Deep Learning of Sparse Patterns in Medical IoT for Efficient Big Data Harnessing

Abstract

The long-term and continuous streaming of big data from medical Internet of Things (IoT), poses a great challenge for the battery-limited tiny devices. To address this challenge, we propose a novel framework for medical IoT data sparsification, leveraging both deep learning and optimal space searching. More specifically, the deep sparsification networks are designed to learn to extract key sparse patterns in the medical IoT data, by projecting the original data stream to a sparsified data representation. Further, the principles for designing deep encoding networks have been analyzed by an optimal space searching strategy, aiming to determine the best deep sparsification architecture that meets the energy constraint or sparsification error constraint. Compared with state-of-the-art approaches, our deep learning-based and space search-optimized framework shows a dramatic capability to tackle the power hungriness problem on medical IoT big data. This novel study, by enabling energy-efficient medical IoT big data sparsification, is expected to boost the continuous and long-term medical IoT applications, such as cardiac monitoring, thereby advancing precision medicine.