Heart rate variability with circadian rhythm removed achieved high accuracy for stress assessment across all times throughout the day

FIGURE 1. The workflow for obtaining the classifier that identify stress across all times throughout the day. The XGBoost model is used as a classifier and trained on detrended HRV feature data from the SCWT experiment.

FIGURE 2. Details of data collection. (A) The age distribution and work status of volunteers participating in the SCWT experiment. (B) The process of the SCWT task. The RR interval data for each volunteer was collected over three consecutive days (marked by blue blocks). A Stroop Color Word Test was conducted at 8:30 a.m., 2:00 p.m., and 10:30 p.m. each day to induce stress (indicated by red vertical lines within the blue blocks).

FIGURE 3. Diagram illustrating the removal of outliers in RR intervals using confidence ellipses. The blue points represent the distribution of the transformed two-dimensional RR interval data, where the values of each point’s two dimensions correspond to the lengths of two consecutive RR intervals. The red area indicates the 95% confidence ellipse interval.

FIGURE 4. Effect of outlier removal. Comparison of 72-h RR interval data (RRI) before and after outlier removal for a volunteer. The blue part represents the data distribution before outlier removal, and the red part represents the data distribution after outlier removal.

FIGURE 5. Comparison of CV coefficients before and after outlier removal for the RR interval data of all volunteers. After outlier removal, the average CV coefficient decreased from 0.28 to 0.2.

FIGURE 6. Distribution of meanNN and HR during the SCWT tasks at three time points each day: (A) meanNN , (B) HR . The blue box plots represent the data distribution for the resting state, with blue dots indicating the median. The orange box plots represent the data distribution for the stress state, with orange stars indicating the median.

FIGURE 7. Period spectrum plots of HRV features for all volunteers. The blue line represents the average amplitude, and the shaded area corresponds to the interval estimate.

FIGURE 8. Results of the second-order sine function fitting for HRV features: (A) meanNN , (B) HR , (C) PNN50 , (D) ST1 .

FIGURE 9. Comparison of feature distributions between rest and stress states before and after removing circadian rhythm trends. For each module, the left panel shows the feature distribution before trend removal, and the right panel shows the distribution after trend removal.

FIGURE 10. Ranking of the importance of the 10 HRV features obtained using XGBoost.

FIGURE 11. Comparison of the accuracy of stress assessment for different time periods using XGBoost classifiers trained on SCWT experimental data for each time period. This includes classifiers trained using only morning (A), afternoon (B), or evening (C) experimental data.

FIGURE 12. Comparison of the accuracy of stress assessment for different time periods using XGBoost classifiers trained on a dataset that randomly samples one-third of the integrated data from different SCWT tasks.