Sleep, a fundamental physiological process, plays an important role in various biological functions. It aids in the regulation of hormones, strengthens immune system, and facilitates memory consolidation — retaining vital memories and purging extraneous information (11, 12, 13). Beyond being a distressing ailment on its own, insomnia has been linked with a range of medical complications. These include cardiovascular diseases and diabetes, as well as mental health disorders such as major depression and Alzheimer's disease (14, 15, 16, 17). Consequently, this interplay between sleep and various health issues has captured the keen interest of the clinical community.
Recent studies have spotlighted RBD as an early warning sign for neurological conditions like Parkinson’s disease, dementia with Lewy bodies, and other α-synucleinopathies (18, 19). Despite its importance, RBD is commonly diagnosed based on self-reports (20). However, such self-reports can be unreliable, as individuals suffering from RBD often omit or are unaware of their symptoms. Therefore, there's an impending need to rely on more objective diagnostic methods. PSG stands out as an essential tool in this regard, offering direct insights into sleep patterns and disturbances. Not just for RBD, but for diagnosing other sleep disorders too, there's an evident shift required - from subjective methods like questionnaires and actigraphs to more definitive ones like PSG.
In our investigation, the consistency between a novel portable 2-channel sleep staging apparatus, SleepGraph, and the standard laboratory PSG was impressively robust, as evidenced by our findings. The 2007 AASM manual advocates the display of three distinct EEG channels — F4-A1 for δ waves, C4-A1 for spindle waves, and O2-A1 for α waves — on a CRT monitor for accurate sleep stage identification. The Polymate follows this recommendation by displaying these three channels in addition to Fp2-A1. In contrast, SleepGraph limits its display to a singular channel, Fp1-A2. However, the overarching concordance between SleepGraph and Polymate was commendably strong, underscored by a κ value of 0.84.
Drawing a parallel to earlier research wherein a single EEG positioned beyond the hairline, complemented by two independent EOG channels, was assessed, the derived κ value was an analogous 0.87 (21). Conversely, a separate study scrutinizing the efficacy of a portable sleep apparatus equipped with only a singular forehead EEG channel recorded a κ value of 0.67 in its epoch-by-epoch comparison with PSG (7). This variation suggests that omitting the EOG might lead to a diminished agreement, represented by a decreased κ value.
Further compounding this theory, several past investigations employing a lone EEG in conjunction with an automated scoring system reported moderately reduced κ values, oscillating between 0.46 and 0.75 (22, 23, 24). Such findings reinforce the notion that the trio of EEG, EOG, and EMG encapsulates sleep-stage specific attributes. Consequently, the integration of all three parameters appears indispensable for the refinement and accuracy of automated scoring systems.
The sensitivity observed in stage N1 was notably diminished. This might be attributed to the absence of distinctive brain waves characterizing this stage. Often, transitions between stage N1 to either stage W or N2 appeared fluid and indistinct. On the other hand, the specificity in stage N2 was the least pronounced. A potential explanation could be that apart from the sleep spindle, stage N2 lacks other distinctive brain waves. Moreover, the presence of sleep spindles exhibited considerable variability across individuals.
Conversely, both sensitivity and specificity for stages W, N3, and R were commendably high. This can be ascribed to the pronounced features unique to each stage: the elevated levels of EMG and EEG α waves in stage W, the dominant EEG δ waves in stage N3, and the reduced EMG levels coupled with rapid eye movements in stage R. From this perspective, the confluence of EEG, complemented by EMG and EOG, becomes indispensable to bolster the precision and reliability of sleep stage identification.
The almost impeccable concordance in sleep stage classifications between SleepGraph and Polymate mandates solid empirical justification. This study illustrated that specific EEG power signatures, combined with EMG integrals and EOG variances, bore striking similarities when measured by both SleepGraph and Polymate, reinforcing the credibility of their agreement.
This study heralds the introduction of five innovative technical strategies, each aimed at refining the process of sleep analysis. The inaugural attempt was focused on discerning EEG artifacts. By employing this technique, we managed to effectively filter out substantial motion artifacts from the otherwise quiet recordings, as illustrated by the solid lines in Fig. 2B.
The subsequent innovation dealt with the method employed to ascertain EEG peak frequencies. We harnessed the 3-point-moving-average approach on the spectrum data, enabling us to clearly define the peak frequency. In scenarios where a peak is sandwiched between two neighboring frequency bins, the representative peak might exhibit a discrepancy - either by 0.25 Hz in SleepGraph or by 0.39 Hz in Polymate. Additionally, the utilization of the 3-point moving average potentially broadens the accuracy threshold of the peak to encompass an area spanning 2 frequency bins. This translates to an accuracy range of 0.50 Hz for SleepGraph and 0.78 Hz for Polymate. Any deviation in the peak beyond the benchmarks of 0.50 Hz in SleepGraph and 0.78 Hz in Polymate could raise concerns regarding accuracy. Nonetheless, it's heartening to note that no participant in our study exhibited a difference in peak frequency exceeding 0.78 Hz when comparing readings from SleepGraph and Polymate. The subsequent technical innovations will be keenly anticipated and will likely further underscore the robustness of this study's methodologies.
Our study's third technical endeavor centered around the spectral analysis of EEG. For an effective sleep evaluation, EEG brain waves ranging from δ to β are indispensable. The optimal sampling rate must range between 80–120 points/second, which results from the multiplication of 4 data points by 20–30 Hz. Since 2007, with the transition of the scoring rule from R&K to AASM (1), there has been a mandate to score sleep stages every 30 seconds. Within the framework of SleepGraph, EEG data spanning 32 seconds – comprising the original 30-second EEG data followed by a 2-second blank data – can be earmarked for EEG spectral analysis. This selection is particularly apt, considering that the aggregate of 128 x 32 data points adheres to the FFT stipulation of being in integral powers of two. On the other hand, Polymate allows for a selection of EEG data amounting to 30.72 seconds – the original 30-second EEG data complemented by an additional 0.72 seconds of blank data. This selection, which totals 200 x 30.72 data points, aligns seamlessly with the FFT criterion.
Conventional wisdom dictates that the EEG spectral analysis utilizing FFT should be executed with a 20-second page length at a 128-Hz sampling rate (25, 26, 27). Our study ventured to test the validation of EEG spectral analysis within SleepGraph. Although a comprehensive analysis of the results has not been encapsulated within this paper, a brief synopsis reveals that the EEG data in SleepGraph was dissected and analyzed independently within the 20-second page span. The subsequent EEG power spectrum, analyzed in this manner, mirrored trends seen in the 30-second page duration in Supplementary Fig. S2b of the representative subject (ID16393). A vast majority of the EEG parameters remained consistent across both durations in Supplementary Table S3 online.
In our study, we observed non-significant correlation coefficients for EEG powers in δ and β bands in two subjects (p > 0.05). A potential reason for this is that band-pass filters often exhibit greater accuracy closer to the central range of their operation. Consequently, at the extremities of the band-pass range, where δ and β are situated, larger variations may emerge.
However, an interesting observation was made upon subjecting the EEG power at the δ peak to a common-log transformation: the pronounced variation was effectively diminished. As evidenced in Table 1, a significant correlation for log10δ was established between SleepGraph and Polymate. Furthermore, this transformed parameter exhibited distinct stage-dependent characteristics, as highlighted in Table 2. This transformation may offer a methodological enhancement to address the inherent challenges posed by the extremities of the band-pass range.
The fourth technical approach we employed pertained to the utilization of designated EMG metrics. Our findings revealed that the designated EMG integral demonstrated more pronounced stage-dependent characteristics compared to the traditional EMG integral, as can be observed in Table 2.
A noteworthy aspect of designated EMG lies in its electrode placement; one of the electrodes is positioned on the lower eyelid, directly above the masseter muscle. The other side of electrodes is placed on the mentalis, over the orbicularis muscle. This strategic location allows the designated EMG to concurrently monitor both the orbicularis and the masseter muscles. Consequently, the designated EMG can capture a broader spectrum of muscular activity. This might explain the enhanced sensitivity of the designated EMG in delineating sleep-stage dependent features. This dual monitoring capability holds promise for more detailed and comprehensive analyses of sleep patterns and muscle activity during sleep.
Our fifth technical initiative was the exploration of designated EOG metrics. A pivotal observation from our study, as illustrated in Fig. 1b, was that the EEG slow waves during stages N2 and N3 displayed more contamination in the EOG as opposed to the designated EOG.
The derivation pathway of the EOG intrinsically traverses through the brain. This is attributed to the placement of its electrodes – one on the upper eyelid and the other on the earlobe of the opposite side. In contrast, the designated EOG's electrode placement, with one electrode on the lower eyelid and the other on the opposing side of the chin, ensures that its derivation does not pass through the brain. This strategic positioning of the designated EOG electrodes resulted in significantly reduced EEG contamination. Consequently, the designated EOG outperformed the EOG in showcasing sleep-stage dependent features. The EOG is important for checking inverse eye movements between two EOG channels but is inappropriate as an objective quantitative parameter.
The evidence of this distinction was clearly manifested in our data. The variance in the designated EOG displayed more pronounced stage-dependent characteristics compared to the EOG, as evidenced in Table 2. This suggests the potential of designated EOG as a more reliable metric in sleep stage assessment.
The journey of SleepGraph began with its prototype creation in 2007, which came with an innovative automatic sleep stage scoring function with single-channel EEG. This pioneering work was showcased at the annual meetings of the Japanese Sleep Research Society in both 2007 and 2009 (28, 29). Initially, the staging algorithm employed relied on discerning the presence or lack thereof of EEG power across various frequency bands: δ (1–4 Hz), θ (4–8 Hz), α (8–12 Hz), σ (10–16 Hz), and β (16-25Hz). The average agreement rates from these early evaluations were 80.8% from 9 participants in 2007 and 81.5% from 6 participants in 2009, respectively. Further study was held with single-channel EEG, 58 normal subjects were tested. The average agreement rates and κ values were 86.9% and 0.75, respectively (30). Recording channels were increased to the latest version of SleepGraph, and 16 normal subjects were examined with two-channel recordings. The average agreement rates and κ values were 83.0% and 0.75, respectively (31).
Fast forward to the present research, we observed a commendable improvement in the mean agreement rate, which surged to 88.9% (n = 8). This was not just serendipity but a result of refining the parameters. EEG power peaks were more accurately delineated: δ (0.75–1.17 Hz), α1 (7.00-9.76 Hz), α2 (8.20-10.15 Hz), spindle (10.93–12.89 Hz), and β (19.96-20Hz). Additionally, our introduction of designated EMG and designated EOG brought forth clearer sleep-stage distinctions.
The granularity and precision of these parameters position them as potent tools in the evolution of AI-driven sleep-stage scoring algorithms (28). The integral nature of SleepGraph, especially with its automatic scoring system, holds promising potential as a foundational tool in the burgeoning domain of AI-enhanced sleep diagnostics.
While our research provides valuable insights, it is necessary to acknowledge the limitations inherent in our approach. The most significant limitation stems from our small sample size (n = 8), which inherently restricts the generalizability of the findings. However, despite this limited sample, the agreement rate and κ value for subjective sleep scores were quite promising. Generally, subjective sleep scores often show inconsistency, both among different scorers and within repeated evaluations by the same scorer.
Notably, the objective parameters we measured, such as EEG, EOG, and EMG – which aren't prone to subjective alteration – presented significant correlations between SleepGraph and Polymate (p < 0.001). This suggests that even with a small sample size, our devices' consistency and precision are noteworthy.
Furthermore, the practical application of SleepGraph extends beyond the scope of this paper. It's been deployed in clinical trials, evaluating conditions such as RBD, Parkinson’s disease (32), dementia with Lewy bodies, other α-synucleinopathies, bruxism, and clenching. Additionally, it's been used in foundational research to understand the aging effects on EEG slow waves. While these studies are yet to be published, their usage indicates that SleepGraph is steadily gaining traction in both clinical and research contexts.
In conclusion, the SleepGraph, a 2-channel portable telemetry PSG system, displayed an impressive congruence with the standard laboratory PSG system, Polymate. This robust alignment is underscored by the nearly identical sleep-stage dependent EEG, EMG, and EOG components exhibited by both devices. Beyond mere agreement in sleep-staging, SleepGraph stands out due to its ability to generate objective parameters from EEG, EOG, and EMG readings. These objective metrics pave the way for enhanced precision in automated sleep scoring and hold promise for the refinement of AI-driven sleep diagnostic algorithms.