PCA-ANN Contactless Multimodality Sensors for Body Temperature Estimation

Abstract

The use of the infrared (IR) forehead thermometer is a convenient way to measure the body temperature; however, it exhibits the concerns of accuracy, reliability, and repeatability because of improper measurement poses such as distance and angle to forehead, and other variations of subject and ambient conditions. According to the advantages of quick response and contactless measurement of using IR thermometry, this article additionally considers the factors of distance to IR sensors, subject’s facial skin color, and forehead pose, as well as ambient temperature and light intensity, to propose an accurate regression for body temperature estimation. As a consequence, the proposed multimodality sensor (MMS) system consists of a set of thermopile IR sensors, a distance sensor, a light intensity sensor, and a web camera. Furthermore, a principal component analysis (PCA) approach combining the artificial neural network is utilized for the body temperature regression. In order to reduce the measurement bias, the digital axillary temperature data was used as the ground body temperature for the machine learning training and validation. The MMS system is set randomly close to the forehead at a distance of 2–30 cm to emulate the human’s operation bias. We have totally 30 subjects for this study, and we divided them into four groups as: training dataset (14 subjects, IDs: 1–14 with 300 cases), validation dataset (three subjects, IDs: 15–17 with 64 cases), concurrent-time test dataset (three subjects, IDs: 18–20 with 64 cases), and three-months-elapsed test dataset (ten subject, IDs: 21–30 with 240 cases). In the training process, the $k$ -fold cross-validation technique ( $k = 5$ ) was applied to reduce the model overfitting. In addition, multiple linear regression (MLR) and stepwise regression (SR) models were also developed for performance comparison. The result shows that the PCA-ANN temperature estimation model outperformed other predictive models with the lowest mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE) values, and highest $R$ values and can optimize the low-cost MLX90614 sensor accuracy. The three-months-elapsed test group achieves MAE $0.1~^{\circ }\text{C}$ with a standard deviation of the error $0.03~^{\circ }\text{C}$ when compared to the state-of-the-art with low cost and contactless IR sensors.