New approximation tech cuts arrhythmia detection power by 64.9%

A team led by Floriaan Bulten from the arXiv paper (2607.14747) tackles the chronic power bottleneck in wearable arrhythmia monitors. Current devices rely on bulky hardware or cloud offloading because deep learning models consume too much energy for continuous edge inference. The researchers systematically apply approximation techniques—specifically reduced data precision and approximate multipliers—to a state-of-the-art DL architecture for ECG classification.

Testing on the MIT-BIH Arrhythmia Database, the optimized hardware implementation achieves 93.7% classification accuracy and 92.1% sensitivity. At a low-frequency 12 kHz operation, power drops from the reference 8.75 μW to just 3.07 μW (64.9% reduction). At 100 MHz, energy consumption falls 61.5% to 0.8 μJ. These gains enable long-term, comfortable monitoring without clinical manual review, directly addressing the 64.9% power reduction and 92.1% sensitivity targets for practical wearables.