ECG-Afibx:a low-power addernet accelerator for atrial fibrillation detection in embedded systemsOA

Background:Long multi-channel ECG recordings that capture both time and space fluctuations are necessary for the accurate identification of atrial fibrillation(AF)in wearable or edge devices.These devices must use low-power computing and effective,real-time compression to manage massive volumes of data.An energy-efficient AdderNet-based method for AF detection utilizing compressed multi-channel ECG signals was presented in this paper.Methods:A Lossless Multi-channel Adaptive Compression Engine(L-MACE)with training,spatial,and temporal compression units is proposed in order to control bandwidth and storage constraints.Instead of employing sophisticated predictors,it uses a basic Sum-Predictor(SP)and a minimum spanning tree to decrease inter-channel redundancy.Compression is made even easier with a Booth-Encoded Logarithmic Quantized(B-LQ)multiplier.In order to save resources,a Quantized AdderNet(Q-ANet)uses sum-of-absolute-differences(SAD)rather than multiply-accumulate to detect AF after compression.Memory use is decreased by an activation-guided quantization technique,while hardware and energy efficiency is increased with SP adders.Results:The proposed method offers an accuracy of 99.07%,a precision of 99.45%,a sensitivity of 99.33%,a specificity of 98.22%,and an F1 score of 99.39%.The proposed B-LQ multiplier achieves 316 LUT and 1.758 ns of latency,while the proposed SP adder achieves 40 LUT,74 IO,and 1.610 ns of delay.Conclusion:Overall,by combining hardware-optimized computation and compression-aware signal processing,the proposed architecture shows an incredibly efficient solution for real-time AF identification in edge and wearable devices.