Comparing stress prediction models using smartwatch physiological signals and participant self-reports

Area of Science:

• Physiological computing
• Machine learning
• Wearable technology

Background:

• Wearable technology enables non-obtrusive monitoring of physiological signals for stress prediction.
• Machine learning models often use objective stressor tasks or subjective self-reports to measure stress.

Purpose of the Study:

• To compare the stress prediction performance of models using objective versus subjective stress markers.
• To develop personalized stress prediction models accounting for inter-individual differences.

Main Methods:

• A laboratory study with 32 volunteers using commercial smartwatches to collect physiological data during stressor tasks.
• Development and comparison of objective (stressor tasks) and subjective (self-reports) stress models using machine learning algorithms (SVM, RF, GBT, AdaBoost, LR).
• Evaluation of model performance using AUROC and F-1 scores, including personalized threshold models.

Main Results:

• Support Vector Machine (SVM) achieved the best performance for objective models (AUROC 0.790, F-1 0.623) and subjective models (AUROC 0.719, F-1 0.520).
• Personalized threshold models improved overall performance (AUROC 0.751, F-1 0.599).
• Subjective stress models showed lower performance, indicating a need for further research.

Conclusions:

• Stress prediction models using commercial smartwatches show performance comparable to state-of-the-art laboratory studies.
• Personalized threshold models offer a promising direction for enhancing stress prediction accuracy.
• Further research is needed to improve the efficacy of self-reports in stress-related studies.