Stressalyzer: Convolutional Neural Network Framework for Personalized Stress Classification.

Area of Science:

• Biomedical Engineering
• Machine Learning for Health

Background:

• Stress detection is crucial for well-being but current methods are complex and resource-intensive.
• Existing techniques often require multiple sensors and hand-crafted features, limiting real-world application.

Purpose of the Study:

• To introduce Stressalyzer, a novel framework for stress classification using single-modality sensor data.
• To eliminate the need for feature computation and selection in stress detection systems.
• To demonstrate the effectiveness of personalized stress models.

Main Methods:

• Developed Stressalyzer, a neural network-based framework utilizing only Electrodermal Activity (EDA) sensor data.
• Employed a single-channel neural network architecture for stress classification.
• Conducted leave-one-subject-out analysis to evaluate model personalization.

Main Results:

• Achieved 92.9% classification accuracy and 0.89 f1 score for binary stress classification using EDA data alone.
• Demonstrated competitive performance compared to multi-modal, feature-intensive state-of-the-art methods.
• Showcased significant performance improvement with personalized stress models, with a 40% decline in performance without personalization.

Conclusions:

• Stressalyzer offers an efficient and effective approach to stress detection and classification using single-modality EDA data.
• Personalization of stress models is essential for accurate and reliable stress monitoring due to individual variability.
• The findings highlight the potential of Stressalyzer for practical, real-world stress management applications.