Definition, outputs, and diagnostic boundary
A respiratory event detection model examines airflow, respiratory effort, oxygen saturation, EEG arousals, snoring, or proxy sensor signals during sleep to localize candidate breathing abnormalities. It may distinguish apneas, hypopneas, respiratory effort-related arousals, periodic breathing, or another predefined event set. Typical outputs include onset and end time, type, confidence, evidence channels, and indeterminate intervals, from which an hourly index can be calculated or reviewable annotations can be created.
Event detection, event typing, severity indices, and diagnosis are separate levels. A model may estimate whole-night AHI or REI well while making errors in individual boundaries, hypopneas, or obstructive-versus-central classification. Conversely, strong local overlap does not ensure that sleep time or valid monitoring time used as the denominator is correct. Diagnosis also depends on symptoms, history, recording type, technical quality, and professional interpretation.
This entry includes automatic and assisted respiratory-event recognition in PSG and home sleep monitoring. It is not synonymous with oxygen-desaturation detection, snore analysis alone, sleep staging, respiratory-rate measurement, or the internal control algorithm of a ventilatory device. When input channels are absent, output must remain within what the available evidence supports. Without thoracoabdominal effort, for example, a model cannot assume equivalence to full-PSG classification of obstructive, central, and mixed events.
Input channels and temporal alignment
Laboratory PSG commonly displays nasal and oral airflow, thoracic and abdominal effort, oxygen saturation, sleep stages, and arousals, with possible position, snoring, ECG, video, or carbon-dioxide channels. Each sensor answers a different question: flow shows ventilation change, effort helps determine whether drive persists, oximetry shows a delayed consequence, and EEG establishes sleep and arousal context. A model should declare required, optional, and substitute channels rather than treat arbitrary combinations as equivalent.
Multichannel time synchronization is fundamental to attribution. Desaturation may lag a flow change, while device averaging and internal filters alter its waveform. PAP flow, nasal pressure, thermal airflow, and substitute effort signals have different dynamic ranges. Import, resampling, and cross-device fusion must preserve the original time base, delay corrections, packet losses, and clock restarts so that unrelated desaturations or arousals are not attached to the same event.
Quality gating should detect displaced cannulas, oral breathing, loose or inverted effort belts, oximeter motion, low perfusion, clipping, baseline drift, PAP leak, and sensor saturation. Bad intervals need affected-channel and duration labels and a defined degradation, rejection, or manual-review response. Smoothing, interpolation, or substitution from another channel may serve a particular algorithm, but it must remain distinguishable from measurement and must not manufacture an apparently complete overnight record.
Event logic and model implementation
Rule systems can form candidates from baseline, amplitude reduction, duration, recovery, oxygen change, and arousal. Machine-learning systems may learn event probability from raw waveforms, time-frequency features, or sequential windows. Practical implementations often separate quality assessment, candidate generation, boundary refinement, type classification, and conflict resolution. Whether rules, trees, or deep networks are used, each final event should be traceable to its inputs and rule version.
Respiratory events occupy continuous time rather than independent fixed windows. The model must handle neighboring events, brief recovery, overlapping desaturations, long events, periodic patterns, and recording boundaries, with explicit rules for merging and splitting candidates. Predicting “event present” per window and joining positives can duplicate counts or lose boundaries. Event-matching evaluation must likewise define allowed overlap and timing tolerance before results are calculated.
Type decisions depend on cross-channel relationships. A marked flow reduction with continuing effort, simultaneous loss of effort, or a changing pattern within an event supports different types, while hypopnea may additionally depend on oxygen or arousal criteria. A model should not force a class from population priors when essential evidence is missing; unclassified or low-confidence output is appropriate. When channels conflict, sensors fail, or rules disagree, preserving the candidate and evidence is safer than emitting a certain but unauditable label.
Reference rules, labels, and data governance
Training and validation must identify the scoring manual and version, epoch conventions, adult or pediatric rules, diagnostic or titration setting, hypopnea definition, sensor definitions, and event denominator. Different rules can change event counts and AHI, RDI, or REI for the same night, so labels cannot be reused independently of their rule version. Home monitoring often uses recording or monitoring time as the denominator, which should not be presented as the same index as PSG based on sleep time.
Human references are not perfectly consistent, with disagreements concentrated in hypopneas, event boundaries, and apnea type. Higher-quality datasets use scorer training, blinded independent scoring, calibration sets, adjudication, or multiscorer consensus and retain each revision. When labels come from routine clinical reports, the initial technologist score, physician-finalized score, and later correction should be distinguished so that versions from one night are not treated as independent truths.
Splits should be participant-based and should check repeat studies, pre- and post-treatment records, duplicate exports, and overlap across databases. Documentation should cover age, sex, body size, severity, central events, cardiopulmonary and neuromuscular disease, supplemental oxygen, PAP, medication, position, center, device, and signal missingness. Selecting only high-risk referrals changes event prevalence and positive predictive value, so performance cannot be assumed in a general screening population.
Performance validation and clinical indices
Event-level evaluation should report sensitivity, positive predictive value, F1, false positives per hour, misses, and type confusion, with the event-matching method stated. Because negative windows greatly outnumber events, accuracy or specificity may appear excellent while providing little information. Correlation alone can also hide systematic bias. Record-level mean difference and limits of agreement for AHI, RDI, or REI, severity-category confusion, and errors near decision thresholds are also required.
Validation data should be independent of training and tuning and include external centers, different acquisition systems, and the intended workflow. Beyond aggregate results, analyses should stratify signal quality, position, sleep stage, event density, disorder type, age, sex, body size, skin-tone-related oximetry performance, supplemental oxygen, and PAP when relevant. Pediatric versus adult rules and laboratory PSG versus home monitoring cannot be interchanged without corresponding validation.
Clinical-impact evaluation should examine final accuracy after human review, review time, proportions of deleted and added events, and changes in severity classification or management. A small number of events can move a person across an index threshold, making boundary cases a separate review group. Performance near the average of one scorer does not equal multiexpert consensus and does not establish that diagnostic reports can be generated without supervision.
Human review, failure, and monitoring
A workstation should let reviewers see original flow, effort, oxygen, sleep stage, arousals, and enough surrounding context, and edit onset, end, type, or delete a candidate. Low confidence, channel conflicts, long missing intervals, index values near thresholds, and suspected central events can be prioritized, while high-confidence output remains available for sampling. Automatic annotations, human edits, user, time, and final sign-off need separate audit records.
Common failures include interpreting movement or cough as reduced flow, oral breathing as apnea, a detached effort belt as a central event, or oximetry artifact as support for hypopnea, and choosing a wrong baseline during dense events. Arrhythmia, chronic lung disease, neuromuscular disease, periodic breathing, low perfusion, supplemental oxygen, and PAP leak also alter features. The system should expose quality and applicability instead of hiding causes behind one risk score.
Postdeployment monitoring should track missingness by channel, rejection, event counts and type distributions, human edits, AHI or REI bias, center and device differences, and subgroup performance. Model, threshold, preprocessing, oximetry-delay correction, sensor, and scoring-rule changes require impact analysis, regression testing, and controlled release. Real changes in referral patterns or population can also shift event distributions and must be distinguished from software degradation.
Selection, regulation, and development
Selection should verify intended use, eligible ages and conditions, required channels, compatible devices, scoring rules, event types, denominator, indeterminate policy, external validation, manual editing, export, audit, and update support. Suppliers should provide a version fingerprint, test plan, known limitations, signal-quality requirements, and reproducible examples. One whole-night index, a single F1 score, or an “automatic diagnosis” claim cannot replace event-level and workflow evidence.
If the model supports diagnosis, clinical reporting, treatment adjustment, or a closed-loop device, regulatory status depends on the exact function, claims, risk, and jurisdiction. Original sleep signals, oximetry, diagnostic labels, and schedules are sensitive health information requiring purpose limitation, minimization, encryption, access audit, retention periods, and deletion. Connected deployment also needs cybersecurity, rollback, and incident response. Open-source or cloud delivery does not automatically remove these responsibilities.
Development directions include multimodal end-to-end event localization, probabilistic boundaries, rule-aware learning, cross-device domain adaptation, weakly supervised annotation, and uncertainty ranking for human review. These methods may improve efficiency and transfer, but must preserve rule versions, original evidence, and independent external validation. Related entries include Apnea-Hypopnea Event Detection Algorithms, Oxygen Desaturation Event Detection Algorithms, EEG Microarousal Detection Algorithms, Sleep Apnea Risk Screening Models, and Nondiagnostic Snoring and Breathing Risk Detection Algorithms.