A Low-Power Opamp-Less Second-Order Delta-Sigma Modulator for Bioelectrical Signals in 0.18 µm CMOS

This article reports on a compact and low-power CMOS readout circuit for bioelectrical signals based on a second-order delta-sigma modulator. The converter uses a voltage-controlled, oscillator-based quantizer, achieving second-order noise shaping with a single opamp-less integrator and minimal analog circuitry. A prototype has been implemented using 0.18 μm CMOS technology and includes two different variants of the same modulator topology. The main modulator has been optimized for low-noise, neural-action-potential detection in the 300 Hz–6 kHz band, with an input-referred noise of 5.0 μVrms, and occupies an area of 0.0045 mm2. An alternative configuration features a larger input stage to reduce low-frequency noise, achieving 8.7 μVrms in the 1 Hz–10 kHz band, and occupies an area of 0.006 mm2. The modulator is powered at 1.8 V with an estimated power consumption of 3.5 μW.


Introduction
Electrogenic cells, such as neurons, cardiac cells, or retinal cells, generate ionic currents across their membrane owing to the different ion channels that populate the cellular membranes. Transmembrane ionic currents produce voltage variations in the extracellular medium that can be detected by miniaturized sensors in close proximity to the cells. These voltage signals occur in different frequency bands and feature different amplitudes depending on their nature: neural action potentials (APs) manifest as spikes with amplitudes ranging from a few tens of µV to 1 mV and most signal power between 300 Hz and 6 kHz, whereas cardiac field potentials may reach tens of mV with most signal power in the 1 Hz-1 kHz band. The development of sensors, capable of simultaneously detecting action potentials of multiple cells, enables advanced electrophysiology studies, improving the understanding of complex signaling and opening paths to restoring lost functions. CMOS technology allows for low-noise recording from thousands of electrodes in parallel, either in vitro [1] or in vivo [2]. In vitro studies can be performed with CMOS microelectrode arrays (MEAs), the electrode array of which is co-integrated with the readout electronics [3]. In vivo interfaces commonly make use of passive probes connected to external readout arrays [2,4,5], although a variety of monolithic silicon probes combining electrodes and electronics on the same die have been developed [6,7].
In order to detect APs with acceptable signal-to-noise ratios, recording front-ends require input-referred noise values below 10 µV rms . This is especially relevant for neural interfaces, where small action potentials generated by different neurons need to be separated and assigned to the respective signal sources. CMOS technology allows for simultaneously recording from thousands of densely packed electrodes at a high spatiotemporal resolution, and it allows for conditioning and digitizing signals on-chip with low noise. Advanced signal post-processing (e.g., spike-sorting [8]) is frequently performed off-chip, due to the limited computing power of on-chip processors. This relieves the specifications of Figure 1 shows a simplified block diagram of the proposed readout circuit. The input stage transforms the electrode voltage (V el ) into a current (I IS ) by means of a high-pass filter and an inverting transconductor. This current is injected into a capacitor (C int ), which acts as the first integrator of the modulator. The capacitor voltage (V C ) drives a VCO-based quantizer consisting of a VCO and a 1-bit frequency-to-digital converter, which is sampled at 1 MHz. The output bitstream drives the 1-bit current DAC that generates the feedback current (I FB ) and closes the loop. The modulator resembles a first-order closed-loop ∆Σ architecture with an opamp-less integrator, using a VCO-based quantizer to achieve an additional order of noise shaping [19,21]. The sampling frequency is high enough to achieve an input-referred quantization noise below 2 µV rms in the 300 Hz-6 kHz band with a 1-bit, second-order modulator, which keeps the structure of the converter very simple. Furthermore, the continuous-time topology obviates the need for an anti-aliasing filter due to the inherent low-pass filtering before sampling.
The modulator is stable, provided that the feedback current can rapidly counterbalance I IS for any possible input signal so that the average current through C int is kept at zero. In normal operation, the capacitor voltage V C is an irregular triangular wave resembling the output of the first integrator of a continuous-time ∆Σ modulator. The average values of V C and Y are defined by the transfer functions of the feedback I DAC and the VCO-based quantizer. Under nominal conditions, (I IS ) = (I FB ) = 1.4 µA and (V C ) = V DD /2 = 0.9 V, which sets the average VCO oscillation frequency close to f s /2 = 500 kHz and Y = 0.5.  Figure 2 shows the schematic of the input stage, which consists of a high-pass filter followed by a single-ended transconductor. The high-pass filter is based on a metal-insulator-metal (MIM) capacitor (C0) and a PMOS pseudo-resistor (M0). C0 must be as large as possible since this capacitor forms a voltage divider with the input capacitance of the next stage (M1). However, given the relatively large size of MIM capacitors, the value of C0 is limited to either hundreds of femtofarads or a few picofarads, depending on the application. The modulator implemented can be programmed with either C0 = 4.25 pF (for fullband low-noise applications) or C0 = 350 fF (for compact action-potential readout). Given the small input capacitance, a very high-ohmic pseudo-resistor is needed to set the cut-off frequency of the high-pass filter well below 1 Hz. These filter characteristics are required to avoid signal attenuation or phase shifts in the band of interest but also to reduce the effect of the thermal noise generated by M0 since, as shown in Figure 3, the noise is low-pass-filtered by the RC circuit. As a consequence, the pseudo-resistor was tuned in the TΩ range.

Input Stage
The transconductor is based on a single PMOS transistor (M1) in weak inversion. The main transistor is complemented with a cascode (M2) to increase the output impedance and keep the output current independent from the output voltage VC, which can oscillate up to ±200 mV around 900 mV. Assuming small input voltages, the current generated by this transconductor follows  Figure 2 shows the schematic of the input stage, which consists of a high-pass filter followed by a single-ended transconductor. The high-pass filter is based on a metalinsulator-metal (MIM) capacitor (C 0 ) and a PMOS pseudo-resistor (M 0 ). C 0 must be as large as possible since this capacitor forms a voltage divider with the input capacitance of the next stage (M 1 ). However, given the relatively large size of MIM capacitors, the value of C 0 is limited to either hundreds of femtofarads or a few picofarads, depending on the application. The modulator implemented can be programmed with either C 0 = 4.25 pF (for full-band low-noise applications) or C 0 = 350 fF (for compact action-potential readout).  Figure 2 shows the schematic of the input stage, which consists of a high-pass filter followed by a single-ended transconductor. The high-pass filter is based on a metal-insulator-metal (MIM) capacitor (C0) and a PMOS pseudo-resistor (M0). C0 must be as large as possible since this capacitor forms a voltage divider with the input capacitance of the next stage (M1). However, given the relatively large size of MIM capacitors, the value of C0 is limited to either hundreds of femtofarads or a few picofarads, depending on the application. The modulator implemented can be programmed with either C0 = 4.25 pF (for fullband low-noise applications) or C0 = 350 fF (for compact action-potential readout). Given the small input capacitance, a very high-ohmic pseudo-resistor is needed to set the cut-off frequency of the high-pass filter well below 1 Hz. These filter characteristics are required to avoid signal attenuation or phase shifts in the band of interest but also to reduce the effect of the thermal noise generated by M0 since, as shown in Figure 3, the noise is low-pass-filtered by the RC circuit. As a consequence, the pseudo-resistor was tuned in the TΩ range.

Input Stage
The transconductor is based on a single PMOS transistor (M1) in weak inversion. The main transistor is complemented with a cascode (M2) to increase the output impedance and keep the output current independent from the output voltage VC, which can oscillate up to ±200 mV around 900 mV. Assuming small input voltages, the current generated by this transconductor follows Given the small input capacitance, a very high-ohmic pseudo-resistor is needed to set the cut-off frequency of the high-pass filter well below 1 Hz. These filter characteristics are required to avoid signal attenuation or phase shifts in the band of interest but also to reduce the effect of the thermal noise generated by M 0 since, as shown in Figure 3, the noise is low-pass-filtered by the RC circuit. As a consequence, the pseudo-resistor was tuned in the TΩ range.
The transconductor is based on a single PMOS transistor (M 1 ) in weak inversion. The main transistor is complemented with a cascode (M 2 ) to increase the output impedance and keep the output current independent from the output voltage V C , which can oscillate up to ±200 mV around 900 mV. Assuming small input voltages, the current generated by this transconductor follows where I0 is the DC biasing current, and G is the transconductance of M1 multiplied by the attenuation due to the capacitive voltage divider. The biasing current I0 chosen for all the measurements reported in this article is 1.4 μA, although the integrated circuit allows for tuning this current in the [200 nA, 1.6 μA] range to define different levels of power consumption and noise. Figure 3. Simulation of the output noise of an RC low-pass filter for C = 350 fF and two different resistor values: R = 2 GΩ (black) and R = 2 TΩ (blue). Although a higher resistance value yields higher power spectral density at low frequencies, the cut-off frequency is significantly reduced, which results in lower noise at high frequencies. In this example, the integrated noise in the actionpotential band (300 Hz-6 kHz) is 58 μVrms for R = 2 GΩ and only 2.0 μVrms for R = 2 TΩ.
Two variants of this circuit have been implemented in the prototype chip: one optimized for compact neural interfaces and a second variant optimized for low noise. The compact transconductor is coupled to the 350 fF input capacitor and consists of M1 = 30 μm/1.2 μm and M2 = 10 μm/1.2 μm. The voltage divider, formed by the input capacitance (350 fF) and the gate capacitance of M1 (95 fF), attenuates the input signal by 2 dB and prevents the use of a larger transistor. The resulting transconductance of the input stage is G ≈ 20 V −1 · I0. The alternative low-noise transconductor is coupled to the 4.25 pF input transistor to avoid any signal attenuation in the capacitive voltage divider. Therefore, the size of M1 was increased up to 70 μm/1.5 μm (with M2 = 10 μm/1.5 μm), which reduces flicker noise at low frequencies. The resulting transconductance is G ≈ 25 V −1 · I0.
Although this transconductor topology is sensitive to power-supply noise and process variations, inverts the input signal, and is inherently nonlinear, it can be very compact and potentially feature low noise. Process variations can cause minor gain variations, which can be coped with through calibration. Nonlinearity may cause the distortion of very large input signals, but action potentials are expected to be smaller than 1 mV. The power-supply rejection ratio (PSRR) is nearly 0 dB since the output current directly depends on the VSG of M1 and, therefore, on the supply voltage. Low-noise external voltage regulators are required to minimize the power-supply noise, and data post-processing can be included to attenuate powersupply noise (50/60 Hz harmonics) and other types of predictable noise. Note that in the case of a MEA with multiple copies of the same converter, the power-supply noise is common to all of them, which makes the extraction and subtraction of common noise during data postprocessing possible. The feasibility and robustness of compact, single-ended input stages for MEAs have also been recently proven in [12]. Simulation of the output noise of an RC low-pass filter for C = 350 fF and two different resistor values: R = 2 GΩ (black) and R = 2 TΩ (blue). Although a higher resistance value yields higher power spectral density at low frequencies, the cut-off frequency is significantly reduced, which results in lower noise at high frequencies. In this example, the integrated noise in the action-potential band (300 Hz-6 kHz) is 58 µVrms for R = 2 GΩ and only 2.0 µVrms for R = 2 TΩ. Two variants of this circuit have been implemented in the prototype chip: one optimized for compact neural interfaces and a second variant optimized for low noise. The compact transconductor is coupled to the 350 fF input capacitor and consists of M 1 = 30 µm/1.2 µm and M 2 = 10 µm/1.2 µm. The voltage divider, formed by the input capacitance (350 fF) and the gate capacitance of M 1 (95 fF), attenuates the input signal by 2 dB and prevents the use of a larger transistor. The resulting transconductance of the input stage is G ≈ 20 V −1 · I 0 . The alternative low-noise transconductor is coupled to the 4.25 pF input transistor to avoid any signal attenuation in the capacitive voltage divider. Therefore, the size of M 1 was increased up to 70 µm/1.5 µm (with M 2 = 10 µm/1.5 µm), which reduces flicker noise at low frequencies. The resulting transconductance is G ≈ 25 V −1 · I 0 .
Although this transconductor topology is sensitive to power-supply noise and process variations, inverts the input signal, and is inherently nonlinear, it can be very compact and potentially feature low noise. Process variations can cause minor gain variations, which can be coped with through calibration. Nonlinearity may cause the distortion of very large input signals, but action potentials are expected to be smaller than 1 mV. The power-supply rejection ratio (PSRR) is nearly 0 dB since the output current directly depends on the V SG of M 1 and, therefore, on the supply voltage. Low-noise external voltage regulators are required to minimize the power-supply noise, and data post-processing can be included to attenuate power-supply noise (50/60 Hz harmonics) and other types of predictable noise. Note that in the case of a MEA with multiple copies of the same converter, the power-supply noise is common to all of them, which makes the extraction and subtraction of common noise during data post-processing possible. The feasibility and robustness of compact, single-ended input stages for MEAs have also been recently proven in [12].

Feedback IDAC
The current DAC consists of two current sources and two transmission gates. As shown in Figure 4, M 4 and M 6 generate I M4 = 0.9·I 0 while M 5 and M 7 produce I M5 = 0.2·I 0 . Therefore, the instantaneous output current is given that-depending on the feedback signal Y-transmission gates M 8 -M 11 control whether the current I M5 is connected to V C , the output, through M 8 -M 9 or discarded via M 10 -M 11 . As for the input stage, the biasing current I 0 was fixed at 1.4 µA.

Feedback IDAC
The current DAC consists of two current sources and two transmission gates. As shown in Figure 4, M4 and M6 generate IM4 = 0.9·I0 while M5 and M7 produce IM5 = 0.2·I0. Therefore, the instantaneous output current is given that-depending on the feedback signal Y-transmission gates M8-M11 control whether the current IM5 is connected to VC, the output, through M8-M9 or discarded via M10-M11. As for the input stage, the biasing current I0 was fixed at 1.4 μA. The modulator is stable only if any possible IIS current ranges between the two possible feedback currents. According to Equations (1) and (2), this condition is met if |V el | < 5 mV for C0 = 350 fF (G ≈ 20 V −1 · I0) and if |V el | < 4 mV for C0 = 4.25 pF (G ≈ 20 V −1 · I0). Nevertheless, the practically available full scale is considered to be 3 mVp in order to limit distortion and to avoid saturation and excessive quantization noise [22,23]. Moreover, this full-scale reduction relaxes requirements in terms of matching and robustness against process variations, since deviations from nominal parameters would not saturate the converter and could be corrected during digital-signal post-processing.

Integrator
The difference between input and feedback currents, IIS-IFB, flows through capacitor Cint, which acts as an integrator. The value of this capacitor defines the integration constant and, along with the biasing current and the voltage-controlled oscillator (VCO)-quantizer gain, the modulator transfer functions. The nominal capacitance for I0 = 1.4 μA is Cint = 775 fF, but a 5-bit programmable capacitor has been implemented to allow for capacitances from 25 fF to 775 fF in order to accommodate different biasing currents without significant changes in the state variable VC.

VCO-Based Quantizer
A VCO-based quantizer has been used to achieve second-order noise shaping without the need for a second analog integrator. The quantizer is the combination of a VCO whose frequency is modulated by voltage VC and a frequency-to-digital (F2D) converter whose output is a logic '1' when a pulse from the VCO is detected during the preceding sampling period. A VCO-based quantizer can be modelled as a frequency integrator (the The modulator is stable only if any possible I IS current ranges between the two possible feedback currents. According to Equations (1) and (2), this condition is met if |V el | < 5 mV for C 0 = 350 fF (G ≈ 20 V −1 · I 0 ) and if |V el | < 4 mV for C 0 = 4.25 pF (G ≈ 20 V −1 · I 0 ). Nevertheless, the practically available full scale is considered to be 3 mV p in order to limit distortion and to avoid saturation and excessive quantization noise [22,23]. Moreover, this full-scale reduction relaxes requirements in terms of matching and robustness against process variations, since deviations from nominal parameters would not saturate the converter and could be corrected during digital-signal post-processing.

Integrator
The difference between input and feedback currents, I IS -I FB , flows through capacitor C int , which acts as an integrator. The value of this capacitor defines the integration constant and, along with the biasing current and the voltage-controlled oscillator (VCO)quantizer gain, the modulator transfer functions. The nominal capacitance for I 0 = 1.4 µA is C int = 775 fF, but a 5-bit programmable capacitor has been implemented to allow for capacitances from 25 fF to 775 fF in order to accommodate different biasing currents without significant changes in the state variable V C .

VCO-Based Quantizer
A VCO-based quantizer has been used to achieve second-order noise shaping without the need for a second analog integrator. The quantizer is the combination of a VCO whose frequency is modulated by voltage V C and a frequency-to-digital (F2D) converter whose output is a logic '1' when a pulse from the VCO is detected during the preceding sampling period. A VCO-based quantizer can be modelled as a frequency integrator (the instantaneous VCO phase is the result of integrating the VCO frequency over time), Sensors 2021, 21, 6456 6 of 14 followed by a phase quantizer and a discrete-time derivative [20,21]. The spectral properties of the resulting signal can be analyzed by modelling the VCO-based quantizer as a pulsefrequency modulator (PFM) [23]. Figure 5a shows the schematic of the VCO. The core of this circuit is a 3-stage voltagecontrolled ring oscillator, whose frequency depends on V SF , as shown in Figure 5b. For V SF = 1.5 V, the oscillation frequency and gain are f VCO = 500 kHz and K VCO = 1.6 kHz/mV with a current consumption of 250 nA. Transistor dimensions are 5 µm/6 µm for PMOS and 2 µm/6 µm for NMOS. Capacitors C 1 -C 3 (70 fF) have been used to reduce the oscillation frequency, which would otherwise be too high or require too low of a current to bias M 12 . V RO is controlled by M 12 (400 nm/20 µm), which acts as a source follower using the oscillator current for its own biasing. The gate of M 12 cannot be directly driven by V C since the target V C (0.9 V) is lower than the target V SF (1.5 V). M 13 is a second source follower, used to adapt the DC level of V C to the 1.5 V required at V SF to set the oscillation frequency to around 500 kHz. Finally, M 14 and M 15 act as a level shifter, adapting the 620 mV pp oscillation at ϕ 1 to the rail-to-rail levels demanded by digital circuitry.
The proposed VCO is sensitive to process variations, and the exact relationship between the input voltage and the output frequency is difficult to predict. Fortunately, since the VCO operates in a closed-loop system, any deviation from the nominal behavior (e.g., the VCO being slower than expected at V C = 0.9 V) would be compensated by the loop (e.g., higher V C , correcting the average oscillation frequency). The VCO was optimized to minimize the impact of phase noise (shown in Figure 5c) and distortion in the performance of the converter [24], which is also mitigated by the closed-loop architecture. Time-domain simulations were used to verify that the performance of the modulator is not limited by phase noise or VCO distortion.
A frequency-to-digital converter is required to transform the asynchronous VCO oscillation into a synchronous pulse-frequency-modulated signal [21,23]. Figure 6a shows a classical F2D converter circuit that is based on two D-type flip-flops (FF) and an XORgate. Ideally, the output of this F2D is a logic '1' if a VCO transition-either rising or falling-has been registered during the last sampling period. However, when transitions occur faster than the sampling frequency (i.e., f VCO > 0.5·f s ), a fraction of the transitions is missed during sampling, and the frequency of the output pulses decreases for faster VCO frequencies. Figure 6b shows that the average output is a function of the normalized VCO frequency and that it is periodic. This F2D converter topology is frequently used in other modulator architectures for which the oscillation frequency is guaranteed to be lower than 0.5·f s . However, the closed-loop modulator presented in this work is intended to operate around f VCO ≈ 0.5·f s . The F2D converter of Figure 6a would not be suitable for this application since the modulator could find undesirable metastable operation points at higher oscillation frequencies, especially at f VCO ≈ 1.5·f s .
The F2D converter used in this design is a variation of the classical exclusive OR (XOR)based approach, depicted in Figure 6c. When the oscillation frequency is slower than the sampling frequency (i.e., for f VCO < f s ), each pulse at V OSC toggles FF1, and this change is then registered by FF2, producing a logic '1' at the output of the converter. When no pulses are received during a sampling period, Q 3 = Q 2 , which renders Y = 0. For f VCO > fs, the F2D converter saturates, and the output is constantly '1' since FF1 would toggle once every sampling period. This saturation at high frequencies, illustrated in Figure 6d, improves the stability of the system since only a specific range of frequencies around f s /2 is possible during normal operation.

Electrical Characterization
The proposed readout circuit has been prototyped in 0.18-μm CMOS technology (1P6M). Figure 7 shows the 3 × 3.8 mm 2 chip, on which the highlighted 130 × 330 μm 2 area was used for testing different ∆Σ configurations. Excluding the biasing and auxiliary circuitry required for testing, the building blocks of the compact modulator (C0 = 350 fF) occupied 0.0045 mm 2 , while the low-noise modulator occupied 0.006 mm 2 due to the larger capacitor (C0 = 4.25 pF). The system was first characterized by applying a 200 μVp sinusoidal input signal at 1 kHz. Figure 8 shows the spectra of the output bitstreams for both C0 = 350 fF and C0 = 4.25 pF. The spectrum in gray is the result of single measurements, while the black plot represents the average magnitude of 20 consecutive measurements. Second-order noise shaping is visible at high frequencies, and the input-referred noise was 5.0 μVrms in the 300 Hz-6 kHz band (C0 = 350 fF, Figure 8a) and 8.7 μVrms in the 1 Hz-10 kHz band (C0 =

Electrical Characterization
The proposed readout circuit has been prototyped in 0.18-µm CMOS technology (1P6M). Figure 7 shows the 3 × 3.8 mm 2 chip, on which the highlighted 130 × 330 µm 2 area was used for testing different ∆Σ configurations. Excluding the biasing and auxiliary circuitry required for testing, the building blocks of the compact modulator (C 0 = 350 fF) occupied 0.0045 mm 2 , while the low-noise modulator occupied 0.006 mm 2 due to the larger capacitor (C 0 = 4.25 pF).

Electrical Characterization
The proposed readout circuit has been prototyped in 0.18-μm CMOS technology (1P6M). Figure 7 shows the 3 × 3.8 mm 2 chip, on which the highlighted 130 × 330 μm 2 area was used for testing different ∆Σ configurations. Excluding the biasing and auxiliary circuitry required for testing, the building blocks of the compact modulator (C0 = 350 fF) occupied 0.0045 mm 2 , while the low-noise modulator occupied 0.006 mm 2 due to the larger capacitor (C0 = 4.25 pF). The system was first characterized by applying a 200 μVp sinusoidal input signal at 1 kHz. Figure 8 shows the spectra of the output bitstreams for both C0 = 350 fF and C0 = 4.25 pF. The spectrum in gray is the result of single measurements, while the black plot represents the average magnitude of 20 consecutive measurements. Second-order noise shaping is visible at high frequencies, and the input-referred noise was 5.0 μVrms in the 300 Hz-6 kHz band (C0 = 350 fF, Figure 8a) and 8.7 μVrms in the 1 Hz-10 kHz band (C0 = The system was first characterized by applying a 200 µV p sinusoidal input signal at 1 kHz. Figure 8 shows the spectra of the output bitstreams for both C 0 = 350 fF and C 0 = 4.25 pF. The spectrum in gray is the result of single measurements, while the black plot represents the average magnitude of 20 consecutive measurements. Second-order noise shaping is visible at high frequencies, and the input-referred noise was 5.0 µV rms in the 300 Hz-6 kHz band (C 0 = 350 fF, Figure 8a) and 8.7 µV rms in the 1 Hz-10 kHz band (C 0 = 4.25 pF, Figure 8b). Unexpected noise is present in the 80-500 Hz band and is especially visible in Figure 8b due to lower flicker noise. This noise is attributed to the measurement setup; however, its contribution to the total integrated in-band noise is minor. 4.25 pF, Figure 8b). Unexpected noise is present in the 80-500 Hz band and is especially visible in Figure 8b due to lower flicker noise. This noise is attributed to the measurement setup; however, its contribution to the total integrated in-band noise is minor. Figure 9 depicts the signal-to-noise ratio (SNR) and the signal-to-noise-and-distortion ratio (SNDR) for different input amplitudes at 1 kHz. Each point represents the average of 20 consecutive measurements. Signals larger than 1-1.5 mVp are limited by the distortion of the input-stage transconductor.    Figure 9 depicts the signal-to-noise ratio (SNR) and the signal-to-noise-and-distortion ratio (SNDR) for different input amplitudes at 1 kHz. Each point represents the average of 20 consecutive measurements. Signals larger than 1-1.5 mV p are limited by the distortion of the input-stage transconductor. The performance of the proposed readout is summarized in Table 1 and compared to the prior art, including readout circuits integrated into arrays for in vitro [3,9,11,12] platforms, in vivo implants [5,10,15,17], and standalone converters [13,16]. Our work features low noise characteristics (<6 μVrms), low power consumption (<5 μW/ch), and a compact footprint (<0.01 mm 2 /ch), which is in line with the state of the art. The overall performance of the converter presented in [13] appears superior; however, the converter reported here was implemented in 0.18-μm CMOS and provided synchronous output, which may be advantageous depending on the application. Finally, it is noteworthy that the converter reported in [17] achieves comparable metrics while it includes on-chip decimation filters.  The performance of the proposed readout is summarized in Table 1 and compared to the prior art, including readout circuits integrated into arrays for in vitro [3,9,11,12] platforms, in vivo implants [5,10,15,17], and standalone converters [13,16]. Our work features low noise characteristics (<6 µV rms ), low power consumption (<5 µW/ch), and a compact footprint (<0.01 mm 2 /ch), which is in line with the state of the art. The overall performance of the converter presented in [13] appears superior; however, the converter reported here was implemented in 0.18-µm CMOS and provided synchronous output, which may be advantageous depending on the application. Finally, it is noteworthy that the converter reported in [17] achieves comparable metrics while it includes on-chip decimation filters.

In Vitro Validation
As a proof of concept, the implemented modulator was used to sense the field potentials of cardiac tissue. The CMOS prototype was wire-bonded to a custom-made PCB, where a 0.24 × 1.54 mm 2 pad, covered with electroless nickel immersion gold (ENIG), was used as the electrode. Another ENIG electrode was used as a reference electrode and was connected to V DD = 1.8 V. As shown in Figure 10, a plastic ring was glued to the PCB, and the ASIC and bonding wires were covered with epoxy to leave only the selected electrodes exposed. These materials were selected in order to simplify the fabrication process and were stable enough for system characterization measurements but would not be suitable for long-term recordings.
connected to VDD = 1.8 V. As shown in Figure 10, a plastic ring was glued to the P the ASIC and bonding wires were covered with epoxy to leave only the selected el exposed. These materials were selected in order to simplify the fabrication pro were stable enough for system characterization measurements but would not be for long-term recordings.
The chip was cleaned and sterilized, and the electrodes were coated with h bronectin to facilitate cell attachment. Human induced pluripotent stem cells (hIP rived from a healthy donor were purchased from FUJIFILM Cellular Dynamics human induced pluripotent stem cell (hIPSC) line, CW30318CC1, was obtained CIRM hPSC Repository funded by the California Institute of Regenerative M (CIRM). The hIPSCs were then differentiated into spontaneously and synchronou ing cardiomyocytes. The cardiomyocytes were carefully lifted from the cell cult to not disrupt the cell-cell connections. The cardiomyocyte tissue was then transfe allowed to adhere to the surface with the coated electrodes.  Figure 11 shows an electrical recording taken inside an incubator one h transferring the tissue to the chip. The input capacitor was set to C0 = 4.25 pF, output bitstream was band-pass filtered with a fourth-order Butterworth band-p between 5 Hz and 200 Hz. Cardiac action potentials with an amplitude of appro 100 μVpp and a beating rate of 95 beats per minute are visible. This beating rate the rate measured by observing the contractions of the tissue through the micros The chip was cleaned and sterilized, and the electrodes were coated with human fibronectin to facilitate cell attachment. Human induced pluripotent stem cells (hIPSCs) derived from a healthy donor were purchased from FUJIFILM Cellular Dynamics Inc. The human induced pluripotent stem cell (hIPSC) line, CW30318CC1, was obtained from the CIRM hPSC Repository funded by the California Institute of Regenerative Medicine (CIRM). The hIPSCs were then differentiated into spontaneously and synchronously beating cardiomyocytes. The cardiomyocytes were carefully lifted from the cell culture dish to not disrupt the cell-cell connections. The cardiomyocyte tissue was then transferred and allowed to adhere to the surface with the coated electrodes. Figure 11 shows an electrical recording taken inside an incubator one hour after transferring the tissue to the chip. The input capacitor was set to C 0 = 4.25 pF, and the output bitstream was band-pass filtered with a fourth-order Butterworth band-pass filter between 5 Hz and 200 Hz. Cardiac action potentials with an amplitude of approximately 100 µV pp and a beating rate of 95 beats per minute are visible. This beating rate matches the rate measured by observing the contractions of the tissue through the microscope. Figure 11. Measurement results using human IPSC-derived cardiomyocytes. The output bitstream of the modulator (C0 = 4.25 pF) was band-pass filtered between 5 Hz and 200 Hz, unveiling cardiac potentials and a beating rate of 95 beats per minute, which matches the beating rate measured optically.

Conclusions
This paper proposes a readout circuit for bioelectrical signals based on a ∆Σ mod lator with a VCO-based quantizer, which achieves second-order noise shaping with m imal analog circuitry. The size of the input capacitor plays a fundamental role in the sign, as it defines the signal attenuation and the size of the input stage. A large P-ty transistor serves as the input transconductor, which minimizes flicker noise without need for chopping. Another capacitor was used for the integration of both input and fe back currents, eliminating the need for operational amplifiers. A novel frequency-to-d ital converter was developed to improve the stability of the VCO-based quantization. T performance of the converter matches that of state-of-the-art devices in terms of no power, and area and constitutes a competitive solution for extracellular action-poten detection in large-scale electrode arrays and neural interfaces.
The resulting converter architecture is very simple and avoids conventional pro lems of more complex modulators, such as feedback DAC non-linearity. However, proposed circuit is not inherently robust against power-supply noise, process variatio or input-stage non-linearity. Therefore, the modulator relies on low-noise external volta regulators to minimize power-supply noise, and off-chip digital filtering can be used attenuate noise at specific frequencies. Process variations may result in unexpected g deviations, making calibration necessary if the absolute amplitude of signals is of intere The non-linearity of the input stage can be neglected for small input signals, and dist tion-compensation methods should be explored if large input signals are expected.   . Measurement results using human IPSC-derived cardiomyocytes. The output bitstream of the modulator (C 0 = 4.25 pF) was band-pass filtered between 5 Hz and 200 Hz, unveiling cardiac potentials and a beating rate of 95 beats per minute, which matches the beating rate measured optically.

Conclusions
This paper proposes a readout circuit for bioelectrical signals based on a ∆Σ modulator with a VCO-based quantizer, which achieves second-order noise shaping with minimal analog circuitry. The size of the input capacitor plays a fundamental role in the design, as it defines the signal attenuation and the size of the input stage. A large P-type transistor serves as the input transconductor, which minimizes flicker noise without the need for chopping. Another capacitor was used for the integration of both input and feedback currents, eliminating the need for operational amplifiers. A novel frequency-to-digital converter was developed to improve the stability of the VCO-based quantization. The performance of the converter matches that of state-of-the-art devices in terms of noise, power, and area and constitutes a competitive solution for extracellular action-potential detection in large-scale electrode arrays and neural interfaces.
The resulting converter architecture is very simple and avoids conventional problems of more complex modulators, such as feedback DAC non-linearity. However, the proposed circuit is not inherently robust against power-supply noise, process variations, or inputstage non-linearity. Therefore, the modulator relies on low-noise external voltage regulators to minimize power-supply noise, and off-chip digital filtering can be used to attenuate noise at specific frequencies. Process variations may result in unexpected gain deviations, making calibration necessary if the absolute amplitude of signals is of interest. The non-linearity of the input stage can be neglected for small input signals, and distortion-compensation methods should be explored if large input signals are expected.