World Sleep Industry ENCYCLOPEDIA

Sleep Respiratory Risk Prediction Models

4.4.4

1. Definition, prediction targets, and classification boundaries

A sleep respiratory risk prediction model uses information available before its prediction time—demographics and history, symptom questionnaires, and nighttime airflow, respiratory effort, oxygen saturation, pulse, sound, position, or movement—to estimate the probability that a person has sleep-related breathing abnormalities, falls within a specified severity range, or needs further testing. Outputs may include a participant-level screening probability, an interval for one night's respiratory-event burden, a risk of deterioration over time, or a flag that a recording requires human review. The model must define its target, threshold, time frame, intended population, and data sufficiency rather than offer an unexplained “breathing health score.”

Risk prediction differs from event detection and from clinical diagnosis. Event detection marks possible apneas, hypopneas, or desaturations within a recording. Risk prediction combines evidence into a future or person-level probability. Diagnosis additionally integrates symptoms, history, test quality, professional interpretation, and differential diagnosis. A consumer-device alert for possible signs of moderate-to-severe obstructive sleep apnea does not by itself establish a diagnosis and must not directly determine treatment.

Sleep-related breathing problems include obstructive apnea and hypopnea, central sleep apnea, sleep-related hypoventilation, sleep-related hypoxemia, and primary snoring. Their mechanisms, required signals, and care pathways differ. Cyclic oxygen changes alone usually cannot reliably separate obstructive from central events and may miss events terminated by arousal without substantial desaturation. A model should restrict its claim to the labels it can support rather than grouping every abnormality under “sleep apnea.”

2. Physiology, event labels, and expressions of severity

Obstructive events arise from repeated narrowing or collapse of the upper airway during sleep; respiratory effort may persist or increase as airflow falls. Central events involve a temporary reduction of ventilatory drive, so airflow and effort may disappear together. Events can cause intermittent hypoxemia, carbon-dioxide changes, autonomic activation, and electroencephalographic arousal, but those consequences do not occur together in every event. Obesity, craniofacial anatomy, supine position, rapid-eye-movement sleep, nasal obstruction, alcohol, sedatives, altitude, heart failure, opioids, and neuromuscular disease can alter risk through different pathways.

Polysomnography calculates the apnea–hypopnea index from the number of apneas and hypopneas per hour of actual sleep. Home tests without electroencephalography usually cannot determine total sleep time precisely and often use monitoring time or estimated sleep time as the denominator, producing a respiratory-event index. Including wake in that denominator can underestimate event frequency. Labels and severity also change with the airflow-reduction, desaturation, or arousal definition for hypopnea and with use of a 3% versus 4% desaturation rule. Training and deployment must therefore use compatible definitions.

Event frequency does not capture the full physiological impact. The same index can represent different event duration, oxygen nadir, time below an oxygen threshold, event-related hypoxic burden, arousal intensity, dependence on rapid-eye-movement sleep or supine position, and daytime symptoms. A model may estimate these dimensions separately but should not claim that one alternative metric replaces clinical interpretation. Sustained hypoxemia can also result from chronic lung disease, obesity hypoventilation, or altitude rather than typical cyclic upper-airway obstruction; the model must preserve uncertainty about such differential explanations.

3. Data sources, reference standards, and quality control

Laboratory polysomnography can record electroencephalography, eye movements, muscle activity, airflow, thoracoabdominal effort, oxygen saturation, electrocardiography, snoring, and position, making it an important reference for detailed event labels. Home sleep apnea testing generally uses fewer channels and fits selected adult settings after clinical assessment, but it does not cover every patient. Lighter inputs include finger or wrist oximetry, photoplethysmography, electrocardiography, microphones, phones, radar, mattress pressure, accelerometry, and questionnaires. Each observes only part of the physiology; multimodal fusion cannot restore a missing physiological definition.

Oximetry is affected by low perfusion, motion, loose wear, ambient light, skin temperature, smoking, carboxyhemoglobin, nail products, and device algorithms. Evidence indicates that some pulse oximeters can show systematic error in darker skin tones and at lower saturation, so evaluation should cover objectively measured skin-tone ranges, devices, and saturation ranges rather than treating race as a physiological correction factor. Snoring is not synonymous with obstruction, and audio can include a bed partner, television, traffic, or air conditioning. Radar and mattress sensors may similarly assign a partner, pet, or turn in bed to the target person's breathing.

Collection pipelines should retain original sampling rate, clock, time zone, valid recording time, placement, firmware, and algorithm version, with flags for disconnection, saturation, clipping, sensor loss, and manual edits. Qualified scorers blinded to model output should establish training labels under a prespecified rule, with inter-scorer agreement reported. Missing channels, short recordings, or poor quality should yield an “unable to assess” state and a reason. Filling the gap with age, body weight, or prior diagnosis to produce an apparently precise nightly result conceals data insufficiency.

4. Modeling, multinight information, and personalization

Questionnaires and clinical rules support inexpensive initial screening, but high sensitivity often comes with many false positives and cannot establish diagnosis. Signal-based models may use oxygen-desaturation counts and morphology, pulse-wave and heart-rate-variability features, audio time–frequency features, breathing cycles, and position, then apply logistic regression, gradient boosting, convolutional networks, temporal models, or probabilistic graphical models. A structured model can jointly output whether a threshold is reached, an index range, hypoxic burden, and data quality, while supporting segments make professional review easier.

Models must distinguish event-level classification, whole-night severity estimation, and participant-level screening. Randomly placing adjacent segments from the same person in training and test sets leaks device signatures and personal breathing patterns. Using the night's final oxygen nadir to predict an earlier event is temporal leakage. A bedtime forecast of that night's risk can use only pre-sleep information. A model updated during the night must run in actual signal-arrival order and cannot borrow a post-event recovery breath or later oxygen rebound.

Sleep-breathing severity varies from night to night. Supine time, rapid-eye-movement sleep, nasal congestion, alcohol, medication, and altitude can move one recording across a classification boundary. A multinight model should retain nightly estimates and uncertainty before estimating an individual distribution rather than simply taking the maximum or mean. Longitudinal baselines can identify sustained change, but increasing weight, progressing cardiopulmonary disease, and device drift must not be absorbed as a new normal. When a special population or signal pattern lies outside the training distribution, the model should abstain and direct the user to an appropriate test.

5. Validation, metrics, and fairness

Development data should be divided by participant, followed by external validation across time, centers, home settings, and devices. The index and reference tests should ideally be recorded simultaneously so that genuine night-to-night change is not counted as algorithm error; separate multinight studies should quantify natural variability. Enrollment limited to highly suspected sleep-clinic patients creates spectrum bias, and positive predictive value can fall sharply in primary care or general populations. Reports should disclose exclusions, invalid studies, and the proportion ultimately analyzed rather than presenting performance only for ideal usable samples.

Participant-level thresholds require sensitivity, specificity, positive and negative predictive values, likelihood ratios, calibration, decision analysis, and confidence intervals. Severity estimates also require absolute error, Bland–Altman agreement, confusion across clinical boundaries, and systematic over- or underestimation. Event models need false alarms per hour, matching tolerance, and performance by event type and sleep stage. With class imbalance, total accuracy and area under a curve can look strong while concealing missed central events, mild disease, or clinically important hypoxemia.

Performance should be stratified by age, sex, body size, objectively measured skin-tone range, pregnancy, cardiopulmonary comorbidity, device, sleep position, and home environment. Children, pregnant people, those with neuromuscular or chronic lung disease, heart failure, opioid exposure, suspected hypoventilation, or central apnea need dedicated evidence; an adult obstructive-apnea model cannot simply be transferred. Fairness means more than similar area under the curve: investigators should test whether one threshold creates unequal missed-case rates or invalid recordings and whether downstream confirmatory care is actually accessible.

6. Clinical workflow, applications, and safety limits

Appropriate applications include previsit screening, prioritization for testing, multinight home trends, prompts for reassessment after treatment, and clinical cohort management. An output should present the probability, target definition, data quality, supporting evidence, and recommended next step together. A positive screen usually calls for professional evaluation or standardized testing under the applicable pathway. A negative screen cannot exclude disease when symptoms persist, clinical suspicion is high, or the study is invalid. Automated interpretation should permit raw-data review and documented human correction.

The model must not independently prescribe or adjust positive-airway pressure, oxygen, medication, surgery, or an oral appliance, nor determine fitness to drive, undergo anesthesia, or perform safety-sensitive work. Screening asymptomatic populations, managing a diagnosed condition, and providing acute monitoring are separate intended uses that require different evidence and oversight. If a consumer function is authorized only to flag certain signs in a defined population, it must not be marketed as a diagnosis of every apnea type, real-time respiratory monitoring, or an emergency alarm.

Severe breathing difficulty, chest pain, cyanosis, fainting, altered consciousness, or another emergency requires immediate action through local emergency services rather than waiting for an overnight analysis. Any isolated home oxygen reading must be interpreted with symptoms, signal quality, and device instructions. Systems should predefine who responds to invalid recordings, critical findings, repeated failure, and loss to follow-up, and prevent overnight alerts from disrupting sleep unnecessarily. A risk score is a triage aid, not a reason to transfer responsibility for care to the user.

7. Privacy, governance, and future directions

Nighttime audio, video, bedroom activity, and physiological signals may capture the user, a bed partner, and the home environment. Products should limit collection to the stated purpose, favor on-device extraction of necessary features, minimize raw audio and video retention, and define partner consent, access, export, deletion, and secondary-research rules. Clinical systems must also protect identity, test results, and treatment information, with encryption, authentication, and audit trails covering transmission, offline caches, and model updates.

A model used for screening or clinical decisions may qualify as software as a medical device. Its intended use, risk class, quality system, cybersecurity, and clinical-performance evidence should follow local requirements. A model card should identify the training population, reference standard, event definition, devices and versions, invalid-study rules, known limitations, and contraindicated settings. Postdeployment monitoring should cover prevalence shifts, sensor firmware, scoring rules, home connectivity, referral completion, and subgroup calibration. Material updates require versioning and revalidation; historical results must not be silently recomputed.

Future work should test models prospectively across multiple nights, devices, and lower-prevalence real-world settings. Combining airflow, effort, oxygen, autonomic, and audio information may better distinguish obstruction, central pauses, hypoventilation, and nonrespiratory artifact. More clinically useful systems will present event frequency, hypoxic burden, arousals, symptoms, and comorbidity risk as separate dimensions and test whether forecast-guided triage shortens time to diagnosis without widening inequity. Progress means not more alerts, but auditable evidence that connects the right person to the right test and care safely.