Anomaly Identification using Multimodal Physiological Signals on the Edge

TL;DR: In this paper, the use of multiple autoencoders was proposed to detect anomalies using multimodal data (Electrocardiography and Electrodermal Activity signals) in an unsupervised manner.

Abstract: Anomaly detection is a challenging data mining task in the field of healthcare. The majority of the proposed models generally rely on an abundance of well-labeled data, which is not always available outside a laboratory environment. This paper proposes the use of multiple autoencoders to detect anomalies using multimodal data (Electrocardiography and Electrodermal Activity signals) in an unsupervised manner. The proposed method provides an insight into features of the input signals that make the anomalous segments stand out and provide better explainability for the anomalous behaviours. The proposed method was able to classify anomalies with an accuracy of 91.5% as well as identify the most commonly changing features during anomalous behaviour. The method was tested and evaluated on a low resource device to be run as an edge application and found to have an overall accuracy of 86.3%.