Introduction
The benefits of sleep on physical and mental health are evident. However, getting a restful night of sleep can be a difficult endeavor. The quality of sleep is a critical parameter for the restfulness of sleep. Current wearable sleep technologies, such as smartwatches and wrist-, arm-, and headbands, aim at assessing sleep quality by providing details of the macrostructure of sleep, i.e., the temporal organization of the night into sleep stages. However, sleep quality depends on both, the macro- and microstructure of sleep, the latter being the identification and quantification of sleep-specific neurophysiological events (Malinowska et al., 2006).
While the macrostructure of sleep can, to a limited extent, be monitored with movement-based wearables and plotted in a hypnogram on a timescale of hours, detailed information about the macro- and microstructure can only be obtained by recording the electroencephalogram (EEG) of the sleeping brain. Consequently, the EEG is part of every scientific and clinical assessment of sleep. However, this assessment typically involves lab visits, stationary amplifiers, as well as a challenge for participants to sleep in an unfamiliar environment, which can lead to the well-known first night effect (Toussaint et al., 1997). This first night effect together with a considerable night-to-night variability of sleep, even under controlled conditions (Buckelmüller et al., 2006), limits the validity of a single night assessment in a research or clinical setting.
Therefore, it is highly desirable to record sleep EEG with a simple, user-friendly, and low-cost mobile device in free-living conditions for extended periods. Dry electrodes could constitute a major component of such a device, as they are reusable, low-cost, and able to establish sufficient electrical contact with the skin without the necessity of electrode gel. However, the use of alternative electrodes for sleep recordings is not evident, as the limited positioning options and altered contact properties might affect derived sleep parameters, including the sleep micro- and macrostructure. Consequently, at each introduction of novel electrode types for sleep monitoring, it is essential to test and characterize these for the specific application.
The Importance of Assessing Sleep Structure
Even though the overall macrostructure of sleep shows high variability between nights even within the same individual, one can affirm that, night after night, the sleeping brain cycles through various sleep stages in a repetitive manner. If this peculiar sleep pattern turns abnormal, it is indicative of a variety of adverse health conditions (Luyster et al., 2012), such as coronary heart disease (Ayas et al., 2003) or obesity and diabetes mellitus type 2 (Tan et al., 2018).
The EEG is the basis for organizing a night of sleep, epoch by epoch, into different sleep stages (Iber et al., 2007). Sleep is commonly classified into rapid-eye movement (REM) sleep and non-rapid-eye movement (NREM) sleep. NREM sleep is further subdivided into sleep stages N1, N2, and N3, which reflects increasing sleep depth, i.e., N1 represents light and N3 deep sleep. During NREM sleep, distinct EEG events with typical frequencies occur: slow waves (0.5–4 Hz), K-complexes (single slow waves) and sleep spindles (10–16 Hz). The occurrence of either of these are the hallmark for sleep stage N2. Slow waves and sleep spindles continue to occur in the deepest sleep stage N3, where slow waves with a peak-to-peak amplitude of at least 75 μV dominate at least 20% of the evaluated epoch.
Slow waves and sleep spindles are of particular interest, as they are tightly linked to memory consolidation (Rasch and Born, 2013) and restorative functions (Vyazovskiy and Harris, 2013; Tononi and Cirelli, 2014). Certain characteristics of their morphology change depending on prior cognitive challenges during wake, i.e., the amplitude and slope of slow waves are increased when preceded by specific learning experiences (Huber et al., 2004; Molle et al., 2004) and decreased when the encoding of information was prevented (Huber et al., 2006). In addition, slow waves are the primary biomarker for sleep pressure, i.e., the drive to fall asleep (Dijk et al., 1993; Borbély and Achermann, 1999). Together with sleep spindles, they account for an important part of the microstructure of sleep.
Usually, the macrostructure of sleep is depicted by a hypnogram, which is determined by scoring three EEG (frontal, central, and occipital), two EOG and one chin EMG derivation. However, a single frontal electrode referenced to the contralateral mastoid alone can capture a large proportion of ongoing neurophysiological events during sleep, as slow waves are most pronounced over frontal areas in both younger and older adults (Landolt and Borbély, 2001), and as sleep spindles are typically found over fronto-central areas (Cox et al., 2017). However, alpha activity (spectral power between 8 and 12 Hz), an important marker for the onset of sleep stage N1, is most pronounced in occipital electrodes. Accordingly, when a scoring based on a single frontal EEG derivation is compared to a three derivation scoring, the agreement of N2 and N3 is high, whereas the agreement of N1 is lower.
Type of Electrodes Used in Wearable EEG Systems
In order to obtain a high-quality EEG with a wearable device, a substantial requirement is the use of high performance electrodes. The electrodes need to ensure a good and constant electrical contact with the skin and therefore need low impedance properties. The electrode-skin contact can either be ensured by adding a conductive gel between the electrode and the skin or by using a conductive material with a high contact surface that ensures electrical contact.
Pre-gelled electrodes have previously been used in wearable EEG systems to measure overnight sleep EEG with high signal quality, but a replacement after each measurement is necessary, rendering them not economical in case of prolonged use. Therefore, replacing pre-gelled electrodes with re-usable dry electrodes at fixed positions in an integrated device would reduce costs and improve usability of wearable EEG systems.
EEG measurements can be performed with different types of re-usable dry electrodes. Pin-type electrodes are designed to reach the scalp through dense hair, but because they are not attached to the skin directly, they have the disadvantage of being subject to strong motion artifacts and can change position throughout the night (Li et al., 2016). Therefore, they require high pressure on the scalp to ensure high signal quality which causes discomfort and even pain during prolonged use (Gao et al., 2018). Bristle electrodes with softer pins are perceived more comfortable, but still require high contact pressure, especially after long-term use without recoating (Grozea et al., 2011). Unlike the pin-type electrodes, flat types do not cause pain or discomfort as they are soft, bendable, and ensure low contact impedance with a high contact surface. On the downside, due to their larger size, they require hair-free or prepared skin to create sufficient contact to the skin. This limits their application to frontal (i.e., on the forehead) electrode positions.
Methods to Evaluate Electrodes Used in Wearable EEG Systems
To enable sleep scoring and analysis, dry electrodes should be sufficiently robust to artifacts that might occur during sleep (e.g., movement and sweating) and not introduce any additional interferences. Furthermore, they need to offer electrical and physical properties that enable the recording of important sleep characteristics such as slow waves and spindles. In general, EEG electrodes should have low skin contact impedance to prevent signal attenuation and impedance mismatch, the main cause for ineffective common mode rejection, i.e., the ability of the differential amplifier to cancel out the signals that are common to both electrodes (Ferree et al., 2001). Particular to wearable applications, the electrode specifications should include a certain level of tolerance toward imperfect placement by inexperienced users that could cause additional artifacts.
The importance of assessing data quality of wearable EEG systems has been identified, but only few studies exist that examine signal quality with respect to electrodes (Radüntz, 2018) and no standardized methodology is available (Casson, 2019). Previous work that evaluated the suitability of electrodes specifically for sleep applications strongly focuses on comparing the macrostructure of the derived sleep patterns. For example, characteristics necessary to perform sleep scoring are evaluated to test a novel dry electrode array around the ear (Sterr et al., 2018). Of main interest are the Bland-Altmann agreement and Pearson correlations of macrostructure parameters obtained after scoring such as the duration of sleep stages. In addition, parameters are compared epoch-by-epoch (Griessenberger et al., 2013; Sterr et al., 2018). In a feasibility study of a tattoo-based electrode setup for sleep, four nights were recorded at the subjects’ home, and sleep is scored by an expert to qualitatively evaluate the EEG and to visually determine whether typical sleep patterns (e.g., spindles and slow waves) can be distinguished (Shustak et al., 2019). Introducing additional quantitative measures, Ferster et al. (2019) compare the correlation of the mean square power in the delta (0.5–4 Hz) and sigma (10–15 Hz) bands during NREM sleep. This comparison uses two separate portable amplifiers that are designed for home-based sleep screening, of which the reference system is a clinically established device. The challenge of using two completely separate systems is the time synchronization between the amplifiers, which leads to only visual and qualitative comparisons or large comparison windows (Ferster et al., 2019). More often, comparative electrode studies rely on a single amplifier system that shares a common reference (and ground) of either electrode type, which enables correlation analysis in the time domain, but may introduce unwanted distortions in the opposite channel (Casson, 2019). Sequential testing of different single-type electrode configurations is not possible in physiological monitoring due to the strong time dependency of the signal. Even in lab-based studies, the quantitative assessment of sleep microstructure is rarely evaluated during electrode testing. Furthermore, differences between various EEG systems and electrodes can also be matched to factors other than technological variability, such as subject and session variability (Melnik et al., 2017). Therefore, it is essential to control for these effects in electrode comparisons. To our best knowledge, the evaluation of EEG electrode quality outside of controlled laboratory conditions and comprising a detailed analysis of sleep micro- and macrostructure as well as a comparison with a reference electrode type has not yet been reported. In summary, no established methodology exists to objectively evaluate and compare electrodes for sleep applications.
Our aim was to establish a reproducible electrode testing protocol that would enable the comparison of essential features to characterize the macro- and microstructure of sleep and highlight differences and limitations that occur when used in a natural setting. In particular, we evaluated the suitability of electrodes for scoring sleep from at home overnight recordings. Furthermore, we investigated the signal quality and the sensitivity to artifacts to evaluate whether the electrodes would be reliable enough for unsupervised recordings of sleep EEG.
Discussion
In this study, we present a comprehensive testing protocol comparing essential features of both the macro- and, more importantly, the microstructure of sleep in EEG signals derived from two different kinds of electrodes. With this testing protocol, we evaluated whether a new generation of dry electrodes with a biocompatible, low impedance contact surface are in principal suitable for sleep EEG assessments in a proof-of-concept study. We demonstrated that the signal quality can be quantified by signal coherence, SNR of SWA, and HSP and that two types of electrodes configured in a frontal-mastoid electrode setting can be compared. Sleep specific features, such as slow waves and sleep spindles, as well as their individual characteristics, were discriminable. In addition, visual sleep scoring was performed on single EEG derivations for each of the two electrode types and did not lead to any substantial differences in the corresponding hypnograms. An increased level of artifacts in the form of slow-sinusoidal, high-amplitude artifacts was observed in the dry electrode EEG. The methodology developed for this evaluation is one of the most detailed published to date and enables an objective evaluation of sleep micro- and macrostructure characteristics obtained from dry electrodes for wearable sleep monitoring outside the lab.
Validation of electrode performance, in particular for novel wearable EEG, is challenging. Casson divided performance factors into four levels: (1) functional testing, (2) technical performance, (3) manufacturing performance, and (4) variability in performance (Casson, 2019). While level 1 and 2 investigations have been conducted previously on the electrode type of interest (Stauffer et al., 2018), our aim was to conduct a proof-of-concept study for a level 4 investigation to gain deeper insight of the technical performance for a particular application, i.e., the monitoring of sleep in a home environment. This application limits the types of tests that can be performed, e.g., the use of a head phantom is excluded (Casson, 2019). However, in such a setting, application specific features from the EEG waveform can be assessed more realistically. Former studies that investigated sleep specific use of electrodes have primarily focused on macrostructure features that characterize sleep, such as the comparison of sleep staging (Griessenberger et al., 2013; Sterr et al., 2018) and sleep timing parameters (Casson, 2019) against an established standard. In addition, the visual comparison of power spectral density (Debener et al., 2015; Stauffer et al., 2018) and/or individually selected, exemplary signal traces (Stauffer et al., 2018; Sterr et al., 2018; Shustak et al., 2019) is common. While these assessments give an overall picture of the suitability to use the wearable device for sleep monitoring, the signal quality and the suitability to assess the sleep microstructure remain unknown. This limits the findings to only basic sleep applications, but leaves the question whether the electrodes are suitable for use in research and clinical applications unanswered. We have introduced sleep microstructure specific measures to obtain a set of parameters that can be used to assess and compare in detail EEG and quality thereof. Many of these features are independent of spatial placement of electrodes and therefore suitable for comparisons under free-living conditions.
The detection of presence and characterization of shapes of sleep microstructure elements in an EEG are essential to assess the nature and quality of sleep. Our signal quality analyses clearly showed that the EEG signal quality of the electrodes is sufficient to study microstructures of brain activity during sleep. Specifically, the SNR of SWA is a measure that indicates the discriminatory power of slow waves and therefore is a good indicator for how easy it is to classify slow wave sleep. Our analyses elucidated that there is no distinct difference in bias and limits of agreement between EEGDrDr and EEGPgPg. Slow-wave characteristics, such as their number, maximum negative amplitude, duration, or descending and ascending slope were similar in both electrodes. Larger differences between electrodes were primarily an artifact of the automated detection algorithm used for the analysis. For example, the period of slow waves was determined by the time in between two consecutive negative zero crossings of the low-pass filtered EEG signal. Occasionally, the EEG signal of one derivation marginally crossed the zero line, whereas in the other derivation the zero line was not crossed, which resulted in a large period difference. Therefore, the investigated slow wave characteristics might be more similar between the two types of electrodes than our results would suggest. Similarly, the HSP is an important sleep biomarker for quantifying the presence of spindles. Spindles are the second key electrophysiological characteristic of NREM sleep and a very sensitive feature for reduced sleep quality due to environmental, nutritional or hormonal factors (Driver et al., 1996; Borbély et al., 1999). Moreover, together with slow oscillations, they are critically involved in memory consolidation during sleep (Rasch and Born, 2013). Again, no unexpected difference in HSP between electrodes was observed. Spindles with similar characteristics could be equally identified in the Dr and Pg EEG signal. Spindle number, their maximum amplitude, frequency, and power did not show any differences between the Dr and Pg EEG derivations. The duration of spindles showed more variability between Dr and Pr EEG derivations. Similar to slow waves, spindles were represented with a slightly higher amplitude and power, as well as a longer duration in the Dr EEG derivation, especially in the recording with the most spindles.
Dry electrodes are prone to a various artifacts types (Guger et al., 2012). The electrodes showed high resistance to electrical artifacts such as electrode pops emerging from abrupt impedance changes, visually recognizable in the EEG as very fast, abrupt, sharp artifacts. However, the Dr electrodes suffered significantly more from another type of artifact, which was identified as slow-sinusoidal, high-amplitude artifacts. The amplitude and the period of such artifacts are similar to slow waves. Therefore, especially when sleep scoring is based on a single EEG derivation, those artifacts could easily be mistaken for slow waves, which biases the scorer toward scoring deep sleep. However, this bias was marginal as no significant differences in scoring deep sleep were observed. One possible source for those artifacts could have been active sweat glands. They by themselves produce slowly changing electrical potentials and release electrolytes, which can change the impedance between the electrode and the skin. Our current method for attaching the dry electrodes with skin-adhesive tape could have facilitated this activity and artifacts.
The sleep macrostructure is represented best with a hypnogram where sleep is temporally organized into wake, sleep stages N1, N2, N3, and REM sleep. Sleep scorings performed on a single EEGDrDr and EEGPgPg derivation showed good agreement with an overall accuracy of 0.78, which was comparable to current automatic sleep scoring algorithms using a single EEG derivation compared to expert scoring (Fiorillo et al., 2019). We can conclude that both types of investigated electrodes are suitable to determine the macrostructure of sleep. However, as no standard sleep montage was available, a comparison of the two single EEG derivation scorings to a reference scoring is missing. Future studies should compare the sleep scoring between a full polysomnography measured with Dr electrodes and a full polysomnography performed with conventional electrodes.
It is important to note that this study does not intend to serve as a validation study; the low number of participants and their generally good health status, as well as the fact that only a single night was recorded prevent any conclusions related to the performance in a clinical setting. We limited the analysis to four participants as the aim was to establish a reliable testing protocol to evaluate electrodes for advanced sleep applications and to test the general feasibility of this protocol.
Though not the primary goal of our proof-of-concept study, our results nevertheless allow some conclusions related to the ongoing technical development of the Dr electrodes. For example, at this prototype stage of the tested Dr electrodes, it remains unclear how the electrode can be reliably fixated at the desired derivations, particularly for the standard reference electrodes at the limited space behind the ears (mastoids). The fixation will need to ensure sufficient contact pressure and low movement throughout the night. We temporarily resolved this with skin-adhesive tape, which likely caused more sweating, but more importantly, reduced comfort and required additional effort during placement. A smaller reference electrode optimized for the space behind the ear would most likely increase data quality and stability.
The sleep electrode analysis could be further expanded with additional, application-specific characteristics of sleep micro- or macrostructure. As such, it would be interesting to investigate how very low-amplitude sleep phenomena, such as high-frequency oscillations are represented in the EEG signal collected with dry electrodes. However, this is rarely assessed in surface electrodes, but more often in intracranial electrode recordings. Furthermore, sleep timing parameters characterizing sleep behavior that is important for clinical use, such as total sleep time, sleep onset latency or wake after sleep onset, could be added to the comparison when the focus of the application is of diagnostic nature. These statistics should only be assessed and compared when a high number of participants is available, as the inter-subject variability is high. To complement the technical performance evaluation in completely free-living conditions in future studies, it would be crucial to evaluate the usability and human-device interaction specific to electrode application and resulting quality. For example, self-administration, changing environmental conditions such as humidity and temperature as well as the variation of external noise sources specific to individual’s bedrooms would require studies with much larger population sizes. This evaluation step was currently not possible as the electrodes were not yet integrated into a single system.
A reliable detection of sleep EEG markers to characterize sleep micro- and macrostructure is essential for sleep research and many clinical applications. However, detailed clinical and scientific sleep assessments are usually performed in a lab environment. This approach is mainly due to the need for high-quality sleep EEG to assess sleep-specific neurophysiological events and the high level of manual configurations needed to operate such systems. Needless to say that such a procedure is cumbersome for both the participant and researcher and, as primarily good sleepers are selected for studies, leads to results biased toward single nights with good sleep characteristics. With dry electrode technologies continuously improving and wearable EEG system becoming more and more available, sleep assessments will likely move from the sleep lab to a home setting where multiple consecutive nights can be assessed in a familiar setting. This allows the long-term recording of natural sleep behavior in more representative populations, which is of high relevance for clinical populations with an increased risk for a sleep disorder. Critical requirements for future electrodes will be that they are reusable, easy to apply, and capable of being combined with low-cost mobile EEG amplifier systems.