Energy-Efficient Spectral Analysis of ECGs on Resource Constrained IoT Devices.

Abstract

Power spectral analysis (PSA) is one of the most popular and insightful methods, currently employed in several biomedical applications, aiming to identify and monitor various health related conditions. Among the most common applications of PSA is heart rate variability (HRV) analysis, which allows the extraction of further insights compared with conventional time-domain methods. Unfortunately, existing PSA approaches exhibit high computational complexity, hindering their execution on power-constrained embedded internet of things (IoT) devices. Such IoT devices are increasingly used for monitoring various conditions mainly by processing the input signals in the less complex time-domain. In this paper, a new low-complexity PSA system based on fast Gaussian gridding (FGG) is proposed, which can be used to calculate the Lomb-Scargle periodogram (LSP) of a non-uniformly spaced RR tachogram. The proposed approach is implemented on a popular ARM Cortex-M4 based embedded system, which is widely used in common wearables, and compared with conventional LSP-based approaches. Utilizing this experimental setup, a meticulous analysis is performed in terms of power, performance and quality under different operational settings, such as the total input/output samples, precision of computations, computer arithmetic (floating/fixed-point), and clock frequency. The experimental results show that the proposed FGG-based LSP approach, when specifically optimized for the targeted embedded device, outperforms existing approaches by up-to 92.99% and 91.70% in terms of energy consumption and total execution time respectively, with minimal accuracy loss.