Development of an algorithm for stress estimation based on wearable data

Burnout syndrome, a condition resulting from chronic work-related stress, is increasingly prevalent among workers, underscoring the critical need for effective stress management. Physiologically, chronic stress triggers hyperactivation of the sympathetic nervous system, leading to elevated heart rates. This correlation between stress and Heart Rate (HR) is well-documented in numerous studies. Heart Rate Variability (HRV), quantifiable through various metrics derived from beat-to-beat intervals within specific time windows, has also emerged as a dependable indicator of stress, commonly employed in assessing the balance between the two branches of the autonomic nervous system. Consequently, the development of algorithms for continuous stress detection in wearable devices has gained momentum, enabling the extraction of long-term stress trends potentially associated with burnout. The objective of this study is to develop a general-purpose algorithm for stress estimation using cardiac features obtained from wearable devices’ data. Data from 23 employees were collected in real-world settings and analyzed over a one-month period. HRV metrics were extracted and compared with literature references to ensure a reliable sample. Tree-based machine learning (ML) models, such as Decision Trees, Random Forest, AdaBoost, LightGBM, and XGBoost, were employed to estimate stress levels, categorized into four classes. Subsequently, regression models, specifically XGBoost, were used, integrating additional HR statistical features to enhance performance. Despite meticulous hyperparameter tuning, ML classifiers demonstrated comparable performances, failing to achieve satisfactory results (best balanced accuracy: 53%) due to a significant class imbalance. Conversely, regressors, assessed with various combinations of HRV or HR features, exhibited lower errors, reaching a Mean Absolute Error (MAE) of 9.11. This could be attributed to the nature of regressors: trained to minimize mean squared error, they penalize larger errors more heavily compared to classifiers, trained to minimize cross-entropy loss. Furthermore, exclusively modelling stress using data during sleep significantly enhanced regression performance (MAE: 5.60). Interpretability methods (SHAP) were employed to gain insights into the functioning of the models, revealing a shift in feature importance between daytime and nighttime conditions: HR features dominated during the day, while HRV features became more crucial during sleep. Overall, in the context of this study, the results emphasize the potential benefits of regression models over classification counterparts for stress estimation, highlighting the importance of integrating both HRV and HR features to achieve accurate estimations. However, some limitations emerged during the study, including the presence of uncleaned motion artifacts due to the absence of accelerometer data in the considered dataset. In addition, the small dataset size and short duration of the study might limit the generalizability of the results. Despite these challenges, the study aligns with existing literature and sets the stage for further algorithm refinement and validation across diverse datasets. Starting from the present findings, future research may also involve new strategies, such as exploring neural networks, alternative features combinations, dataset expansion, and experimenting with different smartwatch brands.