Approximate Computing in Critical Applications: ECG Arrhythmia Classification.

Approximate computing is a technique which designers are fallen back on considering the low-power requirements of applications that are resilient to errors. However, the utilization of approximate computing has been limited to the consideration of the trade-off between the quality of output and the hardware costs. Particularly, the approximate computing has been disregarded in critical health monitoring applications, where a certain level of accuracy has to be ensured. In this paper, we show that despite such a trade-off, employing approximate computing properly, even an increase in the final accuracy is attainable, thanks to regularization that approximate processing introduces. By employing approximate multipliers implementing in the convolution layers of the electrocardiogram (ECG) arrhythmia classification, we show that with even an stringent approximation, we can ensure more than 95% accuracy in ECG arrhythmia detection.