Precise and low-power closed-loop neuromodulation through algorithm-integrated circuit co-design

Jie Yang, Shiqi Zhao, Junzhe Wang, Siyu Lin, Qiming Hou, Mohamad SawanFrontiers in Neuroscience 2024 · DOI

Abstract: Implantable neuromodulation devices have significantly advanced treatments for neurological disorders such as Parkinson's disease, epilepsy, and depression. Traditional open-loop devices like deep brain stimulation (DBS) and spinal cord stimulators (SCS) often lead to overstimulation and lack adaptive precision, raising safety and side-effect concerns. Next-generation closed-loop systems offer real-time monitoring and on-device diagnostics for responsive stimulation, presenting a significant advancement for treating a range of brain diseases. However, the high false alarm rates of current closed-loop technologies limit their efficacy and increase energy consumption due to unnecessary stimulations. In this study, we introduce an artificial intelligence-integrated circuit co-design that targets these issues and using an online demonstration system for closed-loop seizure prediction to showcase its effectiveness. Firstly, two neural network models are obtained with neural-network search and quantization strategies. A binary neural network is optimized for minimal computation with high sensitivity and a convolutional neural network with a false alarm rate as low as 0.1/h for false alarm rejection. Then, a dedicated low-power processor is fabricated in 55 nm technology to implement the two models. With reconfigurable design and event-driven processing feature the resulting application-specific integrated circuit (ASIC) occupies only 5mm<sup>2</sup> silicon area and the average power consumption is 142 μW. The proposed solution achieves a significant reduction in both false alarm rates and power consumption when benchmarked against state-of-the-art counterparts.

Introduction

Neurological disorders represent a massive global health burden. According to the World Health Organization, conditions such as depression, epilepsy, and Alzheimer’s affect hundreds of millions of people worldwide. Managing these conditions effectively requires precise intervention, yet traditional methods often fall short of the adaptive needs of the human brain.

Global brain disease statistics highlighting the widespread impact of neurological disorders.

The Challenge of Open-Loop Systems

Traditional treatments often rely on implantable neuromodulation devices. However, many current clinical standards use "open-loop" architectures. These devices deliver fixed electrical stimulation regardless of the patient's immediate brain state. This "dumb" approach often leads to overstimulation, causing unnecessary side effects and draining batteries quickly, which necessitates frequent surgical replacements.

Limitations of traditional open-loop neuromodulation, including fixed stimulation and associated side effects.

Closed-loop systems represent the next generation of care, designed to monitor brain signals and only stimulate when an event, such as a seizure or tremor, is detected. However, current closed-loop technologies struggle with high false alarm rates (FAR) and high power consumption, as running complex AI models on a tiny, battery-operated implant is computationally expensive.

Methods: Algorithm-Circuit Co-Design

To address these hurdles, we developed an algorithm-integrated circuit co-design. Instead of designing the AI and the hardware in isolation, we optimized both simultaneously to operate in an "event-driven" fashion.

The Dual-Model Strategy

Our system utilizes a hierarchical "two-tiered" AI approach:

  1. Binary Neural Network (BNN): This is a highly compressed, "always-on" model where weights are simplified to 1s and -1s. It monitors brain signals with high sensitivity to ensure no potential events are missed while consuming minimal power.
  2. Convolutional Neural Network (CNN): This more precise 8-bit model remains in a low-power "sleep" state until triggered by the BNN. Once activated, it performs a high-fidelity analysis to confirm the event and reject false alarms.

Hardware Innovation

We fabricated a dedicated Application-Specific Integrated Circuit (ASIC) using 55nm CMOS technology. To minimize the silicon footprint and power draw, we implemented reconfigurable Processing Elements (PEs). These elements use Booth encoding and shift-add operations, effectively replacing bulky traditional multipliers. This allows the same hardware to efficiently process both BNN and CNN operations.

Key Results

The resulting system demonstrates significant improvements over existing technologies:

  • Ultra-Low Power Consumption: The chip consumes just 142.9 $\mu$W at a 300 kHz clock frequency.
  • High Precision and Sensitivity: Validated on the CHB-MIT dataset, the 8-bit CNN achieved 99.6% sensitivity with a false alarm rate of only 0.01/h.
  • Compact Design: The entire processor, including 160 KB of on-chip memory, fits into a tiny 5.06 $mm^2$ silicon area.
  • Efficiency: The BNN operations require only 0.99 $\mu$J per inference, making continuous monitoring feasible for long-term implantation.

Conclusion and Future Work

This co-design approach proves that high-precision, low-power neuromodulation is achievable within the strict constraints of implantable hardware. While the current results are promising, future work will focus on addressing "catastrophic forgetting" through meta-learning and moving toward real-time clinical trials. Ultimately, this technology paves the way for smarter, safer, and more durable brain implants that can adapt to the unique needs of every patient.

Summary updated

Abstract

Implantable neuromodulation devices have significantly advanced treatments for neurological disorders such as Parkinson's disease, epilepsy, and depression. Traditional open-loop devices like deep brain stimulation (DBS) and spinal cord stimulators (SCS) often lead to overstimulation and lack adaptive precision, raising safety and side-effect concerns. Next-generation closed-loop systems offer real-time monitoring and on-device diagnostics for responsive stimulation, presenting a significant advancement for treating a range of brain diseases. However, the high false alarm rates of current closed-loop technologies limit their efficacy and increase energy consumption due to unnecessary stimulations. In this study, we introduce an artificial intelligence-integrated circuit co-design that targets these issues and using an online demonstration system for closed-loop seizure prediction to showcase its effectiveness. Firstly, two neural network models are obtained with neural-network search and quantization strategies. A binary neural network is optimized for minimal computation with high sensitivity and a convolutional neural network with a false alarm rate as low as 0.1/h for false alarm rejection. Then, a dedicated low-power processor is fabricated in 55 nm technology to implement the two models. With reconfigurable design and event-driven processing feature the resulting application-specific integrated circuit (ASIC) occupies only 5mm<sup>2</sup> silicon area and the average power consumption is 142 μW. The proposed solution achieves a significant reduction in both false alarm rates and power consumption when benchmarked against state-of-the-art counterparts.
BibTeX
@inproceedings{yang2024precise,
  author = {Jie Yang and Shiqi Zhao and Junzhe Wang and Siyu Lin and Qiming Hou and Mohamad Sawan},
  title = {{Precise and low-power closed-loop neuromodulation through algorithm-integrated circuit co-design}},
  booktitle = {Frontiers in Neuroscience},
  year = {2024},
  doi = {10.3389/fnins.2024.1340164},
}
Figure