Maternal Voice and Infant Sleep in the Neonatal Intensive Care Unit

This study was approved by a Michigan Medicine Institutional Review Board. A parent of every enrolled infant provided written informed consent. We prospectively recruited late-preterm and term neonates who required intensive care in our single-patient room NICU. Inclusion criteria were: gestational age ≥33 to 41 weeks, need for NICU admission, and stable temperature regulation in an open crib. Exclusion criteria were: congenital brain anomalies, known or suspected genetic syndromes that could be expected to result in cerebral dysfunction, severe encephalopathy that precluded sleep-wake cycling, significant airway anomalies that were likely to result in sleep-disordered breathing, and abnormal hearing screening (brainstem auditory evoked responses). Score for Neonatal Acute Physiology with Perinatal Extension-II (0 = normal; >37 = elevated risk of mortality^11,12 ) and standardized neurologic examination scores (Thompson scores; 0 = normal; >10 = elevated risk for adverse neurodevelopment¹³ ) were recorded for each infant.

A 12-hour attended polysomnogram was recorded after the infant was considered by the clinical team to be medically stable (no respiratory support and able to tolerate a bedside polysomnogram). Polysomnography generally spanned the overnight and morning hours (∼10 pm–10 am). Parents were permitted but not required to remain in the patient room during the polysomnogram. A registered polysomnographic technologist at the infant’s bedside recorded detailed behavioral observations during the entire polysomnogram. In addition to the 9-channel neonatal-montage EEG (expanded from the standard 6-channel montage to enable screening for subclinical seizures, which might influence sleep regulation), other channels included bilateral electrooculogram, chin surface electromyogram, chest and abdominal excursion (inductance plethysmography), nasal pressure, nasal and/or oral airflow (thermocouples), snoring sensor, oxygen saturation, transcutaneous CO₂, electrocardiogram, bilateral anterior tibialis surface electromyogram, and digital video. Polysomnograms were scored off-line by an experienced, registered polysomnographic technologist and reviewed by a board-certified sleep medicine physician. Procedures and scoring followed contemporaneously current published standards.¹⁴  All infants were cared for in an open bassinet. Their care regimens and feeding schedules were maintained throughout the 12-hour polysomnogram. Neonates were generally fed every 3 hours. Infants generally remained in the bassinet for feeding and care.

During the polysomnogram recording, a digital language processing device was placed on the arm of the infant’s bassinet to record the acoustic environment. By using proprietary software (Language Environment Analysis [LENA] Research Foundation, Boulder, CO; www.lenafoundation.org), the audio recording was analyzed to quantify language (eg, adult word count from caregivers speaking at bedside), nonlanguage noise (eg, monitor alarms, which are coded by LENA as “television/electronic sounds”), and periods of silence.¹⁰  This method has been validated for evaluation of the NICU acoustic environment⁹  and can recognize both English and Spanish conversation, although all parents and clinical staff of included infants spoke English at bedside.

We provided the infants’ mothers with 2 children’s books^15,16  from which to read while their voices were digitally recorded. The digital voice recording was played continuously during half (6 hours) of the polysomnogram via a device placed near the foot of the infant’s bassinette. Because the infants’ behavioral state could be influenced by the application of the polysomnogram monitoring electrodes, we randomized the order of 6 hours usual care acoustic environment vs 6 hours enriched maternal voice exposure such that half of the infants had their mother’s voice recording played during the first 6 hours of the polysomnogram and the other half during the second 6 hours. A 1:1 randomization was performed by using an online random number table. There was no washout period between the 2 6-hour segments. The sleep technician and physician who scored and reviewed the polysomnograms, respectively, were masked to the randomization order. The LENA software classified the maternal audio recording as “television/electronic sounds” or “uncertain/fuzzy” and not as adult word counts; this allowed comparison of bedside conversation (classified as adult word counts) with versus without the maternal voice exposure.

On the basis of preliminary data that demonstrated a Spearman coefficient of ρ ≈ 0.7 for correlations between gestational age and durations of bouts of REM sleep and wakefulness,¹⁷  we determined prospectively that a sample size of N = 47 would provide 80% power to detect ρ = 0.4 with ɑ = 0.05.

We quantified standard sleep metrics¹⁴  (ie, proportions of sleep-wake stages, apnea-hypopnea index [number of apneas and hypopneas per hour], and arousal index [number of arousals and awakenings per hour of sleep]) as well as transition probabilities, EEG power, and entropy measures, as previously described^2,17  for the 6-hour maternal voice playback and 6-hour usual care environment epochs of the polysomnograms. The calculated sleep variables included the proportion of each sleep-wake stage and the probability of changing from 1 specific sleep-wake stage to another (transition probabilities). The Walsh spectral entropy method^18,19  was used to measure the entropy of the sequence of sleep-wake stage transitions; increased entropy values are suggestive of decreased predictability of the sleep-wake stage pattern. The power spectra for each 30-second polysomnogram epoch were computed from the C4-M1 channel of the EEG portion of the polysomnogram by normalizing the total periodogram power averaged across all polysomnogram epochs via the Welch method for a fast Fourier transform.²⁰  Quantitative sleep measures were regressed on gestational age at birth and postmenstrual age with adjustment for neurologic examination (Thompson) scores and average ambient noise level. Data were compared for epochs with versus without the recorded maternal voice playback exposure. Our a priori plan was to evaluate sleep measures across gestational ages. Visual inspection of the data suggested logical subgroups of 33 to 34 vs ≥35 weeks’ gestation. LENA-measured adult word counts were also regressed on overall noise level and sleep-wake stage data.

For each 30-second polysomnogram epoch, the ambient noise was quantified from the digital audio file, and the sleep-wake stage was scored. The data were pooled over the participating neonates, and multinomial logistic regression was used to examine the likelihood of remaining asleep or awakening in the subsequent epoch as a function of environmental noise in the presence or absence of exposure to the maternal voice playback. Multivariable logistic regression was used to evaluate the likelihood of being asleep versus awake as a function of within-epoch peak ambient noise and exposure to the maternal recording adjusted for gestational age and neurologic examination (Thompson) scores. Relationships between gestational age, postmenstrual age, environmental noise, maternal voice playback, and sleep were then explored separately for term and near-term (≥35 weeks’ gestation at birth) versus premature infants (33–34 weeks’ gestation at birth). Bivariate analyses were conducted (Wilcoxon paired signed rank test) to compare acoustic profiles during usual care versus during maternal voice playback. All statistical models were constructed by using MATLAB (MathWorks, Natick, MA), and P < .05 was used to define statistical significance.

Forty-seven newborns were enrolled in the study. Clinical and demographic details are presented in Table 1. Sound levels overall were ∼1 dBA higher during maternal speech exposure versus usual care (mean 56.3 ± 4.6 vs 55.3 ± 5.2 dBA, respectively; P < .0001). The amount of language exposure, classified by LENA as adult word count, was low in both settings and did not change when the maternal voice was playing (mean 178 ± 213 vs 127 ± 154 words per hour; P = .16 [Table 2]).

The peak ambient noise during individual 30-second polysomnogram epochs was generally higher during versus without the maternal voice playback exposure (mean peak noise level 72.7 ± 2.7 vs 74.4 ± 2.9 dBA; P = .0004). However, overall, the infants’ sleep bout lengths (durations of uninterrupted sleep) were not strongly associated with the level of background noise. In the usual care setting, shorter sleep bout length was not meaningfully predicted by louder mean noise levels (NREM sleep: linear regression adjusted R² = 0.002; P = .7; REM, sleep: adjusted R² = 0.004; P = .04). During maternal speech exposure, only a limited association emerged between shorter NREM sleep bout lengths and louder average sound levels (adjusted R² = 0.025; P < .001), and REM sleep bout lengths were not affected (adjusted R² = 0.002; P = .14).

The probability was higher during versus without the maternal voice exposure for the neonate to stay asleep in the epochs after elevated peak noise levels (model R² = 0.62; Table 3, Fig 1). Similarly, the probability of the infant being asleep during any given polysomnogram epoch was higher during maternal voice exposure in comparison with usual care (multivariable logistic regression: β = .07; P = .001) after adjustment for peak noise (β = −.11; P < .001), gestational age (β = .1; P < .001), and Thompson score (β = .03; P < .001).

The arousal index during sleep tended to decline only modestly with advancing gestational age (linear regression model adjusted R² = 0.03; P = .1) and postmenstrual age (adjusted R² = 0.08; P = .03). Arousal index increased slightly with versus without the maternal voice playback (19.7 ± 8.5 vs 18.0 ± 7.4, respectively; P = .01). Similarly, the apnea-hypopnea index declined with advancing gestational age (adjusted R² = 0.06; P = .047) and postmenstrual age (adjusted R² = 0.07; P = .03) but was not altered by exposure to the maternal voice recording (21.9 ± 15.8 during maternal voice exposure vs 22.2 ± 16.2 without maternal voice exposure; P = .9).

Associations of quantitative sleep measures with increasing gestational age varied with maternal voice exposure for the 20 neonates born at ≥35 weeks’ gestation but not the 27 born preterm at 33 to 34 weeks’ gestation (Table 4, Fig 2). During the voice exposure, among neonates born at ≥35 weeks’ gestation, increasing gestational age was associated with increased percent time awake (R² = 0.52; P < .001) and more clearly decreased REM sleep (R² = 0.2; P < .001) than NREM sleep (R² = 0.19; P = .04). Similarly, advancing gestational age was associated with increased wakefulness bout duration (R² = 0.42; P < .001) but not REM or NREM sleep bout duration, fewer REM sleep bouts per hour (R² = 0.35; P = .003), and increased sleep-wake entropy (R² = 0.52; P < .001). These associations remained significant after adjustment for Thompson scores and average ambient noise level (adjusted model R² = 0.30–0.58; each P < .004). Without the voice playback, none of these associations were significant. EEG power at 2 to 4 and 4 to 8 Hz increased with gestational age in both age groups (R² = 0.14–0.47; P = .01 to <.001), and this was not changed by the voice playback. For infants born at <35 weeks’ gestation, no other associations emerged between sleep measures and gestational age with or without maternal voice exposure.

Parallel analyses were conducted between postmenstrual age and sleep parameters adjusted for Thompson scores and average noise level. Among neonates born at ≥35 weeks’ gestation, during the maternal voice exposure, the proportion of wakefulness rose with increasing postmenstrual age (R² = 0.34; P = .002) but no other associations emerged. Without the voice exposure, this association was not significant. There were no statistically significant associations between postmenstrual age and sleep-wake variables among infants born at <35 weeks with or without the maternal voice playback.

Results of these quantitative, gold-standard recordings of sleep, ambient noise, and maternal voice exposure suggest that sounds can be remarkably loud in the NICU (even in a single-patient-room design) and that these sounds can influence neonatal sleep. Furthermore, exposure to the mother’s voice (here by recording, but presumably in person as well) may insulate newborns from some of the impact of NICU noise by reducing the likelihood of wakefulness during and just after the highest noise levels. The impact of maternal voice in our data was not uniform across all gestational ages studied; starting at ∼35 weeks’ gestation at birth, in contrast to earlier gestational ages, newborns of advancing ages showed steadily increased amounts of wakefulness during maternal voice exposure. After 35 weeks’ gestation (and aside from periods of loud ambient noise during sleep), newborns may become progressively more alert for their mothers’ voice as they approach term. Finally, our results also suggest that patterns of sleep-wake cycle development and the relationship of sleep to the sensory environment are altered by preterm birth because the neonatal sleep variables were more strongly associated with gestational age at birth than postmenstrual age at the time of the polysomnogram. Overall, these remarkable results could have important implications for care of NICU patients during the earliest days after birth and advance our understanding of newborn dependence on parental interaction.

Moreover, even beyond the first days of life, we speculate that the NICU sensory environment, which differs dramatically from the in utero milieu, may disrupt the stimulus-sensitive plasticity of the immature brain and contribute to abnormal developmental outcome, at least in some vulnerable infants. As a result of physiologic dysmaturity, development of sleep patterns may differ for term versus late-preterm infants at equivalent postmenstrual ages.^21,22  This could explain why our results reveal that the influence of the NICU acoustic environment on neonatal sleep variables is more highly associated with gestational age at birth than with postmenstrual age at the time of the polysomnogram. Animal models have demonstrated the critical, time-specific role of the acoustic environment during auditory cortex organization.^23,24  Others have reported that when played to preterm infants, recorded maternal sounds are associated with improvements in apnea of prematurity²⁵  and increased time in the quiet alert state.²⁶  Among preterm infants born at 25 to 32 weeks, exposure to audio recordings of the mother’s voice and heartbeat for 45 minutes 4 times per day in the first month of life in comparison with usual NICU care was associated with larger auditory cortex size 30 ± 3 days after birth.²⁷  For late-preterm infants, exposure to a lullaby recording (not sung by the infant’s mother) was associated with improvements in the qualitative cycling patterns thought to represent sleep-wake cycles on amplitude-integrated EEG.²⁸  Our study now adds objective evidence, through quantitative polysomnographic and sound analyses, that changes in the NICU acoustic environment (such as enriched maternal voice exposure) can influence sleep physiology for newborn infants. Although this study did not include long-term follow-up, previous work by our group and others has suggested that better quality and efficiency of sleep during the newborn period is associated with improved neurodevelopmental outcomes in cognitive, motor, and language domains.^2,29 

Compelling recent data revealed that, unexpectedly, neonates cared for in a single-room NICU with low parent visitation rates versus an open-bay NICU design were at much higher risk for abnormal language development.⁵  Linguistic outcomes were also better for formerly preterm toddlers whose parents reported noisier rather than quieter NICU environments after adjustment for relevant clinical and socioeconomic variables.³⁰  Increasing language exposure during NICU admission is also associated with better long-term language development.⁶  Recently published data reveal that aiming for silence in the NICU is unrealistic,^4,31  and simply providing a quiet NICU environment may not be an ideal therapeutic approach. We theorize that an ideal balance between opportunity for sleep-wake cycling and appropriate language exposure could help to optimize outcomes.

Importantly, our current and previous results⁸  as well as data from others^6,32  reveal that adult word count is low in the NICU. Although the present findings reveal an association between increased adult word count and higher overall sound levels, this does not mean that conversation was the major driver of overall NICU noise (eg, sound might be higher because the infant was crying, and the caregiver may be more likely to speak if an infant is crying). It is necessary for infants to have language exposure to optimize language development, so it is reasonable to encourage clinicians and families to speak at bedside without excessive concern that they are disrupting the infant’s ability to sleep. However, the potential advantages of necessary language exposure as well as ambient NICU noise required for care targeted selectively to periods of wakefulness rather than sleep remains to be tested. Such care is often provided to older hospitalized patients who, in the absence of such courtesy, could complain about the sleep disturbance and consequent daytime sleepiness. Although infants will not complain, the possibility exists that their developmental trajectory could reflect the impact of sleep disruption.

In the current study, we used gold-standard attended bedside polysomnography to characterize objective measures of neonatal sleep in association with the NICU acoustic environment and maternal voice exposure. However, this study does have some limitations. We evaluated the immediate impact of sound exposure on the infant and cannot know the long-term effect of intermittent language enrichment on longitudinal maturation of sleep-wake cycling or neurodevelopment. Infant responsiveness to sound and language might be hypothesized to be associated with immediate brain function and potentially to predict neurodevelopmental outcomes; long-term follow-up of the present cohort is planned. We analyzed the impact of the acoustic environment, but additional factors such as handling of the infant³³  or neuroactive medications also can influence sleep. Comprehensive approaches to optimize sleep for NICU patients will need to account for these manifold elements of intensive care. Although the overall sound intensity was high during the polysomnograms and tended to be slightly louder during the maternal voice exposure, we found no evidence that the maternal voice playback increased abnormal respiratory events. Of note, even with peak noise as high as 90 dB, the probability for most neonates to remain asleep was high. Epoch-to-epoch sleep scoring in our analyses evaluated for awakenings and not for arousals; loud noises could have a physiologic impact, short of waking the neonate in the subsequent epoch. Whether exposure to live spoken words has a different effect on language development than exposure to the recorded maternal voice is not known. The LENA software classified the maternal voice recording as “uncertain/fuzzy” rather than adult word counts. Still, we show that exposure to the voice recording was associated with measureable changes in sleep physiology.

Given the results of the current study in combination with other published data that suggest consequential long-term impact of healthy early-life sleep or its disruption, a pressing need exists for additional high-quality research to define optimal outcome-relevant conditions for sleep in the NICU and to identify any simple opportunities for intervention. Our findings suggest that to be effective, such interventions will need to be tailored to the infants’ gestational and perhaps postmenstrual age. Moreover, part of the strategy may need to provide the newborn with basic parental voice exposure or other experiences that would have been taken for granted in utero. We anticipate that a greater quantitative understanding of sleep-wake patterns and their associations with the NICU acoustic environment will lead to intervention studies that are focused on improvements in neonatal sleep regulation and ultimately achievement of better neurodevelopmental outcomes for these highly vulnerable patients. More broadly, such research also could open opportunities to improve long-term outcomes for all infants aside from those who require a NICU stay. Conceivably, the timing and volume of noise and parental language exposure and their influence on sleep obtained during the first days of life outside the uterus could have important implications for every new family.

We thank the participating infants and parents. We also thank the research assistants Stephanie Rau, CCRP, and Shannon Lester, CCRP, and the sleep technologists, especially Mark Kingen, RPSGT, and Laura Merley, RPSGT, whose tireless and careful work made this project possible.

LENA

Language Environment Analysis

NREM

nonrapid eye movement

REM

rapid eye movement