Analysis of Cardiac Activities for Stress Quantification Using Machine Learning Approaches
Wonjun Choi
Asheville School, City, Asheville, United States
Publication date: November 20, 2025
Asheville School, City, Asheville, United States
Publication date: November 20, 2025
DOI: http://doi.org/10.34614/JIYRC2025II31
ABSTRACT
Stress is one of the main causes of morbidity in the fields of neurological and cardiovascular diseases, but its quantification has been difficult to quantify due to heterogeneous physiological changes and limitations in the sensors. This present study systematically compares widely recorded biophysical signs to determine the most reliable markers for physiological stress and to determine the suitability for data-driven prediction. Adhering to the directions given in the PRISMA statement, we used Google Scholar and PubMed and summarized results across biomedical engineering, clinical, and device-based research. Heart rate (HR), heart rate variability (HRV), blood pressure (BP), skin conductance (SC), and respiration rate (RR) signal classes were considered. Meta-analytic comparisons show that skin conductance, in particular Galvanic Skin Response (GSR), has the greatest and most uniform separation between the stressed and control states. HRV parameters, most notably the low-frequency to high-frequency (LF/HF) ratio, also showed marked modulation in response to mental effort. As a representative machine-learning analysis, a multi-feature logistic regression model on HRV features gave 94.9% classification performance, highlighting the potential for data-driven approaches for real-time recognition of psychological stress. The results confirm SC/GSR as the main indicator, with HRV providing ancillary information, and collectively provide a basis for next-generation wearable devices to allow for non-invasive, real-time tracking of stress for remote health monitoring, with future research focusing on validation across heterogeneous cohorts and the standardization of acquisition procedures to improve model generalizability.
Stress is one of the main causes of morbidity in the fields of neurological and cardiovascular diseases, but its quantification has been difficult to quantify due to heterogeneous physiological changes and limitations in the sensors. This present study systematically compares widely recorded biophysical signs to determine the most reliable markers for physiological stress and to determine the suitability for data-driven prediction. Adhering to the directions given in the PRISMA statement, we used Google Scholar and PubMed and summarized results across biomedical engineering, clinical, and device-based research. Heart rate (HR), heart rate variability (HRV), blood pressure (BP), skin conductance (SC), and respiration rate (RR) signal classes were considered. Meta-analytic comparisons show that skin conductance, in particular Galvanic Skin Response (GSR), has the greatest and most uniform separation between the stressed and control states. HRV parameters, most notably the low-frequency to high-frequency (LF/HF) ratio, also showed marked modulation in response to mental effort. As a representative machine-learning analysis, a multi-feature logistic regression model on HRV features gave 94.9% classification performance, highlighting the potential for data-driven approaches for real-time recognition of psychological stress. The results confirm SC/GSR as the main indicator, with HRV providing ancillary information, and collectively provide a basis for next-generation wearable devices to allow for non-invasive, real-time tracking of stress for remote health monitoring, with future research focusing on validation across heterogeneous cohorts and the standardization of acquisition procedures to improve model generalizability.
