Exploring the Feasibility of Data-Driven Emotion Modeling for Human Digital Twins

Human Digital Twins (HDTs) are specifically designed to represent humans virtually. Despite their recency, HDTs are considered to be powerful tools for enhancing personalized healthcare, early disease detection, improving patient outcomes, and optimizing lifestyle. Building HDTs is challenging especially due to the complexity of modeling human behavior. To unravel such complexity, in this paper we focus on specific aspects of human behavior and emotions (valence and arousal) to enable a more informed HTD behavior modeling. We assessed the feasibility and performance of our data collection infrastructure with N=112 participants. During a science festival, our participants wore a smartwatch to self-report their emotions while simultaneously their physiological data were collected through smartwatch sensors. We explored predictive modeling possibilities once with all the collected data and then with only the sensor-related data. The former performed best with 25 features, modeled with the Quadratic Discriminant Analysis classifier, resulting in 72.4% accuracy, however, with only 4 features derived from the self-reports, the K-Nearest Neighbors Classifier was able to achieve an accuracy of 71.6%. The latter used 17 sensor-related features and the Quadratic Discriminant Analysis classifier estimator reaching a 51.3% accuracy.