Definition, Uses, and Scope
An HST annotated dataset combines sleep-testing signals collected in a home or another setting outside a sleep laboratory with time-aligned labels for respiratory events, oxygen changes, recording quality, body position, or other observations. In clinical contexts, HST commonly refers to home sleep apnea testing, also abbreviated HSAT. Such data can support research on portable testing, scorer training, device validation, algorithm development, and analysis of real-world technical quality. Its defining feature is the coexistence of raw data, acquisition context, and traceable annotations.
HST is not one device class and is not synonymous with laboratory polysomnography, any wearable sleep report, or an isolated night of oximetry. Different systems measure substantially different physiological variables and therefore support different events and clinical questions. A file containing vendor-generated result fields does not necessarily constitute an independently reviewable annotated dataset. Users should establish whether raw signals, label provenance, device configuration, valid intervals, scoring rules, and any reference examination are available.
Devices, Signals, and the Home Acquisition Context
Home systems may record nasal pressure or other airflow, thoracic and abdominal effort, oxygen saturation, and pulse rate; some use peripheral arterial tone, actigraphy, position, snoring, or limited electrocardiography. Certain systems estimate sleep and wake, while others lack electroencephalography, electrooculography, and electromyography and cannot directly determine sleep stages or arousals in the same manner as PSG. A dataset must list sensors, placement, sampling rate, units, filters, firmware, and algorithm version for each device rather than using “HST” as a complete channel description.
The home setting reduces laboratory intervention but introduces self-application, sensor displacement, leaving bed, temporary device removal, use by the wrong household member, and interrupted data transfer. Documentation should capture equipment issue and return, user instruction, start and stop actions, number of attempted nights, reasons for repeat testing, and technical support. Recording start, time in bed, lights out, estimated sleep onset, valid monitoring, and rising time are different concepts. Retaining only cropped waveforms without these landmarks limits denominator calculation and analysis of failures.
Annotation Types, Metric Denominators, and Meaning
Common annotations include obstructive or central respiratory events, hypopneas, oxygen desaturations, pulse or position changes, snoring, signal artifact, and uninterpretable intervals. A label may describe an exact onset and duration, a fixed-window class, whole-night severity, or a technical-quality grade; these are not interchangeable without a documented transformation. Human scores, device-generated results, physician review, labels transferred from paired PSG, and research-team reannotations should remain separate layers, with a stated method for resolving conflicts.
When reliable sleep staging is unavailable, event rates often use valid recording or monitoring time as their denominator, whereas a PSG apnea-hypopnea index generally uses sleep time. A longer recording denominator can dilute the value. The data dictionary should define each metric's events, denominator, required signals, treatment of invalid intervals, and rounding, and should distinguish REI, AHI, ODI, and proprietary indices. No label, no event, channel failure, and not scored are four different states and must not all be encoded as zero.
Scoring Protocols, References, and Data Harmonization
Respiratory-event definitions vary with scoring-manual edition, sensor type, desaturation or arousal association, and study protocol. A dataset should identify the rule set and version, scoring software, visible channels, automated prescoring settings, scorer qualifications, review, and adjudication. If PSG is the reference, documentation should say whether the tests were simultaneous or performed on different nights, how timelines were aligned, and whether home setting, body position, or natural night-to-night variation could explain disagreement. Different-night results cannot be treated simply as label error.
Harmonization across devices or centers should preserve source fields and document channel mapping, unit conversion, resampling, filtering, time correction, event conversion, and non-comparable items. Conversion from a proprietary format into EDF, CSV, or another open form should be checked for loss of calibration, missing-value meaning, event boundaries, and device status flags. Every release needs a version, file checksums, data dictionary, device inventory, inclusion criteria, change log, and known-issues list so that an analysis can identify the exact snapshot used.
Technical Quality, Label Uncertainty, and Selection Bias
Quality review should examine every channel and interval for missing signal, clipping, drift, low perfusion, motion artifact, detached airflow or effort sensors, clock offsets, and depleted batteries. Technically inadequate, excessively short, or indeterminate recordings must not be silently assigned to the negative class, nor should a device summary alone make a record valid. A dataset should report failure rates, repeat-test rules, exclusions, manual repairs, and quality grades, while retaining appropriate failed examples for research on robustness in real deployment.
Home-testing data generally come from people already selected to receive a particular examination and may underrepresent those who were not referred, could not self-apply sensors, had complex comorbidities, or presented atypically. Keeping only successful uploads and scorable records creates survivorship bias, while reliance on one device, center, payment system, or age range limits generalization. Limited channels and borderline rules add label uncertainty. Studies should report the unscored fraction, scorer agreement, and results stratified by device, population, and signal quality.
Machine-learning Splits, Evaluation, and Reproducibility
Training, validation, and test sets must be divided by participant, with a person's repeat nights, source PSG, cropped segments, and derived features retained in the same split. If device serial numbers, site codes, report templates, or file duration correlate with labels, a model may learn the acquisition workflow instead of respiratory physiology. Preprocessing, imputation, quality filtering, normalization, and threshold tuning must use training information only. The test set should remain independent, with external validation on different devices, centers, and target populations where feasible.
Evaluation should match label granularity and separately report performance for event detection, window classification, whole-night severity, and technical-failure recognition. Beyond aggregate scores, useful results include sensitivity, specificity, calibration, temporal overlap, error distributions, and the uninterpretable rate, all tied to the stated reference. The data release, sample manifest, random seeds, software dependencies, signal processing, and exclusion decisions should be reproducible. A firmware or device-algorithm update requires renewed evaluation; an unchanged field name does not guarantee unchanged behavior.
Privacy, Access Governance, and Clinical Limits
Home records can contain exact dates, household routines, location, device identifiers, free text, and linkable physiological patterns. Publishers should apply data minimization, date handling, de-identification, and re-identification risk assessment consistent with consent and applicable rules, using open, application-based, or controlled access according to sensitivity. Users must protect downloaded files, credentials, and derived data; obey license, no-reidentification, and redistribution restrictions; and follow required correction or deletion procedures when consent is withdrawn, access expires, or errors are discovered.
Annotated data can support research and performance evaluation, but do not automatically make a trained model an authorized medical tool, and a single automated result cannot replace a complete clinical assessment. Appropriate candidates and follow-up depend on device capability, patient circumstances, local standards, and professional judgment. In applicable care pathways, a negative, inconclusive, or technically inadequate home test may still require PSG or another evaluation. A mature dataset presents successes and failures, human and automated labels, and raw and derived information so that its evidentiary limits remain visible.