💻 computer science

A 32-channel event-based bio-signal analog front-end with adaptive delta and pulse frequency encoding

This paper presents a 32-channel event-based analog front-end ASIC fabricated in 180 nm CMOS that utilizes dual-mode Pulse Frequency Modulation and adaptive Asynchronous Delta Modulation to enable highly configurable, low-power, and compressible biomedical signal acquisition for neuromorphic brain-computer interfaces.

Original authors: Narayanan Shyam, Saptarshi Ghosh, Giacomo Indiveri

Published 2026-07-15

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Narayanan Shyam, Saptarshi Ghosh, Giacomo Indiveri

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to listen to a whisper in a crowded, noisy stadium. If you use a standard microphone that just records everything, you get a wall of static that drowns out the whisper. Now, imagine a super-smart microphone that only records the whisper and ignores the crowd, but also knows how to turn the volume up or down instantly depending on how loud the crowd gets. That is essentially what this team of researchers built: a tiny, 32-channel "smart microphone" chip designed to listen to the brain's electrical signals without getting overwhelmed by noise.

The Problem with Old Microphones
Most devices that listen to brain signals (like those used in Brain-Computer Interfaces) use a "fixed" setting. Think of it like a security camera with a motion sensor set to a specific sensitivity. If the setting is too sensitive, a leaf blowing in the wind (noise) triggers the camera, flooding your storage with useless photos. If the setting is too loose, the camera misses the actual person walking by because it's waiting for a bigger movement. The paper argues that this "one-size-fits-all" approach is a dead end for long-term monitoring because real-world signals are messy and change over time. A fixed setting either creates too much data or misses important details.

The New "Smart" Chip
The researchers created a new chip, fabricated in a 180 nm CMOS process, that acts like a 32-channel smart microphone. Each channel has two special ways of listening:

Pulse Frequency Modulation (PFM): This is like a drummer who beats a drum faster when the music gets intense and slower when it's calm. It turns the signal into a stream of "spikes" or events.
Adaptive Asynchronous Delta Modulator (aADM): This is the real star of the show. Instead of using a fixed rule, this circuit has a built-in "envelope tracker." Imagine a surfer who constantly adjusts their board to match the changing shape of the wave. This circuit follows the "envelope" (the overall shape) of the incoming signal in real-time.

How the "Surfer" Works
The aADM circuit is designed to ignore the "noise floor"—the constant background hiss of the signal. It does this by looking at the slow, low-frequency changes in the signal's volume. If the background noise gets louder, the chip automatically raises its "threshold" (the level required to trigger a recording) so it doesn't get triggered by the noise. If the noise drops, it lowers the threshold to catch quieter details.

However, it's smart enough to know the difference between background noise and a real event. If a sudden, sharp spike happens—like a neuron firing an action potential—the circuit ignores the slow noise changes and captures that fast event immediately. This allows the chip to compress data massively, sending only the "interesting" parts of the signal and ignoring the rest.

What They Actually Measured
The team didn't just simulate this on a computer; they built the physical chip and tested it.

The Chip Size: It measures 3.22 mm × 5.67 mm.
The Channels: It has 32 independent channels that can be programmed individually.
The Noise: They measured the noise of the signal chain (from the first amplifier to the end) and found it integrated to 68.72 𝜇Vrms over a bandwidth of 1.6 kHz.
The Results: In their experiments, they fed the chip a synthetic signal that had a "staircase" of changing noise levels. The chip successfully adapted its threshold to the changing noise, as shown in their measurements. When they introduced a specific event (like a simulated action potential) near the 1.7-second mark, the chip reliably recorded it even while the background noise was high.

What It's Not (Yet)
The paper is careful to note that while the chip works, it is a prototype. The authors mention that to make the noise even lower, they could have made the first amplifier (LNA) bigger, but they chose to give more space to the adaptive circuit instead. This means the noise level they measured (68.72 𝜇Vrms) is a trade-off they made to prove the adaptive concept works. They also note that while the PFM part of the chip was tested and found functional, the main focus and the "breakthrough" of this specific work is the adaptive delta modulation.

The Bottom Line
This chip suggests a new way to talk to computers using brain signals. By letting the sensor itself decide what is important and what is just noise, it solves the problem of data overload. It's a step toward wireless systems that can listen to the brain for a long time without needing to send gigabytes of useless static data back to a computer. The authors successfully demonstrated that this adaptive approach works in real hardware, paving the way for more efficient brain-computer interfaces.

Technical Summary: A 32-Channel Event-Based Bio-Signal Analog Front-End with Adaptive Delta and Pulse Frequency Encoding

Problem Statement
Existing bio-signal processing systems often rely on event-based encoding circuits with fixed thresholds to convert analog sensory signals into Address-Event Representation (AER) for Spiking Neural Networks (SNNs). While effective for low-power, low-latency processing, fixed-threshold strategies face significant limitations in real-world, long-term Brain-Computer Interface (BCI) scenarios. Specifically, setting thresholds too closely generates excessive events from noise, while setting them too far apart to avoid noise reduces signal encoding accuracy and degrades downstream feature extraction. Furthermore, noise profiles in biomedical signals can drift over time, rendering static thresholds non-ideal for massive multi-channel arrays or long-term monitoring.

Methodology and Architecture
The authors present a 32-channel Analog Front-End (AFE) Application-Specific Integrated Circuit (ASIC) fabricated in a 180 nm CMOS process. The chip is designed to interface directly with neuromorphic processors using the AER protocol. The architecture consists of 32 independently programmable channels, each containing a signal conditioning chain and a dual-mode event encoding block.

Signal Conditioning Chain: Each channel includes:
- Low-Noise Amplifier (LNA): Uses a current-mirror Operational Transconductance Amplifier (OTA) with a tunable gain (0–22 dB) controlled by a 4-bit Capacitor DAC (CDAC).
- Band-Pass Filter (BPF): A 4th-order filter based on a Flipped Voltage Follower (FVF) topology to improve noise performance via current reuse. It features tunable center frequency ( $\omega_0$ ) and quality factor ( $Q$ ) via an 8-bit CDAC.
- Programmable Gain Amplifier (PGA): Provides additional gain, also controlled by a 4-bit CDAC.
- All analog parameters are programmed via a digital Serial Peripheral Interface (SPI) using an on-chip temperature-compensated bias generator.
Dual-Mode Encoding:
- Pulse Frequency Modulation (PFM): Implements an Adaptive Exponential Integrate-and-Fire (AdExp-IF) neuron to convert voltage signals into spike trains.
- Adaptive Asynchronous Delta Modulator (aADM): This is the core innovation. Unlike classic ADMs that use fixed thresholds, the aADM dynamically adjusts its delta thresholds in real-time based on the input signal envelope.
  - Adaptive Mechanism: The circuit extracts the signal envelope using a subthreshold Source Follower. This envelope is processed by two Differential Pair Integrators (DPIs) with significantly different time constants (a 99:1 ratio) to separate low-frequency envelope changes from high-frequency transients (e.g., action potentials).
  - Threshold Adjustment: A Winner-Take-All (WTA) circuit compares the DPI outputs to detect rapid envelope changes. The resulting current adjusts the baseline threshold voltage (generated by a 7-bit Voltage DAC) for the 'UP' and 'DN' comparators. This allows the system to raise the threshold during high-noise periods to suppress noise events while maintaining sensitivity to high-amplitude, high-frequency signal features.

Key Contributions

First Realization of aADM in Bio-Signals: The paper presents the first ASIC implementation of an adaptive delta modulator specifically tuned to the noise levels of biomedical signals, enabling real-time threshold tuning.
32-Channel Scalability: The chip integrates 32 fully parallel, configurable channels, addressing the scaling challenges of multi-channel bio-signal acquisition.
Sensor-Level Compression: By adapting the encoding threshold to the signal envelope, the aADM provides a trade-off between raw signal reconstruction accuracy and data compression ratio at the sensor level. This reduces data bandwidth requirements for wireless transmission in BCI applications.
Neuromorphic Compatibility: The output is natively compatible with asynchronous SNN processors via the AER protocol, facilitating end-to-end neuromorphic signal processing.

Measurement Results

Noise Performance: The input-referred noise power spectral density, integrated from 10 Hz to 1.6 kHz, was measured at 68.72 $\mu V_{rms}$ . The authors note that noise levels were slightly elevated in the LNA to prioritize area allocation for the aADM circuit, though further optimization is possible.
Adaptive Encoding Validation:
- Simulation: Circuit simulations demonstrated that the aADM successfully tracked a changing noise floor (simulated via a staircase function) while preserving the encoding of high-frequency events.
- Silicon Measurements: Experimental results on the fabricated chip confirmed the adaptive behavior. When a synthetic stimulus with a changing noise floor was applied, the chip adjusted its event generation rate. Crucially, when an event of interest (simulating an action potential) was introduced near $t=1.7$ s, the system reliably encoded it despite the previously adapted high-noise floor, validating the hardware implementation.

Significance
The paper claims that this work paves the way for adaptive wireless communication of neural signals suitable for on-line processing in BCIs. By resolving the issue of data saturation in long-term monitoring through sensor-level adaptive compression, the ASIC serves as a versatile companion front-end. It allows for the direct interfacing of low-frequency biomedical signals with asynchronous SNN processors, balancing the need for high-fidelity signal reconstruction with the constraints of low-power, low-bandwidth transmission. The authors position this as a step toward efficient, end-to-end neuromorphic signal processing systems capable of handling streaming data on the edge.

Technical Summary: A 32-Channel Event-Based Bio-Signal Analog Front-End with Adaptive Delta and Pulse Frequency Encoding

More like this