NeuroCARE: A generic neuromorphic edge computing framework for healthcare applications

Fengshi Tian, Jie Yang, Shiqi Zhao, Mohamad SawanFrontiers in Neuroscience 2023 · DOI

TL;DR

NeuroCARE is a novel neuromorphic framework designed to bring high-performance biomedical AI to power-constrained wearable devices. By combining a unique sparse Gaussian spike encoder with an adaptive weight-mapping strategy, the system converts standard CNNs into energy-efficient Spiking Neural Networks (SNNs), slashing energy consumption by over 80% while maintaining state-of-the-art accuracy across EEG, ECG, and EMG tasks.

What Problem Does This Paper Address?

Deep learning has revolutionized biomedical signal processing, enabling automated seizure prediction and heart arrhythmia detection. However, there is a major "deployment gap." Standard Convolutional Neural Networks (CNNs) are computationally heavy and power-hungry, making them difficult to run on the tiny batteries of wearable or implantable medical devices.

While Spiking Neural Networks (SNNs) offer a biologically inspired, low-power alternative by processing data as discrete "spikes" (1-bit signals), they are notoriously difficult to train from scratch. Previous attempts to convert pre-trained CNNs into SNNs often resulted in a significant drop in accuracy or required excessively long processing times. NeuroCARE tackles this by creating a generic framework that bridges the gap between the high performance of CNN training and the extreme efficiency of SNN inference.

Methods

The NeuroCARE framework introduces three key innovations:

  1. Sparse Gaussian Spike Encoder: Instead of traditional rate-encoding (which requires many time steps), this encoder uses a Gaussian random discretization method. It converts raw, continuous biomedical signals into sparse 1-bit spike sequences. This preserves the statistical distribution of the data while significantly reducing the amount of information the network needs to process.
  2. Adaptive Weight Mapping: The authors developed a mathematical "shadow training" approach. They train a standard CNN first, then use a specialized scaling factor ($f$) to map those weights onto an SNN. This mapping is specifically co-designed with the spike encoder to ensure that the SNN "understands" the encoded data without needing a performance-degrading retraining phase.
  3. Channel-wise Spatial-Temporal Structure: Biomedical signals often have many time points but few physical channels (electrodes). NeuroCARE uses 1-bit, single-dimension kernels to extract features from each channel independently, preventing the "mixing" of data that can lead to lower classification accuracy in multi-electrode setups.

Datasets

The framework was validated on three distinct, publicly available benchmarks to prove its versatility:

  • CHB-MIT EEG Dataset: Used for epileptic seizure prediction (22 patients, scalp EEG).
  • MIT-BIH Arrhythmia Database: Used for electrocardiography (ECG) classification into four distinct heart-beat classes.
  • Nina Pro DB1: A surface electromyography (sEMG) benchmark used for hand gesture recognition (10 electrodes, 27 subjects).

Key Insights and Ideas

  • The 1-Bit Advantage: By moving from multi-bit (32-bit or 16-bit) floating-point math to 1-bit spikes, the hardware can replace expensive "multiplication and accumulation" (MAC) operations with simple "addition" (ADD) operations. This is the primary driver behind the massive energy savings.
  • Sparsity is Efficiency: The Gaussian encoder allows the user to tune "sparsity." By mimicking the randomness of biological neurons, the system can rebuild feature distributions using far fewer time steps than previous neuromorphic approaches.
  • Generic vs. Task-Specific: Unlike many SNN papers that focus on a single medical condition, NeuroCARE is designed as a "generic" framework. The same architecture, with minor configuration changes, outperformed specialized models across three completely different physiological signals (brain, heart, and muscle).

Key Results

The authors compared NeuroCARE against traditional CNN topologies and other state-of-the-art models:

  • Accuracy: Achieved overall scores of 92.7% for seizure prediction, 96.7% for arrhythmia detection, and 85.7% for hand gesture recognition.
  • Complexity: Reduced computational complexity by over 80.7%.
  • Hardware Footprint: On a simulated 45nm process, the spike-based processing element (SPE) reduced energy consumption by over 80% and silicon area occupation by over 64.8% compared to conventional CNN processing elements.
  • Speedup: The framework demonstrated an "efficiency acceleration" of up to 79.7x in specific configurations.

Limitations and Future Work

While NeuroCARE significantly reduces the number of time steps required for SNNs, there is still an inherent trade-off: more time steps (latency) generally lead to higher accuracy. Finding the absolute "sweet spot" for real-time, zero-latency applications remains an area for optimization.

Additionally, while the paper demonstrates the framework on a simulator based on the Eyeriss architecture, the next logical step—and a direction suggested by the authors—is the fabrication of a dedicated physical ASIC (Application-Specific Integrated Circuit) to verify these power savings in a real-world wearable prototype.

Summary updated

Abstract

Highly accurate classification methods for multi-task biomedical signal processing are reported, including neural networks. However, reported works are computationally expensive and power-hungry. Such bottlenecks make it hard to deploy existing approaches on edge platforms such as mobile and wearable devices. Gaining motivation from the good performance and high energy-efficiency of spiking neural networks (SNNs), a generic neuromorphic framework for edge healthcare and biomedical applications are proposed and evaluated on various tasks, including electroencephalography (EEG) based epileptic seizure prediction, electrocardiography (ECG) based arrhythmia detection, and electromyography (EMG) based hand gesture recognition. This approach, NeuroCARE, uses a unique sparse spike encoder to generate spike sequences from raw biomedical signals and makes classifications using the spike-based computing engine that combines the advantages of both CNN and SNN. An adaptive weight mapping method specifically co-designed with the spike encoder can efficiently convert CNN to SNN without performance deterioration. The evaluation results show that the overall performance, including the classification accuracy, sensitivity and F1 score, achieve 92.7, 96.7, and 85.7% for seizure prediction, arrhythmia detection and hand gesture recognition, respectively. In comparison with CNN topologies, the computation complexity is reduced by over 80.7% while the energy consumption and area occupation are reduced by over 80% and over 64.8%, respectively, indicating that the proposed neuromorphic computing approach is energy and area efficient and of high precision, which paves the way for deployment at edge platforms.
BibTeX
@inproceedings{tian2023neurocare,
  author = {Fengshi Tian and Jie Yang and Shiqi Zhao and Mohamad Sawan},
  title = {{NeuroCARE: A generic neuromorphic edge computing framework for healthcare applications}},
  booktitle = {Frontiers in Neuroscience},
  year = {2023},
  doi = {10.3389/fnins.2023.1093865},
}
Figure