Timing Considerations for Noninvasive Vagal Nerve Stimulation in Clinical Studies

Abstract

Noninvasive vagal nerve stimulation (n-VNS) devices have the potential for widespread applicability in improving the well-being of patients with stress-related psychiatric disorders. n-VNS devices are known to affect physiological signals, and, recently, they have been employed in various protocols involving both acute and longitudinal applications. However, questions regarding response time, “dosage,” or optimal treatment paradigms remain open. Prior work evaluated noninvasively obtained biomarkers that quantify the stimulation efficacy based on the changes in autonomic tone in a randomized double-blind study. In this work, we extend the state-of-the-art by investigating the onset of action for n-VNS in these same physiological biomarkers through a three-day clinical trial, including 233 administrations on 24 human participants, with and without immediately preceding acute traumatic stress. Determining n-VNS latency serves as a substantial step toward optimizing stimulation delivery with higher temporal resolution for personalized neuromodulation.

Introduction

The vagus nerve is a major component of the autonomic nervous system, consisting of afferent (sensory) and efferent (motor) connections to regulate autonomic tone and maintain homeostasis¹. Electrical stimulation of the vagus nerve is hypothesized as a “bottom-up” therapy that acts on the vagal afferents through the nucleus tractus solitarus (NTS), which projects to the brain areas that regulate mood and emotion, such as the amygdala, hippocampus and prefrontal cortex^1-3 (see Figure 1 for the representation of a simplified mechanism of action). Through the modulation of vagal afferents, efferent connections carry information to regulatory organs that modulate autonomic state, such as the heart. Surgical implants for vagal nerve stimulation (VNS) appear effective in the treatment of several psychiatric and cardiovascular disorders, such as epilepsy, depression, migraines, and heart failure^4-12. Though high variability in treatment response has been observed with open-loop approaches, various closed-loop algorithms have been proposed for implantable VNS devices using measures related to the heart rate^13-18 or neural response¹⁹. Indeed, compared to their open-loop counterparts, the closed-loop implementations of these algorithms on VNS implants in the market have proven effective in improving clinical outcomes²⁰. However, widespread adoption is limited due to the potential for adverse events during implantation and the normal cost and practical challenges associated with implantable devices21, ^22.

Figure 1.

Simplified representation of n-VNS mechanism of action. The understanding of n-VNS kinetics on noninvasively obtained physiological parameters may enable optimization of n-VNS delivery in unsupervised settings. NTS: nucleus tractus solitarus.

Noninvasive VNS (n-VNS) devices provide advantages in terms of usability and low-cost, with proven effects on central brain regions^23-26, cytokines, and patient physiology^27-29. Therefore, n-VNS devices have the potential to address many shortcomings of their implantable counterparts, potentially improving patient adherence and broadening use³⁰. Their operation, however, is currently limited to open-loop applications prescribed by the medical professional. Translating implantable closed-loop technologies to the noninvasive neuromodulation realm is not straightforward; the physiological measurements should also be obtained noninvasively, leveraging wearable sensing modalities. Moreover, the effects of n-VNS on human physiology would likely differ from the transient responses to implantable VNS observed and quantified in preclinical studies due to the skin-electrode impedance, the filtering effects of the skin as a barrier between the stimulation site and the vagus nerve, and the time it takes to adjust the stimulation amplitude per subjective tolerance³¹. For instance, implantable VNS has sub-second effects in anesthetized animals^32,33 and has been modeled for electrodes that are in direct contact with the nerve^34,35. Latency for the physiological effects of n-VNS would likely be greater than implantable stimulation due to the mentioned differences regarding the electrode-skin-tissue barrier. Determining the onset of action for n-VNS provides pertinent information on n-VNS kinetics in continuous physiological signals germane to design specifications for wearable sensing systems; this, in turn, advances us one step further toward closing the loop for improved n-VNS treatment quality.

Prior work introduced and evaluated downstream cardiovascular and peripheral biomarkers of cervical n-VNS that quantify the physiological response to acute treatment using noninvasively obtained signals; the effects of n-VNS when applied after laboratory-induced psychological stress were also explored^36,38. Statistics involving said biomarkers were presented as static outcomes based on average changes from a baseline state upon acute stress application. These biomarkers included heart rate (HR) as a measure of both sympathetic and parasympathetic influences, the pre-ejection period (PEP) of the heart as a measure of cardiac contractility and sympathetic activity, and the amplitude of peripherally measured photoplethysmogram (PPG) signals as a peripheral sympathetic measure. In this work, we extend the state-of-the-art by making the following contributions: (i) We analyze when these biomarkers change during stimulation on 24 human participants in a double-blind study, half undergoing active n-VNS and half undergoing sham stimulation; (ii) We propose a method to identify the latency between n-VNS initiation and the onset of n-VNS related physiological changes determined by the outset of expected changes in the instantaneous biomarkers; and (iii) To differentiate the physiological effects of n-VNS from possible placebo effects or other means of modulation, we applied the same latency identification method on the data obtained from the half who received sham stimulus at the same treatment area on the neck; we then compared both groups’ results. Our findings from 233 administrations suggest pertinent timing considerations germane to the design of effective clinical studies involving physiological effects of n-VNS.

Methods

Human Subjects Experiments

The study examines the physiological effects of n-VNS following acute traumatic stress. Investigating acute traumatic stress has clinical importance, as exposure to traumatic events can cause strongly encoded intrusive memories, or in contrast, impairments in memory function. These intrusive thoughts may persist in vulnerable individuals, possibly leading to posttraumatic stress disorder (PTSD)³⁹. As traumatic stress is a trigger for PTSD symptoms, the protocol was designed around laboratory-induced traumatic stress exposure followed by stimulation. Note that stimulation also occurred on its own to understand possible physiological differences. The study was approved by the Institutional Review Boards of Georgia Institute of Technology, Emory University School of Medicine, SPAWAR Systems Center Pacific, and the Department of Navy Human Research Protection Program. A total of 24 adults who have experienced prior psychological trauma (ages mean ± SD: 31 ± 9 years, 12 females) were recruited, and written, informed consent was obtained. Each participant was asked to write his / her personal traumatic memories; later, voice recordings based on these memories were prepared for delivery as acute traumatic stress during the protocol. The traumatic stressors (each lasting for approximately one minute) were delivered audibly through headphones. Upon randomization for double-blind protocol, each participant was assigned to an active n-VNS or sham device having identical appearance, placement, and operation (GammaCore, ElectroCore). Participants, clinical staff, and researchers were blinded to the devices. Active n-VNS devices produce a 25 Hz voltage signal with 5kHz sine wave bursts. Sham devices produce a slow AC biphasic voltage signal with 0.2 Hz square pulses. The stimulation intensity ranges from 0 to 5 arbitrary units (AU), with a corresponding peak output ranging from 0 to 30V for active n-VNS, and from 0 to 14 V for the sham device. During each application, intensity was increased to a tolerable level determined by the participant. The amplitude levels participants received were 3 AU (± 0.8 SD) for active n-VNS and 4.5 AU (± 1 SD) for sham stimulus.

The protocol spanned three days for each participant (see Figure 2 for the protocol summary). On the first day, participants listened to their traumatic stress scripts as acute traumatic stressors (a total of six scripts per participant), and stimulation was applied immediately after each script. To understand the physiological changes for the stress-free (baseline) condition, this day also included two separate n-VNS/sham applications without traumatic stress scripts read immediately prior to stimulation. On both the second and third days, participants received one stimulation without a preceding stress script. Overall, the participants received six stimulations following six traumatic stress prompts and four stimulations without stress. Five participants did not complete all the traumatic stress prompts in the protocol. Excluding the missing data from these participants, there was a total of 137 traumatic stress prompts, followed by 137 stimulation administrations (72 active, 65 sham), and 96 administrations without stress (48 active, 48 sham). Each participant received approximately 10 administrations over the protocol days (233 administrations in total, 120 active, 113 sham).

Figure 2.

Protocol diagram. The protocol included three days: the first day included six traumatic stress prompts followed by immediate n-VNS or sham stimulation and two stimulation administrations without stress. Each of the second and third days included one stimulation administration without stress.

Physiological Sensing

Noninvasive cardiovascular and peripheral signals were concurrently collected during the protocol. To measure the electrical activity of the heart, three-lead electrocardiography (ECG) signals were collected. As a blood volume pulse measure, finger-based transmissive photoplethysmography (PPG) signals were collected. Chest-wall vibrations (indicative of the mechanical activity of the heart) were measured using seismocardiogram (SCG). ECG and PPG signals were acquired using wireless amplifiers (Bionomadix RSPEC-R and PPGED-R, Biopac Systems). SCG signals were collected using a low-noise accelerometer placed on the mid-sternum (345A32, PCB Electronics, Depew, NY). All data were transmitted to a 16-bit data acquisition system (MP150, Biopac Systems) at a 2kHz sampling rate.

Signal Processing and Feature Extraction

Figure 3 summarizes the signal processing and feature extraction steps completed in MATLAB (R2017b, Natick, MA), also detailed in³⁶. All signals were first bandpass filtered with finite impulse response filters for noise reduction using 0.6-25Hz, 0.6-40Hz, and 0.4-8Hz frequency ranges for SCG, ECG, and PPG signals, respectively, to maintain consistency with prior literature. The continuous physiological features extracted were HR, PEP, and PPG amplitude. First, R-peaks of ECG signals were located using thresholding. The R-R intervals were then calculated and converted to instantaneous HR in beats per minute (bpm). Second, SCG and PPG signals were segmented (ensemble averaged) referenced to the R-peaks of ECG signals. Third, SCG beats were exponentially moving averaged with a time constant of three beats to reduce the effect of motion artifacts. After segmentation, fiducial points of the beats were located by i) labeling the second peak of SCG beats as aortic opening (AO) point⁴⁰ and ii) by finding the global maximum and minimum of PPG beats to extract PPG amplitude⁴¹. The time interval from ECG R-peak to SCG AO point was extracted as instantaneous PEP (in milliseconds, ms), and the amplitude difference between global maximum and global minimum of each PPG beat was extracted as instantaneous PPG amplitude (in AU).

Figure 3.

Signal processing and feature extraction steps. BPF: bandpass filter; exp. MA: exponential moving average; EA: ensemble average; AO: aortic opening point; PEP: pre-ejection period; HR: heart rate.

As the instantaneous variables are prone to corruption by motion artifacts, arrhythmias, and other complex mechanisms, they should not directly be used, as suggested in multiple works examining closed-loop physiological control18, 42, 43. Therefore, we smoothed the instantaneous signals using local regression, assigning lower weight to the outliers and using a span of 5% of the total number of data points in a 25-minute interval. Smoothing provides new variables that reflect the transient changes associated with n-VNS.

Onset of Action Annotation

The onsets of action following device initiation were annotated manually using the smoothed variables. Manual annotation was completed by three researchers. The researchers were instructed to peruse all participants’ data prior to manual labeling. Based on our previous investigations36,³⁷ and related implantable VNS literature³³, the anticipated changes following stimulation compared to the period before stimulation were: i) decrease in HR (i.e., decrease in sympathetic tone or increase in parasympathetic tone), ii) increase in PEP (i.e., decrease in cardiac contractility and cardiac sympathetic activity), and iii) increase in PPG amplitude (i.e., decrease in peripheral sympathetic activity). As multiple sensing modalities were used, occasional noisy measurements existed. Therefore, observing the occurrence of all three of these changes following each n-VNS administration was not expected. Thus, as indicated in in Figure 4, we located the first datapoint that satisfied two of the three aforementioned criteria and marked this time point as the onset time. These criteria were established and agreed upon prior to any manual labeling to ensure guidelines were not contorted to match desired results a posteriori. To find the corresponding onset of action, we subtracted the n-VNS start time from the labeled onset time. If only one of the three signals experienced noticeable change, the onset of action was marked as “absent.” Likewise, if no changes at all were observed or if all changes were relatively insignificant compared to noise and normal variation, the corresponding onset of action was marked as “absent.” To allow for independent sampling, we averaged the counts of absent onsets for each participant. Additionally, all annotations were independently performed by the three researchers to later assess the inter-annotator agreements for validation purposes. These were calculated as follows: pairwise agreements between the annotators were calculated in seconds (absolute difference between each onset time annotation) and in counts (agreement percentage for absent onsets). The final agreement results reported are the overall average absolute difference in annotated onset times and the average percent agreement in absent onsets.

Figure 4.

Annotation diagram. The smoothed instantaneous biomarkers (HR, PEP, PPG amplitude) were plotted from pre-stimulus to post-stimulus. If at least two of the three mentioned changes in the biomarkers occurred, the onset time was marked at the onset of the second change. If no eligible change was observed, the annotation was marked as “absent.”

Statistical Analysis

To understand whether baseline characteristics and baseline biomarker measures were comparable between the active n-VNS and sham groups, demographic characteristics (age, gender, weight, height, body mass index) and baseline measures of the biomarkers (HR, PEP, PPG amplitude) were compared using two-sample t-tests for normally distributed continuous variables, Wilcoxon rank-sum tests for non-normal continuous variables, and chi-squared tests for categorical variables. The Shapiro-Wilk test was used to assess normality. Next, the Wilcoxon rank-sum test was used for absent onset count comparison between the groups. A two-sided threshold of p<0.05 was used to indicate statistical significance. All statistical analyses were performed using MATLAB Statistics and Machine Learning Toolbox.

Findings

Table 1 lists the baseline characteristics of the device groups and associated p-values, proving there are no statistical differences in either demographics or baseline biomarkers. The continuous physiological parameters from two representative participants in the presence of traumatic stress are shown in Figure 5, one participant undergoing sham and the other undergoing active n-VNS. The pre-stimulus values for each of the biomarkers are shown with dashed lines to provide a reference for the predicted deviations.

Table 1.

Participant demographics and baseline biomarker values in each device group. P: p-value for the comparison of participants’ characteristics between groups. Values represent mean ± SD. BMI: body-mass index. F: female

Parameter	Active	Sham	P
Age [years]	29 ± 7	32 ± 11	0.23
Sex [F, %]	5F, 41.6%	7F, 58.3%	0.41
Weight [kg]	77 ± 14	79 ± 13	0.71
Height [cm]	175 ± 11	172 ± 6	0.49
BMI [kg / m2]	25 ± 3	27 ± 5	0.34
HR [bpm]	67 ± 14	61 ± 9	0.22
PEP [ms]	67 ± 18	75 ± 37	0.48
PPG Amplitude [V]	0.3 ± 0.3	0.2 ± 0.1	0.37

Figure 5.

Continuous physiological parameters showing n-VNS without traumatic stress, for one participant undergoing sham (left) and one participant undergoing active n-VNS stimulus (right). Markers represent the extracted data, lines represent the smoothed data. Shaded regions represent stimulus delivery. Dashed lines show the pre-stimulus averages of the measures. The onset time is marked as soon as two of the three annotation criteria mentioned in Figure 4 are detected.

Table 2 lists the onset of action for active n-VNS for both with and without traumatic stress, as well as the absent onset counts comparing the active and sham device groups. Based on the physiological biomarkers used, the effects of n-VNS were observed 18 ± 7 seconds from the start of n-VNS without stress. When n-VNS was applied after traumatic stress (six traumatic stress scripts followed by six n-VNS per participant, n=12 participants), effects were observed on the biomarkers in a similar latency, resulting in 16 ± 9 seconds. As for the absent onset counts (n-VNS administrations that had not met the criteria), a significant difference exists between the device groups: there were 5 ± 2 absent onsets per participant for the sham group, significantly higher than the active group’s 2 ± 2 absent onset counts (p=0.006). There were 24 and 65 absent onset counts for 120 active n-VNS and 113 sham administrations, respectively. The overall average of inter-annotator agreements resulted in a 4 ± 1 seconds difference between the labeled onset times and a 90 ± 5.5% agreement in absent onset counts.

Table 2.

Onset of action and absent onset counts. Values represent mean ± SD.

n-VNS administrations	Onset of Action [seconds]
n-VNS with traumatic stress (n=72 administrations)	16 ± 9
n-VNS without stress (n=48 administrations)	18 ± 7
Device groups (p=0.006)	Absent Onsets [counts per subject]
Active n-VNS (n=12 subjects)	2 ± 2
Sham (n=12 subjects)	5 ± 2

Discussion

This work investigated the onset of action for n-VNS as observed by noninvasive physiological signals. The time durations reported herein serve to estimate the onset of action for n-VNS, which can be used for clinical study design and to provide engineering considerations of closed-loop n-VNS therapy. The continuous signals from representative subjects in Figure 5 show the broad expected changes in the entire sample. The participant undergoing sham stimulus did not experience notable change in HR, PEP, and PPG amplitude during the stimulus, compared to the period before the stimulus. In contrast to the sham participant, the active participant experienced notable physiological modulation during the stimulus: a sudden drop in HR, keeping it below the pre-stimulus range during the stimulation. Additionally, PEP first recovered to the pre-stimulus range during the stimulation, then it exceeded this range by the end of the stimulation, indicating decrease in cardiac contractility and cardiac sympathetic activity. A latent increase in PPG amplitude compared to pre-stimulus was also observed, indicating decrease in peripheral sympathetic activity. These transient changes (decrease in HR, increase in PEP and PPG amplitude) could be interpreted as decreased sympathetic tone during the stimulation. Note that in this active participant, the PPG amplitude remained at a relatively increased level after the stimulus was removed, indicating a sustained decrease in peripheral sympathetic activity even after application. This finding was quite typical in active participants, supporting post-stimulus findings from our previous static analysis study³⁶. Another noteworthy set of differences between the active and sham responses deals with changes in these biomarkers prior to stimulus delivery, which can be seen in Figure 5. In particular, we observe a pre-stimulus spike in HR and decrease in PEP. These momentary increases in sympathetic tone prior to stimulus delivery for active n-VNS could be related to anxiety and was quite typical in the overall sample set. During application, stimulation intensity is increased to as high a level as the subject can tolerate without pain. The active device delivers higher power than the sham device for the same intensity level, which might be perceived as a higher perturbation to the physiology.

As the vagus nerve is a complex structure that affects both sympathetic and parasympathetic activity, the timing of n-VNS effects will vary depending on the type of autonomic response. We observe that the effects in sympathetic measures, PEP and PPG amplitude, for the representative active subject are more latent than HR, another typical occurrence in the overall sample^41,44. Also notice that the peak effects in these biomarkers are observed towards the end of the stimulus period. In contrast, the physiologic response due to parasympathetic activation as seen in HR would occur sooner, as it is influenced by both sympathetic and parasympathetic changes. Sympatho-inhibitory phenomena (as observed in PEP and PPG amplitude in the context of this study) are regarded as slower than parasympathetic activation (seen in HR), which goes along with our observations^45,46.

Timing Considerations for n-VNS Studies

Mental stress studies are particularly important for psychiatry and the cardiovascular disease domain^47,50. Physiological responsivity to mental stress has been shown to carry cues on the identity and severity of several conditions such as PTSD, depression, anxiety, and coronary artery disease. Therefore, various types of stressors serve as clinical tools that could be induced in the laboratory under controlled conditions for diagnosis and prognosis purposes. Pairing stressors with vagal stimulation in attempts to improve mood, performance, or plasticity has been complementary to mental stress research and extensively studied in vagal stimulation studies with implants^51,52, and recently with noninvasive vagal stimulation tools^53,54. Due to differences in stressors, their duration, or the application time of stimulation, the physiological outcomes have been variable, specifically in noninvasive studies 27-29, 54. The onset time analysis from this randomized, double-blind trial could be instrumental to the design of clinical studies.

The findings of this study are also salient in the design of next-generation engineering tools to improve quality of life based on n-VNS. Recent advances in noninvasive technologies pave the way towards closed-loop systems that combine wearable sensing and stimulation. Real-time decision making through noninvasively measured parameters stands as the key mechanism for such technologies. This study lays the groundwork for closed-loop n-VNS systems that could provide “dosage” recommendations, personalized optimal delivery, or determination of response time.

Limitations

The sample size was relatively low; however, the two groups had comparable baseline characteristics. The onsets of action were manually annotated, as the sample size was low. To ensure reproducibility of results, multiple annotators independently annotated, and the inter-annotator agreement was quantified. The current study is also limited by left vagus stimulation per the study design. The findings should be re-evaluated for trials that include right vagus stimulation, as it is predicted to have different effects on cardiovascular and peripheral function⁵⁵.

Conclusion and Future Work

Wearable cardiovascular and peripheral biomarkers—HR, PEP, and PPG amplitude—were modulated in 18 ± 7 seconds for n-VNS without stress and in 16 ± 9 seconds for n-VNS following traumatic stress with a significant difference in the absent onset counts between active and sham groups. The utility of this work is to eliminate ambiguity of the physiological outcomes for n-VNS studies. The findings of this study have applications in clinical studies that use n-VNS in tandem with or without stress, as well as in the design of wearable systems that combine sensing and stimulation. For instance, stimulation timing appears to be important in the outcomes of multiple studies that test the differences in cognitive functioning, memory functioning, psychomotor functioning, or executive functioning with VNS application^56,62. The outcomes of this study could be used to design effective clinical studies for n-VNS devices. From the wearable sensing standpoint, the results could be instrumental for decision making algorithms, such as the determination of stimulation timing or the effectiveness of the stimulation. The inclusion of the sham group and multi-day protocol presented in this study provides unique, continuous wearable sensing data to consider for physiological outcomes of similar studies.

Future work should assess approaches to automate this process as more patients are recruited who have been diagnosed with psychiatric conditions. Further investigation into translating these findings from laboratory settings to wearable devices will help generalize this work, and confirming the reproducibility of these results in larger sample sizes is also pertinent. Of particular importance is the sexually dimorphic nature of cardiac and neurological responses in men and women; as the recruitment process continues, we will leverage larger sample sizes in future work to separate based on sex and quantify any gender differences. The multiparametric determination presented herein may answer questions related to identifying physiological changes during n-VNS and the optimal time of delivery in response to a detected event.

Acknowledgments

This work is based on material supported by the Defense Advanced Research Projects Agency (DARPA), Arlington, VA, under Cooperative Agreement N66001-16-2-4054.