Deep learning for stress oriented human activity recognition.

Area of Science:

• Biomedical Engineering
• Computer Science
• Artificial Intelligence

Background:

• Human Activity Recognition (HAR) using sensor data is crucial for understanding mental and physical states, particularly for behavioral disorders.
• Previous studies have explored deep learning for HAR, but stress detection accuracy needs improvement.

Purpose of the Study:

• To enhance stress detection accuracy in Human Activity Recognition (HAR) using advanced deep learning models.
• To evaluate the performance of Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM) networks, and Transformer models for stress-related HAR tasks.

Main Methods:

• Utilized benchmark datasets for stress-related activities.
• Implemented and compared RNN, LSTM, and Transformer deep learning architectures for feature extraction and classification.
• Investigated the impact of window size and overlap ratio on classification accuracy.

Main Results:

• Transformer models achieved the highest classification accuracy at 97.83%, outperforming LSTM (97.36%) and RNN (92.4%).
• The proposed Transformer-based approach demonstrated significant improvement over existing deep neural networks on the Stressense dataset.
• Model performance was sensitive to variations in window size and overlap ratio.

Conclusions:

• Transformer architectures are highly effective for HAR tasks, particularly in stress detection.
• The enhanced accuracy in stress detection using HAR shows promise for developing non-intrusive mental health monitoring systems.
• Wearable sensor data combined with advanced deep learning offers a pathway to seamless behavioral health assessment.