Improving Stress Detection Using Weighted Score-Level Fusion of Multiple Sensor

Abstract

Work-related stress is now a widespread issue in our modern life. Consequently, stress management is crucial. Automated stress detection is one of the innovations in helping people manage their well-being better by providing information about their stress levels. The advancement of sensor technology and artificial intelligence has made this task easier. The assessment of stress levels by using a variety of sensors from a smartwatch and machine learning algorithms has been very popular in recent years. The use of multiple sensor data enables richer information to train the machine learning algorithm so that the trained model can be more robust. However, if we use a feature-level fusion, since it is the most popular fusion strategy, generally, each sensor data will have the same significance. In fact, stress is personal and each subject can have a different reaction to the stress so that a particular sensor may be an effective stress indicator for some subjects but it might not be for others. Therefore, we propose a personalized stress detection system based on a weighted score-level multiple sensor fusion strategy. For each individual, this strategy gives different weights to each sensor based on the performance of the sensor on the individual's data. The experiment results show that both feature-level and weighted score-level fusion models obtained better performance than models from the individual sensor strategy. The weighted score-level fusion strategy achieved a better performance than the feature-level strategy with accuracy, precision, recall, and F1-measure of 0.931, 0.824, 0.939, and 0.868, respectively.