A machine-learning approach for stress detection using wearable sensors in free-living environments

Area of Science:

• Physiology
• Computer Science
• Health Informatics

Background:

• Stress negatively impacts physical and mental health, necessitating early detection.
• Wearable sensors offer non-intrusive, continuous monitoring of vital signs for stress detection.
• Existing research predominantly focuses on stress data from controlled environments.

Purpose of the Study:

• To propose and evaluate a machine learning-based approach for stress detection in free-living environments using wearable sensor data.
• To compare the performance of various machine learning models for stress classification.
• To investigate the impact of data balancing techniques (SMOTE) on stress classification accuracy.

Main Methods:

• Utilized the SWEET dataset comprising electrocardiography (ECG), skin temperature (ST), and skin conductance (SC) from 240 subjects.
• Assessed K-Nearest Neighbors (KNN), Support Vector Classification (SVC), Decision Tree (DT), Random Forest (RF), and XGBoost (XGB) models.
• Evaluated models in binary and multi-class classification scenarios, both with and without the Synthetic Minority Over-sampling Technique (SMOTE).

Main Results:

• Random Forest achieved 98.29% accuracy and 97.89% F1-score for binary classification without SMOTE.
• K-Nearest Neighbors (KNN) performed best in binary classification with SMOTE (95.70% accuracy and F1-score).
• Random Forest excelled in three-level classification without SMOTE (97.98% accuracy, 97.22% F1-score), while XGBoost led with SMOTE (98.98% accuracy and F1-score).

Conclusions:

• Machine learning models effectively detect stress in free-living environments using wearable sensor data.
• Model selection and preprocessing techniques significantly influence stress classification performance.
• The study highlights the potential of wearable technology and AI for real-time stress management.