Classification of Mental Stress from Wearable Physiological Sensors Using Image-Encoding-Based Deep Neural Network

Area of Science:

• Physiology
• Computer Science
• Machine Learning

Background:

• Chronic stress negatively impacts health, leading to conditions like cancer and cardiovascular disease.
• Existing machine learning methods for mental stress detection often use raw sensor data, leading to errors and poor performance.
• Wearable sensor data frequently contains corrupt values, further hindering accurate stress detection.

Purpose of the Study:

• To develop a deep learning-based method for accurate mental stress detection.
• To address limitations of previous methods that used raw, unprocessed sensor data.
• To improve the performance of stress detection models using wearable sensor datasets.

Main Methods:

• Encoding time-series raw data into Gramian Angular Field images.
• Utilizing a deep learning approach for stress level detection.
• Conducting experiments on the WESAD and SWELL benchmark datasets using chest sensor data (ACC, ECG, TEMP, RESP).

Main Results:

• Achieved high testing accuracies: 94.8% for the WESAD dataset and 99.39% for the SWELL dataset.
• Demonstrated the effectiveness of Gramian Angular Field encoding for stress detection.
• Showcased promising accuracy in detecting individual stress levels.

Conclusions:

• The proposed deep learning method, utilizing Gramian Angular Field images, offers a robust solution for mental stress detection.
• This approach mitigates issues associated with raw and corrupt sensor data, enhancing model reliability.
• The high accuracies achieved on benchmark datasets highlight the potential of this method for real-world applications.