A Method for Stress Detection Using Empatica E4 Bracelet and Machine-Learning Techniques.

Area of Science:

• Physiology
• Computer Science
• Biomedical Engineering

Background:

• Continuous stress monitoring is vital for preventing stress-related health issues.
• Wearable devices enable real-time data collection for personalized stress assessment.
• Identifying stress requires analyzing physiological signals effectively.

Purpose of the Study:

• To evaluate machine learning algorithms for differentiating stress levels using wearable sensor data.
• To determine the efficacy of Random Forest, SVM, and Logistic Regression in stress detection.
• To assess the performance of machine learning models with pre-processed physiological signals.

Main Methods:

• Collected photoplethysmographic and electrodermal activity data from 29 subjects using the Empatica E4 bracelet.
• Extracted 27 features from the physiological signals for analysis.
• Applied Random Forest, SVM, and Logistic Regression algorithms for binary stress classification, with feature importance ranking using chi-square and Pearson's correlation.

Main Results:

• The Random Forest model achieved the highest accuracy of 76.5% when utilizing all extracted features.
• Feature selection using the chi-square test improved consistency for the Random Forest model.
• Precision, recall, and F1-measure for Random Forest with feature selection were 71%, 60%, and 65%, respectively.

Conclusions:

• Machine learning, particularly the Random Forest algorithm, is effective for stress detection using wearable sensor data.
• Wearable devices combined with machine learning offer a viable approach for continuous, personalized stress monitoring.
• Feature selection can enhance the stability and reliability of machine learning models in physiological signal analysis for stress evaluation.