Cross Dataset Analysis for Generalizability of HRV-Based Stress Detection Models.

Area of Science:

• Psychology
• Computer Science
• Biomedical Engineering

Background:

• Stress is a global health concern with significant economic impact.
• Current stress monitoring relies heavily on self-reporting, limiting objective assessment.
• Artificial Intelligence (AI) shows promise for stress detection from physiological signals, but generalizability is a challenge.

Purpose of the Study:

• To develop a proof-of-concept tool for stress monitoring.
• To evaluate the generalization capabilities of Heart Rate Variability (HRV)-based machine learning models for stress detection.
• To compare Logistic Regression and Random Forest models in classifying stress across different datasets.

Main Methods:

• Utilized two machine learning models: Logistic Regression and Random Forest.
• Analyzed two distinct stress datasets differing in protocol, stressors, and recording devices.
• Performed leave-one-subject-out cross-validation within datasets and cross-dataset validation between datasets.

Main Results:

• Both models achieved good classification results on independent datasets.
• The Random Forest model demonstrated superior generalization capabilities.
• A stable F1 score of 61% was achieved by the Random Forest model in cross-dataset validation.

Conclusions:

• Random Forest models can effectively generalize HRV-based stress detection.
• Improved stress detection models can advance mental health and medical research.
• Integrating diverse models enhances the analysis of physiological stress indicators.