The present investigation sheds new light on the dynamics of brain activity during sleep onset. We discovered that the earliest manifestations of drowsiness during SOP detected by eye closure dynamic, coincide with an increase in the power spectral density of BOLD oscillations at a frequency of 0.05 Hz. These changes became evident upon participants entering the 'likely drowsy' state of the ‘drowsigram’, with the most marked differences between awake and sleep states. Crucially, our findings underscore a specific spatial distribution of such BOLD oscillations within somatomotor and visual networks.
The current findings are consistent with previous studies indicating that the BOLD signal may fluctuate during rest periods, potentially indicative of sleep onset (Fukunaga et al., 2006; Horovitz et al., 2009; Stern et al.,2011; Jahnke et al., 2012; Song et al., 2022). The work of Fukunaga and colleagues (2006) was one of the first to demonstrate that the evolution of the BOLD signal during extended rest periods, leads to large amplitude variations that authors associated with sleep events. While most of these studies considered sleep onset within an AASM framework, our study offers valuable insights by introducing a more temporally refined classification. Henceforth, we identified an increase in PSD of low-frequency BOLD oscillations during the transition from wake to sleep with the emergence of changes as early as the 'likely drowsy' state. This is a significant discovery as the ‘likely drowsy’ state of the ‘drowsigram’ represents one of the earliest manifestations of sleep-related vigilance fluctuations, often encountered even during activities like driving (Abe, 2023). We were thus able to showcase changes in the BOLD signal well before conventional indicators of sleep onset. This corroborates the well-known lack of clear-cut delineation of N1, which is the early known sleep stage, partly because of the misclassification of drowsiness during SOP (Carskadon et al., 2011; Andrillon et al., 2024).
Subsequently, the increase in PSD of low-frequency BOLD oscillations across the ‘drowsigram’, peaked at 0.05 Hz which is at first a valuable argument in favor of the relevance of the rs-fMRI technique to improve knowledge on the SOP. Indeed, in sleep studies leveraging EEG and AASM framework, the lowest frequency of oscillation of brain electrical activity (Delta activities: 0.5 Hz; Ogilvie, 2001, Carskadon et al., 2011) is ten times higher than the frequency (0.05 Hz) of BOLD oscillations in the present study. In addition, reports have interestingly argued that the oscillation of brain activity at a frequency of 0.05 Hz, is a dynamic coupling between the electrical activity of neurons and the hemodynamic responses in the brain (Fultz et al., 2019). These authors observed a temporal pattern wherein neural activities precede hemodynamic oscillations ultimately resulting in changes in cerebrospinal flow, suggesting that BOLD oscillations at 0.05 Hz during sleep would originate from neural activity (Fultz et al., 2019). This hypothesis particularly corroborates our results, as we meticulously isolate the BOLD signal, and thus neurovascular coupling, from the physiological confounds in data extraction, thereby amplifying the neural origins of the outputs.
Notably, BOLD oscillations at 0.05 Hz closely match the cyclic alternating patterns (CAPs) identified during sleep as an indicator of sleep instability caused by internal and external factors (Terzano et al., 1986; Parrino et al., 2012). Interestingly, our participants were sleep-deprived and were specifically instructed to avoid falling asleep. Thus, sleep onset under such conditions, especially in the disruptive environment of a noisy MRI, would likely be precarious favoring the occurrence of CAPs (Terzano et al., 1986; Parino et al., 2012). Therefore, it could be assumed that CAPs were among the physiological modifications underlying this low-frequency BOLD oscillation during SOP.
The increase in PSD0.05 during SOP was specific to some brain regions. In the ROI-wise analysis, the power increased progressively throughout the ‘drowsigram’ first in the postcentral and cerebellum (awake vs likely drowsy), then on occipital frontal, and temporal regions (likely drowsy vs drowsy), and finally in parietal regions (sleep vs drowsy). In addition, this progressive increase in PSD0.05 across the ‘drowsigram’ states, was replicated in the visual and somatomotor functional networks. This latter observation supports the notion that the BOLD oscillation at 0.05 Hz may have a sensory origin. This is also supported by prior studies, which similarly suggest a sensory explanation for other low-frequency BOLD oscillations observed during sleep (Song et al., 2022). Henceforth, we hypothesize that the BOLD oscillation at 0.05 Hz would be the manifestation of the reduction of sensory processing to favor sleep, particularly during moments with high sleep pressure. This sensory hypothesis gains further traction when considered alongside rs-fMRI studies that include sleep graphoelements like KCs. Indeed, reports have indicated that when a KC is triggered by external stimuli, it results in an augmented BOLD signal within primary sensorimotor cortical areas (Riedner et al., 2011; Stern et al.,2011; Jahnke et al., 2012). Interestingly, KCs are also acknowledged as a crucial component in the sleep continuum, bridging the gap between vertex waves in light sleep and the slow wave cycle in deep sleep (Halasz et al., 2016). Therefore, it wouldn't be surprising if KCs share the same neural activity background with the SOP as both play crucial roles in the downward progression of the sleep continuum, despite the latter occurring very early.
During SOP, it's noteworthy we found an absence of the BOLD signal oscillations at 0.05 Hz in the default mode network (DMN) or attention networks. This finding might seem odd, considering numerous studies have highlighted significant changes within the DMN during sleep (Horowitz et al., 2009; Spoormaker et al., 2012; Tagliazucchi and van Someren 2017). However, the absence of DMN involvement could be attributed to several factors. Firstly, it could suggest that BOLD oscillations at 0.05 Hz might be specific to sensory networks, thereby reinforcing the validity of the sensory hypothesis to explain this variation during SOP. Secondly, since we are examining SOP and the earliest stages of sleep, networks such as the DMN and attention networks may only show changes in oscillations at 0.05 Hz once sleep is fully established. This observation is a strong argument in favor of using PSD0.05 to delineate SOP.
The main limitation of the present study would be the light sleep deprivation applied the night before the experiments. This could have led to the process of falling asleep not following its usual course. However, since limited sleep duration in the MRI environment was detrimental in previous studies for a comprehensive analysis of BOLD changes (Stevner et al.,2019), we applied light sleep deprivation to ensure that a significant amount of sleep was collected. Furthermore, people spent approximately 20% of the duration of this experiment asleep which is similar to previous studies of sleep in an MRI environment (Tagliazzuchi et al., 2014; Joliot et al., 2024).
In conclusion, this work offers valuable insights into brain activity during the critical period of sleep onset detected by the innovative ‘drowsigram’. We discovered SOP-related oscillations in the BOLD signal at 0.05 Hz which reflects neurovascular coupling with a neural origin as outlined by previous research. The most important contribution of this research lies in linking the BOLD signal to an innovative and precise temporal classification of SOP, achieved through the pioneering use of the eye closure dynamic in sleep-related neuroimaging studies. This breakthrough supports using ocular activity via camera surveillance to accurately detect drowsiness during SOP, especially in MRI environments where implementing techniques like EEG can be challenging. Moreover, the association between the BOLD signal and the ‘drowsigram’, carries significant implications for the field of rs-fMRI. Indeed, it underscores the possibility of underestimating the known confounding effect of drowsiness when relying on current sleep scoring systems. Future investigations should build upon the ‘drowsigram’ framework to delve deeper into the relationship between BOLD signal and drowsiness, refine existing correction methods, and enhance the rs-fMRI signal quality.