World Sleep Industry ENCYCLOPEDIA

Respiratory Event Annotated Datasets

4.4.3

Definition, Uses, and Scope

A respiratory-event annotated dataset organizes sleep-period airflow, respiratory effort, oxygen saturation, and other synchronized physiological signals together with time-aligned labels for apneas, hypopneas, respiratory-effort-related arousals, desaturations, periodic breathing, or related events. It can support sleep-breathing research, scorer education, device and algorithm validation, quality control, and reproducible methods. The essential content is not only an event name but also the supporting signals, scoring rules, temporal boundaries, and recording context.

This resource differs from a summary table containing only a nightly AHI, diagnostic conclusion, or severity class. Event-level data usually require an onset and end or duration and must distinguish no event, not scored, indeterminate, and signal failure. A label is not absolute truth independent of sensors and rules: the same interval can receive a different result under another scoring system, device capability, or population-specific rule. Users must establish the target event and evidentiary limits first.

Signal Channels, Sensor Evidence, and Time Synchronization

Event decisions may use nasal pressure, thermal airflow, positive-airway-pressure flow, thoracic and abdominal effort, oximetry, EEG arousal, heart rate, body position, or snoring. Detecting reduced airflow and classifying an event as obstructive, central, or mixed require different evidence. Without effort signals, some types generally cannot be separated reliably; without EEG, an arousal association cannot be established in the same way as PSG. A dataset should list channels actually present in each record rather than merely saying “PSG” or “HSAT.”

Every channel also needs sensor type, placement, sampling rate, units, filters, calibration, device and firmware, and valid or invalid intervals. Separate device clocks, oximeter processing delay, resampling, and format conversion can misalign an event with its physiological consequence. Publishers should describe the synchronization reference, clock corrections, allowed uncertainty, and handling of events that cross file boundaries, and should inspect examples to confirm that annotations actually align with the source waveforms.

Event Classes, Layers, and Annotation Meaning

A vocabulary may include obstructive, central, and mixed apneas; hypopneas; respiratory-effort-related arousals; periodic breathing; desaturations; and flow limitation, but not every dataset can support every class. Events may overlap or be associated, as when a respiratory event is followed by desaturation or arousal. They should remain separate, linkable layers rather than one ambiguous label. Start, end, duration, peak change, and identifiers for associated events all need explicit definitions.

Human scores, automated prescores, physician review, consensus results, and algorithmically regenerated labels should occupy different fields with an edit history. Fixed-window presence labels, epoch event codes, and precise interval annotations have different temporal resolution and cannot be exchanged directly. If an event spans several epochs, the derivation of window labels must be stated; if several events occupy one interval, a table limited to one class must not silently discard the others.

Scoring Rules, Versions, Indices, and Harmonization

Respiratory scoring changes with manual edition, adult or pediatric population, diagnostic or treatment setting, sensor, and study protocol. Hypopnea requirements concerning airflow reduction, duration, desaturation, or arousal association particularly need explicit documentation; “scored by standard rules” is insufficient. Datasets should retain the original rule and edition, software, display settings, scorer qualifications, blinding, review, and adjudication. Cross-dataset harmonization should preserve source labels and non-convertible cases.

Nightly indices must define their event set and denominator. A PSG AHI generally uses sleep time, while home testing without reliable staging may use valid monitoring time to produce an REI; an ODI further depends on the desaturation definition. Values produced under different rules should not be merged merely because field names look alike. Version management needs a sample manifest, data dictionary, checksums, transformation scripts, change log, and known issues so that the exact annotation release can be reproduced.

Signal Quality, Ambiguous Events, and Selection Bias

Common quality problems include displaced airflow sensors, detached or inverted effort belts, low-perfusion or artifactual oximetry, baseline drift, clipping, clock offsets, and long missing intervals. A technically inadequate interval must not be encoded as no event, and a generated nightly summary does not make an incomplete record valid. Datasets should provide channel-level quality flags, unscorable intervals, repeat or exclusion reasons, and enough failed examples to evaluate robustness in real deployment.

Borderline events, incomplete recovery between repeated events, and limited channels create scorer disagreement. Annotation guidance, multi-rater agreement, adjudication, and uncertainty rates should be reported. Referral cohorts may contain many obstructive events but few normal nights, central events, children, or people with complex comorbidities; retaining only high-quality scorable records creates further selection bias. A large event count does not guarantee enough independent participants, so distributions should be described by population, center, device, sleep stage, and position.

Machine-learning Splits, Matching Evaluation, and Reproducibility

Training, validation, and testing should be partitioned by participant, keeping a person's repeat nights, related PSG and HSAT, neighboring segments, and derived features together. Random window splitting leaks individual, device, and nightly background information. Preprocessing, imputation, quality selection, class sampling, and thresholds must use training data only. The test set should remain independent, with external validation across centers, devices, condition spectra, and populations where feasible.

Event evaluation cannot rely only on window accuracy. The temporal matching of predictions to reference events, overlap criterion, one-to-many pairing, and boundary tolerance should be predefined, followed by event sensitivity, positive predictive value, false-alarm rate, duration error, and bias in nightly indices. Rare classes need separate reporting, and uninterpretable outputs must count. Reproducibility materials should preserve data and annotation releases, partition manifests, dependencies, random seeds, preprocessing, postprocessing, and final thresholds.

Privacy, Access Governance, and Clinical Limits

Whole-night physiological signals and event sequences can reveal conditions, treatment status, routines, and distinctive physiology, while files may also contain dates, device identifiers, sites, or free text. Publishers should perform data minimization, de-identification, and re-identification risk assessment consistent with consent and applicable rules, assigning open, application-based, or controlled access according to sensitivity. Users must protect downloaded copies, credentials, and derived labels; comply with license and restrictions on re-identification or redistribution; and handle withdrawal, correction, and access expiry.

Event labels can support research and algorithm evaluation but cannot by themselves diagnose, exclude disease, or determine treatment for an individual after one automated analysis. Diagnostic meaning depends on examination completeness, applicable population, sleep evaluation, and professional interpretation; a negative, indeterminate, or technically inadequate record is not equivalent to normal. Clinical systems additionally require validation of failure detection, human review, device changes, and applicable regulation. A mature dataset presents event evidence, quality limitations, and unanswered questions together.