World Sleep Industry ENCYCLOPEDIA

Infant Sleep Datasets

4.4.3

Definition, scope, and developmental specificity

An infant and toddler sleep dataset organizes sleep-wake state, physiological signals, activity, caregiving events, sleep environment, and health context collected from newborns, infants, or young children. Data may originate in sleep laboratories, neonatal or pediatric wards, outpatient cohorts, or longitudinal home studies. They can support research on sleep development, emergence of circadian organization, associations between sleep and health, device performance, and algorithm generalization. Every resource should state its own age boundaries rather than using “pediatric” or “infant” as a substitute for precise age information.

Sleep architecture, cycle length, distribution across day and night, and observable behavior change rapidly during this period. A preterm infant, a term newborn, an older infant, and a toddler are not one statistical population. Hospitalized patients, clinically referred children, and healthy home cohorts also address different questions. Adult night boundaries, stage definitions, and algorithm thresholds cannot be transferred without age-appropriate validation. The value of the dataset therefore depends not only on signal volume but on preservation of developmental stage and recording context.

Age, cohort, and caregiving context

Core variables should include gestational age at birth, a protected basis for birth date or age, chronological age at recording, and postmenstrual or corrected age where applicable to preterm children. Units and calculation rules must be explicit. Birth and hospitalization history, major conditions, medications, feeding, naps, overnight feeds, soothing, room arrangements, and unusual events on the recording day may also be important. These variables can change between visits; a baseline value should not be copied silently across every later night.

Cohort documentation should describe recruitment setting, inclusion and exclusion criteria, health or clinical status, planned and completed follow-up visits, nights per child, withdrawals, and loss to follow-up. Family resources, housing, care practices, cultural context, caregiver work schedules, and access to devices may influence both participation and recorded sleep, but an observed association is not automatically causal. Children, caregivers, households, visits, and recording nights should be represented as linked but distinct levels so that longitudinal analyses do not count repeated observations as independent people.

Modalities, synchronization, and recording conditions

Data may include EEG, EOG, chin EMG, ECG, airflow and respiratory effort, oxygen saturation, heart rate, movement, actigraphy, skin temperature, audio, infrared or color video, mattress pressure, and contactless sensing. Each channel needs a device name, hardware and software version, sensor placement, sampling rate, units, reference scheme, filtering, timestamp definition, and valid interval. An actigraph on the ankle, wrist, or trunk produces different observations. Hospital polysomnography and home video are likewise not interchangeable measures.

Infant recordings are often interrupted by feeding, diaper changes, soothing, examination, electrode replacement, and caregiver entry. Event logs should mark these actions as well as lights-off or placement in bed, observed sleep onset, removal from bed, sensor loss, and reconnection. Synchronization across devices requires a shared time source or a documented offset, time zone, clock-drift assessment, resampling method, and logic across midnight. When video or diary times are only supporting evidence, this should be stated; coarse caregiver recall must not conceal timing error in physiological signals.

Sleep labels, references, and annotation quality

Labels may use infant states such as active sleep, quiet sleep, indeterminate or transitional sleep; standard stage categories may be appropriate at particular ages and under suitable recording conditions; some resources label only sleep and wake. Publishers must identify the applicable age range, scoring manual and version, epoch duration, signal montage, event definitions, and treatment of unscorable periods. Technician-scored PSG or EEG, actigraphy algorithms, video behavior ratings, caregiver diaries, and consumer-device outputs have different evidentiary properties and should not all be called a gold standard.

Annotation metadata should preserve scorer qualifications, training, blinding, single-scorer or consensus procedure, adjudication rules, and agreement by age stratum. Labels created under different rules should remain distinguishable instead of being harmonized by name alone. Apnea, hypopnea, desaturation, arousal, movement, and feeding events each need a definition and timing convention. Using the output of the device under evaluation as its own training reference creates circular validation. An uncertain interval should receive an explicit unknown or unscorable code rather than being forced into the nearest stage.

Quality control, missingness, and representation

Delicate skin, changing head circumference, clothing, turning, perspiration, low perfusion, caregiving, and multiple people in a video frame can all affect quality. Quality control should distinguish absent signal, detached sensors, severe artifact, nonwear, care interruption, removal from bed, and algorithm rejection. Valid duration should be reported for each channel and night. Zero filling, smoothing, or carrying observations forward cannot transform missing data into true sleep. Exclusion, manual repair, realignment, relabeling, and corrected files require traceable versions.

Retaining only families able to complete multi-night wear, maintain network connectivity, or share home video creates adherence and access bias. Clinical PSG, by contrast, can overrepresent symptomatic children and comorbid conditions. Reports should separately count recruited children, recorded children, independent households, total nights, valid nights, and observations within each developmental stratum. Missingness and error should be examined by prematurity, illness, sex, socioeconomic conditions, site, and device where justified and available. Millions of windows do not imply millions of independent children.

Machine-learning evaluation and reproducibility

Training, validation, and test partitions should be separated by child. Repeated nights, visits at different ages, overlapping windows, and multimodal views from one child must stay in the same partition. Siblings, twins, and children sharing a care environment can also leak household-specific information, so household-level grouping is often the safer design. Normalization, imputation, quality filters, feature selection, and thresholds must be learned from training data only. Useful evaluations test transfer across age bands, sites, devices, or home environments rather than applying one developmental model to all infants and toddlers.

Metrics should match the intended task. Sleep-wake and stage models need class-specific sensitivity, specificity, confusion patterns, and agreement. Sleep duration, nap, and awakening estimates need bias, boundary error, error distributions, and limits of agreement. Alerting systems also require missed-event rates, false alarms per hour, failure-to-output rates, and caregiver alarm burden. Results should be stratified by age, health status, signal quality, and setting with uncertainty calculated at the child level. Reproduction packages should fix dataset and label versions, household-level splits, preprocessing, software environment, random seeds, final thresholds, and a locked test set.

Children's rights, household privacy, and safety boundaries

Infants and toddlers cannot provide legally effective research consent. Collection and reuse require applicable ethics review and parental or guardian permission, with later continued use, withdrawal, and age-appropriate assent considered as the child matures under the governing rules. Audio and video may incidentally capture caregivers, siblings, conversations, and private household spaces. Collection scope and retention should be minimized, identity keys separated, facial display and commercial reuse restricted, cross-purpose linkage controlled, and access matched to sensitivity. Permission and licensing must address derived features and models as well as raw files.

Accuracy on infant monitoring data does not show that a device can prevent, detect, or reduce sudden infant death syndrome or other sleep-related deaths, and monitoring does not replace safe-sleep guidance. Research instrumentation, consumer home monitors, and medical monitoring prescribed for a particular condition have different intended uses and evidence requirements. Respiratory, oxygen, or heart-rate alerts require validation in the target age and clinical population, documented failure modes, a clear escalation path, and qualified interpretation. Persistent breathing abnormalities, cyanosis, feeding difficulty, or another urgent concern calls for timely medical assessment rather than independent reliance on a dataset model.