Introduction

Exposure to the hyperbaric environment, such as in scuba diving, results in the uptake of dissolved inert gases into the blood due to the elevated partial pressure of the inspired gases, which then diffuses into the tissues of the body. Inert gases, most notably nitrogen (N2), subsequently diffuse back out of tissues during the ascent, or decompression phase of the dive, and may result in inert gas bubble formation within the blood vessels and tissues1,2,3,4,5,6,7. Decompression sickness (DCS) occurs when these bubbles produce clinical symptoms which range from mild (skin itching, muscle pain, nausea, and tingling) to serious including neurological dysfunction, spinal cord injury, cardiopulmonary collapse and disseminated intravascular coagulation8,9,10,11. Many divers, even recreationally, try to reduce DCS risk by increasing the oxygen partial pressure (PO2) in the breathing gas, effectively decreasing the nitrogen partial pressure (PN2) and reducing the amount of absorbed nitrogen. Less nitrogen absorption during the dive reduces the risk of DCS during decompression12,13,14,15,16. In maximizing the potential benefit of this, 100% oxygen breathing is used in some underwater diving, functionally eliminating nitrogen absorption all together. For example, each breath of a 100% oxygen closed circuit rebreather (CCR) diving apparatus utilizes a relatively small amount of oxygen that is “rebreathed” by a diver after carbon dioxide (CO2) is filtered from the exhaled gas by CO2 adsorbent17,18. These devices are often used for military clandestine operations as no bubbles escape from the unit which could reveal a diver’s position, and due to lack of inert gas in the breathing mix, DCS risk eliminated15,16.

Breathing increased PO2 in elevated pressure environments, also known as hyperbaric oxygen (HBO2), carries a risk of unwanted side effects. One of which is the development of central nervous system oxygen toxicity (CNS-OT), a set of complex and incompletely understood physiologic responses which may lead to a generalized seizure. Mild symptoms of CNS-OT such as visual disturbance, headache, diaphoresis, nausea, tinnitus, twitching and/or tingling of the face or limbs often, but do not always precede more severe symptoms such as seizure or loss of consciousness19,20. Symptoms can be variable among individuals and even within the same individual21 making CNS-OT difficult to predict. In the undersea and hyperbaric medical community, it is widely accepted that oxygen-induced seizures are not life-threatening in the controlled, dry environment of a hyperbaric chamber. However, losing consciousness or convulsing under water could result in the dislodgement of the diver’s air supply from his or her mouth, and lead to drowning19,21,22,23. Adherence to established depth and time limits24 has made severe CNS-OT a rare event, but imposes operational limitations. Hence, a means to detect the impending onset of CNS-OT would significantly improve diver safety and potentially expand current operational limitations if the diver could be alerted to implement countermeasures. No such warning mechanism currently exists.

Seizure activity is physiologically associated with elevated sympathetic nervous system activity25. Studies involving animal and human models breathing increased PO2 in pressurized hyperbaric chambers have demonstrated increased sympathetic activity related to CNS-OT seizures26,27,28,29,30. Therefore, a sensitive measure of sympathetic activity may be useful as a surrogate physio marker for seizure prediction.

Electrodermal activity (EDA) has gained popularity in recent years as a measure of sympathetic activity31. The sympathetic nervous system controls sudomotor activity, which regulates sweat secretion32,33,34. EDA measurement techniques use skin electrodes to measure conductance, which is proportional to sweat secretion33, and is therefore a measure of sympathetic activity31. Furthermore, EDA has been linked to the central nervous system sympathetic activity as demonstrated by a significant surge in EDA amplitude preceding epileptic seizures in humans35,36. It is not yet known if this technique can be used to detect impending CNS-OT.

Traditionally, analysis of EDA has been confined to the time domain, decomposing the signal into tonic and phasic EDA37. The tonic EDA is typically quantified using the mean value of the skin conductance level (SCL). The phasic EDA is comprised of the skin conductance responses. The skin conductance responses are the rapid transient events contained in the EDA signal. However, several studies have reported low reproducibility of these time-domain measures38,39,40. To this end, we have developed a more sensitive and consistent measure based on time-varying spectral analysis of EDA, which we termed TVSymp41.

We have recently shown that dynamics of EDA, typically in the 0.08-0.25 Hz frequency band for humans, are significantly altered by cognitive stress under water42. Moreover, we have used TVSymp to quantitatively assess EDA data under various scenarios including stress during sleep deprivation43,44,45,46. Our quantitative method was more sensitive in determining stress-induced changes than were other previously used methods in the literature41,44,45. Specifically, TVSymp exhibited a lower coefficient of variation and a higher intra-class correlation and sensitivity for the detection of orthostatic, cognitive, and physical stresses41. TVSymp also had higher correlation with pain-inducing stimulation level45 and more sensitivity to cognitive performance deterioration caused by sleep deprivation43,44 compared to time-domain and time-invariant spectral indices of EDA and heart rate variability.

We have recently published a study showing a large increase in phasic components of EDA represented with TVSymp of 1.9 min preceding HBO2-induced seizures in rats47. As expected, exposures to 100% oxygen at 5 ATA pressure led to seizures in all rats which demonstrates proof of concept for detection of epileptic activity resulting from CNS-OT. For operational relevance to military 100% O2 CCR diving, EDA would ideally be able to predict the onset of any (even milder) CNS-OT symptoms well before seizures occur. Since the time-varying spectral index of EDA is highly sensitive and consistent, and EDA can be measured in subjects submerged in the water as discussed above, we hypothesize that it can be used as a reliable physio marker to predict the symptoms of CNS-OT in simulated working divers, in a hyperbaric chamber. This study had two aims; to assess the feasibility of collecting EDA data from human subjects exercising in an immersed hyperbaric environment, and to determine the ability of TVSymp analysis to predict the onset of non-epileptic CNS-OT symptoms. Preliminary results were presented previously in a conference proceeding paper48.

The results of this study demonstrate a significant increase in the TVSymp index prior to CNS-OT symptoms by analyzing EDA data from hyperbaric oxygen exposure and control groups. This suggests that TVSymp may serve as an early warning signal of impending CNS-OT, as it can predict symptoms consistent with non-convulsive CNS-OT approximately 1 min prior to onset. However, further research is needed to refine detection methods and address practical challenges, such as developing robust algorithms for motion artifact detection and real-time processing of EDA data in underwater environments.

Materials and Methods

Subjects and Experiment

The primary protocol and subsequent addition of EDA data collection were approved by the Duke University Institutional Review Board (protocol Pro00070499), and all volunteers provided written informed consent to participate in the study. This study was part of a larger ketogenic diet-based interventional trial registered on ClinicalTrials.gov with ID NCT03433261. The results of this study are presented in this manuscript in advance of the publication of the clinical trial data. We have received permission to publish these results in advance of the release of the full study results. The experiments were carried out in the “Foxtrot” chamber pool at the Duke University Hyperbaric Center, Duke University Hospital, Durham, North Carolina, USA. The multiplace chamber was compressed with regular air (21% oxygen) to 35 feet of seawater (FSW), 2.06 atmospheres absolute (ATA). Subjects were immersed in 28 ± 1 °C water to the shoulders. For safety purposes, subjects were seated in a head-out position secured by a harness to prevent head submersion in the event of a convulsion or loss of consciousness. The start of the experiment was marked when the subject began breathing 100% FiO2 via mouthpiece (PO2 2.06 ATA), while exercising on an underwater cycle ergometer at approximately 100 W total output (approximately 50 W from water resistance plus 50 W added by the cycle ergometer) and performing the NASA Multi-Attribute Task Battery-II (MATB-II) cognitive test. The subjects performed the NASA MATB-II cognitive test until the exposure was terminated. Subjects were supervised in the chamber and pool by medical and study personnel. Nose clips were applied to the subjects. Subjects were educated on the signs and symptoms of CNS-OT on multiple occasions, including immediately before starting the experiment. The 100% FiO2 exposure was continued for a maximum of 120 mins or terminated immediately upon signs or symptoms of CNS-OT as subjectively reported by the subject or objectively observed by the study team, at which point the subject’s mouthpiece was removed and the subject resumed breathing regular chamber air (21% FiO2). Raw EDA data was visible to personnel outside the chamber but was not used as an indicator of CNS-OT to discontinue the protocol. Challenges in maintaining the EDA signal related to the water environment and movement did not allow for the researchers to be blinded from the EDA signal, thus we can not completely exclude confirmation bias.

This study was not intended to induce convulsions. The endpoint was the latency, or time to onset, of any non-convulsive symptom of CNS-OT as determined by in-chamber study personnel or the test subject. While CNS-OT-related convulsions or loss of consciousness are easily identifiable by an observer, other symptoms such as tinnitus or visual disturbance are not. No confirmatory test for CNS-OT exists. The symptoms of non-convulsive CNS-OT are often subjective and difficult to determine in real time. Symptoms for each case were reviewed by experienced, board-certified undersea and hyperbaric medicine physicians; the PI who was present for every experiment, and 3 co-investigators who later reviewed the description of the signs/symptoms in an effort to classify the likelihood of CNS-OT. The final determination was made by the PI, generally erring to the lesser classification when not all investigators agreed. The subjective nature of attempting to classify the encountered symptoms is a notable limitation. Each case was labeled as “definite” (CNS-OT) or “probable” (pCNS-OT). Signs or symptoms classically associated with CNSOT such as tingling, twitching, altered mentation, motor control/coordination issues, sudden nausea, visual or auditory changes and diaphoresis when associated with other symptoms was generally assigned as “definite.” A “probable” designation was assigned when symptoms were more questionable such as lightheadedness, mild headache, mild nausea or diaphoresis. The designation of “not CNSOT” was assigned for symptoms like fatigue, mild headache and mild concentration difficulty. The case was classified as “finished” if the subject remained asymptomatic during the entire 120-min protocol, when CNS-OT did not occur.

EDA and ECG signals were collected simultaneously. The sampling frequency was set to 100 Hz for both signals. EDA was collected using two sets of stainless-steel electrodes (ADInstruments MLT118F) placed on the index/3rd and middle/5th fingers of subjects’ left hand (out of the water) using ADInstruments PowerLab 16/35 PL3516 Data Acquisition System and ADInstruments Galvanic Skin Response (GSR) Amp FE116 (with fully isolated AC excitation and automatic zeroing low voltage amplifier, 22 mVrms at 75 Hz). ECG signals were collected using 3 M Red Dot Ag/AgCl electrodes and a Hewlett-Pâckard M1094B/M1176A monitoring system. ECG and EDA data were recorded in ADInstruments LabChart8 Pro® software. ECG signals were band-pass filtered (0.05–40 Hz) to reduce noise and motion artifacts. From the filtered ECG, the R peaks were detected using a widely used peak detection algorithm49,50. The detected R peaks were manually corrected to avoid misdetections. Heart rate (HR) was computed from the R-R intervals.

Exercise dataset

We aimed to evaluate if a large elevation in EDA or HR predicts the occurrence of CNS-OT. Exercise increases sympathetic activity and subsequently several physiological parameters driven by it, including HR and EDA; therefore, we needed to evaluate if exercise caused a similar elevation that could confound CNS-OT symptom detection and prediction. For this reason, we incorporated pre-existing data from eighteen subjects exercising in a non-hyperbaric, normoxic environment. In the exercise test, the subjects performed an incremental intensity test on a motorized treadmill which began with an initial warm-up, followed by walking at 3 mi/h (~4.82 km/h). The speed was increased to 5 mi/h (~8 km/h) and increased 0.6 mi/h (about 1 km/h) every subsequent minute until the subjects reached 85% of their HRmax (HRmax = 206.9–0.67*age). These data was previously published in a study looking at the shift of spectral content of the EDA toward higher frequencies due to physical activity51.

EDA signal processing

Pre-processing

Motion artifacts can produce large surges in the EDA data and TVSymp, which could produce similar signals experienced prior to CNS-OT. To avoid false positives, we used a simple method for automatic EDA data quality assessment52. This method uses a set of rules to label invalid EDA data: EDA data out of range, abnormally fast changes in the signal, misconnection detection based on a temperature sensor (not implemented in this study), and corruption of neighboring 5-second segments from each data sample. A total of 1,695,673 (4.65%) of the samples were identified as motion artifacts and labeled invalid. The EDA labeled as valid data was used for subsequent analysis. A median filter (1-second width) and a low-pass finite impulse response filter with cut-off-frequency of 1 Hz were applied to further remove noise from the selected EDA data.

Time domain decomposition of EDA

The traditional analysis of EDA is based on the decomposition of raw signals into tonic and phasic components33,37. Along with HR and TVSymp, this study evaluated the sensitivity of tonic and phasic components of EDA for CNS-OT detection. We decomposed the EDA signal into its tonic and phasic components in the time domain, using the convex optimization approach53. Note that the Tonic EDA component contains the slow changes of the signal, and the Phasic EDA contains the rapid changes of the signal. From the phasic EDA, the phasic driver was estimated. The phasic drivers are an estimation of the location and the strength of the underlying sympathetic activation that causes each individual skin conductance response37. Tonic EDA, phasic EDA, and the phasic driver are illustrated in Fig. 1 for a sample EDA signal (all three have the same units as the raw EDA, microsiemens, µS).

Fig. 1: Decomposition of raw electrodermal activity (EDA) and heart rate.
Fig. 1: Decomposition of raw electrodermal activity (EDA) and heart rate.The alternative text for this image may have been generated using AI.
Full size image

Top: Tonic EDA, Phasic EDA, and phasic driver. Bottom: time-varying spectral analysis of EDA (TVSymp) and heart rate obtained from the ECG.

Time-variant spectral analysis of EDA, TVSymp

We previously developed time-invariant and time-variant (TVSymp) spectral analyses of EDA signals, and validated them for sympathetic tone assessment in different applications including stress, fatigue, dehydration, and pain detection39,41,44,45,46. TVSymp has lower intra-subject variability and higher day-to-day consistency (coefficient of variation) and sensitivity compared to the time-invariant spectral index and the traditional time-domain measures of EDA (tonic and phasic components) in response to orthostatic and cognitive stress40. To compute TVSymp, the time-frequency representation of EDA was computed using variable frequency complex demodulation (VFCDM), which enabled more accurate amplitude estimation along with obtaining one of the highest time-frequency resolutions54,55. At a sampling frequency of VFCDM decomposition at 2 Hz, the second and third components, comprising the approximate frequency range 0.08–0.25 Hz, were used to compute TVSymp as defined in a previous study41. Amplitudes of the time-varying components in this band at each time point were summed together to obtain an estimated reconstructed EDA signal, X’(t). X’(t) was normalized to unit variance (making TVSymp a dimensionless quantity), and its instantaneous amplitude was computed using the Hilbert transform56, as follows:

$$Y^{\prime} (t)=\frac{1}{\pi }P{\int }_{\!\!\!-\infty }^{\infty }X^{\prime} (\tau )/(t-\tau ){{\rm{d}}}\tau$$
(1)

where P indicates the Cauchy principal value. X ’(t) and Y ’(t) form the complex conjugate pair, to define an analytic signal, Z(t), as

$$Z(t)=X^{\prime} (t)+{iY}^{\prime} (t)=a(t){e}^{\,j\theta (t)}$$
(2)

in which

$$a(t)={{\left[X^{\prime 2}(t)+Y^{\prime 2}(t)\right]}^{1/2},}$$
(3)
$$\theta (t)=arctan {\left(Y^{\prime} (t) / X^{\prime} (t)\right)}{\cdot}$$
(4)

The resulting a(t) is the instantaneous amplitude of Z(t) and corresponds to the TVSymp time series. Note that TVSymp is in non-dimensional units (Fig. 1).

Statistics and reproducibility

For each group (CNS-OT, non-CNS-OT, exercise, and pCNS-OT), the normality of baseline and end-of-study values of maximum Tonic EDA, phasic driver, TVSymp, and HR was tested for each group using the one-sample Kolmogorov–Smirnov test57,58,59. All indices were normally distributed. The paired t test was used to compare differences between end-of-study and baseline maximum values for each group, and P  <  0.05 was considered significant. Values are mean (MM) and standard deviation (STD) unless stated otherwise.

Receiver operating characteristic analysis

The ROC analysis was performed to test the feasibility of HR and EDA signals as classifiers for the detection of CNS-OT symptoms from baseline values. All baseline measures for CNS-OT cases, non-CNS-OT cases, exercise subjects, and pCNS-OT cases, as well as end-of-study values for non-CNS-OT cases and exercise subjects, were considered as negative values of CNS-OT. End-of-study values for CNS-OT cases were considered as positive values of CNS-OT, for each index. Only definite CNS-OT were considered positive values of CNS-OT. For all the indices (Tonic EDA, Phasic driver, TVSymp, and HR) we compiled the ROC curve60, obtained the area under the ROC curve (AUC)61, and calculated Youden’s index (J = Sensitivity + Specificity – 1, an index of the overall performance of the detector)62. We have set the optimal threshold for each index as the value that maximizes J, and computed the sensitivity and specificity of the classifier for such threshold value using the definite group. Probable CNS-OT cases were classified using the threshold for each index.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Results

Data from 50 HBO2 exposures (26 individual subjects, 24 males and 2 females, aged 27.6 ± 5.4 years) was recorded and combined with 18 exercise only cases (11 males and 7 females, aged 21 ± 3 years, no HBO2 exposure) from previously published data, for a total of 68 cases. Baseline values for maximum Tonic EDA, phasic driver, TVSymp, and HR were established from the 68 cases within the first 5 min of the experiment (start of 2.06 ATA O2 breathing for the HBO2 exposures or start of exercise for the non-HBO2 exposures. The same values from within the last 5 min of the experiment were analyzed in 4 groups: (1) those labeled as non-CNS-OT (non-CNS-OT, N = 18) which included cases where subjects finished the protocol without symptoms (N = 12 cases from 9 different subjects) and cases where the subject stopped the protocol due to symptoms which were not felt to be CNS-OT (N = 6 cases from four different subjects); (2) those labeled as definite CNS-OT cases (CNS-OT, N = 18 cases from 12 different subjects); (3) the exercise only subjects (exercise, N = 18); and (4) those labeled as probable CNS-OT (pCNS-OT, N = 14 cases).

TVSymp was able to be calculated and graphically represented for each analyzed case. As an example, Fig. 2 depicts data from one subject who developed CNS-OT and another subject who did not. In the bottom panel, the subject exhibited symptoms suggestive of CNS-OT, described by the subject as a sensation of rapid onset warmth and flushing which coincided with visible facial diaphoresis noted by in chamber personnel. In this case, a large increase in the TVSymp amplitude preceded the occurrence of symptoms of CNS-OT, reaching a TVSymp value greater than 8 (marked with a black circle). This increase occurred 100 s prior to the 100% FiO2 oxygen being discontinued due to CNS-OT symptoms, indicated with an arrow and marked with a red line at the end of the HBO2. The non-CNS-OT case in the top panel consistently showed low values of TVSymp throughout the HBO2 exposure and after oxygen was discontinued. In contrast, the CNS-OT case showed several prominent peaks of TVSymp even after the end of HBO2, possibly because the symptoms continued even after the oxygen was discontinued.

Fig. 2: Comparison of TVSymp values between a subject who developed CNS-OT and a subject who did not.
Fig. 2: Comparison of TVSymp values between a subject who developed CNS-OT and a subject who did not.The alternative text for this image may have been generated using AI.
Full size image

The plots show data for the last 10 min of breathing 2.06 ATA oxygen, and 5 min after switching to breathing hyperbaric air. Subject A (top panel) did not exhibit symptoms of CNS-OT, and subject B (bottom panel) presented symptoms of CNS-OT (diaphoresis). The vertical red line represents the end of breathing 100% O2. The black circle represents the maximum of TVSymp in a 5-min window preceding the end of the study.

EDA indices for each group (1) non-CNS-OT case group, (2) CNS-OT case group, (3) the exercise only data group (no HBO2 case), and pCNS-OT are depicted in Fig. 3. Tonic EDA exhibited a significant increase in the end-of-study values for the non-CNS-OT case, CNS-OT case and pCNS-OT case groups, as did the exercise only group compared to baseline values. The phasic driver values did not exhibit significant differences between groups. There was a significant increase in the maximum value of TVSymp in the CNS-OT case and pCNS-OT case groups for the end-of-study values when compared to baseline and the other groups. A difference in HR was seen between baseline and end-of-study values in the exercise only data group, but not the other groups.

Fig. 3: EDA indices for non-CNS-OT, CNS-OT, exercise, and pCNS-OT groups.
Fig. 3: EDA indices for non-CNS-OT, CNS-OT, exercise, and pCNS-OT groups.The alternative text for this image may have been generated using AI.
Full size image

Tonic EDA, phasic driver, TVSymp, and HR values for baseline and end-of-study for non-CNS-OT (N = 18), CNS-OT (N = 18) exercise (N = 18), and pCNS-OT (N = 14) subjects. The number of samples for each group is shown in parentheses. In each box, the central mark is the median, and the edges of the box are the 25th and 75th percentiles. * represents significant difference to baseline.

Figure 4 shows the receiver operating characteristic (ROC) curves for CNS-OT symptom detection. The closer a ROC curve is to the upper left corner, the more efficient is the classification because the model reaches a higher combination of sensitivity and specificity. In this study the TVSymp ROC curve is a better classification index because its curve is more pronounced towards the left upper corner, followed by Tonic EDA. HR and the phasic drivers of EDA exhibit poor performance. Table 1 includes the AUC, sensitivity, specificity, and J values for all the indices as detectors of CNS-OT symptoms. Overall, TVSymp showed better performance as a detector of CNS-OT than did the other measures, exhibiting the highest AUC (0.79), sensitivity (1.0), and J (0.48). The specificity for TVSymp (0.48) was lower than for HR and Phasic drivers. If the pCNS-OT case group is included as positive CNS-OT, TVSymp a higher AUC (0.83), sensitivity specificity (0.68), and J (0.68), with a very high sensitivity (0.95).

Fig. 4
Fig. 4The alternative text for this image may have been generated using AI.
Full size image

ROC curves for the detection of CNS-OT symptoms for Tonic EDA, phasic drivers, TVSymp, and HR (N = 53).

Table 1 Table for ROC analysis of detection of CNS-OT symptoms using measures of EDA and HR
Full size table

Finally, we measured the time between the maximum elevation of TVSymp and the onset of symptoms of CNS-OT for the definite CNS-OT group (Fig. 5). On average the TVSymp peak occurred 57 s prior to the observance of symptoms, with a standard deviation of 79 s, and a median of 10 s. If the group of pCNS-OT cases is included as cases of CNS-OT, on average the TVSymp peak occurred 84 ± 76 s prior to the observance of symptoms.

Fig. 5: Time between the maximum elevation of TVSymp and the observance of true CNS-OT symptoms as evaluated by the diving physicians (N = 18).
Fig. 5: Time between the maximum elevation of TVSymp and the observance of true CNS-OT symptoms as evaluated by the diving physicians (N = 18).The alternative text for this image may have been generated using AI.
Full size image

The central red line is the median, the edges of the box are the 25th and 75th percentiles, and the whiskers extend to the most extreme datapoints.

Discussion

Regarding our first aim, collecting EDA data from human subjects exercising in an immersed hyperbaric environment was challenging, but possible. The finger electrodes needed to be kept dry, and data needed to be transmitted in real time to a recording station out of the chamber as no hyperbaric-rated EDA device is available.

Regarding our second aim, we found a significant elevation of TVSymp values preceding the onset of non-convulsive CNS-OT symptoms, which suggests it can potentially be used as a physio marker for impending CNS-OT. CNS-OT symptoms can vary, and EDA captures the changes in skin conductance which is directly affected by sweat secretion33. It is yet to be determined if EDA is more sensitive to CNS-OT presentations where sudden diaphoresis is the predominant symptom as opposed to others such as twitching or visual disturbance which may not have as much associated sweat production.

Symptomatic CNS-OT subjects exhibited elevated dynamics in the phasic component of EDA, which were better captured by TVSymp41, before or around the time the symptoms of CNS-OT. Although phasic drivers showed an apparent increase in value, it was not significant. A significant increase was also observed in Tonic EDA values before the symptoms of CNS-OT were observed, however such elevation is also observed as a result of physical activity without HBO2 exposure (exercise only dataset). The phasic driver was significantly higher for exercise subjects, but the phasic components were partially shifted outside of the frequency range used to compute TVSymp51. Importantly, the elevation of TVSymp was observed in CNS-OT cases but not in exercising subjects who were not exposed to HBO2 and therefore could not have had CNS-OT. This indicates the utility of TVSymp in predicting the symptoms of CNS-OT in conditions of physical activity.

Comparison with the animal model

This study complements our recent findings in an animal model that hypothesized large spikes in EDA values will precede seizures in rats breathing 5 atmospheres absolute (ATA) HBO247. The results of that study showed that a significant increase in the amplitude of a modified TVSymp (to account for the spectral content of rats’ EDA) values derived from EDA recordings occur on average 1.9 min before HBO2-induced seizures. There are several differences between the animal and human experimental models. Although the same device was used to collect EDA in both experiments, the rats (N = 10) were sampled using a pair of parallel subcutaneous needle electrodes implanted in the back. In the human model we collected the EDA from cutaneous electrodes on the fingers, which is the best location for collecting EDA non-invasively for signal quality and amplitude. The rats were compressed to 5 ATA breathing 100% FiO2 (PO2 = 5.0) to intentionally induce CNS-OT seizures for all rats. The PO2 for the human study was lower (2.06 ATA) with the intent to avoid generalized seizures, but rather elicit non-convulsive symptoms of CNS-OT with close monitoring. In other words, the end point for rats was the onset of the seizure, whereas in humans it was any symptom of CNS-OT, typically mild in comparison. For this reason, the EDA reaction is more visible and precedes the end point with more anticipation (1.9 min in rats vs. 1 min in humans). It is possible that the rats experienced non-convulsive symptoms preceding the seizure, and it is also possible human subjects would have experienced a seizure had the oxygen not been removed in a timely manner, but this remains speculative. Finally, it is worth mentioning that the computation of TVSymp using time-frequency analysis of EDA was adjusted for rats, since the dynamics of EDA in rats are faster (higher frequencies). For rats, TVSymp was computed using the range 0.16–0.48 Hz, whereas in humans the range is 0.08–0.25 Hz47.

In this study, we analyzed human EDA and ECG signals in both dry (non-HBO2 exercise) and immersed conditions at increased PO2 in a hyperbaric chamber. Some data was not included as it was non-usable due to either low signal fidelity or technical issues at the onset of the study. The former was most likely due to motion artifact of the wires or subject movement related to exercise, EDA electrode skin contact issues or inadvertent submersion. Nevertheless, most data from this study could be processed and analyzed. Monitoring EDA in a completely submerged, exercising diver presents practical challenges moving forward. We have, however, developed hydrophobic electrodes63 and have been able to collect EDA data under water with insulated electrodes42 which could be tested in the future.

CNS-OT occurrence is highly variable among individuals, and even within the same individual in different exposures, making it difficult to predict the onset of symptoms20,21. Making the determination that a diver experienced CNS-OT is also based on clinical assessment and not a definitive test, making the diagnosis of CNS-OT relatively subjective. This creates an inherent limitation for this study where physicians made a best effort differentiation between “definite,” “probable” and “not” CNS-OT cases. Based on these results, it is possible that many of the cases labeled as “probable” were in fact true cases CNS-OT. Hence, having an autonomous and quantitative approach to predict CNS-OT using EDA would be beneficial for diver safety. The results of this study preliminarily demonstrate the feasibility of TVSymp index to predict symptoms consistent with non-convulsive CNS-OT approximately 1 min prior to onset. This may provide enough advanced warning for a diver to implement countermeasures such as switching to an alternate breathing gas (lower PO2). There could also be a potential to lengthen safe exposure limits in certain circumstances, and improve disabled submarine escape capability12,14,64. While there are many current limitations, this study shows potential for using TVSymp as a detection method to provide early warning for a diver to implement countermeasures such as gas switching, ascent to a shallower depth, etc.

For the automatic detection of motion artifacts and corruption of EDA signals collected in this study, we used the simple, transparent, and flexible method recently reported52. Although this tool was able to identify several instances of poor data quality, many times it was not able to properly identify data corruption clearly observed on first inspection. As we envision eventual automated CNS-OT detection using a wearable EDA device65, we will need to develop a more robust approach for motion artifact detection. Improving the robustness may require a sequence of algorithms to automatically and accurately determine if a data segment to be analyzed is either clean or motion-corrupted. This task is crucially important to minimize sudden spikes in TVSymp values (due to abrupt motion artifacts) being falsely identified as true sympathetic responses to CNS-OT. As even in a controlled lab environment, we have experienced motion artifacts in EDA data. It is highly likely that this problem will be even more prevalent when divers perform their tasks in natural environments. Hence, we need an intelligent algorithm to discern clean EDA automatically and accurately from motion- and noise-corrupted data, which should be achievable. Furthermore, no device exists that can run the necessary algorithm for EDA at pressure.

For future studies, the use of tri-axial accelerometers to discern sudden movements is also a possibility. Tri-axial accelerometers detect the combined magnitude and direction of linear, rotational, and gravitational acceleration, which can be used to compute statistical features for machine learning models to detect and remove motion artifacts. In addition, we have already published a method to detect motion artifacts even without accelerometer data66 followed by removing them using a novel deep learning approach67. We are also working on the development of a wireless device to collect EDA and accelerometer data. All data will be processed in near real-time using a smartphone application. Note that motion artifacts in other physiological signals including ECG, and other techniques to obtain HR such as photoplethysmographic signals, are especially pronounced in data from wearable devices, thus, many methods have been developed to overcome them68,69,70,71,72. We foresee that both machine learning and deep learning approaches will be suitable for this purpose. Deep learning has been shown to provide more accurate results than machine learning, albeit at higher computational cost73.

We face challenges related to the practical application of the detection of CNS-OT symptoms based on TVSymp and EDA. The detection of CNS-OT symptoms using a threshold approach for TVSymp values was not perfect (sensitivity = 1.0 and specificity = 0.48) as some CNS-OT cases showed a lower increase in the phasic component of EDA, possibly because their symptoms were not linked to autonomic control. The specificity of TVSymp is better if the group of pCNS-OT cases is included (specificity = 0.68). Similarly, some non-CNS-OT cases may have exhibited an elevation in TVSymp due to other underlying conditions like stress and anxiety. With the available data, we can set a threshold value for the increase in TVSymp values when CNS-OT occurs, but this value is tuned from previously collected data. Hence, when we have new subjects, this arbitrarily derived threshold value may no longer be valid. Thus, to resolve this issue, we need a machine learning approach to automatically classify (without an arbitrarily set threshold value) whether a given increase in TVSymp values is truly due to increased sympathetic nervous system activity related to CNS-OT. In future studies, we will aim to improve the performance of the detection of CNS-OT symptoms using machine learning to differentiate between autonomic and non-autonomic induced elevation of the phasic component of EDA.

In this study, using HR as a physio marker showed poor performance for the prediction of CNS-OT symptoms as shown in Fig. 4 (ROC curve) and Table 1. With a sensitivity of 0.11 and a specificity of 1.0, HR as the sole physio marker is insufficient to detect impending CNS-OT symptoms. However, HR could bear relevant information for use in machine learning algorithms to increase the specificity of the prediction models based on EDA.

Ketogenic vs. control

The data for this analysis were collected during a study for which the subject exposures occurred while the subject was on a normal (control) diet or a dietary change (ketogenic). The researchers that processed the data did not have access to the diet information (normal or ketogenic) for each exposure, to avoid bias, as we intended to predict CNS-OT symptoms independent of diet. Although the analysis of the effect of diet on the development of CNS-OT is beyond the scope of our paper, we have evaluated the influence of diet on the parameters of EDA and HR. For doing this, the diet data was unblinded and a three-way analysis of variance (ANOVA) was performed to test if the significant differences were produced solely by the CNS-OT or if the diet also contributed to it. The Bonferroni method was used for correction of multiple comparisons and P  <  0.05 was considered significant.

Table 2 includes the results of the three-way ANOVA. The analysis confirmed the significance of the effects of CNS-OT symptoms on the different indices of EDA. Importantly, the diet (Keto vs. Control) had a significant effect only on HR. The diet did not have any significant effect in combination with other factors (symptoms of CNS-OT or BL vs. end-of-study). This confirms that the elevation of spectral components of EDA preceding CNS-OT symptoms in divers is independent of the diet.

Table 2 Three-way ANOVA table
Full size table

Conclusion

Based on this preliminary work, analyzing the phasic component of EDA captured by TVSymp, may offer a means of predicting onset of CNS-OT but needs further investigation and refinement. TVSymp elevation occurred with a mean and median of 57 and 10 s respectively, prior to CNS-OT manifestations seen in the “definite” CNS-OT case group. A simple threshold approach can detect the impending onset of CNS-OT with a high sensitivity (1.0) but lower specificity (0.48) using TVSymp. Based on this sensitivity, the risk of missing an approaching CNS-OT event appears low, but there is a risk of unnecessarily aborting a dive early due to the lower specificity. Further research is necessary to observe the interplay between CNS-OT-induced symptoms and TVSymp to automatically quantify the increase seen in TVSymp attributable to CNS-OT as opposed to motion artifact, and to develop technology and methods for monitoring real-time EDA in the austere underwater environment.