said low power amplifier is implemented as a micropower amplifier array for neural signals.

3. The low-power amplifier of claim 1, wherein the adaptive power biasing scheme is based on statistical properties of the recorded signals.

4. A micropower neural amplifier with adaptive power biasing for use in multi-electrode arrays comprising: a low noise gain stage, said low noise gain stage implemented using amplifier and pseudoresistor elements; anda band pass filter stage to shape a pass band of the amplifier.

5. The micropower neural amplifier of claim 4, wherein said micropower neural amplifier of claim 1 is able to record LFPs and neural spikes.

6. The micropower neural amplifier of claim 4, wherein said amplifier is implemented using a modified version of a standard folded-cascode topology, the modification involving biasing an operational transconductance amplifier such that only currents in transistors of a folded branch are only a fraction of the current in the input differential pair transistors.

7. The micropower neural amplifier of claim 6, wherein reduction in currents is accomplished using a plurality of source-degenerated current mirrors.

8. The micropower neural amplifier of claim 4, wherein said amplifier is implemented in conjunction with a neuromotor prosthetic system.

9. A micropower neural amplifier with adaptive power biasing for use in multi-electrode arrays comprising: a low noise gain stage, said low noise gain stage implemented using an amplifier and pseudoresistor elements, said pseudoresistor element's noise being in the low frequencies, thereby allowing said low frequency noise to be filtered prior to pass band and does not appear in a frequency of interest; anda band pass filter stage to shape a pass band of the amplifier.

10. The micropower neural amplifier of claim 9, wherein said micropower neural amplifier of claim 1 is able to record LFPs and neural spikes.

11. The micropower neural amplifier of claim 9, wherein said amplifier is implemented using a modified version of a standard folded-cascode topology, the modification involving biasing an operational transconductance amplifier such that only currents in transistors of a folded branch are only a fraction of the current in the input differential pair transistors.

12. The micropower neural amplifier of claim 11, wherein reduction in currents is accomplished using a plurality of source-degenerated current mirrors.

13. The micropower neural amplifier of claim 9, wherein said amplifier is implemented in conjunction with a neuromotor prosthetic system.

14. A micropower neural amplifier with adaptive power biasing for use in multi-electrode arrays comprising: a low noise gain stage, said low noise gain stage implemented using an amplifier and pseudoresistor elements, said amplifier implemented using a modified version of a standard folded-cascode topology, the modification involving biasing an operational transconductance amplifier such that only currents in transistors of a folded branch are only a fraction of the current in the input differential pair transistors; anda band pass filter stage to shape a pass band of the amplifier.

15. The micropower neural amplifier of claim 14, wherein reduction in currents is accomplished using a plurality of source-degenerated current mirrors.

16. A method of forming a micropower neural amplifier with adaptive power biasing for use in multi-electrode arrays comprising: implementing a low noise gain stage using an amplifier and pseudoresistor elements, said pseudoresistor element's noise being in the low frequencies, thereby allowing said low frequency noise to be filtered prior to pass band and does not appear in a frequency of interest; andimplementing a band pass filter stage to shape a pass band of the amplifier.

17. The method of claim 16, wherein said micropower is able to record LFPs and neural spikes.

18. The method of claim 16, wherein said amplifier is implemented using a modified version of a standard folded-cascode topology, the modification involving biasing an operational transconductance amplifier such that only currents in transistors of a folded branch are only a fraction of the current in the input differential pair transistors.

19. The method of claim 18, wherein reduction in currents is accomplished using a plurality of source-degenerated current mirrors.

20. The method of claim 16, wherein said amplifier is implemented in conjunction with a neuromotor prosthetic system.

21. A method of performing operations of a micropower neural amplifier with adaptive power biasing for use in multi-electrode arrays comprising: utilizing an amplifier and pseudoresistor elements in a low noise gain stage, said pseudoresistor element's noise being in the low frequencies, thereby allowing said low frequency noise to be filtered prior to pass band and does not appear in a frequency of interest; andshaping a pass band of the amplifier in a band pass filter stage.

22. The method of claim 21, wherein said micropower is able to record LFPs and neural spikes.

23. The method of claim 21, wherein said amplifier is implemented using a modified version of a standard folded-cascode topology, the modification involving biasing an operational transconductance amplifier such that only currents in transistors of a folded branch are only a fraction of the current in the input differential pair transistors.

24. The method of claim 23, wherein reduction in currents is accomplished using a plurality of source-degenerated current mirrors.

25. The method of claim 21, wherein said amplifier is implemented in conjunction with a neuromotor prosthetic system.