Construction of Facial Skin Temperature-Based Anomaly Detection Model for Daily Fluctuations in Health Conditions

Abstract

A method for estimating health conditions is required to monitor daily health conditions. Various types of data have been used in healthcare studies; however, imaging data are superior because they allow quick and remote measurements. Thermal face images can be measured safely and economically using infrared thermography. Many physiological and psychological states have been evaluated based on the data from these images. A previous study, using short-term experiments, confirmed that an anomaly detection model constructed using a variational autoencoder enables the detection of anomalous states of thermal face images. A long-term experiment is essential to estimate long-term fluctuating human states, such as health conditions. The purpose of this study is to construct a facial skin temperature-based anomaly detection model for human health conditions. The authors obtained thermal face images with health condition questionnaires for approximately a year. Based on the questionnaire responses, the thermal images in good and poor health conditions were labeled “normal state” and “anomaly state,” respectively. The facial skin temperature-based anomaly detection model for health conditions was constructed using a variational autoencoder with only thermal face images in the normal state. The AUC, which represents anomaly detection performance, was 0.70. In addition, an increasing trend of the performance of the model by learning a wider area of skin temperature was confirmed.