UAED: Unsupervised Abnormal Emotion Detection Network Based on Wearable Mobile Device

With the development of the internet-of-medical-things, health monitoring through physiological signals has become a critical task. Given this opportunity, research on personal healthcare systems for abnormal emotion detection using physiological signals brings significant benefits to the field of digital healthcare and human-computer interaction. However, it is a challenging task because of the diverse patterns of time series and the lack of labels. In this work, we present a novel model for Unsupervised Abnormal Emotion Detection (UAED) combining Gaussian mixture variational autoencoder (VAE) and convolutional neural networks (CNNs), whose core idea is to reconstruct the input by learning its latent representation thus capturing the normal patterns and applying whitening distance as the anomaly score to detect outliers. In addition, UAED uses stacking operation to transform one-dimensional time series into high-dimension to help the model capture the periodic features and reflect diverse normal patterns. We conduct extensive experiments on four public datasets to demonstrate that our UAED obtains the best performance in various metrics. Furthermore, we deploy UAED in a real environment using a low-cost wearable sensor developed by us to collect electrocardiogram signals and run UAED on mobile terminals with an accuracy of 85%, validating the feasibility of our healthcare system for detecting abnormal emotions.