Shuffled ECA-Net for stress detection from multimodal wearable sensor data.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Psychophysiology

Background:

• Stress is a significant factor in individual and societal problems.
• Existing stress detection methods using sensors are often impractical or subjective.
• A need exists for objective, reliable stress monitoring.

Purpose of the Study:

• To develop a novel deep neural network (DNN) model for psychological stress detection.
• To leverage multimodal sensor fusion for improved accuracy.
• To overcome limitations of existing subjective and impractical methods.

Main Methods:

• Developed a novel DNN, the shuffled efficient channel attention network (ECA-Net), for advanced feature-level sensor fusion.
• Acquired multimodal bio-signals (ECG, respiratory, EGG) and salivary cortisol data from 26 participants under relaxed and stressed states.
• Optimized and evaluated the model using a generated training dataset via five-fold cross-validation.

Main Results:

• The proposed shuffled ECA-Net achieved high accuracy (0.916), sensitivity (0.917), specificity (0.916), and F1-score (0.914).
• Demonstrated a high area under the receiver operating characteristic curve (AUROC) of 0.964.
• Confirmed that fusing multiple bio-signals with the shuffled ECA module enhances psychological stress detection accuracy.

Conclusions:

• The developed DNN model shows significant potential for accurate psychological stress detection.
• Multimodal sensor fusion using the shuffled ECA-Net is a viable and effective approach.
• This technology could substantially aid in addressing stress-related issues.