Low-Power FPGA-Based Quantized Neural Network Accelerator for Real-Time Arrhythmia Detection from Single-Lead ECG

Arrhythmia detection from long-term electrocardiogram (ECG) recordings is essential for early diagnosis of cardiovascular diseases in portable and wearable monitoring systems. This paper presents a low-power hardware accelerator for real-time beat-level arrhythmia classification using an ۸-bit quantized ۱D convolutional neural network (QNN) implemented on a Xilinx Artix‑۷ FPGA. The network is trained and evaluated on the MIT-BIH Arrhythmia Database under an inter-patient scheme and classifies beats into three clinically relevant categories: normal (N), ventricular ectopic (V), and supraventricular ectopic (S). After quantization-aware training, the ۸-bit fixed-point model achieves ۹۸.۴% overall accuracy, with ۹۹.۱% sensitivity and ۹۸.۷% positive predictive value (PPV) for N beats, and ۹۶.۶% sensitivity with ۹۵.۹% PPV for V beats on the held-out test set. The hardware accelerator exploits parallel multiply–accumulate units, on-chip BRAM reuse, and a streaming dataflow to process each ۲۵۶-sample beat segment in ۸.۵ µs at ۱۰۰ MHz, corresponding to a throughput of approximately ۱۱۸ k beats/s while consuming ۰.۹۲ W power. Compared with a ۳۲-bit floating-point implementation on an ARM Cortex-A۹ processor, the proposed design provides about ۹.۳× speedup and ۲۷× lower energy per beat, making it suitable for wearable and home monitoring systems.

کلیدواژه ها:

ECG ، Arrhythmia Detection ، Quantized Neural Network ، FPGA Accelerator ، Low Power Design