Abstract
Mental stress detection using wearable technologies is important for preventive healthcare; however, many existing approaches rely on expensive laboratory equipment or lack validation under realistic conditions. This thesis proposes a cost-effective, wearable Internet of Things (IoT)–based framework for automatic stress detection using physiological signals.
The system integrates non-invasive sensors, including photoplethysmography (PPG), galvanic skin response (GSR), and skin temperature, and employs signal preprocessing, handcrafted feature extraction, and supervised machine learning for binary stress classification (baseline versus stress). The framework is designed for edge-based operation to support real-time inference and data privacy.
Experimental evaluation was performed on a custom dataset and externally validated using a subset of the WESAD dataset. On the custom dataset, ensemble-based models achieved the best performance, with XGBoost and Random Forest reaching classification accuracies of 0.89 and 0.87, respectively. External validation demonstrated strong generalization, with the Random Forest model achieving a maximum ROC AUC of 0.92 on the WESAD dataset.
These results demonstrate that reliable stress detection can be achieved using affordable wearable sensors and lightweight machine learning models, supporting the practical deployment of privacy-preserving, real-world stress monitoring systems.
