Fetal sleep: a cross-species review of physiology, measurement, and classification

Abstract

Study Objectives

Fetal sleep is a vital yet underexplored aspect of prenatal neurodevelopment. Its cyclic organization reflects the maturation of central neural circuits, and disturbances in these patterns may offer some of the earliest detectable signs of neurological compromise. This is the first review to integrate more than seven decades of research into a unified, cross-species synthesis of fetal sleep. We examine: (1) Physiology and Ontogeny—comparing human fetuses with animal models; and (2) Methodological Evolution—transitioning from invasive neurophysiology to non-invasive monitoring and deep learning frameworks.

Methods

A structured narrative synthesis was guided by a systematic literature search across four databases (PubMed, Scopus, IEEE Xplore, and Google Scholar). From 2925 identified records, 169 studies involving fetal sleep-related physiology, sleep-state classification, or signal-based monitoring were included in this review.

Results

Across the 169 studies, fetal sleep states become clearly observable as the brain matures. In fetal sheep and baboons, organized cycling between active and quiet sleep emerges at approximately 80%–90% gestation. In humans, this differentiation occurs later, around 95% gestation, with full maturation reached near term. Despite extensive animal research, no unified, clinically validated framework exists for defining fetal sleep states, limiting translation into routine obstetric practice.

Conclusions

By integrating evidence across species, methodologies, and clinical contexts, this review provides the scientific foundation for developing objective, multimodal, and non-invasive fetal sleep monitoring technologies—tools that may ultimately support earlier detection of neurological compromise and guide timely prenatal intervention.

Introduction

Sleep plays a crucial role in brain development, synaptic plasticity, and metabolic regulation [1–3]. In neonates and infants, consolidated sleep–wake cycles have been shown to support neurodevelopmental milestones and long-term cognitive outcomes [1, 4–6]. However, the nature and function of sleep before birth—during fetal life—remain poorly understood. This lack of understanding poses a significant gap in prenatal care, as fetal sleep may reflect underlying brain maturation and help identify early signs of neurological compromise or neurodevelopmental disorders. Developing objective and non-invasive fetal sleep monitoring tools could empower clinicians to assess neurodevelopmental trajectories, detect early signs of complications such as antepartum hypoxia or fetal growth restriction, and implement timely interventions to improve perinatal outcomes.

While the importance of sleep for postnatal development is well established, systematic studies of sleep before birth only began in the 1950s and 1960s. Early research during this period primarily focused on fetal heart rate patterns and behavioral states, laying the groundwork for later investigations. In 1967, studies focusing on the relationship of intrauterine fetal activity to maternal sleep [7] and evidence of fetal-sleep cycles [8] examined fetal movements in relation to maternal sleep, suggesting that distinct sleep cycles may exist in utero. The 1970s marked a significant shift, as technological advancements allowed for more precise monitoring of fetal brain activity in large animal models such as sheep and calves. A pivotal advance came in 1971, when Ruckebusch systematically recorded alternating high-voltage (HV) slow activity and low-voltage fast activity in fetal lambs and calves, associating HV slow activity with NREM sleep and low-voltage fast activity with REM sleep or alert wakefulness [9]. His work laid the foundation for interpreting fetal electroencephalogram (EEG) activity in relation to behavioral states. Building upon this, a 1974 study confirmed the presence of EEG-based fetal sleep stages and linked them to cardiovascular regulation in fetal sheep [10]. Further validation came in 1977, as alternating HV slow activity and low-voltage fast activity states were again observed in fetal lambs, reinforcing their similarity to neonatal and adult sleep patterns [11]. Finally, a 1979 study on human fetuses identified short-term cyclic patterns of fetal activity [12], supporting the hypothesis that early sleep–wake cycles emerge before birth.

A major conceptual breakthrough occurred in 1982, when Nijhuis et al. introduced the fetal behavioral states (FBS) framework [13]. This classification system defined distinct fetal states based on physiological and behavioral markers, drawing parallels between fetal and neonatal behavioral state patterns [13]. Nijhuis further expanded the framework in a 1986 study [14], providing a more detailed characterization of these states and their relevance to neurodevelopment and clinical assessments [14]. These findings confirmed that human fetuses exhibit four behavioral states (1F to 4F), analogous to neonatal states [13, 14]. States 1F and 2F correspond to quiet (NREM-like) and active (REM-like) sleep, respectively, while 3F is characterized by continuous eye movements without body movement, and 4F by vigorous body activity with unstable heart rate [14]. Fetal breathing, heart rate, and movement patterns were shown to serve as key indicators of neurodevelopment [14]. These insights reinforced the potential of FBS as a clinical tool for assessing fetal brain function and detecting potential developmental abnormalities [14].

Despite extensive behavioral classifications such as the FBS framework [13], the presence of true wakefulness in the fetus remains controversial. In sheep, baboons, and humans alike, episodes of increased activity or arousal-like features often lack the sustained neural and behavioral markers associated with postnatal wakefulness [15–17]. Some have suggested that the apparent wakefulness state reflects a transitional phase between sleep states, similar to indeterminate sleep in the newborn [1], rather than true wakefulness [15].

These foundational studies laid the groundwork for modern research, which continues to refine our understanding of fetal sleep across different species. However, despite eight decades of progress, there remains no comprehensive literature review that provides a structured synthesis of findings on fetal sleep physiology, measurement, classification, and disruption under pathological conditions. To address this gap, we synthesize and structure the state of the art in fetal sleep research. Specifically, we address the following key research questions:

• What are the species-specific similarities and differences in fetal sleep ontogeny and physiology across humans, sheep, and baboons?

• How do different measurement modalities—ranging from invasive techniques in animal models to non-invasive approaches in human studies—compare in monitoring fetal sleep?

• What methods have been developed to classify fetal sleep states, including rule-based and deep learning approaches, and how do they incorporate multimodal signal integration?

• What abnormal intrauterine conditions—particularly hypoxia, fetal growth restriction, and intrauterine inflammation—are known to disrupt fetal sleep–wake cycling, and how are these disruptions reflected in fetal physiological signals (e.g., EEG, electrocardiogram (ECG), heart rate variability (HRV), and movements)?

By bridging physiology, engineering, and clinical relevance, this review provides a foundation for future work in fetal neurodevelopment and the advancement of objective, fetal sleep-state monitoring tools.

Search Strategy and Study Selection

This review is not intended as a full systematic review or meta-analysis; rather, it is a structured narrative synthesis guided by a systematic literature search.

To capture the breadth of fetal sleep research, we constructed search strings around three thematic domains: (1) fetal population and developmental context, (2) sleep states and behavioral classification, and (3) signal-based monitoring and classification modalities. The search was applied across four databases (PubMed, Scopus, IEEE Xplore, and Google Scholar) using Boolean operators (AND, OR), without restrictions on publication date. The search was conducted on June 16, 2024.

Studies were included if they involved fetal or in utero sleep-related physiology, sleep-state classification, or signal-based monitoring using modalities such as EEG, ECG, or electromyography (EMG). Both human and animal studies were considered. Studies were excluded if they:

• Focused exclusively on maternal, neonatal, or postnatal conditions

• Studies unrelated to fetal sleep states, or those lacking systematic behavioural (e.g., ultrasound-based movement/eye movement, heart rate patterns) or physiological (e.g., EEG, ECG, EMG) assessment, were excluded.

• Lacked sufficient methodological detail

• Were not written in English

Figure 1 presents the full PRISMA flow diagram.

Figure 1.

PRISMA flow diagram showing article selection with exclusion counts at each step.

More details of PRISMA flow diagram including exclusion counts by category, is presented in the Supplementary Material.

Evolution of fetal sleep research across decades

As illustrated in Figure 2, decadal trends across study design, species, signal modalities, and analytical approaches show a clear methodological evolution in fetal sleep research. Early studies (1960s–1970s) were predominantly observational and human-based, relying on a limited set of classical electrophysiological and ultrasound techniques. The 1980s–2000s saw a substantial rise in experimental studies—largely in sheep—accompanied by broader multimodal electrophysiological recordings. In the 2010s–2020s, the field shifted toward increased longitudinal and cross-sectional work, wider species representation, diversified non-invasive monitoring, and the rapid adoption of signal-processing and machine-learning methods. These patterns reflect a transition from early descriptive investigations to increasingly data-rich and computationally advanced research paradigms. Numerical counts for all categories are provided in Supplementary Tables S1, S2, S3, and S4. To facilitate interpretation of these trends, we structure the literature using four complementary classification dimensions:

Figure 2.

Evolution of fetal sleep research across decades: (A) study designs, (B) species studied, (C) signal modalities, and (D) methodological approaches. Detailed numerical counts are provided in the supplementary materials (Tables S1–S4).

(A) Study design—Observational: descriptive recordings without intervention; Experimental: active manipulation or controlled perturbation; Longitudinal: repeated measurements across time; Cross-sectional: single-time recordings.

(B) Species—Human, Sheep, Baboon, and other mammals.

(C) Signal modality—EEG/electrocorticography (ECoG), electrooculogram (EOG), Doppler-based fetal heart rate (cardiotocography (CTG)/1D Doppler), fetal magnetoencephalography (FMEG), fetal ECG (FECG), EMG, and fetal magnetocardiography (FMCG).

(D) Methodological approach—Observation/Manual: visual or rule-based staging; Statistical: classical inferential analyses; Signal processing: features derived from signal power, variability, or complexity; Machine learning/AI: data-driven or deep-learning models for automated classification.

Background

The significance of fetal sleep

Fetal sleep is essential for neurodevelopment, supporting the maturation of both the central (CNS) and autonomic nervous systems (ANS) [5, 18–23]. Cycling through distinct FBS promotes synaptogenesis, brain plasticity, and neuronal differentiation, contributing to postnatal cognitive and behavioral development [5, 24]. Fetal sleep, unlike postnatal sleep, cannot be characterized through behavioral observation [5, 25, 26] and is therefore defined primarily by physiological markers such as HRV, body movements, and EEG (in animal models).

The emergence of distinct FBS such as quiet and active sleep during the third trimester has been interpreted as a sign of increasing functional brain complexity, driven by coordinated neural activity across developing circuits [5, 27, 28]. This maturation is paralleled by the progressive development of the ANS, which plays a key role in supporting physiological regulation during fetal life [19, 21, 22]. In particular, the vagus nerve—a central component of the parasympathetic system—has been implicated in vital processes such as anti-inflammatory signaling and metabolic regulation across fetal, perinatal, and postnatal periods [29, 30]. From around 25 weeks of gestation, increasing vagal tone and myelination have been reported [22, 31, 32], potentially contributing to the regulation of heart rate and state-dependent behaviors. As such, the emergence of distinguishable sleep states in the third trimester—detectable via patterns of fetal heart rate, eye movements, and body activity [13]—may reflect both neural and autonomic maturation.

REM-like sleep is characterized by spontaneous fetal movements, including fetal breathing movements (FBM), which are thought to play a critical role in preparing vital systems for postnatal life [33]. When fetal movements are pharmacologically suppressed—such as through anesthesia or neuromuscular blockade—there is a marked reduction in oxygen consumption [34], suggesting that fetal activity itself significantly contributes to metabolic demand. While this observation does not directly implicate REM sleep in energy conservation, it highlights the physiological cost of fetal activity and the potential adaptive role of sleep states in regulating energy expenditure.

Moreover, FBM—often observed during REM-like states—are believed to contribute to lung growth and maturation. Disruption of these movements, whether experimentally or in pathological conditions such as prolonged oligohydramnios, can result in pulmonary hypoplasia [35–40]. Although these disruptions are not exclusive to REM sleep, the strong association between FBM and REM-like states suggests that fetal sleep behavior may indirectly support pulmonary development. Altogether, these findings imply that FBS contribute not only to neurodevelopment, but also play a broader role in maintaining metabolic balance and promoting organ system maturation [28].

Current research status and gaps

Advances in fetal monitoring techniques have enabled detailed characterization of sleep states using EEG, integrating ECG, and HRV analysis [41, 42]. Non-invasive methods such as FMEG complement traditional measures by providing insights into fetal cortical activity [43]. Current research has primarily focused on automated classification of fetal sleep states based on heart rate and movement patterns [44, 45]. Studies have also explored EEG-based analysis of REM and NREM sleep across species [26, 46]. Additionally, researchers have investigated neural and autonomic markers of fetal brain maturation through spectral EEG and HRV analysis [41, 47], and recent studies have also evaluated EEG spectral power and sleep state cycling to assess maturation after hypoxia–ischaemia in fetal sheep [48].

Despite these advances, current fetal sleep research remains constrained by the limited availability of direct neural recordings and the lack of harmonized protocols across studies.

Physiological Characteristics of Fetal Sleep

Fetal sleep cycle in different species

Since direct recordings of fetal EEG and other neural activity are not feasible in humans, many fetal sleep studies rely on animal models (particularly sheep and, to a lesser extent, nonhuman primates) to infer developmental physiology and staging rules [33, 49]. Understanding cross-species similarities and differences is therefore essential for interpreting findings and assessing translational relevance.

Comparative analyses across species provide insight into how fetal sleep develops and its evolutionary significance. Species-specific features of fetal sleep cycling are summarized below, highlighting convergent and divergent patterns relevant to translation.

Fetal sheep and baboons are widely used models for studying sleep state maturation due to their well-characterized sleep architecture and physiological similarities to humans [46, 49–54]. For example, fetal sheep exhibits a transition from disorganized to structured sleep patterns around 80% gestation, mirroring human fetal sleep development at approximately 32-36 weeks gestation [13]. To better compare the sleep states of different species, gestational length and birth weight are presented in greater detail, as shown in Table 1.

Table 1.

Comparison of gestational length and birth weight across species

Species (Fetus)	Gestation Length	Birth Weight (kg)
Human	36–41 weeks (testing 36–38; birth 39–41) [43] 37–41 weeks (259–294 days) [57]	3.02–3.80 [43]
Sheep	13–20 weeks (90–140 days) [52] 15–19 weeks (106–136 days) [50] 16–21 weeks (112–144 days) [24] 18–19 weeks (128–132 days) [119] 18–20 weeks (128–141 days) [155] 18–21 weeks (127–144 days) [10]	2.5–4.2 [155] 4.3–5.0 [78]
Baboon	16–22 weeks (112–144 days) [62] 20–21 weeks (143–148 days) [72] 20–22 weeks (143–153 days) [61] 20–22 weeks (144–152 days) [121]	0.462–0.9 [62] 0.63-0.71 [121]

Multiple rows per species reflect gestational length and birth weight values reported across different studies. Separate lines correspond to different literature sources.

Table 1 shows that sheep fetuses have a higher birth weight than human fetuses, while baboon fetuses are the lightest. Interestingly, this pattern does not follow the order of gestational length: human fetuses have the longest gestation, baboons fall in between, and sheep have the shortest. This dissociation between gestation length and birth weight suggests that longer gestation may not be solely for somatic growth. Instead, it may reflect species-specific neurodevelopmental priorities. In humans, for instance, the prolonged gestation supports a brain growth spurt that begins in mid-gestation and extends well into early postnatal life, enabling greater cortical and synaptic development [55]. This implies that cerebral complexity, rather than body weight, may better explain interspecies differences in gestation length—at least among medium-sized mammals. Moreover, human fetal birth weight is influenced by factors such as fetal sex [56], maternal weight [2], and gestational diabetes [57], while such influences have not been extensively studied in sheep or baboon models.

Comparison of sleep state development across species is summarized in Table 2. This table outlines key developmental milestones—ranging from the onset of physiological rhythmicity to the emergence and maturation of distinguishable sleep states—in humans, sheep, and baboons.

Table 2.

Comparison of sleep state differentiation across species

Sleep State Differentiation	Human Fetus	Sheep Fetus	Baboon Fetus
Initial Physiological Rhythmicity	80–90% of gestation: FHR, eye, and body movements cycle independently; chance overlaps lack the synchrony and stability required for true behavioral states [13].	79–83% of gestation: 35.1 ± 2.5% NREM, 52.7 ± 2.4% REM, 11.2 ± 1.7% Wake like activity were observed, marking the transition from disorganized to organized behavioral states [24].	78–86% of gestation: EEG power and coherence cycles (1 h) observed in fetal baboons, even in the absence of full behavioral state measures, may reflect early sleep-like rhythmicity [53]
REM/NREM Differentiation	90–95% of gestation: In most fetuses, heart rate, movement, and eye activity are only partially synchronized, preventing reliable state classification [13].	83–90% of gestation: 38.4 ± 1.6% NREM, 49.9 ± 1.7% REM, 11.8 ± 1.6% Wake like state [24]	82–87% of gestation: 20.9% NREM, 58.3% REM, 20.9% Transition [121]
Emergence of Stable Behavioral States	95% of gestation: NREM: 32% (range: 9–53.5%) REM: 42.5% (range: 23–64%) Wake-like state (combined 3F and 4F): 14.75% (range: 6.5–33%) No state identified: 11.5% (range: 3–53.5%) [13]	90–97% of gestation: 38.2 ± 1.8% NREM, 46.1 ± 2.0% REM, 15.1 ± 1.9% Wake like state [24]	82–87% of gestation: 36.8% NREM, 63.2% REM [156]
Full Maturation of Sleep Cycles	99% of gestation: NREM: 38% (range: 24.5–52.5%) REM: 42.5% (range: 22–73.5%) Wake-like state (combined 3F and 4F): 13.5% (range: 2.5–38%) No state identified: 5% (range: 0–26.5%) [13]	97–99% of gestation: 43.8 ± 2.5% NREM, 37.7 ± 2.8% REM, 18.3 ± 2.9% Wake like state [24]	73–90% of gestation: 48% NREM, 32% REM, 20% transition [62]

Abbreviations: FHR, fetal heart rate; EEG, electroencephalogram; REM, rapid eye movement sleep; NREM, non-REM sleep; 3F, quiet awake; 4F, active awake.

Among the three species, fetal sheep appear to exhibit organized sleep states earliest, with REM/NREM-like differentiation observable as early as 79% of gestation. Fetal baboons also show distinct REM- and NREM-like cycling by 82–87% of gestation, although data are limited to a narrow gestational window and lack a comprehensive developmental trajectory. In contrast, humans typically show consistent REM/NREM differentiation only after 90% of gestation, suggesting a relatively delayed maturation process.

These cross-species comparisons provide insight into sleep state ontogeny; however, differences in methodology, available data, and sampling windows, particularly in the fetal baboon limit direct comparisons. The table should therefore be interpreted as an approximate alignment across species rather than a definitive staging framework.

Comparison of sleep-like states across species

FBS are broadly classified into sleep-like states (REM and NREM), transitional or indeterminate states, and potential wakefulness. However, the nature and classification of fetal wakefulness remain a subject of debate. In this section, sleep patterns in fetal sheep, baboons, and humans are compared based on available research.

Sleep-like states

All three species exhibit two primary sleep-like states—Quiet sleep and Active sleep—along with transitional periods. Their electrophysiological and behavioral correlates are summarized in Table 3.

Table 3.

Comparison of sleep-like state definitions across species

State	Sheep	Baboon	Human
Quiet sleep (NREM / 1F)	HV, low-frequency EEG; absent REMs; variable nuchal EMG tone; intermittent fetal breathing [24, 157, 158].	Trace alternant pattern (bursts of HV EEG); reduced high-frequency activity; fetal breathing present but less frequent [61, 72].	Low fetal movement; stable FHR with low variability; absent EOG activity; analogous to NREM sleep [14, 159].
Active sleep (REM / 2F)	LV, high-frequency EEG; continuous REMs; absent sustained EMG tone, with phasic contractions; frequent fetal breathing [24, 158].	Increased high-frequency EEG; frequent fetal breathing, resembling human REM [72].	Frequent fetal movement; continuous EOG; variable FHR; analogous to REM sleep [14, 159].
Transition (TR/Indeterminate)	Mixed EEG features during REM–NREM shifts; spectral intermediates between HV slow activity and LV fast activity; 23% of recording time; excluded if 3 min [58, 59].	Graded transitions evident in EEG; sleep as a continuum; up to 60% time in indeterminate states [61, 62].	Indeterminate epochs with mismatched behavioral markers (FHR, movement, EOG); typically brief (3 min); 5–10% prevalence near term [13, 63, 64].

Abbreviations: FHR, fetal heart rate; EOG, electrooculogram; EMG, electromyogram; HV, high-voltage; LV, low-voltage.

In brief, Quiet sleep corresponds to HV/low-frequency EEG with absent eye movements and stable physiology, whereas Active sleep is characterized by low-voltage/high-frequency EEG, frequent eye and body movements, and variable autonomic activity. Transitional or indeterminate epochs show mixed features and can occupy a substantial fraction of fetal recordings. Representative traces from our fetal sheep dataset are shown in Figure 3, highlighting EEG/EMG differences between states.

Figure 3.

Representative physiological signals illustrating sleep states in fetal sheep [154]. The figure includes bilateral EEG (left [L] EEG and right [R] EEG), nuchal EMG obtained from electrodes implanted in the fetal neck muscles, and the SEF derived from the right EEG (R SEF), computed by estimating the windowed power spectral density and identifying the frequency below which 90% of total power is contained. All signals were collected from fetal sheep using chronic invasive instrumentation, including surgically implanted EEG and EMG electrodes, enabling continuous in utero monitoring of physiological and neural activity. Non-rapid eye movement (NREM) sleep is marked by HV, low-frequency EEG patterns, elevated EMG tone, and correspondingly low SEF values. In contrast, rapid eye movement (REM) sleep is characterized by low-voltage (LV), high-frequency EEG activity, EMG atonia, and elevated SEF values reflecting the shift toward faster cortical oscillations. Transitional periods (TR) represents an intermediate state between REM and NREM, capturing the dynamic shift from one state to the other. This state typically exhibits mixed EEG features that do not fully conform to either REM or NREM characteristics. The left-hemisphere SEF is omitted because both hemispheres show highly symmetric spectral profiles in fetal sheep.

Sleep transitions and indeterminate sleep

All three species—sheep, baboons, and humans—exhibit transitional or indeterminate fetal sleep states, reflecting the developmental complexity of sleep organization.

In fetal sheep, these ambiguous states—often termed intermediate sleep—are neither clearly REM nor NREM. Mellor (2005) linked them to immature brain regulation [15]. Quantitative EEG studies identified two spectral intermediates, falling between HV slow activity and low-voltage fast activity, as exemplified by the TR highlighted in black in Figure 3, accounting for 23% of recording time [58]. During late gestation, the intrinsic sleep cycle of fetal sheep lasts approximately 20–40 minutes, typically showing an almost 1:1 ratio between NREM and REM states [11, 24]. It is also worth noting that cycle duration varies with gestational age, reflecting ongoing maturation of sleep–wake regulation. Rao et al. (2009) defined indeterminate sleep as transitional or mismatched EEG/EOG periods, excluded if under 3 minutes [59]. Such states may reflect neural transitions key to sleep maturation [60].

Fetal baboons also show EEG evidence of graded transitions between quiet and active sleep. EEG-ratio analyses reveal sleep as a continuum, with up to 60% of time spent in indeterminate states [61, 62].

In humans, fetal sleep cycles last 70–90 minutes, with state transitions typically within 3 minutes [63, 64]. Indeterminate states, defined by mismatched behavioral markers, occur in approximately 5–10% of recordings near term [13], suggesting lower prevalence compared to baboons.

Fetal arousal and wakefulness

The definition of arousal and wakefulness in the fetus remains contentious. In fetal sheep, brief periods of activity characterized by low-voltage ECoG, EOG, and increased EMG activity have been interpreted as an aroused state [65–68]. Nevertheless, accumulating evidence suggests that such episodes may merely reflect transitional phases between sleep states, rather than true wakefulness [15]. Direct observations of unanaesthetized fetal sheep provide further evidence against the existence of true wakefulness in utero. Rigatto et al. [16] observed a fetal sheep through a Plexiglas window for 5000 hours and found no signs of wakefulness, such as eye opening or coordinated head movements. This suggests that fetal sheep remain in sleep-like states throughout gestation (promoted by placental and CNS derived hormones such as progesterone, allopregnanolone and adenosine) without experiencing a state that would be comparable to postnatal wakefulness.

In fetal baboons, studies indicate that hiccups and gross fetal movements do not necessarily induce sleep-state transitions, implying that fetal activity does not always equate to wakefulness [17]. Consistent with this, some authors have argued that apparent wakefulness in the fetal baboon is exceedingly rare and may in fact reflect misclassified transitional periods between sleep states [61].

In human fetuses, state 3F (quiet awake) is rarely observed, and state 4F (active awake), though defined, is difficult to identify reliably due to obscured eye movements [13]. As inferred from electrophysiological evidence, EEG activity during such periods may resemble spontaneous sleep-state transitions rather than sustained wakefulness [15].

The ability to intentionally wake the fetus is not well established. In fetal sheep, external stimuli such as maternal hormone fluctuations influence fetal sleep patterns, but there is little evidence that these lead to wakefulness [15, 52]. In fetal baboons, experimental challenges such as hypoxia and auditory stimuli have been proposed to investigate fetal behavioural state changes, but the extent to which these lead to true wakefulness is unclear [17]. For human fetuses, vibroacoustic stimulation has been studied as a method to elicit state transitions, but behavioral state organization remains largely resistant to external influences [69]. Nevertheless, early acoustic studies in humans described a low-frequency random intrauterine sound field with rhythmic components time-locked to maternal cardiac activity, indicating that the fetal environment is shaped mainly by internally generated, rhythmically modulated auditory input [70]. Moreover, term fetuses have been shown to display differential heart rate responses to their mother’s voice compared to a stranger’s, indicating a capacity for auditory learning and in utero voice recognition [71]. Although factors such as maternal emotions, Braxton Hicks contractions, and uterine contractions during labor do not significantly alter fetal behavioral state patterns [69], these findings imply that specific types of auditory stimulation may modulate fetal physiology without necessarily inducing full wakefulness. All three species spend the majority of their time in sleep-like states, suggesting that fetal wakefulness, if it exists, is actively suppressed. In fetal sheep, mechanisms such as prostaglandin-mediated regulation (namely prostaglandin E2) and hypoxia-induced depression of breathing, contribute to the maintenance of prolonged sleep-like states, primarily through reduced active (REM) sleep [15, 52]. Baboons exhibit a similar predominance of sleep-like states, with only brief transitions into undefined states [72]. In humans, the fetal nervous system appears to be adapted for continuous sleep-like states, with developing neuronal circuits reinforcing these patterns [28, 73].

To summarize, fetal sheep, baboons, and humans exhibit similar sleep-like states, though true wakefulness remains unclear. Transitional states are frequent, and sleep organization matures with gestation. Across species, sleep state dominates with little evidence of sustained wakefulness before birth.

Maternal and external factors

Maternal physiology and environmental conditions have been shown to influence fetal sleep states [43, 74]. Fetal sleep rhythms begin to develop in utero and are thought to be entrained by maternal melatonin and circadian cues [75]. Various factors, including maternal sleep position, circadian rhythms, sleep disorders, and external stressors, contribute to the regulation of fetal brain activity and behavioral states. External stressors can also influence fetal sleep states. For example, hypoxia and pro-inflammatory stimuli, such as bacterial derived endotoxin (lipopolysaccharide) suppress FBM and modifies EEG activity [76–79], a process likely mediated by elevated adenosine levels, which act as an inhibitory neuromodulator, cytokine mediated suppression of neuronal excitation and by increased neurosteroids such as allopregnanolone, which suppress neuronal excitation and protect the fetal brain [80, 81].

The position a mother adopts during sleep may affect FBS. Supine sleep is known to be associated with reduced uteroplacental perfusion, leading to fetal quiescence [43]. This reduction may be due to altered maternal cardiac output and uteroplacental perfusion, which transiently affect oxygen and nutrient delivery to the fetus, potentially influencing fetal sleep patterns and activity levels [82–84]. Late stillbirth is independently related to the position women adopt during sleep [43]. Vulnerable fetuses, who may already experience chronic hypoxia, have a reduced ability to adapt to maternal sleep position stressors [43]. Another study found that passive maternal movements, such as rocking or swaying, can alter fetal heart rate and potentially the behavioral states, likely through activation of the vestibular system [85].

Fetal sleep patterns are also closely linked to maternal circadian rhythms. Studies have shown that fetal sleep states align with maternal melatonin secretion and activity-rest cycles, indicating that maternal circadian rhythms play a role in regulating fetal brain activity [86]. Furthermore, maternal sleep-disordered breathing (SDB), such as sleep apnea, becomes more common in the third trimester and can disrupt the intrauterine environment by inducing nocturnal hypoxia and heightened maternal autonomic activity. These changes have been associated with alterations in fetal physiological behaviors, including heart rate decelerations and reduced fetal breathing movements—both of which are key indicators of fetal sleep states. While the precise impact on long-term neurodevelopment remains uncertain, these findings suggest that maternal SDB may acutely influence fetal sleep regulation [87, 88].

In summary, maternal physiological and environmental factors have significant effects on fetal sleep states through mechanisms involving hemodynamics, hormonal regulation, and neural modulation. These factors collectively contribute to shaping fetal brain activity and behavioral states.

Prenatal circadian entrainment and maternal–fetal coupling

Before birth, fetal sleep shows organized ultradian cycling (approximately 70–90 min) with canonical state proportions (about 30% NREM, 55% REM, and 15% wake). Maturation is indexed by increasing state stability and longer transition durations: early gestation is intermediate sleep–dominant [89]; quiet sleep rises markedly from 32–40 weeks while active sleep remains relatively unchanged [1]; and by term, quiet and active sleep equalize [63, 64, 73].

Circadian entrainment begins prenatally [1] and is largely maternally driven [75, 90]. The fetal suprachiasmatic nucleus (SCN) has intrinsic rhythmicity yet remains functionally immature postnatally [1, 90]. Maternal melatonin crosses the placenta and the fetal blood–brain barrier [75]; selective melatonin-receptor blockade alters NREM/REM proportions [91]; and because endogenous fetal melatonin rhythms are absent before birth, rhythmic secretion appears only about 9–12 weeks after term (later in preterm infants) [92, 93].

In humans, day–night variation in fetal heart rate synchronizes with maternal rest–activity patterns, heart rate, cortisol, melatonin, and body-temperature rhythms [2, 86], consistent with maternal entrainment [75], and circadian modulation of autonomic (including vagal) control [94, 95]. Yet organization is still emerging near term: at approximately 38 weeks, about 73% of fetuses show a circadian rhythm in basal heart rate, whereas only 30–50% show circadian patterns in HRV or activity [96], likely reflecting SCN immaturity, dominant ultradian oscillations, and substantial inter-fetal variability [96, 97].

Causal evidence comes from sheep. Maternal and fetal melatonin exhibit 24-h peaks (approximately 18:00–06:00) [90, 98]; maternal pinealectomy abolishes melatonin rhythms [90, 99, 100] and removes the nocturnal rise in fetal breathing, while morning feeding unmasks metabolic zeitgebers [33]. Shifting the dark phase shifts prolactin peaks, indicating transfer of maternal circadian phase to the fetus [101]. Maternal behaviors can acutely modulate these dynamics: during late-pregnancy overnight recordings, the supine (vs. left-lateral) position increased 1F and reduced 4F with HRV changes, consistent with reduced uteroplacental perfusion and limited adaptive reserve in vulnerable fetuses [43].

Together, these findings support a model in which the fetal circadian system is present but largely maternally entrained before birth, biasing the timing and autonomic tone of sleep states, while ultradian cycling governs their structure.

Acquisition of Physiological Signals in the Fetus

To understand fetal behavioral and sleep states, researchers have relied on both non-invasive technologies suitable for human fetuses and invasive modalities enabled by animal models such as fetal sheep and baboons. Table 4 summarizes the key measurement techniques across species. In the remainder of this section, we describe each modality in more detail, with a focus on signal types, acquisition methods, and the physiological information derived.

Table 4.

Comparison of fetal sleep measurement techniques

Measurement	Fetal Human (Non-invasive)	Fetal Sheep (Invasive)	Fetal Baboon (Invasive)
EEG/ECoG	Infeasible	Implanted dural electrodes [58, 116, 117]	Dural electrodes [46, 61, 72]
EOG	Ultrasound imaging [69, 114]	Canthus electrodes [24, 60, 117]	Canthus electrodes [62, 121]
FHR	CTG [44, 102, 103] Abdominal ECG [107] Scalp ECG (Invasive) [108] FMCG [109–112]	Arterial pressure [118, 126] ECG [125] Arterial Perivascular Doppler Probe	ECG electrodes [62]
FBM	Ultrasound imaging [69, 114]	Tracheal catheter [122] Diaphragm EMG [123, 124] Laryngeal EMG (PCA) [124]	Tracheal catheter [17, 62, 72]
Body Movements	Actocardiogram [56, 112]	Limb EMG [24, 134] Nuchal EMG [126, 160] Ultrasound [161]	Not Commonly used

Abbreviations: EEG, electroencephalogram; ECoG, electrocorticogram; EOG, electrooculogram; EMG, electromyogram; FHR, fetal heart rate; CTG, cardiotocography; FMCG, fetal magnetocardiography; FBM, fetal breathing movements; PCA, posterior cricoarytenoid muscle.

Non-invasive Technologies for Human Fetus

In human fetal research, ethical and technical limitations necessitate the use of non-invasive techniques. These technologies prioritize safety, cost-effectiveness, and practicality, while attempting to capture physiological signals linked to FBS.

• CTG: A widely used method employing 1D Doppler ultrasound to monitor fetal heart rate (FHR) and uterine contractions through the maternal abdomen [44, 102, 103]. CTG is low-cost and non-invasive [104], but it provides only a smoothed heart rate estimate rather than beat-to-beat intervals, limiting detailed HRV analysis [105, 106].

• FECG: Electrodes on the maternal abdomen record fetal cardiac signals, though maternal ECG interference often degrades signal quality [107]. A scalp electrode applied intrapartum offers improved fidelity but is invasive and limited to labor and rupture of the fetal membranes [106, 108].

• FMCG: A high-resolution modality that uses superconducting quantum interference devices sensors to detect fetal cardiac magnetic fields through the maternal abdomen [109–112]. It offers millisecond temporal resolution and is reported to be less susceptible to artifacts than FECG [113], but it is expensive and technically demanding.

• Ultrasound Imaging: Ultrasound is used to monitor fetal body and eye movements, amniotic fluid volume, breathing, and muscle tone [69, 114]. Recent work has also quantified fetal eye movement activity as a potential behavioral proxy of REM-like states, identifying developmental inflection points around 28–29 and 36–37 weeks of gestation, though such indices remain indirect and cannot substitute EEG-based sleep state definitions [115].

• Actocardiography: Actocardiography combines Doppler-derived FHR and movement data [56, 112], enabling richer behavioral state characterization [45].

Invasive Technologies in Animal Models

Several fetal monitoring techniques used in humans, including FECG and FMCG, are also employed in animal models. In particular, fetal sheep and baboons enable the use of invasive methods that offer high-resolution, direct physiological measurements. These modalities facilitate a more granular analysis of fetal sleep and behavior, including electrocortical activity, eye movements, respiration, muscle tone, and cardiovascular dynamics.

• EEG/ECoG: Electrodes implanted on or beneath the fetal skull record electrocortical activity. In both sheep and baboons, stainless-steel screw or solder-ball electrodes are positioned on the dura mater to record ECoG signals [46, 58, 72, 77, 116, 117].

• Signals are filtered (0.1–40 Hz or up to 100 Hz) and digitized at 50–200 Hz [53, 118]. Sleep states are differentiated by power spectral analysis: quiet sleep shows 1–4 Hz bursts (Trace Alternans), while active sleep exhibits elevated 12–24 Hz power [61]. Common quantitative measures include EEG amplitude and multi-band power, with derived indices such as the spectral edge frequency (SEF) or EEG power ratio (0.03–0.2 Hz vs. 12–24 Hz) used to differentiate sleep states [53, 119, 120].

• EOG: detects eye movements through electrodes implanted near the orbits (sheep) [117] or subcutaneously around the eye (baboons) [121].

• Respiratory Activity Monitoring: FBM are monitored invasively using pressure catheters or EMG electrodes. In sheep, a pressure catheter is placed in the fetal trachea to detect intrathoracic fluid shifts [122], while EMG electrodes can be sewn into respiratory muscles such as the diaphragm or posterior cricoarytenoid [123, 124]. In fetal baboons, FBM are similarly assessed using tracheal catheters to record intratracheal pressure changes [17, 62, 72].

• Breathing is typically intermittent and linked to REM-like sleep, characterized by low-voltage ECoG [122]. These fetal breathing movements are essential not only for lung development but also for training the neural circuits that control respiration and for strengthening respiratory muscles in preparation for breathing after birth. Data are digitized at rates such as 25 Hz for waveform analysis [72].

• Cardiovascular Signal Acquisition: Fetal heart rate is derived from ECG, arterial pressure signals, or Doppler flow probes secured onto major arteries. In both sheep and baboons, ECG electrodes (often silver solder balls or stainless-steel wires) are implanted subcutaneously over the cardiac apex and right atrium to record cardiac activity, while arterial pressure waveforms are obtained via catheters placed in fetal arteries [46, 62, 118, 125, 126]. In sheep, arterial perivascular Doppler probes can also be positioned around the carotid or femoral artery to obtain pulsatile blood-flow waveforms for beat-to-beat FHR monitoring, providing a high-fidelity complement to ECG- or pressure-derived heart rate.

Comparison of EEG and FHR frequency bands

Table 5 compares EEG and FHR frequency band definitions across fetal human, sheep, and baboon studies. This table highlights a key challenge in cross-species comparisons: the frequency band boundaries, especially for EEG rhythms, vary considerably due to both biological differences and species-specific research conventions. For instance, delta and theta bands in fetal sheep span wider frequency ranges than in baboons. Similarly, the definition of FHR bands such as very low frequency (VLF) and high frequency (HF) also differs between species, which complicates the translation of findings from animal models to human contexts. Recognizing these inconsistencies is essential for interpreting spectral analyses and designing cross-species comparative studies.

Table 5.

Comparison of EEG and FHR frequency band definitions across species

Signal Type	Fetal Human	Fetal Sheep	Fetal Baboon
EEG	Not available in fetal human studies	Delta (0–3.9 Hz), Theta (4–7.9 Hz), Alpha (8–12.9 Hz), Beta (13–22 Hz) [133, 154]	Delta (1–4 Hz), Theta (4–7 Hz), Alpha (8–12 Hz), Beta1 (14–18 Hz), Beta2 (22–29 Hz) [53]
FHR	VLF (0.02–0.08 Hz), LF (0.08–0.2 Hz), Intermediate (0.2–0.4 Hz), HF (0.4–1.7 Hz) [162–164]	VLF (0–0.04 Hz), LF (0.04–0.15 Hz), HF (0.15–0.4 Hz) [108, 165, 166]	LF (0.05–0.2 Hz), HF (0.5–2.0 Hz) [62]

Abbreviations: EEG, electroencephalogram; FHR, fetal heart rate; VLF, very low frequency; LF, low frequency; HF, high frequency.

Automatic Classification of Fetal Sleep

Fetal sleep classification has been explored using rule-based and deep learning approaches. Rule-based methods rely on expert-defined thresholds and logic rules, sometimes supported by clustering-based preprocessing such as K-means. In contrast, deep learning enables end-to-end, data-driven modeling from raw physiological signals.

Table 6 summarizes key distinctions across these approaches, including recent developments in multimodal signal integration.

Table 6.

Comparison of FBS classification studies

Study	Species	Size	GA	Signals	Method	States	Performance
Vairavan (2016) [44]	Human	39	30–38w	FMCG (HR + Acto)	Rule-based	36w:1F/2F 36w:1F/2F	36w: ICC = 0.88(1F), 0.65(2F) 36w: ICC = 0.88(1F), 0.41(2F) AUC = 0.99
Semeia (2022) [42]	Human	52	27–39w	FMCG (HRV + Acto)	Rule-based	32w: Active/Passive 32w:1F/2F	32w: AUC1.0(HRV), 0.8(Acto) 32w: AUC1.0(HRV), 0.86(Acto)
Myers (1993) [61]	Baboon	3	20–22w	EEG	Rule-Based	1F vs. 2F	Agreement: 87.1% (Overall), 79.7%(1F), 91.3%(2F)
Grieve (1994) [121]	Baboon	3	23–25w	EEG, EOG, ECG	Rule-Based	1F vs. 2F	Agreement: 81.5% (Overall), 83.7%(1F), 79.4%(2F)
Samjeed (2022) [128]	Human	105	20–40w	NI-FECG	1D CNN	1F vs. 2F	F1: 80.2%(1F), 69.5%(2F) Acc: 76% Sensitivity: 72.7%(1F), 82.6%(2F)
Subitoni (2022) [129]	Human	115	27–39w	FHR	HMM + CNN	1F vs. 2F	HMM + CNN: F1: 87.87%, Acc: 88.37% HMM only: F1: 77.73%, Acc: 83.30%

Abbreviations: 1F, quiet sleep (analogous to NREM); 2F, active sleep (analogous to REM); FMCG, fetal magnetocardiography; GA, gestational age; AUC, area under the curve; NI-FECG, Non-invasive fetal ECG; Acto, Actogram; w, weeks.

Fetal HRV analysis

Physiological basis of fetal heart rate variability in sleep

Sleep states in fetuses are associated with distinct changes in physiological parameters, prominently fetal heart rate variability (FHRV) [47, 62]. In humans, transitions between sleep states are reflected in changes in heart rate patterns, strongly correlating with behavioral states defined by heart rate, body movements, and eye movements [13]. Similar correlations have been observed in fetal baboons, where FHRV measures, combined with EOG and EEG data, have been successfully used to define behavioral state cycles [121]. High EEG-Ratio periods, indicative of quiet sleep, correspond to lower heart rates and reduced FHRV in fetal baboons [61], suggesting that ANS modulation of FHRV is influenced by sleep states [121].

In fetal sheep, physiological studies show distinct differences in FHRV between sleep states. Quiet sleep is typically associated with lower beat-to-beat variability compared to active sleep, indicating varying ANS modulation [47].

Feature extraction for sleep classification

Feature extraction from FHRV typically relies on time-domain and frequency-domain measures derived from RR intervals. In fetal baboons, features such as the standard deviation of RR intervals (SD-RR) and the root mean square of successive differences (RMSSD) are computed on a minute-by-minute basis, provided that at least 90% of RR intervals are artifact-free [62]. In human fetal studies, SDNN, RMSSD, and permutation entropy have similarly been employed to classify sleep states [112]. In fetal sheep, frequency-domain spectral measures—such as low-frequency (LF), HF, and the LF/HF ratio—have been used to differentiate FBS [47]. However, the interpretation of LF/HF as a marker of sympatho-vagal balance is controversial, as LF power reflects a combination of sympathetic and parasympathetic influences, and the ratio can be affected by non-neural factors such as respiration and physiological changes associated with different heart rates [127].

Rule-based approaches

Threshold-based classification using FMCG and actogram signals

Rule-based approaches for FBS classification typically rely on deterministic thresholds derived from physiological signals, such as HRV and actogram-based fetal movement data. These systems apply expert-defined rules to classify states by comparing extracted features with fixed thresholds. While these methods provide interpretable and practical solutions for assessing fetal sleep and wakefulness, they lack the flexibility to adapt to individual variability and gestational changes, limiting their robustness in real-world scenarios.

In 2016, Vairavan et al. [44] developed an early automated pipeline for FBS classification using FMCG recordings from 39 fetuses between 30 and 38 weeks of gestation. They translated the Nijhuis criteria [13] into fixed-threshold rules based on fetal heart rate patterns and actogram-derived movements. The system demonstrated strong agreement with expert annotations, particularly for quiet sleep (intraclass correlation coefficient (ICC) = 0.88), though performance declined for active sleep in later gestation (ICC dropped to 0.41). These findings suggest that while rule-based classification is feasible with FMCG and CTG, behavioral complexity increases with maturation, potentially limiting such approaches.

Building on Vairavan et al.’s work, Semeia et al. [42] refined rule-based FBS classification by introducing gestational age-specific distinctions. Using a large FMCG dataset, they separated younger (32 weeks) and older (32 weeks) fetuses and adapted the classification accordingly—distinguishing active/passive states in early gestation and 1F/2F states later. Their results showed that HRV-derived parameters, especially RMSSD and standard deviation (STD) of HR, achieved near-perfect classification accuracy (AUC 1.0), while actogram-based features were less reliable. These findings reinforce the utility of HRV metrics for FBS classification and highlight the need for developmental stage-specific models.

Both studies demonstrated that rule-based approaches can achieve high classification accuracy for prototypical FBS, but their reliance on fixed thresholds restricts their adaptability across different gestational ages and individual variations. The absence of a temporal component in these models makes it difficult to capture transitional states that naturally occur as fetal development progresses. Moreover, rule-based methods do not account for probabilistic uncertainty, potentially leading to overconfidence in misclassified instances.

Semeia et al. [42] identified a key limitation of rule-based methods: substantial overlap in parameters between quiet and active sleep, which hampers accurate classification—especially during transitional phases with gradual physiological changes. This suggests such methods may oversimplify the complex dynamics of FBS.

As shown in Table 6, Vairavan et al. [44] and Semeia et al. [42] achieved good results using FMCG, but did not assess generalizability to more accessible modalities like CTG. Their methods also struggle with fetal state transitions and individual variability, highlighting the need for more advanced probabilistic and machine learning approaches.

Rule-based classification aided by K-means on EEG/multimodal signals

Clustering techniques have also been explored in relation to FBS classification, primarily as a preprocessing aid rather than a stand-alone unsupervised approach. Myers et al. [61] proposed an early rule-based method using fetal baboon EEG data. They developed the “EEG-Ratio” defined as the power in the 0.03–0.2 Hz band (associated with trace alternant) divided by power in the 12–24 Hz band. This feature correlated with visually scored sleep states, and K-means clustering was applied in a preprocessing step to group the data into two categories—trace alternant (TA, representing quiet sleep) and non-TA (active sleep). The resulting classification achieved an 87.1% agreement with expert scoring.

However, the study had limitations. It included only three fetal baboons, each contributing four EEG recordings, totaling 3694 minutes of usable data. The authors did not use any form of cross-validation or independent testing, as thresholds were optimized and validated on the same dataset, potentially inflating accuracy estimates. Moreover, the EEG-Ratio—being a scalar feature—may not generalize well across subjects or conditions.

To improve robustness, Grieve et al. [121] extended this approach by integrating multimodal signals—EEG, EOG, and ECG—from the same three fetal baboons, each recorded for 16 continuous hours. They applied K-means clustering to extract binary thresholds for three features: EEG ratio, EOG spectral power, and RR interval variability (CVRR). These features were then combined using rule-based criteria to define two sleep states: 1F (quiet sleep) and 2F (active sleep). Transitions and indeterminate states were also identified using temporal continuity rules. Agreement with expert annotations reached 81.5% overall (83.7% for 1F and 79.4% for 2F), demonstrating the feasibility of long-term, automated multimodal classification.

While the use of multimodal features provided a more physiologically grounded framework, limitations remained, including small sample size, absence of gestational stratification, and no direct comparison with unimodal or machine learning-based models. Nonetheless, these early efforts laid the groundwork for automated fetal sleep state detection using interpretable, unsupervised approaches.

Machine learning approaches

Deep learning has emerged as a powerful tool for classifying FBS from physiological signals. Two recent studies—by Samjeed et al. [128] and Subitoni et al. [129]—have proposed deep neural network-based methods leveraging fetal ECG and FHR signals, respectively.

Samjeed et al. [128] proposed a 1D convolutional neural network (1D-CNN) to classify FBS from non-invasive abdominal ECG recordings of 105 fetuses (20–40 weeks gestation, 3–10 min duration). The CNN consisted of three convolutional layers and was trained using stochastic gradient descent with momentum (SGDM) with 5-fold cross-validation. It achieved 76% accuracy, with F1-scores of 80.2% for the quiet state and 69.5% for the active state, suggesting challenges in distinguishing between states. Limitations included dataset imbalance, lack of temporal modeling, and use of a single modality. Future directions included exploring RNNs, multimodal inputs, and transfer learning.

Subitoni et al. [129] proposed a hybrid model combining hidden markov models (HMMs) and a U-Net style 1D-CNN (U-Sleep variant) to classify FBS from 115 manually annotated FHR recordings. The dataset was stratified into three gestational age groups (27–32, 33–36, 37–39 weeks) to account for developmental differences.

The approach used HMMs for unsupervised segmentation to generate pseudo-labels, which were then used to pre-train a CNN. The model was subsequently fine-tuned using expert annotations. This two-stage training allowed the system to leverage both unlabeled and labeled data.

The hybrid model achieved a Macro F1-score of 87.87%, outperforming the HMM alone (77.73%). However, the study did not clarify whether evaluation was subject-wise or sample-wise, and relied on annotations from a single expert, limiting generalizability. Still, the method demonstrates a promising strategy to reduce dependence on annotated data via hybrid learning.

Effects of Abnormal Conditions on Fetal Sleep

Hypoxia

Hypoxia is a major disruptor of fetal sleep and neurodevelopment. Graded hypoxia experiments in fetal sheep have demonstrated significant disruptions in sleep states, including altered ECoG and behavioral activity [47, 130]. Koos et al. [130] reported that mild hypoxia did not alter the incidence of low-voltage ECoG activity, FBM, or REMs, whereas moderate and severe hypoxia markedly suppressed both FBM and REMs. A critical threshold was identified, with a reduction in arterial oxygen content of ml/dl associated with inhibition of both eye and breathing activity. Importantly, ventilatory responses to hypoxemia differ fundamentally across development: in utero, acute hypoxia suppresses FBM, whereas after birth the neonate mounts an increase in ventilation; moreover, prolonged hypoxic exposures exert distinct short- and long-term effects on the maturation of the neural respiratory control network [131].

In fetal sheep models, neuroprotective strategies yield mixed outcomes. Prophylactic creatine supplementation improves EEG power, frequency recovery, and reduces seizures after umbilical cord occlusion [132], whereas magnesium sulfate attenuates seizures and gliosis but fails to restore EEG power or sleep-state cycling [48].

In preterm fetal sheep, chronic or repeated hypoxia causes lasting EEG suppression, reduced delta–theta power, and blunted gestational increases in HRV, indicating impaired sympathetic control and altered respiratory plasticity [47, 131, 133].

Mechanistically, hypoxia elevates adenosine and neurosteroids (e.g., allopregnanolone), promoting sleep-like EEG states [80, 81] and reducing movements and REMs as oxygen-conserving adaptations [134]. Adenosine-mediated inhibition suppresses FBM [135–137], with severe hypoxia inducing atonia [138]. The thalamic parafascicular nucleus mediates this inhibition, typically reversible within 12–16 h, though prolonged hypoxemia can lead to cerebral ischemia and persistent EEG dysmaturation [139–141].

During REM, hypercapnia enhances fetal breathing and reduces apnea through CO₂-dependent stimulation [142].

Fetal growth restriction

Fetal growth restriction (FGR) disrupts the normal organization of fetal sleep states, particularly in late gestation. Growth-restricted fetuses exhibit greater sleep-state instability, often spending more time in quiet sleep than in active sleep [143, 144]. In addition, FGR is commonly associated with impaired oxygenation, reduced breathing and general movements, and an increased number of heart rate decelerations, reflecting ANS dysfunction [2, 145].

FGR fetuses also show diminished motor activity, characterized by slower, monotonous, and lower-amplitude movements, which reflect central nervous system impairment [73, 146, 147]. These disturbances are considered late-stage indicators of fetal compromise, often preceded by abnormalities in HRV and blood flow parameters [73, 148]. Collectively, these findings highlight FBS as important indicators of neurodevelopment and fetal well-being.

Intra-amniotic infection (Chorioamnionitis)

Intra-amniotic infection (clinical or subclinical chorioamnionitis) produces a non-hypoxic CTG pattern with persistently elevated baseline FHR, reduced variability and accelerations, and loss of FHR cycling—implying disruption of fetal sleep–wake organization. Consistently, systemic inflammation induced by lipopolysaccharide in late-gestation fetal sheep suppresses high-frequency EEG activity, particularly in the beta band, reflecting impaired cortical activation and sleep-state transitions. These EEG changes are accompanied by reduced fetal movements [79, 149, 150] and persist beyond resolution of inflammation, indicating sustained disturbance of behavioral state cycling [77].

Although direct evidence linking infection to altered fetal sleep is limited, concurrent abnormalities in movements, HRV, and breathing suggest that sleep-state organization is affected. Fetal sleep may thus offer an additional surveillance parameter, though its clinical application remains constrained by current technology and knowledge.

Conditions with emerging evidence

A number of maternal and fetal conditions with more limited or emerging evidence also appear to influence fetal behavioral or “sleep-like” states. Structural or chromosomal anomalies often lead to abnormal organization of FBS, reflecting impaired central nervous system integration. These disturbances manifest as prolonged episodes of reduced or excessive activity—hypokinesia or hyperkinesia—and poor differentiation of state cycling between quiet and active phases [148]. Fetuses with autosomal recessive neuromuscular or skeletal disorders (e.g., fetal akinesia/hypokinesia sequences, restrictive dermopathy, arthrogryposis) typically show persistent hypokinesia and absent behavioral states, whereas those with central nervous system malformations or chromosomal defects such as anencephaly, trisomy 18, or 1q and 5p abnormalities may exhibit hyperkinesia or uncoordinated state transitions [73, 151]. Case reports describing Smith–Lemli–Opitz and Prader–Willi syndromes similarly note missing or atypical sleep-state organization, further linking congenital brain dysfunction to altered sleep architecture. These abnormalities are often accompanied by prolonged periods of low HRV and portend poor outcomes—only about 14% of affected fetuses in one review survived with neurodevelopmental impairment [151].

Maternal diabetes appears to disrupt the maturation of fetal sleep-state organization. In a longitudinal study using ultrasound and FHR monitoring, Dierker et al. observed that uncomplicated pregnancies exhibited a characteristic developmental shift between 28–32 and 36–40 weeks: a significant reduction in hourly activity–quiescence cycles (from 4.98 to 2.15) alongside a substantial lengthening of both active (16.1 to 36.2 min) and quiet periods (16.2 to 26.0 min). In contrast, these indices remained relatively stagnant in diabetic pregnancies (cycles/hour: 4.94 to 4.75; active: 13.6 to 16.6 min; quiet: 16.8 to 12.5 min), with none of the changes reaching statistical significance, thereby indicating delayed maturation of behavioral state cyclicity [152]. Because behavioral states were defined jointly by long-term FHR variability and fetal movements, these findings suggest impaired central integrative control in diabetic pregnancies, although the absence of postnatal follow-up limits interpretation.

Alcohol exposure provides another example of a condition with emerging but incomplete evidence. Acute maternal alcohol intake near term—roughly two glasses of wine in a controlled study—suppressed fetal rapid eye movements and markedly reduced fetal breathing movements [64]. Since REM (active) sleep dominates late gestation and is closely linked to immature respiratory and arousal control, repeated REM suppression may interfere with the maturation of sleep–wake and brainstem regulatory networks, though current evidence in humans remains inferential [73, 153]. Alcohol-induced reductions in REM and eyeball immobility have been proposed, albeit speculatively, as contributors to neurobehavioral and ophthalmic abnormalities in fetal alcohol spectrum disorders [73]. More broadly, near-term fetal state organization serves as a marker of central nervous system integrity, and more mature, synchronized transitions predict better postnatal self-regulation, underscoring the potential developmental impact of even brief perturbations [73].

Synthesis and evidence grading

To synthesize these heterogeneous findings, Table 7 provides a concise graded summary of abnormal conditions, their observed impact on FBS, and the relative strength of supporting evidence (evidence rubric detailed in the Supplementary Information, Section S1).

Table 7.

Concise graded summary of abnormal conditions and their observed impact on FBS

Condition	Impact on FBS	Evidence
Hypoxia	FBM and REMs reduced; shift toward Quiet (NREM-like); EEG power suppressed; HRV down-regulated [47, 130, 131, 133]	Strong†(sheep), Moderate‡(human)
FGR	Shift toward Quiet (NREM-like); Active (REM-like) and overall activity reduced; HR decelerations increased [144, 145, 148]	Moderate‡(human)
Congenital malformations	Hypokinesia with prolonged low HRV [73, 148, 151]	Moderate‡(human)
Maternal diabetes	Delayed maturation of Quiet/Active (NREM/REM) between 32–40 weeks [152]	Preliminary‡(human)
Maternal sleep position	Transition toward Quiet (NREM-like) consistent with reduced uterine perfusion [43, 167]	Moderate‡(human)
Alcohol consumption	REM (Active) and FBM reduced in controlled maternal intake studies [64, 73]	Moderate‡(human)
Intra-amniotic infection	Loss of FHR cycling; suppression of EEG power; disruption can persist post-inflammation [77, 168, 169]	Strong†(sheep), Emerging‡(human)

Strength of evidence — four levels:

Strong: 2 high-quality, independent studies with causal or mechanistic support, consistent results, and clear effect size.

Moderate: Stable associative evidence but with limitations (e.g., small sample size, residual confounding).

Preliminary: Limited sample, major methodological limitations, or inconsistent findings.

Emerging: Few or very recent studies providing suggestive signals.

Abbreviations: FBS, fetal behavioral states; FBM, fetal breathing movement; REM, rapid eye movement; NREM, non-REM; EEG, electroencephalography; FHR, fetal heart rate; HRV, heart rate variability; FGR, fetal growth restriction. Directness flags: ^†direct; ^‡indirect; ^§ maternal proxy.

Future Directions and Challenges

Limitations of current studies

Technological limitations

While current tools have advanced FBS research, technologies like ultrasound, CTG, and FMCG face limitations such as low signal quality, motion artifacts, and limited applicability in early gestation. Fetal MRI provides better CNS insights but is costly and inaccessible. Overall, existing methods lack the resolution and scope to fully capture early fetal neurodevelopment and sleep transitions.

Analytical challenges

FBS detection often depends on visual inspection or algorithms trained on prototypical segments, overlooking transitional or ambiguous states that may offer important developmental insights. Inconsistent terminology across studies further hinders reproducibility and comparison. Most methods also fail to capture complex dynamics like diurnal rhythms or maternal-fetal interactions.

State-aware modeling and cross-species standardization

Most studies decouple sleep staging from pathology detection, which risks confounding and can mask effect modification by state; circadian or time-of-day terms and causal designs are seldom incorporated. In addition, the absence of a unified cross-species framework and computable ontology linking human behavioral states (e.g., Nijhuis 1F–4F) to electrophysiology-defined states in animal models leads to inconsistent labels and minimum-duration rules across datasets, limiting meta-analysis, external validation, and translational generalizability.

Sample size and Interindividual variability

Small sample sizes limit the statistical power and generalizability of fetal sleep studies, especially in linking FHR to biochemical markers. Variability across fetuses—driven by gestational age, maternal health, and environmental factors—complicates standardizing FBS classification. Broad gestational groupings (e.g., mid and late gestation) may mask critical developmental transitions.

Longitudinal and genetic considerations

Links between prenatal sleep and postnatal outcomes are limited by long assessment gaps and unaccounted genetic influences shared by mother and fetus. This highlights the need for integrated, genetically-informed longitudinal studies to better explain outcome variability.

Future research directions

To overcome current limitations, future research should explore advanced analytical tools—such as point process models and detailed HRV metrics—to uncover biomarkers of fetal brain and ANS development. Although fetal EEG is infeasible in humans, invasive recordings in animal models (e.g., sheep, baboons) can inform the interpretation of non-invasive human data (e.g., FMEG, FMCG, coherence). Cross-modal integration of spatio-temporal and synchrony features may further elucidate fetal CNS maturation and sleep-state transitions, provided that studies report harmonized annotation protocols, inter-rater agreement, and time-of-day metadata on open, multi-site benchmarks with fixed training–validation–test splits.

Improving automated FBS detection remains critical. Machine learning models, validated against tools like fetal MRI and synchronized physiological signals, should undergo prospective multi-site external validation with device and site stratification, probability calibration and uncertainty reporting, and decision-analytic assessment of clinical utility. Consistent terminology and operational definitions—including window length, minimum state duration, and handling of transitional or ambiguous epochs—will also enhance model generalizability and cross-study comparability.

To address small sample sizes and individual variability, transfer learning is a promising solution. Pretrained models on adult sleep data can be fine-tuned on fetal recordings, leveraging shared low-level features while adapting high-level patterns to fetal physiology. This reduces data demands and improves model robustness across gestational ages and maternal-fetal conditions. In addition, adult sleep data can be transformed to better match the spectral characteristics of fetal sleep data using signal processing or generative adversarial networks (GANs). Aligning spectral distributions in this way provides a more compatible source for fine-tuning, further enhancing transfer learning performance. Inspired by reinforcement learning, reward-guided fine-tuning based on physiological plausibility or expert preference can further improve adaptation across gestational stages.

A multidisciplinary approach—linking neuroscience, obstetrics, and neuroimaging—is essential. Longitudinal studies from early gestation to infancy with continuous maternal-fetal monitoring can clarify developmental trajectories, especially sleep-state transitions and maternal influences.

Further exploration of vagal tone and ANS maturation may identify critical periods of vulnerability. Refining measurement tools through multimodal integration will be key to developing clinical guidelines for identifying fetuses at risk of autonomic or neurodevelopmental disorders.

Ultimately, a comprehensive perinatal perspective—recognizing bidirectional maternal-fetal interactions and the continuity of sleep-state development—is vital. Monitoring fetal sleep may enable early detection of FGR, chronic hypoxia, or emerging neurological issues. Early intervention (e.g., optimized delivery, neuroprotective agents, maternal care) is crucial for improving long-term outcomes during this sensitive developmental window.

Conclusion

Fetal sleep is a fundamental yet undercharacterized aspect of prenatal neurodevelopment and a promising window into early central and autonomic nervous system maturation. In this review, we synthesized more than seven decades of work spanning human fetuses and large-animal models, outlining how FBS emerge, how they map onto electrophysiological and cardiovascular signatures, and how they can be monitored using both invasive and non-invasive modalities. By comparing species, we highlighted both conserved features of sleep-state cycling and important differences in timing and expression that shape how animal data can inform human pregnancy. We also summarized rule-based and machine learning approaches for automated classification, together with the evidence that abnormal intrauterine conditions such as hypoxia, fetal growth restriction, and intra-amniotic infection disrupt fetal sleep–wake organization and related physiological signals.

Across these literatures, several themes emerge. First, inconsistent terminology, limited sample sizes, and heterogeneous annotation protocols hinder cross-study comparison and impede the development of robust, generalizable tools. Second, current monitoring technologies and analytical pipelines only partially capture the richness of fetal sleep dynamics, often overlooking transitional states, diurnal modulation, and maternal–fetal coupling. Finally, links between prenatal sleep, postnatal outcomes, and genetic or environmental risk remain fragmentary. Addressing these gaps will require a harmonized cross-species framework, multimodal and longitudinal cohorts, and computational models that undergo rigorous clinical validation and integrate physiology with clinically meaningful end points.

Together, these advances could transform fetal sleep from an obscure developmental phenomenon into a clinically actionable biomarker. Standardized and objective assessment of FBS has the potential to improve detection of fetuses at risk of hypoxia, growth restriction, or inflammatory injury and to guide timely neuroprotective and obstetric interventions. Clarifying how sleep develops before birth therefore offers not only deeper insight into human neurodevelopment, but also a path toward safer and more personalized prenatal care.

Disclosure statement

Financial disclosure: No commercial or industry funding was received.

Non-financial disclosure: The authors declare no non-financial competing interests, such as personal, academic, political, or religious conflicts.