Automatic work-related stress detection among nurses
Nearly half of the nurses leave their jobs due to work-related stress. Surveys are no longer effective in capturing the real-time stress levels. Our project aims to revolutionize stress measurement using innovative technologies .
Factsheet
-
Schools involved
School of Health Professions
School of Engineering and Computer Science - Institute(s) Nursing
- Research unit(s) Innovation in the Field of Health Care and Human Resources Development
- Funding organisation BFH
- Duration 01.06.2024 - 31.12.2025
-
Head of project
Prof. Dr. Christoph Golz
Prof. Dr. Souhir Ben Souissi
Situation
Healthcare organisations face a shortage of health professionals, and almost half of the nurses leave their profession prematurely due to high work-related stress. Previous research on work-related stress relies on surveys, but the response rates are declining due to the high work-related stress among nurses and the increasing number of studies. To assess the effectiveness of measures, it is essential to continuously monitor stress. Therefore, alternative approaches to stress measurement that do not rely on surveys are needed, such as wearables and routine data. These approaches should provide insights into stress levels and predict potential job leaves without burdening the nursing staff with additional tasks. The central research question is whether it is possible to measure work-related stress among nurses accurately using these additional data sources and thus avoid the need for conventional surveys.
Course of action
We use a longitudinal design with repeated measurements involving 24 nurses. The focus is on data triangulation from the following data sources: (a) wearables to measure pulse, (b) text data from the clinical information system authored by the participants, and (c) participants' shift schedules. These data will be linked to the self-assessed work-related stress via the personal ID. After data collection, it will be verified that the information contains no identifying details. This ensures the protection of privacy while maintaining data integrity for accurate analysis. Moreover, we will develop a multimodal machine-learning model capable of synthesising and analysing these various data sources (a, b and c). This model aims to reveal complex patterns between self-assessed stress levels, biomarkers, linguistic indicators, and work schedules.
Result
A total of 24 nurses from two healthcare organizations participated in the pilot study. The analyses show promising results for stress detection using physiological data. While Random Forest performed less well, the CatBoost classifier achieved performance levels above 94% across all evaluation metrics. This suggests that boosting-based models are particularly effective at capturing complex patterns in physiological signals. Given the small sample size and identical stress labels within individual shifts, the findings primarily demonstrate methodological feasibility and provide a solid foundation for larger-scale studies. For text-based stress prediction, appropriate imbalance handling proved essential. Combining Random Forest with SMOTE substantially improved model performance, confirming the feasibility of text-based approaches when supported by suitable preprocessing techniques. The ScanWatch 2 smartwatch was rated as generally user-friendly. However, willingness to use the device on a daily basis was rated lowest, indicating potential challenges for long-term adoption. Over the three-month period, an average of eight shift changes occurred. Early shifts were perceived as the most stressful, while mid-shifts were rated as least stressful. Training activities, shift changes, and covering additional shifts increased work–privacy conflict, whereas late and half shifts were associated with lower conflict.
Looking ahead
Building on these findings, a follow-up project is planned that will collect more extensive and longitudinal data. The aim is to generalize the current results and develop robust models that enable reliable and practice-oriented stress measurement in everyday nursing work. In addition to expanding the sample, further modalities will be integrated and evaluated. By leveraging a multimodal data foundation, the project seeks to develop a model capable of capturing stress in a differentiated way and, in the long term, supporting preventive measures, feedback systems, and organizational improvements.
