Objective:

To improve prediction performance in recognizing stress-related human activities using advanced deep learning architectures, specifically RNNs, LSTMs, and Transformers.

Key Findings:

• Transformer models achieved the highest classification accuracy of 97.83%, outperforming LSTM (97.36%) and RNN (92.4%) models, highlighting the effectiveness of advanced architectures in this domain.
• The proposed approach significantly improved upon previous results from the Stressense dataset study, indicating a step forward in stress detection methodologies.

Interpretation:

Transformer-based architectures are highly effective for Human Activity Recognition tasks related to stress detection, suggesting significant potential for enhanced mental health monitoring through wearable devices in real-world applications.

Limitations:

• The study primarily focuses on benchmark datasets, which may limit generalizability to real-world scenarios, as actual user behavior can vary significantly.
• Variability in sensor data quality and environmental factors could affect model performance, necessitating further research in diverse settings.

Conclusion:

The findings underscore the potential of deep learning techniques, particularly Transformers, in advancing stress detection through non-intrusive monitoring methods, contributing to the growing body of literature in this field.