Autism spectrum disorder (ASD) is a highly genetically heterogeneous1, lifelong neurodevelopmental disorder. Whereas ASD affects only 1 out of 36 children in the United States2, 1 of 5 infant siblings of children with ASD later received an ASD diagnosis3. Early detection and intervention can improve deficits of ASD and long-term outcomes. The Infant Brain Imaging Study Network conducted magnetic resonance imaging (MRI) studies of infants at high risk for ASD and showed that brain imaging markers from the first year of life predicted later ASD diagnosis among infant siblings4,5. Moreover, their recent infant sibling study recognized aberrant visual brain development as a potential MRI marker of ASD6, which prompted this review paper. To improve the accuracy of early detection of ASD, stratify ASD patients, and identify a subtype with aberrant visual brain development4, I suggest the addition of timing rapid eye movements (REMs) in sleep. REMs under closed eyelids can be timed from video recordings in functional MRI (fMRI) studies (see Supplementary Information).
The revised diagnostic criteria (Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition) for autism include atypical sensory experience as a core diagnostic feature, along with social deficits, repetitive behaviors, resistance to changes, and restricted interests7. Atypical sensory experiences affect up to 90% of autistic individuals, in all sensory modalities, and potentially serve as early diagnostic markers; the neural underpinnings of atypical sensory experiences and processing have been theorized8. Employing REMs in sleep as a probe in a functional brain imaging study (“REM-probe” study) may help elucidate the neural underpinnings of the atypical visual sensory processing seen in autism.
Because brain activation time-locked to REMs in sleep is characteristically widespread9, to avoid “cherry picking,” the focus in this paper is on peak activation (P < 0.00005, corrected for multiple comparisons) (Fig. 1). Activation peaks were clearly localized to the primary visual cortex (V1), extrastriate cortex (V2), thalamic reticular nucleus (TRN), claustrum, superior temporal gyrus (STG), retrosplenial cortex in the right hemisphere (RSC-Rt), areas encompassing the cholinergic basal nucleus, and bed nucleus of the stria terminalis (BNST)9. The most robust REM-locked activation in the whole brain occurred bilaterally in the V1. Event-related fMRI (with the events being REMs identified from video recordings) captures consecutive snapshots of fundamental mind/brain events, i.e., hierarchical generation of visual percepts in the dreaming brain9, starting in the V1, then multisensory integration accomplished in the TRN10, claustrum11–13 and STG14,15, and “valence surveillance” occurring in the BNST16. As perception is a temporal process17, time-series analysis of REM-locked changes over time may be suitable to investigate perception.
Our REM-probe fMRI study exhibited superb performance in localizing REM-locked peak activities9,18 and deactivation19, as discussed below. All methods were carried out in accordance with relevant guidelines and regulations9,19. All 24 individual studies, even a short (6-minute) individual study with 43 REMs, showed similar regional patterns9. Statistical efficiency at the individual level is essential for assessing the probability of a given individual developing ASD.
Crucially, fMRI REM-probe studies capture snapshots of fundamental events in an ideal state. As much of the external sensory input to the brain is blocked in REM sleep20, albeit incompletely (for a review, see 18), the brain is unperturbed by environmental sensory signals. For studies of perception during wakefulness, environmental sensory signals irrelevant to the aim of investigation may confound the interpretation of findings9,18. The unique strength of REM-probe studies is that they are performed during REM sleep, an ideal state to study intrinsic organization of large-scale brain networks (including the default mode network [DMN]19) because the brain is relatively isolated from the environment 18.
The strengths of REM-probe studies of infants are manifold. REMs serve as a natural, task-free probe useful for studying infants or animals, who cannot comply with conventional visual activation tasks9,18. As stated by Seth and Bayne (2022) “Infants and animals are…unable to produce introspective reports [of conscious experience]” and “there is a pressing need to identify ‘markers’ of [their] consciousness”21 Furthermore, the preponderance of REM sleep in infants22 makes them ideal participants for REM-probe studies.
REMs in sleep are an ideal, natural, task-free probe for studying both the ontogenetic and phylogenetic development of consciousness, in particular visual perception18. Notably, REM sleep has a special role in neurodevelopment. REM sleep deprivation in immature rats supports a role of REM sleep in neurodevelopment of the visual system23. Extrapolation from observations of premature newborns suggested a marked preponderance of REM sleep in the last trimester of pregnancy22,24. Indeed, ultrasound imaging visualized fetal eye movements and showed that REMs begin in utero at 23 weeks and become more frequent in the third trimester of pregnancy 25. Hobson proposed that REM sleep has a foundational role in the development and maintenance of waking consciousness 24. REMs in sleep occur both in mammals and birds24, and even in Australian dragons26. REM-locked neural changes can be studied in other species than humans, including Australian dragons26. REMs in sleep may allow straightforward comparison of, and translation between, animal and human data. Additionally, differences in the level of attention confound comparison across the lifespan and between individuals with and without autism. A REM-probe study controls the level of attention across the lifespan and across typical and atypical developments. Employing REMs in sleep as a natural probe not only aids the design and establishment of a standardized set of sensory paradigms suitable for cross-species comparison8 but also obviates concern regarding the confounding effect of differences in the level of attention across both the lifespan and individuals with or without autism. Thus, REMs in sleep are a useful probe for studying lifelong neurodevelopmental disorder, like autism, across lifespan.
We showed the feasibility of REM-probe fMRI studies in neonates, as discussed below. fMRI and other brain imaging modalities have been used to study the development of large-scale brain networks in fetus27. REM-probe fMRI studies are now feasible even in the fetus. REMs were timed by identifying the lens center and eye center on serial fMRI images, and fMRI study of the neural correlates of REMs was performed in the human fetus28. Longitudinal REM-probe studies from before birth are now possible in principle. This is relevant to autism research, because “ASD begins in prenatal life” and “most ASD risk genes are expressed prenatally in many ASD-relevant brain regions”29.
REMs are the hallmark of REM sleep, but a substantial proportion of REM sleep does not contain REMs; such sleep is called “tonic REM sleep,” whereas sleep containing bursts of REMs is known as “phasic REM sleep.” We previously analyzed phasic REM sleep9,19; we studied the neural correlates of REMs in sleep but not of REM sleep. Instead of comparing non-REM sleep and REM sleep, or tonic and phasic REM sleep using a block design, we used event-related functional MRI to study precise temporal changes time-locked to REMs9. Notably, REM-locked brain activity is distinct from baseline activity during phasic REM sleep18,19. Taking account of this important subtlety will improve understanding of the utility of REMs in sleep as a probe, given the high capacity to localize REM-locked peak activities. Event-related fMRI for studying the neural correlates of REMs in sleep is discussed below, along with the statistical efficiency and high localizing capacity of the REM-probe study.
To sample visual sensory data and enable waking visual perception, the eyes move quickly from one fixation point to the next. These scanning, ballistic, REMs are called saccades. The brain automatically chooses points to scan that provide salient or precise information, such as another’s eyes and mouth. Crucially, it was demonstrated with both position emission tomography (PET)30 and fMRI9, that REMs in sleep are saccades. In what is widely regarded as the standard textbook of eye movements, REMs in sleep are considered saccades31. Recently, an animal study demonstrated coupling of head direction system with REMs, both in sleeping and awake mice, and confirmed that REMs in sleep are saccades that scan the virtual world of dreams, akin to saccadic eye movements in wakefulness that scan the environment during exploration32. Saccades actively sample the world to infer “the signal source (world)”33. Saccades are a prime example of “active inference” (i.e., predictive coding extended to include action)18,34.
Counterintuitively, the brain is essentially a closed system and consciousness is an internal phenomenon even when awake35–38. The brain has no direct access to the source of sensory signals 33 and is essentially self-evidencing39. In line with this predominant view, nearly 95% of input to the lateral geniculate nucleus is non-retinal40. The brain, enclosed inside the skull, samples sensory signals from the outside world and makes probabilistic knowledge-driven inferences to ascertain their source33,41−43. This view builds on the insight of von Helmholtz who, in the 19th century, posited that perception is an internal, inferential, and constructive process44. Visual percepts are generated in essentially the same way when awake and when dreaming9,30,45,46; this generation process is constrained only by sensory signals from the environment when awake43. This view is in line with the “controlled hallucination” view of waking perception47.
The brain does not perceive the world as it is. Although we believe that what we perceive during wakefulness is the world “as is” (dubbed “naïve realism”), what we actually perceive is a model of the world generated by the brain from external sensory signals collected from the world35. For example, the color we perceive is not “out there” in the real world; instead “It is a property of the internal model … created by your brain”35. The only things “out there” are in fact the reflectance properties of objects and electromagnetic waves of varying wavelengths18. The real world is colorless, but the brain paints the internal model with colors to encode differences in wavelengths of the visual spectrum (400–700 nm)35. The basis of consciousness is generation of the world-model, with the self-model at its center, during both wakefulness48 and dreaming43,49,50. This enables adaptive interactions of the self with the world in simulations18,35,50.
We do not perceive the body merely because it is there but rather because the brain is able to generate a model of the bodily self from interoceptive signals arising from the body; the phenomenal self is not fixed; instead, it is an ongoing process51. Cotard delusion indicates that self-modeling is a brain-based process that can go awry52. Examples of aberrant bodily self-model generation may include body dysmorphic disorder and anorexia nervosa18. My patient with Cotard syndrome (unpublished case) declared repeatedly, “My body feels not real. My skin feels fake.” Multimodal sensory integration is essential for generating a bodily self-model during wakefulness48,53,54 and apparently also in dreaming18. REM-locked multisensory integration suggests temporally precise co-occurrence of REMs in sleep and generation of a bodily self-model while dreaming. The RSC in the left hemisphere (RSC-Lt) plays a key role in the development of Cotard syndrome and other “delusional misinterpretations”55. We proposed that the RSC plays a role in generating the spatial bodily self-model during dreaming and wakefulness18. The RSC’s role in spatial navigation of the bodily self through the environment is well established56. The RSC is “a hub between visual processing streams and the medial temporal lobe regions” and is “at the top of a cortical visual perception hierarchy”56. As we put it previously18, “The world-model, with the self-model at its center, can be viewed as a highly ‘advanced, user-friendly interface design,’ and ‘a wonderfully efficient control device,’ ‘like a total flight simulator’ (but whose ‘virtuality is hidden’) in which our bodily self manipulates the virtual environment in wakefulness35, and, of course, in dreaming.”
We proposed that saccadic rapid eye movements in sleep and wakefulness generate, as well as scan, what we see when we are either dreaming or awake9,18,43,57. In this conception, to generate a model of the world “hidden” behind sensory data, the brain infers the source (i.e., the world) from visual sensory data that are actively sampled, and this inference machine runs in essentially the same way when using endogenous visual sensory data, i.e., when dreaming9,30,45,46. REM-locked generation of visual percepts relates to another insight of von Helmholtz, namely “efference copy”, i.e., a copy of motor commands sent to sensory processing regions of the brain44. This notion of efference copy was subsequently expanded to “corollary discharge”58 and then to the “activation synthesis hypothesis”59. According to this hypothesis, the brain synthesizes visual imagery appropriate to the corollary discharge of eye movement commands. Do REMs in sleep scan dream imagery30,57,60,61 or generate it57,59,62,63? It initially appeared that these functions are contradictory, but we proposed that REMs in sleep may in fact perform both: “… in active inference, eye movements are both cause and consequence of perception”43. fMRI findings regarding video-timed REMs in sleep9 support the premise that REMs generate and scan dream visual imagery, in accordance with the active inference43. Consciousness is a constructive, active, and inferential process both when awake and when dreaming18,43. The brain actively and selectively samples sensory signals to generate and update probabilistic representations of the world64, and of the virtual world when dreaming18,43, using multimodal hierarchical predictive coding17.
Predictive coding accounts for the atypical perceptions that characterize autism, showing its explanatory power42,65−67. In particular, the aberrant precision account of autism, where greater weight (i.e., precision) is assigned to incoming sensory signals (prediction errors) relative to existing predictions (prior beliefs), provides parsimonious explanations of the empirical findings17,65,66. fMRI studies revealed greater extrastriate activation68–71 and less prefrontal activation68,71 during processing of exogenous visual stimuli in autistic people compared with typical controls. These fMRI studies employing exogenous visual information led to the testable hypothesis that REM-locked processing of endogenous visual information recruits the extrastriate visual cortex to a greater extent in individuals with ASD. REM-probe studies in infant siblings of children with ASD may advance our knowledge of typical and atypical development of visual perception by exploiting the explanatory power of predictive coding and active (Bayesian) inference18, a dominant theory of the generation of visual perception and consciousness in general.
How does REM sleep latency change across the life span? This question will arise when designing a REM-probe study across the lifespan. REM sleep latency increases linearly as newborns age, reaches its maximum (around 200 minutes) at age 6–7 years72,73, and then gradually decreases to around 100 minutes by age 18 years73,74 and 60 minutes by age 80 years74. The long REM sleep latency of children will probably require two consecutive overnight assessments for REM-probe studies, as in our REM-probe studies in adults9. In 6 out of 11 adult participants in our study, a sufficient number of REMs occurred only on the second night. Presumably, head restraint, which is required for MRI studies, suppresses REM sleep, and REM sleep deprivation during the first night likely builds “REM pressure” for the second 9.
A group that used ultrasound technology to develop functional ultrasound (fUS)75,76, a new, fully fledged brain imaging technique with far superior spatio-temporal resolution than fMRI, subsequently used fUS to study the time course of cerebral blood volume (CBV) changes during phasic REM sleep in rats77,78. fUS studies of REM sleep in rats revealed interesting similarities with REM-locked findings in humans. They found massive brain-wide CBV spikes lasting 5–30 s during phasic REM sleep and called them “vascular surges”78. Notably, the time course of the widespread cortical “vascular surges” during phasic REM sleep in rats77,78 is similar to that of the widespread brain activation time-locked to REMs occurring during phasic REM sleep in humans9. Thus, the vascular surges seen in rats during phasic REM sleep may also be time-locked to REMs. A close-up camera to time REMs during fUS was then proposed77. As an additional similarity, sensory processing was coupled with deactivation of the RSC both in humans9 and in mice79.
fUS has high sensitivity and temporal resolution of 10–100 ms75. fUS enabled assessment of the top-down propagation of signals in hierarchical processing after only a single trial of visual tasks80. Studying hierarchical REM-locked processing of endogenous visual signals 9 using fUS will expand our knowledge regarding the generation of dreaming and waking consciousness, given the high resolution. Notably, this group performed trans-fontanel (where the skull bone that impedes fUS is absent) fUS imaging in human neonates during sleep81,82. As the portability of fUS enables bedside monitoring of neonates, it has great potential in ASD infant sibling studies.
To motivate REM-probe fMRI and fUS studies of the infant siblings of children with ASD (and typically developing controls), I will analyze the REM-locked brain events, and explore their implications for autism research, in terms of REM-locked processing of visual information, multisensory motor binding, cholinergic activation, cerebral vasodilation, activation of the RSC-Rt and deactivation of the RSC-Lt, activation of the BNST, and possible deactivation of the DMN. REM-locked peak activation sites and neural processing overlap extensively with those reported to be atypical in autism.