Prof. Dr.-Ing. Bert Arnrich

Fighting Label Scarcity: Semi-automated human-in-the-loop label generation for automated stress and mental workload classification for uncontrolled environments

Samik Real Enrı́quez, Supervisors: Sidratul Moontaha, Christoph Anders

Master's Thesis

Mental workload (MWL) is an important field of study in neuroscience due to the impact it has on the whole population. Mental workload can be defined as the amount of brain activity in use in a specific time-window [1]. High amount of brain activity, or mental effort, is required to achieve high task demands. However, excessive mental workload can cause rapid fatigue, increased mistakes, and frustration. Which is why high workload may cause stress symptoms to appear [2]. Stress has a direct impact on people’s performance at their jobs, as well as everyday decision making, since it can be defined as a state where negative emotions are present, and energy is utilized ineffectively. Stress has an even greater impact on people with morbidities such as epilepsy since stress and negative emotions have been shown to be one of the major causes of epileptic seizures [3]. By creating a reliable way to detect MWL, stress and seizures can be prevented which would provide a better quality of life to patients. Electroencephalograms (EEG) have been shown to detect high MWL and stress, but conventional EEG requires devices that are difficult to wear in everyday scenarios, as well as needing trained professionals to operate the devices. Even when using machine learning models, training data requires to be labeled by experts. The use of wearable devices may allow stress to be detected in everyday scenarios, however, there exists another problem with this approach. Wearable devices can produce large amounts of data, which creates a challenge for data annotation. By using wearable devices, other physiological sensors could also be leveraged to detect MWL and stress. This Master's Thesis aims to mainly explore the contribution of multimodality for the detection of MWL with deep learning, as well as to explore a solution for the scarcity of labeled data through semi-supervised learning. Both objectives are in the context of multimodal wearable devices.

[1] GU Hao and Yin Zhong. “Mental workload assessment based on EEG and a hybrid ensemble classifier”

[2] AWK Gaillard. “Comparing the concepts of mental load and stress”

[3] Kristijonas Puteikis et al. “How are results of EEG activation procedures associated with patient perception of seizure provocative factors?”