Stress detection using deep neural networks.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Health Informatics

Background:

• Over 70% of Americans experience stress, which detrimentally impacts physiological and psychological health.
• Chronic stress is linked to severe health conditions including cancer, cardiovascular disease, depression, and diabetes.
• Accurate and rapid human stress detection is crucial for public health.

Purpose of the Study:

• To develop advanced deep neural networks for enhanced stress detection and emotion classification.
• To overcome limitations of traditional machine learning methods requiring hand-crafted features.
• To analyze physiological signals from wearable sensors for improved stress monitoring.

Main Methods:

• Developed a 1-dimensional convolutional neural network (1D CNN) for chest-worn sensor data.
• Developed a multilayer perceptron (MLP) neural network for wrist-worn sensor data.
• Utilized deep neural networks to automatically extract features from raw physiological data for binary (stress/non-stress) and 3-class (baseline/stressed/amused) classification tasks.

Main Results:

• The 1D CNN achieved 99.80% accuracy for binary and 99.55% for 3-class classification.
• The MLP achieved 99.65% accuracy for binary and 98.38% for 3-class classification.
• Demonstrated significant accuracy improvements over previous machine learning methods.

Conclusions:

• Deep neural networks show great potential for robust, continuous, and noninvasive stress detection.
• These methods can aid in emotion classification, ultimately aiming to enhance quality of life.
• The study highlights the efficacy of deep learning in analyzing physiological signals for health monitoring.