Book Description
Pain assessment has remained largely unchanged for decades and is currently based on self-reporting. Although there are different versions, these self-reports all have significant drawbacks. For example, they are based solely on the individual’s assessment and are therefore influenced by personal experience and highly subjective, leading to uncertainty in ratings and difficulty in comparability. Thus, medicine could benefit from an automated, continuous and objective measure of pain. One solution is to use automated pain recognition in the form of machine learning. The aim is to train learning algorithms on sensory data so that they can later provide a pain rating. This thesis summarises several approaches to improve the current state of pain recognition systems based on physiological sensor data. First, a novel pain database is introduced that evaluates the use of subjective and objective pain labels in addition to wearable sensor data for the given task. Furthermore, different feature engineering and feature learning approaches are compared using a fair framework to identify the best methods. Finally, different techniques to increase the interpretability of the models are presented. The results show that classical hand-crafted features can compete with and outperform deep neural networks. Furthermore, the underlying features are easily retrieved from electrodermal activity for automated pain recognition, where pain is often associated with an increase in skin conductance.
