Determination of the Effects of Transcutaneous Auricular Vagus Nerve Stimulation on the Heart Rate Variability Using a Machine Learning Pipeline

Figures

• 

  FIG. 1.. Experimental design. There were two sessions: no stimulation and stimulation. Each lasted 25 min and contained 10 min of baseline, 5 min of stress inducement, and 10 min of recovery. There was also a 20 min break in between the sessions.

• 

  FIG. 2.. Distribution of metric values derived from ECG as described in section S1.6 in Supplementary Material across subjects for each state for both conditions: with (b) and without (a) stimulation. The values do not have units for the frequency domain metrics (LF, HF, and LF/HF), and they represent spectral power for the respective bands. For the time domain metrics (RMSSD and SDNN), the values are in milliseconds. The box and whiskers refer to distribution parameters as described in section S1.7 in Supplementary Material. ECG, electrocardiogram; HF, high frequency; LF, low frequency; RMSSD, root mean squared of successive differences in heartbeat intervals; SDNN, standard deviation of N-N intervals.

• 

  FIG. 3.. Distribution of the changes in values of ECG-derived metrics between neighboring states, that is, baseline to stress (B→S) and stress to recovery (S→R) for both conditions: with (b) and without (a) stimulation. The values do not have units for the frequency domain metrics (LF, HF, and LF/HF), and they represent spectral power for the respective bands. For the time domain metrics (RMSSD and SDNN), the values are in milliseconds. We quantify the statistical significance of the change by testing against the null hypothesis that the population mean of the changes in metric is equal to 0 through the Wilcoxon signed-rank test. The resulting p-values are listed in the title of each panel.

• 

  FIG. 4.. Effect of stimulation on change in HRV metrics between neighboring states, i.e. baseline to stress (a) and stress to recovery (b). Values on the left show if a metric increased (>0) or decreased (<0) in going from one state to the next when no stimulation is applied. Values on the right correspond to the change in the value of the same metric for the stimulated condition. Lines joining the points are colored depending on whether they represent an increase (green), a decrease (orange), or a stationary value (purple). The threshold between increase, stationary, and decrease is set at 0.5×σ′, where σ′ is the standard deviation obtained from all the changes between the two relevant states with and without stimulation. p-Values in the title come from the Wilcoxon signed-rank test between the values on the left (Stim = 0) and on the right (Stim = 1) of each plot. HRV, heart rate variability.

• 

  FIG. 5.. Confusion matrices for discerning stimulated (Stim = 1) and nonstimulated (Stim = 0) conditions. The numbers inside the panels represent percentages of the total number of samples (1760 from 22 individuals) in the test set. Panel (a) is an aggregate across all three states, (b) baseline, (c) stress, and (d) recovery. Here the matrices (b, c, and d) are a decomposition of matrix (a).

• 

  FIG. 6.. (a) Distribution of mismatches between the true and predicted labels in stimulation condition classification across individual subjects, including the source of these mismatches across the three different states. The percentage of mismatches was calculated for each session (2 conditions and 22 subjects) across the 40 cross-validation iterations and placed into histogram bins 5% wide. (b–d) A similar histogram where the data were restricted to those laying in the baseline, stress, and recovery state, respectively.

• 

  FIG. 7.. (a) Confusion matrix for discerning baseline and stress states in a binary classification mode across the stimulated and nonstimulated conditions. In this figure the labels are Baseline=0, Stress=1, Recovery=2. The numbers inside the panels represent percentages of the total number of samples (1760 across 22 individuals) in the test set. (b) Confusion matrix for discerning baseline, recovery, and stress states in a three-way classification mode across the stimulated and nonstimulated conditions.