Stabilizing neuromorphic ECG processors via adaptive fractional fatigue

Area of Science:

• Neuromorphic Engineering
• Biomedical Signal Processing
• Artificial Intelligence

Background:

• Spiking Neural Networks (SNNs) on wearables can overfire with noisy ECG signals, increasing power consumption and instability.
• This hyperactivity stems from processing challenges in real-world, noisy biomedical data.
• Existing methods struggle to balance noise suppression with classification accuracy in SNNs.

Purpose of the Study:

• To introduce the Adaptive Fractional Fatigue (AFF) SNN to suppress redundant firing in noisy ECG signal processing.
• To maintain high ECG classification performance while reducing energy consumption.
• To improve the stability and predictability of SNNs for low-power neuromorphic applications.

Main Methods:

• Developed the AFF-SNN, incorporating a power-law weighted memory buffer for fractional activity trace.
• Implemented an adaptive threshold mechanism driven by accumulated neuron activity to suppress spikes.
• Evaluated AFF-SNN on MIT-BIH ECG records using spikes per sample as the energy proxy, with accuracy and AUROC as secondary metrics.

Main Results:

• Spikes per sample reduced by 62.62% (from 2384.990 ± 537.719 to 891.481 ± 227.654), a statistically significant decrease (p=0.000015).
• Test accuracy was preserved (78.923 ± 1.725% vs. 79.206 ± 1.541%, p=0.227761).
• Cross-seed spike variance decreased by 5.58×, indicating improved predictability.

Conclusions:

• AFF-SNN effectively suppresses noise-driven hyperactivity in ECG signals without compromising classification accuracy.
• The adaptive threshold mechanism provides intrinsic negative feedback, stabilizing event traffic and reducing dynamic power.
• AFF-SNN offers a promising solution for efficient and reliable low-power neuromorphic ECG analysis on wearable devices.