Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2023 Jul 1.
Published in final edited form as: Psychophysiology. 2022 Feb 12;59(7):e14013. doi: 10.1111/psyp.14013

Validation of a new impedance cardiography analysis algorithm for clinical classification of stress states

Shafa-at Ali Sheikh 1,2, Nil Z Gurel 3, Shishir Gupta 5, Ikenna V Chukwu 5, Oleksiy Levantsevych 5, Mhmtjamil Alkhalaf 5, Majd Soudan 5, Rami Abdulbaki 5, Ammer Haffar 5, Gari D Clifford 1,4, Omer T Inan 2, Amit J Shah 5,6,7
PMCID: PMC9177512  NIHMSID: NIHMS1772283  PMID: 35150459

Abstract

Pre-ejection period (PEP) is an index of sympathetic nervous system activity that can be computed from electrocardiogram (ECG) and impedance cardiogram (ICG) signals, but sensitive to speech/motion artifact. We sought to validate an ICG noise removal method, three-stage ensemble-average algorithm (TEA), in data acquired from a clinical trial comparing active versus sham noninvasive vagal nerve stimulation (tcVNS) after standardized speech stress. We first compared TEA’s performance versus the standard conventional ensemble-average algorithm (CEA) approach to classify noisy ICG segments. We then analyzed ECG and ICG data to measure PEP and compared group-level differences in stress states with each approach. We evaluated 45 individuals, of whom 23 had posttraumatic stress disorder (PTSD). We found that the TEA approach identified artifact-corrupted beats with intraclass correlation coefficient >0.99 compared to expert adjudication. TEA also resulted in higher group-level differences in PEP between stress states than CEA. PEP values were lower in the speech stress (vs. baseline rest) group using both techniques, but the differences were greater using TEA (12.1 ms) than CEA (8.0 ms). PEP differences in groups divided by PTSD status and tcVNS (active vs. sham) were also greater when using the TEA versus CEA method, although the magnitude of the differences was lower. In conclusion, TEA helps to accurately identify noisy ICG beats during speaking stress, and this increased accuracy improves sensitivity to group-level differences in stress states compared to CEA, suggesting greater clinical utility.

1. Introduction

Psychological stress is a common risk factor for heart disease that is difficult to quantify through objective physiologic measures of cardiac risk (Dimsdale, 2008; Menezes et al., 2011; Chaddha et al., 2016; Bremner et al., 2018). Psychological stress causes activation of the sympathetic nervous system. One candidate autonomic measure is the pre-ejection period (PEP), which is inversely related to myocardial contractility and has been validated as an index of cardiac sympathetic beta-adrenergic activity (Newlin & Levenson, 1979; Cacioppo et al., 1994; Mezzacappa et al., 1999). Therefore, PEP is an important non-invasive measure of sympathetic nervous system (SNS) activity and can help examine a wide range of neuropsychiatric and heart diseases, including post-traumatic stress disorder (PTSD) (Shah et al., 2013; Peters et al., 2018; Cohen et al., 2020). Laboratory-induced mental stress paradigms help researchers understand the autonomic response to various conditions. Research in recent years unveiled that the participants with PTSD or relevant animal models exhibit pathological autonomic activity - SNS reactivity is likely higher, while parasympathetic nervous system (PNS) activity is likely lower (Shah et al., 2013; Cohen et al., 2007; Meyer et al., 2016; Bremner et al., 2018; Park et al., 2017). Hence, neuromodulation treatments that target the autonomic imbalance in PTSD have gained interest (Gurel, Huang, et al., 2020; Bremner et al., 2019; Peña et al., 2014; Lambet et al., 2017). Robust PEP computation is a critical step toward quantification of autonomic activity for assessing either the effectiveness of potential treatments or extracting information about the neurobiology of the disease.

Impedance cardiography is a non-invasive and cost-effective method used to measure PEP with the simultaneously obtained electrocardiogram (ECG). PEP is measured by the time interval between the onset of the ECG Q or R point (onset or peak of ventricular depolarization) and the impedance cardiogram (ICG) B point (representing the opening of the aortic valve) (Seery et al., 2016). Accurate measurement of PEP relies on the precise measurement of the timing of these points, which can be difficult under certain conditions. This problem is particularly important in the application of PEP measurements to settings in which muscular activity related to speech or movement occurs (Van et al., 2015).

Respiration, speaking, and movement are examples of activities that may reduce the accuracy of PEP measurement (Sherwood et al., 1990). Some authors found a noisy signal even for the recordings where participants were seated, while others do not find it problematic, and observed little noise. In general, small movements can induce noisy artifacts in the signal (Czarnek et al., 2021). For example, 22% of the ICG signal were identified as noisy in participants who were reading in a seated position (Cybulski et al., 2011). Bicycle exercise can cause larger distortions in the location of B point (opening of the aortic valve) on the ICG signal and therefore limit the ability to measure sympathetic activation with exercise (Hurwitz et al., 1990). This underlies the need for a suitable ICG noise removal algorithm for real-world PEP computation.

A popular method to mitigate the noise in the ICG signal is ensemble averaging (EA) (Kizakevich et al., 1976; Sornmo & Laguna, 2005; Malcolm et al., 2016; Kibler, 2018; Silvia et al., 2020; Czarnek et al., 2021), which we refer to as the “conventional EA algorithm” (CEA). In our prior work, we have improved upon this approach with an open-source automated ICG noise removal algorithm – three-stage EA algorithm (TEA) (Sheikh et al., 2020). In the CEA, R peaks of the synchronized ECG signals were detected and used as the reference points to average both ECG and ICG signals without removal of noisy beats. However, the ICG signal remained affected by intra-subject variability of signal shape and event latency (Hurwitz et al., 1993; Cybulski, 2011). The TEA addressed these issues by forming the final averaged beat after noisy beat removal in three stages. The first stage of amplitude filtering removed spikes and noisy beats based on amplitude threshold values outside of normal physiological ranges. The second and third filtering stages, based on cross-correlation, removed artifact-corrupted beats with a significantly different morphology from the averaged beat templates. We showed that the TEA outperformed the CEA by identifying and removing noisy beats, but the signal was relatively free from the artifact and the sample consisted entirely of older white men.

In this study, we evaluate TEA and CEA in a more diverse sample undergoing speech and arithmetic speaking tasks, where motion artifacts may lead to significant misclassification. Our study involves an ancillary analysis of 45 participants with and without PTSD randomized to active and sham transcutaneous cervical noninvasive vagal nerve stimulation (tcVNS) (Gurel, Huang, et al., 2020). We hypothesize that the TEA approach, compared to CEA, would better identify noisy versus (vs.) clean ICG beats and more accurately measure PEP. We also hypothesize that this innovative approach results in increased clinical validity of PEP and increases the contrast between groups that likely have different levels of autonomic activation: PTSD (vs. control), a speech/math stressor (vs. baseline), and tcVNS (vs. sham) (Miller & Sita, 1994; Willemsen et al., 1998; Matthews et al., 2003; Kelsey et al., 2007; Kelsey et al., 2010).

2. Method

2.1. Dataset and the Study Protocol

This study was performed as an ancillary project on the data from the project, “Closed Loop Vagal Nerve Stimulation for Patients with PTSD," which examines the “the potency and kinetics of the neurologic, autonomic peripheral, inflammatory, and behavioral responses to tcVNS vs. sham treatment, at baseline and in response to stressful traumatic scripts related to personal traumatic events, as well as a series of other stressors" (Bremner 2016). The work was approved by the institutional review boards of Emory University (IRB00091171), Georgia Institute of Technology (H17126), SPAWAR Systems Center Pacific, and the Department of Navy Human Research Protection Program. The study spanned over three consecutive days for individuals with PTSD and non-PTSD trauma controls to test the kinetics of tcVNS over several days. On the first day, the participants listened to traumatic scripts which have been shown to acutely increase PTSD symptoms (Fani et al., 2011), neutral scripts, which act as a control for the traumatic scripts to be compared against and underwent sessions of active or sham tcVNS. The second and the third days were identical and involved both speech and arithmetic tasks requiring verbal responses. Depending on their assigned randomization, active or sham tcVNS was administered in between the speech and arithmetic tasks. Written informed consent from each participant was obtained before study initiation. Details of the protocol can be found in Gurel, Huang, et al. (2020). Keeping in view the vulnerability of the PTSD subjects, it was ensured that participants shortlisted for this study were not psychiatrically unstable with suicidal/homicidal ideations or psychosis. Also, all participants were closely monitored for distress, and offered immediate counseling as needed by the study coordinators and study psychiatrist if needed.

In this study, we analyzed the speaking stressors which occurred on day two of the protocol. Figure 1 (a) represents the data set of 45 (18 male, 27 female) participants with ages having a mean (standard deviation (SD)) of 33.7 (11.4) years. Out of the 45 subjects, 22 had PTSD, and 23 had prior psychological trauma exposure without PTSD. Table 1 represents characteristics of PTSD (n =22) and non-PTSD (n = 21) participants with active and sham tcVNS. It is pertinent to highlight that the records for two non-PTSD participants in the arithmetic stage were rejected due to inferior quality, therefore, Table 1 presents the data for 21 non-PTSD participants. The gender disproportion in PTSD participants is also the representation of the fact that women have two to three times higher risk of developing PTSD than men (Olff, 2017).

Figure 1:

Figure 1:

Open in a new tab

Data set and study protocol. (a) represents the data set of 45 participants (PTSD: 22, and non-PTSD: 23) in the study protocol depicted in (b). All participants underwent the stages of baseline, preparation to speech, and speech task. Out of 22 participants with PTSD, we randomized 10 to receive active tcVNS, and 12 to receive sham tcVNS. Out of 23 non-PTSD participants, we randomized 11 to receive active tcVNS, and 12 to receive sham tcVNS. All participants regardless of tcVNS type underwent the recovery phase for 8 mins before the arithmetic task. Active and sham tcVNS were also applied after the arithmetic task but not shown in the figure since it was not analyzed. These tcVNS phases were not studied due to the induction of the electrical stimulation artifacts in the ICG signal (Gurel et al., 2020). Image (c) Emory University, CC-BY-SA.

Table 1:

Mean (SD) of characteristics of PTSD and non-PTSD participants with active and sham tcVNS along with PEP measured during baseline stage using the CEA and TEA.

Characteristics PTSD Participants Non-PTSD Participants
Active tcVNS
(n=10)		 Sham tcVNS
(n=12)		 Active tcVNS
(n= 10)		 Sham tcVNS
(n= 11)
Age (years) 35.5 (12.9) 37.6 (13.3) 29.6 (5.8) 32.5 (10.4)
Gender (Male/Female) 1/9 5/7 7/3 5/6
Height (cm) 168.0 (11.0) 172.5 (11.95) 172.0 (15.38) 171.8 (7.48)
Weight (kg) 81.6 (29.1) 88.93 (23.32) 77.1 (15.39) 82.9 (12.60)
BMI (kg/m2) 28.4 (7.9) 29.45 (4.9) 26.3 (5.99) 28.2 (4.88)
Baseline PEP (ms) computed using algorithms
TEA 98.1 (20.7) 89.2 (18.9) 94.6 (13.0) 94.6 (10.7)
CEA 95.4 (22.3) 88.6 (20.7) 92.4 (13.5) 95.4 (10.5)
Open in a new tab

Figure 1 (b) depicts the study protocol. The speech task started with a resting baseline (8 minutes), followed by the speech preparation (2 minutes) and the speaking task (3 minutes). Participants spoke about a scenario in which they were accused of theft. Active or sham tcVNS for two minutes was applied after the speech task. Regardless of the type of tcVNS, all participants rested for 8 minutes in the recovery stage. Then, the arithmetic task consisted of arithmetic questions for two to three minutes with time pressure, negative feedback, and questions of increasing difficulty.

Synchronized ECG and ICG signals were recorded by trained researchers using impedance and electrocardiography equipment (BIOPAC Systems Inc., Goleta, CA). Three-lead ECG signals were recorded with dual wireless respiration and ECG RSPEC-R. Two disposable electrodes (3 M red dot electrodes) were attached to the collar bone, with a ground electrode located on the hip. The ICG signal was recorded using the wireless NICO-R module using four disposable spot electrodes EL507. The outer current electrodes (one at the side of the neck and one at the left lateral side of the thorax below the xiphisternal junction) were used for injection of alternating current of 1 mA (rms) at 50 kHz frequency. Inner voltage electrodes (one at the side of the neck below the current electrode and one at the left lateral side of the thorax at the xiphisternal junction) were used for voltage pick-up. Both ECG and ICG signals were sampled at 2000 Hz with a resolution of 16 bits.

Hand-held tcVNS devices (GammaCore, ElectroCore, Basking Ridge, New Jersey) were used for the application of both active and sham stimuli using collar electrodes on the left side of the neck. An active stimulation device produced an AC voltage signal consisting of sine wave pulses at a frequency of 5 kHz, and a duration of 1 ms (i.e., 5 complete periods), at intervals of 40 ms. A continuous AC biphasic square wave voltage signal of period 200 ms was produced by the sham device. The sham device delivers a mild buzzing sensation like the active device without stimulating the vagus nerve (Gurel, Huang, et al., 2020).

2.2. Preprocessing

Acknowledge 5.0 was used for data storage, display, and export to .mat format. MATLAB R2017b (MathWorks, Inc., Natick, MA) was used for the signal processing, application of the CEA and TEA, and PEP extraction/analysis. To remove the low-frequency drift and high-frequency noise, the ECG signals were filtered with finite impulse response (FIR) band-pass filters, with lower and upper cutoff frequencies of 0.6 Hz and 40 Hz (Sornmo & Laguna, 2005), respectively, and the ICG signal was filtered with a 4th-order Butterworth band-passed filter with the lower and upper cutoff frequencies of 0.5 Hz and 40 Hz, respectively. To avoid phase shift, the FIR and the Butterworth filter were applied in both forward and backward directions (Gustafsson, F., 1996).

2.3. Application of the Conventional and Three-Stage EA Algorithm

EA ECG and EA ICG signals were formed from the respective 60-second non-overlapping windows, which will be referred to as the “analysis window" in this study. The PhysioNet Cardiovascular Signal Toolbox was used to detect R peaks in the ECG signal analysis window using jqrs (Behar et al., 2014; Johnson et al., 2014; Vest et al., 2018). The median RR interval of the ECG analysis window was used to segment the ECG signals. The ICG signal in an analysis window and R peaks detected from the synchronized ECG analysis window were used to form EA ICG beats using the CEA and TEA. In the CEA, R peaks of the synchronized ECG signals were detected and used as the reference points to average both ECG and ICG signals without removal of noisy beats. The TEA formed the final averaged beat after removal of noisy beats in three stages. The first stage of amplitude filtering removed spikes and noisy beats based on amplitude threshold values outside of normal physiological ranges. The second and third filtering stages, based on cross-correlation, removed artifact-corrupted beats with a significantly different morphology from the averaged beat templates. Additional details on the description/application of CEA and TEA are available in our previous publication as referenced (Sheikh et al., (2020)).

2.4. Evaluation Method

2.4.1. Technical evaluation

Figure 2 presents the scheme for technical evaluation of the TEA to classify noisy signals during the speech and arithmetic tasks. After pre-processing, an analysis window was randomly selected from each participant which resulted in the availability of 3700 and 3500 ICG beats from speech and arithmetic tasks, respectively. Two expert physicians (expert-A, and expert-B) with physiology training independently performed the visual inspection of the ICG beats to identify the noisy beats. The inspected beats were then collectively reviewed and noisy beats were rejected with consensus. In the case of TEA, the noisy beats were automatically rejected to obtain the clean beats. The output of the collective expert-inspection and TEA was used to perform confusion matrix analysis for computing accuracy, sensitivity, specificity, positive predictive value (PPV), and F1 score (F1). The ICG beats from both tasks were also used to plot receiver operating characteristic curve for the classification of noisy and clean beats.

Figure 2:

Figure 2:

Open in a new tab

Technical evaluation scheme for the TEA in the speech and arithmetic tasks. The boxes show different steps, while the ellipses show the evaluation parameters in the technical evaluation scheme. After pre-processing of the ECG and ICG signal, a non-overlapping synchronized window of 60-second duration was extracted from all participants. This resulted in the availability of 3500 and 3700 ICG beats from speech and arithmetic tasks, respectively. Noisy ICG beats were identified by 2 experts (Expert- A and Expert- B), which were removed after reaching a consensus via collective inspection. The TEA automatically detected and removed the noisy beats. Noisy and clean beats from the expert-inspection and TEA were used for confusion matrix analysis and plotting receiver operating characteristic curves. After the removal of noisy beats via collective expert inspection (2 experts), all clean beats were averaged to form an expert EA beat. Three-stage EA beats were formed by averaging the TEA-detected clean beats. The noisy beats were not removed for the application of CEA to form the conventional EA beat. R-peak on EA ECG beats was automatically detected, while the B-point was visually marked by collective expert visual inspection on the expert, conventional, and three-stage EA ICG beats. The R-peaks and B points were used to compute PEP values for all three methods. The computation of intraclass correlation coefficients, Bland-Altman plot analysis, and qualitative analysis was performed using EA beats and respective PEP values from the three methods. Image (c) Emory University, CC-BY-SA.

For a given analysis window of ICG beats, the expert EA beat was formed by averaging expert-identified clean beats. The CEA beat was formed by averaging all beats of the analysis window without any noise removal, and the TEA beat was the output of our algorithm. PEP was measured as RB interval (from R peak of EA ECG beat to the B point of synchronized EA ICG beat) (Seery et al., 2016; Kelsey et al., 2010). R peaks were automatically detected (Vest et al., 2018), whereas expert-A and expert-B independently performed manual annotation of the B-point on the averaged beats from all methods as per guidelines (Sherwood et al. 1990; Arbol et al. 2017) using the “Impedance Cardiogram Manual Annotation application (ICMAA)” introduced in open-source ICG toolbox (Sheikh S. A., Shah A, Inan O. T., & Clifford G. D., 2020). For each method, the B-point annotations by expert-A and expert-B were also reviewed collectively and edited to ensure accurate scoring for PEP computation. Using these PEP values, the inter-rater reliability between the expert and algorithms (CEA and TEA) was assessed using intraclass correlation coefficients (ICC) based on 2-way random effects absolute agreement model (Model (2,2)) (Shrout & Fleiss, 1979; McGraw & Wong, 1996; Brownhill, 2020). Agreement between the expert/CEA and the expert/TEA was also assessed using Bland-Altman plots for PEP computed from 45 participants in both speech and arithmetic tasks (Bland & Altman, 1986). Qualitative analysis for different ranges of PEP differences in the Bland-Altman plots was also performed by comparing the morphology of the expert, CEA, and TEA EA beats.

2.4.2. Evaluation scheme for classification of clinically defined stress states

Table 2 lists the three comparisons using PEP analysis that we performed as part of clinical validation of TEA vs. CEA for both PTSD and non-PTSD participants. In the first comparison, PEP intervals were compared between the study stages of the baseline and speech task. We compared the difference in the PEP change from baseline to speech between PTSD and non-PTSD participants in the second analysis. In the third analysis, a PEP comparison was made between the baseline and arithmetic tasks with active and sham tcVNS. The data distributions of PEP values for these groups under different study stages were checked for normality using the Lilliefors test (Lilliefors, 1969). PEP data were compared between groups using paired/unpaired t-test (for normal data distributions) and Wilcoxon signed-rank/rank-sum test (for non-normal data distributions) with 95 % confidence interval (CI). The standardized differences between the means of the groups were also computed using Cohen’s d effect size. The area under the curve (AUC) of the receiver operating characteristic (ROC) curves for classification between distinct groups in different study stages were calculated based on PEP values.

Table 2:

PEP analysis of comparisons for different study groups under different study stages. PEP intervals were computed using both CEA and TEA. The statistically significant difference between the PEP means of these groups under different study stages was evaluated. Also, AUC of ROC curves for classification between distinct groups based on PEP values were calculated.

Study stage Study group
Baseline vs Speech task PTSD participants
Non-PTSD participants
Difference between Speech and Baseline task PTSD vs non-PTSD participants
Arithmetic task (Active tcVNS vs Sham tcVNS) PTSD participants
Non-PTSD participants
Open in a new tab

3. Results

3.1. Technical Validation

The confusion matrix in Table 3 illustrates the performance of the automated technique (TEA) for detecting noisy ICG beats compared to the expert annotation results on 3700 annotated ICG beats selected at random during the speech task. An accuracy of 92.0%, sensitivity of 89.1%, specificity of 93.9%, PPV of 90.5%, and F1-score of 89.8% were achieved by the TEA. For the arithmetic task, 3500 randomly selected ICG beats were evaluated against the expert annotation and results are compiled in the confusion matrix in Table 4. TEA achieved an accuracy of 94.6%, sensitivity of 87.5%, specificity of 97.6%, PPV of 93.7%, and F1-score of 90.5%.

Table 3:

Confusion matrix for performance analysis of the TEA over 3700 ICG beats in the speech task.

Expert
Noisy ICG beats Clean ICG beats Total
TEA Noisy ICG beats 1297 (TP) 159 (FP) 1456
Clean ICG beats 136 (FN) 2108 (TN) 2244
Total 1433 2267 3700
Open in a new tab

Table 4:

Confusion matrix for performance analysis of the TEA over 3500 ICG beats in the arithmetic task.

Expert
Noisy ICG beats Clean ICG beats Total
TEA Noisy ICG beats 899 (TP) 60 (FP) 959
Clean ICG beats 129 (FN) 2412 (TN) 2541
Total 1028 2472 3500
Open in a new tab

Figure 3 depicts the ROC curves for the classification of noisy and clean ICG beats in the speech and arithmetic tasks using the TEA with higher AUCs for speech and arithmetic tasks. For both speech and arithmetic tasks, the ICCs, presented in Table 5, were found to be high, for the expert vs. TEA as compared to the expert vs. CEA. Bland-Altman plots showed closer agreement between the expert and the TEA compared to the expert and the CEA at different PEP values across participants for both speech and arithmetic tasks as depicted in Figures 4 and 5.

Figure 3:

Figure 3:

Open in a new tab

ROC curves for classification between noisy and clean ICG beats in the speech and arithmetic tasks. For the TEA, AUC of ROC curves for classifying noisy and clean ICG beats in the speech and arithmetic tasks are 0.91 and 0.92, respectively. Image (c) Emory University, CC-BY-SA.

Table 5:

Intraclass correlation coefficient analysis between the expert and algorithms (TEA and CEA). Two experts, expert-A and expert-B, performed independent inspections of ICG beats to identify the noisy beats. After collective review, the noisy beats were rejected, and clean beats were averaged to form the “expert” EA beat. Both experts also visually marked the B-point independently on the EA beats generated using expert, TEA, and CEA. For each method, the B-point annotations were also collectively reviewed and edited to ensure accurate scoring. PEP values for each method were computed as RB intervals, which were used to perform ICC analysis. Uniformly consistent higher ICC values have been observed for expert vs. TEA as compared to expert vs. CEA.

Comparison Task Intraclass Correlation Coefficient (ICC)
PTSD Non-PTSD Overall
Expert vs. TEA Speech 0.998 0.998 0.998
Arithmetic 0.999 0.999 0.998
Expert vs. CEA Speech 0.832 0.868 0.841
Arithmetic 0.950 0.923 0.939
Open in a new tab

Figure 4:

Figure 4:

Open in a new tab

Bland Altman plot analysis for the speech task. (a) Plot for PEP difference between the expert and the TEA estimates against the expert PEP with bias = 0.3 ms and limits of agreement = [−2.8, 3.4] ms. (b) Plot for PEP difference between the expert and the CEA estimates against the expert PEP with bias = −4.8 ms and limits of agreement = [−28.4, 18.9] ms. Closer agreement of PEP estimates between the expert and TEA was observed as compared to that between the expert and the CEA. Image (c) Emory University, CC-BY-SA.

Figure 5:

Figure 5:

Open in a new tab

Bland Altman plot analysis for the arithmetic task. (a) Plot for PEP difference between the expert and the TEA estimates against the expert PEP with bias = 0.1 ms and limits of agreement = [−1.9, 2.2] ms. (b) Plot for PEP difference between the expert and the CEA estimates against the expert PEP with bias = 0.0 ms and limits of agreement = [−12.3, 12.2] ms. For different PEP values across all participants, closer agreement of PEP estimates between the expert and TEA was observed as compared to that between the expert and the CEA. Image (c) Emory University, CC-BY-SA.

The time required to perform CEA and TEA was also compared. On a standard laptop with an Intel(R) Core™ i7-8550U CPU @ 1.80GHz processor and 16.0GB RAM, it took 154.3 s and 158.8 s respectively by CEA and TEA to produce 8 averaged beats from 8 minutes raw ECG and ICG signal.

3.2. Classification of Clinically defined Stress States

Since the participants were quiet and relatively motion-free during the baseline stage, the ICG signal was not substantially affected by motion and speaking artifacts. This resulted in similar PEP values measured using both CEA and TEA as shown in Table 1.

Table 6 represents the PEP analysis for PTSD and non-PTSD participants in the study stages of the baseline and speech task. PEP values were computed from EA beats formed via the TEA and CEA. PEP in the speech task was found to be lower than that of the baseline stage for both algorithms, which was in line with the earlier studies (Miller & Sita, 1994; Matthews et al., 2003). Specifically, for PTSD patients, the Wilcoxon signed-rank test with 95% CI revealed that there was a difference between baseline and speech task both when TEA and CEA algorithms were used but a larger effect size was obtained for TEA. Comparison of the change scores (speech task minus baseline) between TEA (Mean (M) = −11.8 ms, Standard error (SE) = 3.7 ms) vs. CEA (M = −8.4 ms, SE = 4.4 ms) were non-significant (Wilcoxon rank-sum test with 95% CI, z = 1.03, p = 0.3). For non-PTSD participants, the comparison was performed for 22 participants as 1 participant had noisy CEA beats in the speech stage. The paired t-test with 95% CI revealed that there was a difference between baseline and speech task both when TEA and CEA algorithms were used but a larger effect size was obtained for TEA. Comparison of the change scores (speech task minus baseline) between TEA (M = −8.2 ms, SE = 1.7 ms) vs. CEA (M = −5.3 ms, SE = 1.5 ms) were non-significant (unpaired t-test with 95% CI, t = 1.29, p = 0.2).

Table 6:

PEP analysis in the study stages of baseline and speech task for PTSD and non-PTSD participants. PEP values, in ms, were computed using the TEA and CEA. For both groups, PEP for the speech task was found to be lower than that of the baseline stage for both algorithms (p < 0.05) with larger effect size and higher AUC in TEA than in CEA. For difference in baseline and speech task, a higher PEP mean difference was observed, but it was not statistically significant in the PTSD (p = 0.3) and non-PTSD (p = 0.2) subgroups.

Algorithm Baseline Speech task p value
(p<0.05)		 Test
statistics		 Cohen’s d
effect size		 AUC Mean difference /
Confidence interval
Mean (SD) Mean (SD)
PTSD (n =22)
TEA 92.8 (18.8) 80.9 (22.3) 0.001 3.3 0.68 0.66 11.8 [4.2, 19.5]
CEA 91.3 (19.9) 82.8 (23.4) 0.046 2.0 0.41 0.57 8.4 [−0.8, 17.6]
Non-PTSD (n =22)
TEA 93.8 (11.2) 85.6 (11.2) <0.001 4.9 1.0 0.68 8.2 [4.7, 11.7]
CEA 93.4 (11.9) 88.0 (10.6) 0.002 3.6 0.77 0.59 5.3[2.3, 8.4]
Open in a new tab

Comparison of the change scores (speech task minus baseline) between PTSD (M = −11.8 ms, SE = 3.7 ms) vs. non-PTSD (M = −8.2 ms, SE = 1.7 ms), using TEA, were non-significant (Wilcoxson rank-sum test with 95% CI, z = 0.0, p = 1.0). Also, via CEA, comparison of the change scores (speech task minus baseline) between PTSD (M = −8.4 ms, SE = 4.4 ms) vs. non-PTSD (M = −5.3, SE = 1.5 ms) were non-significant (Wilcoxson rank-sum test with 95% CI, z = 0.79, p = 0.77).

Table 7 presents the results of the PEP analysis for PTSD participants with active (n = 10) and sham (n = 12) tcVNS during the arithmetic task for the CEA and TEA. The unpaired t-test with 95% CI revealed that there was a difference between active and sham stimulation for TEA algorithm only with a larger effect size than CEA. The mean PEP difference calculated using the TEA among the PTSD group with active and sham tcVNS was approximately 52% greater than the PEP computed using the CEA. Table 7 also depicts the results of the PEP analysis for non-PTSD participants with active (n = 10) and sham (n =11) tcVNS during the arithmetic task. The Wilcoxson rank-sum test with 95% CI revealed no significant difference for either method.

Table 7:

PEP analysis for participants with/without PTSD, with active vs. sham tcVNS in the arithmetic task. PEP values, in ms, were computed using the TEA and CEA. For PTSD participants, the TEA revealed a significant difference for active vs. sham tcVNS (p < 0.05). Also, the mean difference between active vs. sham for the TEA is approximately 52 % greater than that of the CEA, with larger effect size. For non-PTSD participants, no significant differences for active vs sham tcVNS (p < 0.05) were found using either TEA or CEA.

Algorithm Active
tcVNS		 Sham
tcVNS		 p value
(p<0.05)		 Test
statistics		 Cohen’s d
effect size		 AUC Mean difference /
Confidence interval
Mean (SD) Mean (SD)
PTSD (nActive = 10, nsham = 12)
TEA 94.4 (16.0) 83.3 (19.0) 0.034 2.2 0.63 0.84 11.1 [1.2, 21.0]
CEA 92.5 (19.5) 85.2 (21.9) 0.228 1.2 0.35 0.68 7.3 [−4.4, 19.0]
Non-PTSD (nActive = 10, nsham = 11)
TEA 88.1 (13.3) 88.0 (13.0) 0.982 0.02 0.01 0.31 0.1 [−7.5, 7.7]
CEA 87.5 (14.8) 87.4 (13.2) 0.956 0.05 0.01 0.33 0.1 [−8.0, 8.2]
Open in a new tab

3.3. Qualitative Analysis for Select Cases

To investigate the potential reasons for inaccurate PEP estimation by the CEA, and improved performance of the TEA method, a qualitative analysis was conducted over selective cases by inspecting/comparing the RB interval of the averaged beats generated using expert-identified clean beats, CEA, and TEA. Figure 6 presents a case where “Beat Contribution Factor (BCF)" (BCF was the ratio of the number of clean beats to the total number of beats in an analysis window) for the expert EA beat was 0.78, indicating that more than 20% of beats were noisy in this analysis window. BCF for the TEA beat was 0.75. The feature morphology type of B point for the expert and the TEA beat was “Onset of the rise", while for CEA beat it was “Plateau". The PEP difference between the expert EA beat and the TEA beat was 0.5 ms. The difference in the B-point feature types for the CEA and expert EA beat resulted in a higher PEP difference of 20 ms. This comparison revealed that the B-point feature morphology remained the same for the expert and the TEA beats with minor PEP differences. The higher PEP differences between the expert and CEA beats were due to a change in the feature morphology of the B point. Appendix A represents other cases to demonstrate the efficacy of TEA for capturing the feature morphology of the expert beats and more accurate PEP computation than CEA.

Figure 6:

Figure 6:

Open in a new tab

Effect of removal of noise artifacts on feature type and temporal location of B point. (a) RB interval for the expert and CEA ICG beats (b) RB interval for the expert and TEA ICG beats. Feature type for B point in the expert and TEA beat is onset of the rise. For CEA beat, B point feature type is plateau. PEP difference between the expert and CEA beat is 20 ms, and between the expert and the TEA ICG beat is 0.5 ms. Image (c) Emory University, CC-BY-SA.

4. Discussion

In this study, we validated a novel TEA for PEP measurement and showed that, in cases of substantial motion artifact, our algorithm outperforms the CEA method and results in greater separation between groups of varying stress physiology. The method distinguished between clean and noisy ICG beats with high accuracy, as compared to expert adjudication, and our close examination of specific cases shed light on why careful assessment of noise is critical for accurate physiological assessments. This likely drove the superior performance of the TEA vs. CEA in differentiating participants by PTSD status, stress vs. rest states, and active vs. sham tcVNS. This is particularly significant for PTSD participants in active vs. sham tcVNS analysis, where active tcVNS stimulates the afferent vagus, and then, through the mechanisms in the brain stem, decreases the sympathetic nervous system activity resulting in an increase in PEP (Shah et al., 2021; Gurel, Wittbrodt, et al., 2020). Overall, our findings suggest that careful examination and removal of noisy ICG waveforms are critical in psychophysiology research involving PEP and that such methods should be the standard for ICG signal processing of noise-prone ambulatory data moving forward.

TEA helped to overcome several limitations in CEA by addressing its several incorrect signal-averaging assumptions. CEA assumed the ICG signal to be repetitive, invariable, and uniformly aligned (Sornmo and Laguna (2005)). However, this could be violated by motion artifacts, electrode adhesion problems, as well as physiological inter-subject and intra-subject ICG morphological variability (Sherwood et al., 1990; Ermishkin et al., 2014; Arbol et al., 2017). In addition, a time-domain shift added to the ICG signal by breathing also results in the misalignment of beats for EA computation. Violation of the basic assumptions led to artificial morphology variability and event latency in EA ICG beats generated via CEA (Hurwitz et al., 1993). Our algorithm addressed these issues effectively by: (1) removing spikes, (2) discarding noisy beats based on morphological differences from the ICG EA templates, and (3) aligning beats with similar morphology for the formation of the EA ICG beat. The proposed TEA also presented a scientific advancement in comparison to previously published ICG noise removal approaches (Hurwitz et al., 1993; Cybulski et al., 2017; Cieslak et al., 2018; Forouzanfar, Baker, Colrain, & de Zambotti, 2019), because it did not require human intervention for apriori annotation of fiducial points, derivation of hemodynamic parameters, and the creation of an external ICG cycle template before its application. The interested reader may reference our previous publication (Sheikh et al., 2020) for a detailed comparison of TEA and other ICG noise removal methods.

By evaluating specific examples, we were able to better understand the underlying reasons why noise led to misclassification and reduced physiologic classification states. TEA was particularly important for accurately identifying B point feature morphology, which was critical in accurate PEP estimation. Noise detection was critical in B-point estimation because it influenced the classification of feature types, including inflection, plateau, the onset of the rise, and valley. TEA, we found, accurately identified B-point feature types vs. expert identification and resulted in lower PEP differences than CEA (Figure 6, Appendix A).

Our findings have important potential clinical applications and public health implications, which call for more research in this area. Accurate PEP examination with TEA may improve prediction of mortality in the general population (Medina-Lezama et al., 2018), evaluation of early cardiovascular disease, management of hypertension (Silva Lopes et al., 2019; DeMarzo, 2020), and for assessing stress levels in persons with PTSD and monitoring autonomic function (Czarnek et al., 2021; Silvia et al., 2020; Kibler, 2018; Malcolm et al., 2016; Meijer et al., 2011; Kelsey et al., 2007).

The framework used in TEA, which we have made available and adaptable as open-source (Sheikh et al., 2020), may also apply in the signal processing of other physiological signals that are also overly sensitive to noise, including ballistocardiograph, seismocardiograph, and photoplethysmography (Inan et al., 2014; Inan et al., 2018; Moraes et al., 2018). The modular form of the algorithm can also accommodate modifications in its various stages. Future work may look at the combination of TEA and moving EA (Cieslak et al., 2018) which can provide improved temporal resolution for ICG-based analysis. Although some of our findings did not reach statistical significance, they generally confirm previous literature describing higher sympathetic activity during acute mental stress, chronic mental stress (PTSD), and stress reduction treatments (tcVNS). Future applications may include treatment, monitoring, and closed-loop feedback systems in which interventions with a psychophysiological benefit are positively reinforced.

Our findings were subject to several limitations. The ICG signal was filtered at specific cut-off frequencies [0.5, 40] Hz in the pre-processing stage without unambiguous evidence on the validity of these cut-points. The algorithm’s performance needs verification at different cut-off frequencies, which may require modification of the parameters for the second and the third stage of the proposed algorithm. The proposed algorithm was tested under off-line conditions; future work may investigate the real-time application of the TEA to remove ICG noisy beats. This may be possible by analyzing the signal on minute-to-minute basis. Although TEA is an automated noise removal algorithm, user input for the B-point annotation to compute PEP is still required. Therefore, our future effort will be to automate the B point detection for efficient ICG-based PEP analysis (Sheikh et al., 2022). For our clinical comparisons, we assumed that the groups had different levels of sympathetic activity based on prior literature, but we were not able to prove it with an independent metric such as microneurography-based sympathetic measurement.

5. Conclusion

In this study, we confirmed the validity of the TEA for the removal of noisy ICG beats and PEP analysis in data from a randomized controlled clinical trial of tcVNS in PTSD. The TEA outperformed the CEA for noisy beat detection by showing close agreement with the expert annotations. Consistent with this improved noise removal, our algorithm showed superior performance in differentiating groups in which differences in sympathetic activity were expected compared to the conventional method, including acute stress challenge, PTSD, and active tcVNS. Future work should consider applying this open-source algorithm in more settings such as 24-hour ambulatory monitoring, in which there is an even greater amount of motion artifact.

Acknowledgement

Shafa-at Ali Sheikh is funded by Fulbright Scholarship Program. The authors wish to acknowledge the National Institutes of Health (Grant # NIH K23HL127251, R01HL136205, R01HL125246, R01HL130619, and R03HL146879), the National Science Foundation Award 1636933, and Emory University for their financial support of this research. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Institutes of Health, the National Science Foundation, Georgia Institute of Technology and Emory University.

Appendix A. Effects of removal of noisy artifacts on the feature type and temporal location of B point

In this section, the effects of removing noisy artifacts on the feature morphology and temporal location of B point are analyzed by comparing the RB interval for the expert, CEA, and TEA beats. We observed that feature types for B points remained the same for the expert and TEA beats with minor PEP differences. For expert vs. CEA beats, PEP differences were higher between the expert and CEA beats for different feature types of B points.

Figure A1 presents the case where BCF for expert EA beat was 0.45 indicating that more than 50% of beats were noisy in this analysis window. BCF for the TEA beat is 0.50. The feature type of B point for the expert and the TEA beat was inflection, while no feature morphology was observed for the CEA beat. The PEP difference between the expert and TEA ICG beat was 1 ms. Different feature types of B point for CEA, and expert resulted in a higher PEP difference of 54 ms.

Figure A2 presents the case where BCF for expert EA beat is 0.37 indicating that 60% of beats were noisy in this analysis window. BCF for the TEA beat came out to be 0.47. B point feature type for all three types of EA beats was “valley". However, the PEP difference between the expert and CEA beat (6.5 ms) is greater than the PEP difference between the expert and TEA ICG beat (0.5 ms).

Figure A1:

Figure A1:

Open in a new tab

Different B point temporal location with different feature type. (a) RB interval for the expert and CEA ICG beats (b) RB interval for the expert and TEA ICG beats. Feature type for B point in the expert and TEA beat is inflection. No feature for B point is present for CEA beat. PEP difference between the expert and CEA beat is 54 ms. PEP difference between the expert and TEA beat is 1 ms. Image (c) Emory University, CC-BY-SA.

Figure A2:

Figure A2:

Open in a new tab

Different B point temporal location with same feature type. (a) RB interval for the expert and CEA ICG beats (b) RB interval for the expert and TEA ICG beats. For all three cases, the B point feature type is valley. PEP difference between the expert and CEA beat is 6.5 ms. PEP difference between the expert and TEA beat is 0.5 ms. Image (c) Emory University, CC-BY-SA.

References

  1. Árbol JR, Perakakis P, Garrido A, Mata JL, Fernandez-Santaella MC & Vila J (2017). Mathematical detection of aortic valve opening (B point) in impedance cardiography: A comparison of three popular algorithms. Psychophysiology, 54(3), 350–357. 10.1111/psyp.12799 [DOI] [PubMed] [Google Scholar]
  2. Behar J, Johnson A, Clifford GD, & Oster J (2014). A comparison of single channel fetal ECG extraction methods. Annals of Biomedical Engineering, 42(6), 1340–1353. 10.1007/s10439-014-0993-9 [DOI] [PubMed] [Google Scholar]
  3. Bland JM & Altman D (1986). Statistical methods for assessing agreement between two methods of clinical measurement. The Lancet, 327(8476), 307–310. 10.1016/S0140-6736(86)90837-8 [DOI] [PubMed] [Google Scholar]
  4. Bremner JD (2016). Closed loop vagal nerve stimulation for patients with posttraumatic stress disorder. https://clinicaltrials.gov/ct2/show/NCT02992899 Accessed on: Oct 7, 2020. [Google Scholar]
  5. Bremner JD, Campanella C, Khan Z, Shah M, Hammadah M, Wilmot K, … & Vaccarino V (2018). Brain correlates of mental stress-induced myocardial ischemia, Psychosomatic Medicine, 80(6), 515. 10.1097/PSY.0000000000000597 [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Bremner J, Gurel N, Wittbrodt M, Nye J, Alam Z, Herring I, …& Inan O (2019). Non-invasive vagal nerve stimulation paired with stress exposure in posttraumatic stress disorder (PTSD). Brain Stimulation: Basic, Translational, and Clinical Research in Neuromodulation, 12(2), 438. 10.1016/j.brs.2018.12.417 [DOI] [Google Scholar]
  7. Brownhill K (2020). Intraclass correlation coefficients. https://www.mathworks.com/matlabcentral/fileexchange/21501-intraclass-correlation-coefficients, MATLAB Central File Exchange. Accessed on: Apr 2, 2020. [Google Scholar]
  8. Cacioppo JT, Berntson GG, Binkley PF, Quigley KS, Uchino BN and Fieldstone A (1994). Autonomic cardiac control. II. Noninvasive indices and basal response as revealed by autonomic blockades. Psychophysiology, 31(6), 586–598. 10.1111/j.1469-8986.1994.tb02351.x [DOI] [PubMed] [Google Scholar]
  9. Chaddha A, Robinson EA, Kline-Rogers E, Alexandris-Souphis T and Rubenfire M (2016). Mental health and cardiovascular disease. The American Journal of Medicine, 129(11), 1145–1148. 10.1016/j.amjmed.2016.05.018 [DOI] [PubMed] [Google Scholar]
  10. Cieslak M, Ryan WS, Babenko V, Erro H, Rathbun ZM, Meiring W, … and Grafton ST (2018). Quantifying rapid changes in cardiovascular state with a moving ensemble average. Psychophysiology, 55(4), e13018. 10.1111/psyp.13018 [DOI] [PubMed] [Google Scholar]
  11. Cohen H, Kaplan Z, Matar MA, Loewenthal U, Zohar J & Richter-Levin G (2007). Long-lasting behavioral effects of juvenile trauma in an animal model of PTSD associated with a failure of the autonomic nervous system to recover. European Neuropsychopharmacology, 17(6–7), 464–477. 10.1016/j.euroneuro.2006.11.003 [DOI] [PubMed] [Google Scholar]
  12. Cohen JR, Thomsen KN, Tu KM, Thakur H, McNeil S & Menon SV (2020). Cardiac autonomic functioning and post-traumatic stress: A preliminary study in youth at-risk for PTSD. Psychiatry Research. 284: 112684. 10.1016/j.psychres.2019.112684 [DOI] [PubMed] [Google Scholar]
  13. Cybulski G (2011). Ambulatory Impedance Cardiography, The Systems and their Applications. Springer, Berlin, Heidelberg. 10.1007/978-3-642-11987-3 [DOI] [Google Scholar]
  14. Cybulski G, Młyńczak G, Żyliński M, Strasz A, Gąsiorowska A, & Niewiadomski W (2017). The quality of automatic artifact identification in ambulatory impedance cardiography monitoring. In EMBEC & NBC 2017 (pp. 165–168) Springer, Singapore. 10.1007/978-981-10-5122-7_42 [DOI] [Google Scholar]
  15. Cybulski G, Niewiadomski W, Aksler M, Strasz A, Gasiorowska A, Laskowska D, & Palko T (2011). Identification of the sources of artefacts in the Holter-type impedance cardiography recordings, In 5th European Conference of the International Federation for Medical and Biological Engineering (pp. 1272–1274). Springer, Berlin, Heidelberg. 10.1007/978-3-642-23508-5_329 [DOI] [Google Scholar]
  16. Czarnek G, Richter M, & Strojny P (2021). Cardiac sympathetic activity during recovery as an indicator of sympathetic activity during task performance, Psychophysiology, 58(2), e13724. 10.1111/psyp.13724 [DOI] [PubMed] [Google Scholar]
  17. DeCaro JA (2016). Beyond catecholamines: Measuring autonomic responses to psychosocial context. American Journal of Human Biology, 28(3), 309–317. 10.1002/ajhb.22815 [DOI] [PubMed] [Google Scholar]
  18. DeMarzo AP (2020). Clinical use of impedance cardiography for hemodynamic assessment of early cardiovascular disease and management of hypertension. High Blood Pressure & Cardiovascular Prevention, 27(3), 203–213. 10.1007/s40292-020-00383-0 [DOI] [PubMed] [Google Scholar]
  19. Dimsdale JE (2008). Psychological stress and cardiovascular disease. Journal of the American College of Cardiology, 51(13), 1237–1246. 10.1016/j.jacc.2007.12.024 [DOI] [PMC free article] [PubMed] [Google Scholar]
  20. Ermishkin VV, Kolesnikov VA, & Lukoshkova EV (2014). Age-dependent and ‘pathologic’ changes in ICG waveforms resulting from superposition of pre-ejection and ejection waves. Physiological Measurement, 35(6), 943. 10.1088/0967-3334/35/6/943 [DOI] [PubMed] [Google Scholar]
  21. Fani N, Ashraf A, Afzal N, Jawed F, Kitayama N, Reed L & Bremner JD (2011). Increased neural response to trauma scripts in posttraumatic stress disorder following paroxetine treatment: a pilot study. Neuroscience Letters, 491(3), 196–201. 10.1016/j.neulet.2011.01.037 [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Forouzanfar M, Baker FC, Colrain IM & de Zambotti M (2019). Automatic artifact detection in impedance cardiogram using pulse similarity index. In 2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), (pp. 2629–2632). IEEE. 10.1109/EMBC.2019.8856542 [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Forouzanfar M, Baker FC, Colrain IM, Goldstone A & de Zambotti M (2019). Automatic analysis of pre-ejection period during sleep using impedance cardiogram. Psychophysiology, 56(7), e13355. 10.1111/psyp.13355 [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Gurel NZ, Huang M, Wittbrodt MT, Jung H, Ladd SL, Shandhi MMH, … & Inan OT (2020). Quantifying acute physiological biomarkers of transcutaneous cervical vagal nerve stimulation in the context of psychological stress. Brain Stimulation, 13(1), 47–59. 10.1016/j.brs.2019.08.002 [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Gurel NZ, Wittbrodt MT, Jung H, Shandhi MMH, Driggers EG, Ladd SL, … & Bremner JD (2020). Transcutaneous cervical vagal nerve stimulation reduces sympathetic responses to stress in posttraumatic stress disorder: A double-blind, randomized, sham controlled trial. Neurobiology of stress, 13, 100264. [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Gustafsson F (1996). Determining the initial states in forward-backward filtering. IEEE Transactions on Signal Processing, 44(4), 988–992. 10.1109/78.492552. [DOI] [Google Scholar]
  27. Hurwitz BE, Shyu LY, Lu C-C, Reddy SP, Schneiderman N, & Nagel JH (1993). Signal fidelity requirements for deriving impedance cardiographic measures of cardiac function over a broad heart rate range. Biological Psychology, 36(1–2), 3–21. 10.1016/0301-0511(93)90076-K [DOI] [PubMed] [Google Scholar]
  28. Hurwitz BE, Shyu LY, Reddy SP, Schneiderman N, & Nagel JH (1990). Coherent ensemble averaging techniques for impedance cardiography. In [1990] Proceedings. Third Annual IEEE Symposium on Computer-Based Medical Systems (pp. 228–235). IEEE. 10.1109/CBMSYS.1990.109403 [DOI] [Google Scholar]
  29. Inan OT, Baran Pouyan M, Javaid AQ, Dowling S, Etemadi M, Dorier A, … & Klein L (2018). Novel wearable seismocardiography and machine learning algorithms can assess clinical status of heart failure patients. Circulation: Heart Failure, 11(1), e004313. 10.1161/CIRCHEARTFAILURE.117.004313 [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Inan OT, Migeotte P-F, Park K-S, Etemadi M, Tavakolian K, Casanella R, … & Di Rienzo M (2014). Ballistocardiography and seismocardiography: A review of recent advances. IEEE Journal of Biomedical and Health Informatics 19(4), 1414–1427. 10.1109/JBHI.2014.2361732 [DOI] [PubMed] [Google Scholar]
  31. Johnson AE, Behar J, Andreotti F, Clifford GD, & Oster J (2014). R-peak estimation using multimodal lead switching. In Computing in Cardiology 2014 (pp. 281–284). IEEE. [Google Scholar]
  32. Kelsey RM, Alpert BS, Dahmer MK, Krushkal J & Quasney MW (2010). Beta-adrenergic receptor gene polymorphisms and cardiovascular reactivity to stress in black adolescents and young adults. Psychophysiology, 47(5), 863–873. 10.1111/j.1469-8986.2010.01006.x [DOI] [PubMed] [Google Scholar]
  33. Kelsey RM, Ornduff SR and Alpert BS (2007). Reliability of cardiovascular reactivity to stress: internal consistency. Psychophysiology, 44(2), 216–225. 10.1111/j.1469-8986.2007.00499.x [DOI] [PubMed] [Google Scholar]
  34. Kibler JL (2018). An extension of the perseverative cognition hypothesis to posttraumatic stress disorder symptomatology: Cardiovascular recovery in relation to posttraumatic stress disorder severity and cognitive appraisals of stress. Journal of Traumatic Stress 31(1), 25–34. 10.1002/jts.22252 [DOI] [PMC free article] [PubMed] [Google Scholar]
  35. Kim DW, Baker L, Pearce J & Kim WK (1988). Origins of the impedance change in impedance cardiography by a three-dimensional finite element model. IEEE Transactions on Biomedical Engineering, 35(12): 993–1000. 10.1109/10.8683 [DOI] [PubMed] [Google Scholar]
  36. Kizakevich P, McDermott J & Gollan F (1976). An automated system for systolic time interval analysis. Proceedings of the Digital Equipment Users Society, 2, 795–798. 10.1016/j.jcm.2016.02.012 [DOI] [Google Scholar]
  37. Lamb DG, Porges EC, Lewis GF, & Williamson JB (2017). Non-invasive vagal nerve stimulation effects on hyperarousal and autonomic state in patients with posttraumatic stress disorder and history of mild traumatic brain injury: preliminary evidence. Frontiers in Medicine, 4, 124. 10.1080/01621459.1969.10500983 [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Lilliefors HW (1969). On the Kolmogorov-Smirnov test for the exponential distribution with mean unknown. Journal of the American Statistical Association, 64(325), 387–389. [Google Scholar]
  39. Malcolm L, Kibler JL, Ma M, Tursich M, Augustin D, Greenbarg R, & Gold SN (2016). Psychophysiological reactivity and PTSD symptom severity among young women. International Journal of Psychology and Neuroscience, 2(3), 17. [PMC free article] [PubMed] [Google Scholar]
  40. Matthews KA, Salomon K, Brady SS, & Allen MT (2003). Cardiovascular reactivity to stress predicts future blood pressure in adolescence. Psychosomatic Medicine, 65(3): 410–415. 10.1097/01.PSY.0000057612.94797.5F [DOI] [PubMed] [Google Scholar]
  41. McGraw KO & Wong SP (1996). Forming inferences about some intraclass correlation coefficients. Psychological Methods, 1(1), 30. 10.1037/1082-989X.1.1.30 [DOI] [Google Scholar]
  42. Medina-Lezama J, Narvaez-Guerra O, Herrera-Enriquez K, Morey-Vargas OL, Bolaños-Salazar JF, Abugattas JP, … & Chirinos JA (2018). Hemodynamic patterns identified by impedance cardiography predict mortality in the general population: the PREVENCION study. Journal of the American Heart Association, 7(18), e009259. 10.1161/JAHA.118.009259 [DOI] [PMC free article] [PubMed] [Google Scholar]
  43. Meijer JH, Elbertse E, Boesveldt S, Berendse HW & Verdaasdonk RM (2011). Using the initial systolic time interval to assess cardiac autonomic nervous function in Parkinson's disease. Journal of Electrical Bioimpedance, 2(1), 98–101. 10.5617/jeb.216 [DOI] [Google Scholar]
  44. Menezes AR, Lavie CJ, Milani RV, O’Keefe J & Lavie TJ (2011). Psychological risk factors and cardiovascular disease: is it all in your head? Postgraduate Medicine, 123(5), 165–176. 10.3810/pgm.2011.09.2472 [DOI] [PubMed] [Google Scholar]
  45. Meyer T, Albrecht J, Bornschein G, Sachsse U & Herrmann-Lingen C (2016). Post-traumatic stress disorder (PTSD) patients exhibit a blunted parasympathetic response to an emotional stressor. Applied Psychophysiology and Biofeedback, 41(4), 395–404. 10.1007/s10484-016-9341-1 [DOI] [PubMed] [Google Scholar]
  46. Mezzacappa ES, Kelsey RM, & Katkin ES (1999). The effects of epinephrine administration on impedance cardiographic measures of cardiovascular function. International Journal of Psychophysiology, 31(3), 189–196. 10.1016/S0167-8760(98)00058-0 [DOI] [PubMed] [Google Scholar]
  47. Miller SB & Sita A (1994). Parental history of hypertension, menstrual cycle phase, and cardiovascular response to stress. Psychosomatic Medicine. 10.1097/00006842-199401000-00008 [DOI] [PubMed] [Google Scholar]
  48. Moraes JL, Rocha MX, Vasconcelos GG, Vasconcelos Filho JE, De Albuquerque VHC and Alexandria AR (2018). Advances in photoplethysmography signal analysis for biomedical applications. Sensors 18(6), 1894. 10.3390/s18061894 [DOI] [PMC free article] [PubMed] [Google Scholar]
  49. Newlin DB, & Levenson RW (1979). Pre-ejection period: Measuring beta adrenergic influences upon the heart. Psychophysiology 16(6), 546–552. 10.1111/j.1469-8986.1979.tb01519.x [DOI] [PubMed] [Google Scholar]
  50. Olff M (2017). Sex and gender differences in post-traumatic stress disorder: an update. European Journal of Psychotraumatology, 8 (sup4), 1351204. 10.1080/20008198.2017.1351204 [DOI] [Google Scholar]
  51. Park J, Marvar PJ, Liao P, Kankam ML, Norrholm SD, Downey RM, … & Rothbaum BO (2017). Baroreflex dysfunction and augmented sympathetic nerve responses during mental stress in veterans with post-traumatic stress disorder. The Journal of Physiology, 595(14), 4893–4908. 10.1113/JP274269 [DOI] [PMC free article] [PubMed] [Google Scholar]
  52. Peña DF, Childs JE, Willett S, Vital A, McIntyre CK, & Kroener S (2014). Vagus nerve stimulation enhances extinction of conditioned fear and modulates plasticity in the pathway from the ventromedial prefrontal cortex to the amygdala. Frontiers in Behavioral Neuroscience, 8, 327. 10.3389/fnbeh.2014.00327 [DOI] [PMC free article] [PubMed] [Google Scholar]
  53. Peters JR, Eisenlohr-Moul TA, Walsh EC & Derefinko KJ (2018). Exploring the pathophysiology of emotion-based impulsivity: The roles of the sympathetic nervous system and hostile reactivity, Psychiatry Research, 267, 368–375. 10.1177/1077559514538115 [DOI] [PMC free article] [PubMed] [Google Scholar]
  54. Seery MD, Kondrak CL, Streamer L, Saltsman T, & Lamarche VM (2016). Preejection period can be calculated using R peak instead of Q. Psychophysiology 53(8), 1232–1240. 10.1111/psyp.12657 [DOI] [PubMed] [Google Scholar]
  55. Shah AJ, Lampert R, Goldberg J, Veledar E, Bremner JD & Vaccarino V (2013). Posttraumatic stress disorder and impaired autonomic modulation in male twins. Biological Psychiatry, 73(11): 1103–1110. 10.1016/j.biopsych.2013.01.019 [DOI] [PMC free article] [PubMed] [Google Scholar]
  56. Shah AJ, Wittbrodt MT, Bremner JD, & Vaccarino V (2021). Cardiovascular Pathophysiology from the Cardioneural Perspective and its Clinical Applications. Trends in Cardiovascular Medicine. 10.1016/j.tcm.2021.03.001 [DOI] [PMC free article] [PubMed] [Google Scholar]
  57. Sheikh SA, Shah AJ, Levantsevych O, Soudan M, l Alkhalaf M, Rad AB, … & Clifford GD (2020). An open-source automated algorithm for removal of noisy beats for accurate impedance cardiogram analysis, Physiological Measurement, 41(7), 075002. 10.1088/1361-6579/ab9b71 [DOI] [PMC free article] [PubMed] [Google Scholar]
  58. Sheikh SA, Shah A, Inan OT, & Clifford GD (2020). Open-Source Toolbox for ICG Analysis. v1.0 (https://github.com/cliffordlab/ICG_OSToolbox Accessed: 5 April 2020) [Google Scholar]
  59. Sheikh SA, Gurel NZ, Gupta S, Chukwu IV, Levantsevych O, … & Rad AB (2022). Data-Driven Approach for Automatic Detection of Aortic Valve Opening: B-point Detection from Impedance Cardiogram, Manuscript in preparation. [DOI] [PMC free article] [PubMed] [Google Scholar]
  60. Sherwood A, Allen MT, Fahrenberg J, Kelsey RM, Lovallo WR, & Van Doornen LJ (1990). Methodological guidelines for impedance cardiography. Psychophysiology 27(1), 1–23. 10.1111/j.1469-8986.1990.tb02171.x [DOI] [PubMed] [Google Scholar]
  61. Shrout PE, & Fleiss JL (1979). Intraclass correlations: uses in assessing rater reliability. Psychological Bulletin, 86(2), 420. 10.1037/0033-2909.86.2.420 [DOI] [PubMed] [Google Scholar]
  62. Silva Lopes B, Craveiro N, Firmino-Machado J, Ribeiro P & Castelo-Branco M (2019). Hemodynamic differences among hypertensive patients with and without heart failure using impedance cardiography. Therapeutic Advances in Cardiovascular Disease, 13: 1753944719876517. 10.1177/1753944719876517 [DOI] [PMC free article] [PubMed] [Google Scholar]
  63. Silvia PJ, McHone AN, Mironovova Z, Eddington KM, Harper KL, Sperry SH, & Kwapil TR (2020). RZ interval as an impedance cardiography indicator of effort-related cardiac sympathetic activity. Applied Psychophysiology and Biofeedback, 46(1), 83–90. 10.1007/s10484-020-09493-w [DOI] [PMC free article] [PubMed] [Google Scholar]
  64. Sornmo L, & Laguna P (2005). Bioelectrical signal processing in cardiac and neurological applications. Vol. 8, Academic Press. 10.1016/B978-012437552-9/50003-9 [DOI] [Google Scholar]
  65. Vest AN, Da Poian G, Li Q, Liu C, Nemati S, Shah AJ, & Clifford GD (2018). An open-source benchmarked toolbox for cardiovascular waveform and interval analysis. Physiological Measurement, 39(10), 105004. 10.1088/1361-6579/aae021 [DOI] [PMC free article] [PubMed] [Google Scholar]
  66. Van Lien R, Neijts M, Willemsen G, & de Geus E. J. (2015). Ambulatory measurement of the ECG T-wave amplitude. Psychophysiology, 52(2), 225–37. 10.1111/psyp.12300 [DOI] [PubMed] [Google Scholar]
  67. Willemsen G, Ring C, Carroll D, Evans P, Clow A, & Hucklebridge F (1998). Secretory immunoglobulin and cardiovascular reactions to mental arithmetic and cold pressor. Psychophysiology, 35(3), 252–259. 10.1111/1469-8986.3530252 [DOI] [PubMed] [Google Scholar]

RESOURCES