On-Device Multi-Level Signal Quality Aware Compression for Energy-Efficient Wearable PPG Sensing

Abstract

On-device computing in biomedical sensors has become attractive for developing wearable health monitoring systems. The challenge is to make a compromise between the latency and complexity in a resource-constrained implementation. This article describes an on-device implementation of multi-level signal quality aware and quality controlled compression (MSQQCC) that enhances the compression factor while preserving the clinical features in a wearable photoplethysmography (PPG) sensing application. The multi-level quality assessment (QA) provides three eligible PPG qualities, viz., “excellent,” “good,” and “average,” based on which corresponding upper limits are set for further compression using a discrete wavelet transform, while the “corrupted” segments are discarded. A pretrained multilayer perceptron neural network (MLPNN) provides the optimal quantization level of coefficients. The residual data is separately compressed using an autoencoder (AE). MSQQCC was evaluated with 300 min of PPG data from three public data sets and 110 min of data collected at a laboratory. The end-to-end pipeline was implemented in a standalone system with an ARM Cortex A53 controller, requiring 35.51 kB of memory and 1.8 s latency to process 4 s PPG data. The on-device QA achieved 98.45% overall accuracy (Ac), which outperforms published works on PPG QA. The mean deviation of PPG clinical features by 5%, with overall compression ratio (CR) and percentage root mean squared difference (PRD) were 40.85 and 2.52, which are superior to many published works. Real-time transmission over Bluetooth shows an improvement of energy efficiency by a significant factor and a 34% extended battery life for wearable PPG sensors. The results are encouraging for the adoption of MSQQCC in wearable biomedical health monitoring.