Stress is a prevalent aspect of our daily lives and one of the influential drivers of physical and psychological illness. Advancements in wearable computing enable the collection of physiological data in real time. Such physiological data can, in turn, be used to predict cognitive states, such as emotions or stress. Hence, this project aims to build robust systems at the intersection of human-computer interactions, specifically for stress recognition.
Current Research
In the field of affective computing, where emotional and cognitive states such as stress, workload, or relaxation are investigated, standardized data collection is essential. Our frameworks support the systematic acquisition of multimodal sensor data, the annotation of experimental phases, and the integration of self-reports and contextual information. This allows studies to be conducted more transparently, datasets to be compared more effectively, and future research projects to be built more efficiently.
Building on these frameworks, we design and conduct experimental studies in which physiological data are recorded under controlled conditions and during different activities, such as neutral baseline phases, cognitive workload, socio-evaluative stress, and physical activity. Beyond the data collection itself, we place strong emphasis on careful documentation, annotation, and preparation of the data for publication. By making such datasets available to the research community, we aim to improve reproducibility, support comparability across studies, and provide a reliable foundation for future analyses and model development.
Since many existing datasets differ in terms of sensors, sampling rates, study designs, label structures, and data formats, we also work on the integration of heterogeneous datasets. The goal is to transform data from different studies into a common and comparable structure. This includes harmonizing labels, unifying physiological signals, standardizing timestamps, and developing shared preprocessing pipelines. Such integration enables models to be trained on broader and more diverse data sources and allows their generalizability to be evaluated across different studies, participant groups, and measurement contexts.
Building on the collected and integrated datasets, we develop methods for automated stress recognition using physiological time series data. Signals such as photoplethysmography, electrodermal activity, temperature, and accelerometer data are used to identify stress-related patterns. We investigate both machine learning methods based on manually extracted features and deep learning approaches that learn meaningful representations directly from raw or preprocessed signals. A particular focus lies on robust and generalizable models that can account for individual differences between participants as well as differences between datasets and measurement settings.
As part of our research, we published VitaStress, a dataset containing physiological data of 21 subjects, collected during different activities or phases: neutral baseline, public speaking (socio-evaluative stress), arithmetic (cognitive stress), and physical activity. Physiological data include photoplethysmography, skin conductance, temperature, and accelerometer data. Additionally, subjects conducted self-assessment questionnaires throughout their trial. The dataset is rigorously annotated and labeled to increase the usability for other researchers.
Letzte Änderung: 6. August 2026