1. Introduction
Electroencephalography (EEG) records the electrical impulses of the brain. It is an essential tool in the real-time assessment of brain health and the detection of abnormal neurological activity. The accurate measurement of EEG signals is of significant interest to both clinical and research domains. However, EEG monitoring entails the use of expensive equipment and clinical expertise that is only available in tertiary care facilities. Studies have shown the importance of EEG in correctly diagnosing abnormal neurological events, such as seizures in neonates. Without EEG monitoring, only 9% of seizures are correctly identifiable by clinical staff [
1]. Interpreting EEG is a difficult task and must be conducted by experienced neurophysiologists [
2]. However, neurophysiologists are seldom available to review the EEG in real-time. A study examining the delay between the onset of a seizure and the administration of the necessary medication showed that the entire process of preparing the patient, recording and interpreting EEG, and ordering medication takes on average over 2 h. This study was completed in a hospital with the appropriate resources, equipment, and personnel to monitor EEG [
3]. Primary care and low-resource settings often have no means of monitoring neurological health. Abnormal neurological activity is often entirely missed or retrospectively diagnosed due to a lack of EEG resources. Both cases increase the risk of a poor outcome for the neonate [
4].
The task of acquiring high quality EEG signals is non-trivial and requires the expertise of a neurophysiologist or EEG technician. Additionally, it is time-consuming, taking up to an hour in some cases [
5,
6]. EEG signals are very small in amplitude (±100 µV) and are susceptible to many types of artefacts, such as physiological, electrical, and movement artefacts that can often saturate the EEG signals with noise [
7]. Therefore, to record accurate EEG signals, a lengthy patient preparation procedure is currently required, which involves the abrasion of the top layer of the skin and the application of conductive gels to improve the electrical conductivity [
5]. The EEG monitoring process begins by locating the sites for electrode placement based on the international 10–20 system [
8]. For neonates, a reduced electrode montage is often followed, as in
Figure 1a [
9]. The patient’s skin at these sites is then abraded with gel. This removes the top layer of the skin, the stratum corneum (SC), which is the largest contributor to the skin-electrode (SE) interface impedance. The SC impedance is inversely proportional to frequency, ranging from 200 kΩ–200 Ω over a frequency range of 1 Hz–1 MHz [
10]. The most commonly used electrodes are flat or cup shaped metal (Ag/AgCl) electrodes, as in
Figure 1b [
9]. They must be coated with conductive gel to improve the electrical conductivity across the SE interface. Using this process, the SE impedance can be significantly reduced and a stable electrical connection between the electrode and skin can be established to capture the low voltage potentials. The standards of the International Federation of Clinical Neurophysiology (IFCN) state that SE impedances should remain below 5 kΩ throughout EEG recording [
11]. Impedances below 5 kΩ are achievable using this procedure. The opposite end of the electrodes is connected to the amplifier and analog-to-digital converter (ADC) circuitry. The digitized EEG data is then processed by a computer and displayed on a monitor for interpretation by a neurophysiologist. Due to the sensitivity of EEG voltages, highly accurate electronics are required to ensure accurate EEG signals. This partly explains why clinical EEG systems are expensive. The cost of the system, its maintenance, and the clinical personnel required to operate the system significantly reduces its pervasiveness outside of tertiary care [
12].
Alternative patient preparation techniques and electrode designs for adults have been thoroughly researched in recent years to facilitate quicker EEG electrode application. EEG head-caps and headsets provide ready-placed electrode holders positioned, according to the 10–20 system. Such devices are widely used in applications where high-density electrode montages are required [
13,
14]. Screen-printed electrodes are an alternative technology that provide comparable quality to conventional electrodes, while providing quicker and easier application [
15,
16]. Having fixed locations for electrodes can be a disadvantage, as it relies on all patients having similar head sizes. Such designs are of limited use on neonatal patients due to the variability in head sizes [
17].
Dry electrode technologies rely on leveraging the use of novel mechanical designs to achieve improved connection with the skin, without the use of abrasive creams or conductive gels. There is a large variety of designs, as seen in
Figure 1c. These include foam electrodes [
18], polymer electrodes [
19], and spring-based probes [
20]. Arrays of nano, micro, and milli-needle dry electrode designs have been developed to penetrate the SC, significantly lowering impedance [
21,
22]. These technologies drastically reduce the amount of patient preparation and time required to record EEG.
Several studies assessing the performance of dry electrodes have shown comparable results between wet and dry electrodes using multiple evaluation experiments on adults. Such tests include simultaneously recording wet and dry channels and comparing correlation and coherence [
6,
20], or performing tasks such as generating an EEG alpha rhythm [
19] and other neurological tasks [
22,
23].
There has been significant development in improving the cost, size, and usability of EEG monitoring systems. Many systems are moving to battery-powered and wireless solutions, which are usable across a multitude of applications and settings [
24,
25]. These developments further support the envisioned goal of wearable EEG [
26]. The advocacy to provide open-source platforms has significantly aided advances [
27,
28]. Many of these systems have been evaluated with respect to gold standard machines outside of the clinical domain [
29]. Though many novel EEG electrode and system technologies have been widely used in the brain computer interface and related research domains [
30,
31], their use in the clinical domain has been limited due to concerns regarding their accuracy and a lack of previous clinical use [
32,
33]. Clinically validating such systems requires using uncertified electronics on patients, which is often restricted by patient safety and ethical barriers [
34]. In particular for newborns, there has been minimal development and evaluation of dry electrodes [
6].
In this paper, a framework for assessing the quality of neonatal EEG acquisition systems is presented that bypasses the need for human volunteers by using an arbitrary waveform generator (AWG) outputting signals from a database of clinically obtained neonatal EEG. Previous studies using sinusoid signals to test EEG systems have shown that the performance varies depending on the frequency of the input signal [
35]. Therefore, testing the system on actual EEG data provides a more accurate measure of performance [
36,
37]. The EEG simulation framework is used to compute the accuracy of an open-source, wireless EEG acquisition board and compare the performance of various dry electrodes to the gold standard wet electrodes using accurately obtained SE impedance models of each electrode. To verify the effectiveness of the simulation framework, the electrodes were tested on healthy adult volunteers and the results between both testing methods were compared.
The goals, experiments, and contributions of this paper are summarized below:
Investigation and modelling of skin-electrode impedance of dry and wet electrodes on adults.
Development of a neonatal EEG simulation test bench, using above impedance models.
In-vivo assessment of dry versus wet electrodes on adults, and a comparison of in-vivo results versus the proposed simulation framework.
This work forms part of the front-end design of a proposed portable, low-cost, and machine learning-assisted neonatal EEG acquisition and interpretation system [
38,
39]. The proposed system aims to increase the demographic of hospitals and clinicians that have access to EEG monitoring equipment and expertise. The portable hardware, used in collaboration with dry electrodes that are easy to apply, significantly simplifies the EEG acquisition process. In addition, the processing of EEG data using visual, audible, and machine learning algorithms on a smartphone or tablet facilitates intuitive EEG interpretation, leading to quicker diagnosis and treatment of possibly undiagnosed or misdiagnosed neurological abnormalities in neonates.
3. Results
Figure 7,
Figure 8 and
Figure 9 present the mean and 95% CI results of the SE impedance, resistance, and capacitance over a 200-point, 20-1000 Hz frequency sweep for each electrode at each location.
The specific impedance and resistance values for the 31 Hz test signal are presented in
Table 2 for each electrode at both the frontal and occipital location.
The results of the simulation framework are presented in
Table 3. There are 9 channels of data in total, as described in
Figure 5. The simulation experiment was repeated on 15 different 30-sec neonatal EEG segments. The mean and 95% CI over the 15 iterations was calculated for each channel. The correlation and SNR values were computed both on the unfiltered data and filtered data (50 Hz notch filter). The power noise (50 Hz) values were computed before filtering. Without additional filtering, the initial output signal from the generator, in the ±0.5 V range, achieved an SNR of 25 dB. The signal from the output of the resistor divider circuit, in the ±100 µV range, achieved an SNR of 23.4 dB and the signal on the cloth had an SNR of 22.7 dB. The losses in signal quality for the remaining channels were solely due to the electrodes and their respective SE impedance models.
The average SNR results for each channel from the simulation framework are plotted in
Figure 10. The circular and triangular markers represent the unfiltered and filtered results, respectively. In general, for electrodes with larger SE impedances, there is a greater loss in SNR and correlation. The power noise and signal quality losses are inversely related.
The results of the in vivo experiments, comparing wet and dry electrode simultaneous EEG recordings, are presented in
Table 4. The correlation values were calculated after band-pass (1–100 Hz) and notch (50 Hz) filtering. The MicroTIP electrode achieved a correlation of 0.92 at the frontal region. g.tec achieved a correlation of 0.86 at the occipital region.
Figure 11 shows a section of EEG recorded by the electrodes at both the frontal and occipital region after filtering. On visual inspection, the EEG traces at the frontal region were closely matched for all three electrodes. At the occipital region, there were remnants of the 50 Hz artifact on the MicroTIPs and g.tec electrodes in particular, even after filtering.
5. Conclusions
This work forms part of the front-end design of a proposed portable, low-cost, and user-friendly neonatal EEG acquisition and interpretation system. The paper presents an accurate and low-cost platform for assessing the accuracy and quality of EEG recording equipment without the need for human participants. This negates the concerns related to obtaining health/safety and ethical approval, which is required for human testing. Thus, the development iterative revision and improvement of novel EEG systems becomes much quicker and easier. The intrinsic losses in the simulation framework itself due to the digital-to-analog conversion and down-scaling were minimal. The low-cost, portable EEG acquisition system used in this study achieved high accuracy, with respect to the original EEG signal from the database. The study evaluated the use of dry electrodes compared to that of wet electrodes. The use of dry EEG electrodes resulted in higher skin-electrode impedances, as expected. However, micro-machined structures, such as MicroTIPs, effectively reduced the impedance, particularly in regions without hair. The use of larger pins, such as in g.tec-g.SAHARA electrodes, reduced the impedance over the hair. Introducing these impedances into the simulation experiment had a drastic impact on the EEG signal due to an impedance mismatch and signal attenuation. The inclusion of impedance models into the simulation framework developed a more realistic scenario for testing EEG equipment. The results show that, with the use of additional filtering, large impedances do not corrupt the EEG signals enough to significantly affect the intelligibility of the EEG signal.
Providing a quickly applicable EEG recording system to medical staff to assess the brain immediately after birth and during suspected abnormal neurological activity will improve the care and outcomes for neonates. This paper assists the development and testing of portable and user-friendly EEG technologies for the neonatal population.