Network-Disjointed Asymmetric Autoencoders for Data Compression in IoT Based Healthcare Applications

This paper introduces an Asymmetric Autoencoder (AAE) based data compression framework aimed at optimizing communication bandwidth usage and transmitter complexity in healthcare Internet of Things (IoT) systems. Such IoT devices often face constraints in terms of energy, computational resources, and transmission capacity while handling high-dimensional sensor data. The proposed framework is optimally implemented for specific applications, executed at a central receiver using sensor data collected and wirelessly transmitted by IoT sensors. A neural network-based data-compression encoder is deployed at the transmitter, and a decoder at the receiver reconstructs the received compressed data. The AAE architecture supports asymmetric designs to accommodate diverse computational capabilities, bandwidth requirements, and application-specific performance goals. Joint neural network training of the encoder and decoder ensures efficient extraction of application-relevant information from multi-dimensional sensor time-series data. In this paper, the proposed framework is validated for a specific example application of Human Activity Classification, which is a common need in general purpose health monitoring. Experimental results demonstrate that the framework achieves effective feature compression while maintaining acceptable application level performance. A detailed analysis of the trade-offs between transmitter complexity, receiver complexity, and bandwidth efficiency is performed. The findings highlight the interplay between transmitter complexity and bandwidth usage, providing crucial insights for building efficient, reliable, and resource-conscious IoT-based healthcare systems.