ECG Pulse Encoding: Three Bio-inspired Encoders for Converting Electrocardiograms to Spiking Neural Network Inputs

This article introduces an open-source project (ECG-SPIKE-ENCODING) that converts traditional ECG signals into spike sequences using three encoding methods—Bens Spiker Algorithm, Delta Modulation, and Level Crossing—to provide efficient neuromorphic inputs for Spiking Neural Networks (SNNs). The core value of the project lies in solving the SNN input adaptation problem and promoting the deployment of medical AI on edge devices. The source is GitHub (author: Deshwan25boe10077, release date: June 14, 2026).

Traditional ECG analysis relies on deep learning models such as CNN/RNN, which have limitations like high computational intensity, high power consumption, and large latency, making them unsuitable for edge devices. As the third generation of neural networks, SNNs have advantages of event-driven operation, ultra-low power consumption, biological plausibility, and hardware friendliness (compatible with chips like Intel Loihi). However, they require input in the form of spikes, so encoders are needed to convert continuous ECG signals into spike sequences.

The project implements three encoding strategies:

1. Bens Spiker Algorithm (BSA): FIR filtering + threshold comparison, outputs binary spikes, accurately preserves the peak time of R waves, suitable for rate-coded SNNs;
2. Delta Modulation: Tracks the difference between adjacent sampling points, outputs ternary spikes (+1/-1/0), captures waveform shape changes, suitable for temporal SNNs;
3. Level Crossing Encoding: Divides the signal into fixed frequency bands, emits spikes when crossing bands, outputs a sparse event stream, has optimal hardware efficiency, suitable for neuromorphic chips.

The input data uses record 210 from the MIT-BIH Arrhythmia Database (360Hz sampling rate, approximately 30 minutes, MLII/V1 channels), loaded via the wfdb library. The processing flow is: Raw ECG signal → Pulse Encoder → SNN Input. The project structure includes core encoder scripts (ecg_spike_encoder.py), data conversion scripts (convert_to_csv.py), sample data, and an output directory (storing CSV files of spike sequences from the three encoders).

Feature	BSA	Delta Modulation	Level Crossing
Output Type	Binary (0/1)	Ternary (+1/-1/0)	Sparse events
Core Mechanism	FIR filtering + threshold	Temporal difference	Amplitude band crossing
Spike Density	Medium	High	Most sparse
Hardware Efficiency	Good	Good	Optimal
Application Scenario	Rate-coded SNNs	Temporal SNNs	Neuromorphic chips

The project's planned expansion roadmap includes: 1. Training SNN classifiers using snnTorch; 2. Benchmarking the accuracy of the three encoders in arrhythmia classification tasks; 3. Extending to all 48 records in the MIT-BIH database; 4. Deploying on neuromorphic hardware such as Intel Loihi; 5. Comparing performance with CNN-based ECG classifiers.

This project combines traditional signal processing with neuromorphic computing to solve the SNN input bottleneck and provides three trade-off solutions. The target audience includes neuromorphic computing researchers, medical AI engineers, edge computing developers, and students in biological signal processing. As neuromorphic hardware matures, such encoding technologies will play an important role in scenarios like wearable medical devices and implantable sensors.