Artificial intelligence (AI) has progressed rapidly from concept to operative tool across respiratory and sleep medicine. The convergence of large, annotated sleep datasets, expanding home-based diagnostics, and increasing consumer awareness of physiologic wearable data has created new opportunities to deploy algorithmic frameworks for treatment optimization. Simultaneously, the excitement around AI often outpaces its evidence base. Therefore, it is incumbent on sleep clinicians, pulmonologists, and integrated respiratory providers to identify where AI meaningfully augments patient care, and where it remains investigational.
This review synthesizes current evidence on AI applications relevant to sleep-disordered breathing. It also outlines how national home-respiratory providers can pragmatically support responsible deployment of these tools within multidisciplinary sleep care.
AI-ENABLES MASK FITTING: EMERGING TOOLS WITH EARLY VALIDATION
Poor mask fit and interface issues continue to impede initiation and long-term adherence to PAP therapy. Recently, several groups have explored machine-learning and computer-vision approaches to interface selection, using facial scans and anthropometric phenotyping.
One study evaluated an AI-based mask-fit algorithm in a Veterans Affairs population and found a higher mask acceptance rate when the algorithm guided fitting, compared with historical controls1. Another line of research has described the development of custom-fabricated CPAP masks via 3D facial scanning combined with computer-aided design (CAD) manufacturing, raising the possibility of fully individualized mask interfaces soon2.
Despite these advances, limitations remain important. The training datasets used in mask-fitting studies often under-represent patients with facial hair, cranio-facial anomalies, and postoperative anatomy. Moreover, most published studies focus on fit and comfort rather than long-term adherence or clinical outcomes such as residual AHI or cardiovascular benefit. For now, AI-assisted mask fitting is best conceptualized as a workflow adjunct, or a first-pass interface recommendation that accelerates triage but does not replace clinician judgment. In the context of a national home-respiratory provider, integrating AI-guided mask selection can streamline remote starts and reduce early PAP frustration, thereby contributing to improved initiation success.
WEARABLE DETECTION OF OSA: EXPANDING ACCESS WITH VARIABLE ACCURACY
Wearable sensors, including wrist/forearm photoplethysmography, accelerometry, and single-channel oximetry, have become attractive tools for OSA risk detection by virtue of convenience, accessibility, and lower cost compared with full polysomnography. Machine-learning algorithms applied to these sensor streams promise screening or pre-diagnostic triage.
A recent systematic review of wearable AI in sleep apnea (38 studies, >6000 subjects) reported pooled accuracies of approximately 0.89 for apnea vs non-apnea event detection, sensitivity 0.79, specificity ~0.95; when applied to classify OSA severity, performance was lower (accuracy ~0.65)4. These results indicate promise but also clear limitations, particularly in real-world heterogeneous populations. Another review emphasized that algorithmic performance varied significantly by device placement (wrist vs chest), signal types used, and apnea subtype (obstructive vs central) and concluded that wearable AI “is not yet ready for routine clinical use” as a replacement for traditional sleep testing4.
Clinically, this means that wearables should currently be viewed as pre-diagnostic screening tools, useful in patients reluctant to undergo formal sleep testing, or as part of remote monitoring triage pathways, but not as definitive diagnostics. For pulmonologists and sleep specialists, the relevance lies in identifying high-risk patients warranting formal home sleep apnea test (HSAT) or polysomnography (PSG) rather than relying on wearable output alone.
AUTOMATED SLEEP STUDY INTERPRETATION: APPROACHING CLINICAL-GRADE PERFORMANCE
Strong AI evidence in sleep medicine resides in automated sleep staging and respiratory event detection applied to PSG and HSAT data. Large, annotated datasets (e.g., the National Sleep Research Resource) have enabled deep-learning architectures that deliver classification performance within the range of inter-scorer human variability.
For example, one study found that a deep-learning algorithm achieved 87-90% agreement with expert human scorers for staging and respiratory event detection across multiple centers4. A conceptual piece from the American Academy of Sleep Medicine (AASM) committee emphasized strengths and threats of AI in sleep medicine and pointed out that automating segmentation and scoring could free clinician time for interpretation and management5. Even as performance metrics improve, several caveats apply. Training sets frequently derive from adult OSA populations, limiting performance in populations with central sleep apnea, hypoventilation, neuromuscular disease, or pediatric patients. A recent systematic review found accuracy ranges from 67% to 98% across studies but also highlighted issues of transparency and generalizability6.
In clinical practice, the optimal deployment is a hybrid workflow: AI performs initial scoring (stage, respiratory events), flags epochs of uncertainty or atypical events, and the human technologist/sleep physician retains oversight and makes final interpretation.
CLINICAL/EHR-BASED RISK STRATIFICATION AND LOW-FEATURE SCREENING MODELS
Beyond wearables, several groups have developed machine-learning risk models that rely only on basic clinical features (age, sex, BMI, neck circumference, blood pressure, symptom questionnaires) or simple physiologic parameters such as overnight oximetry. Their strength lies in scalability and minimal hardware requirements.
Recent studies report AUC values in the 0.80–0.90 range for classification of moderate-to-severe OSA using just demographics and single-channel physiologic data4. For example, deep-learning models using oximetry and anthropometrics achieved promising discrimination in external validation cohorts. Professional society commentary (AASM) emphasizes the value of triage tools in underserved or resource-limited settings5.
These tools have relevance for pulmonologists and sleep providers who must prioritize limited home-sleep-test capacity or deploy remote programs. They offer a lower cost, lower-friction entry point into AI-based screening. Going forward, incorporating low-feature risk scores into referral workflows could enhance early identification of high-risk patients, streamline testing pathways, and reduce delay to therapy.
IMAGING AND ANATOMY-BASED AI: UPPER AIRWAY PHENOTYPING AND CUSTOM INTERFACE DESIGN
In addition to functional data approaches, there is growing interest in anatomical and imaging-based AI, particularly as it pertains to upper-airway structure, custom mask/interface design, and patient-specific therapy selection. Machine-learning models analyzing craniofacial computed tomography (CT) scan, magnetic resonance imaging (MRI), or cephalometric data can help predict OSA severity and identify patients more likely to respond to non-PAP therapies (oral appliances, surgery) rather than CPAP.
Concurrently, pilot work has demonstrated in-office fabrication of custom-fit CPAP masks via 3D-printed or milled interfaces derived from facial scans; one recent study highlighted this workflow as feasible and acceptable to patients2, though long-term adherence data are lacking.
These developments suggest future conditional workflows: imaging-based phenotyping could help tailor therapy (PAP vs oral appliance vs surgery), and AI-driven mask customization could improve comfort and seal in difficult anatomies. For sleep medicine physicians and DME-partners alike, such tools represent the next frontier of truly personalized sleep therapy.
INSOMNIA, CIRCADIAN DISORDERS, AND DIGITAL PHENOTYPING
Although the majority of sleep AI work centers on OSA and PAP therapy, other sleep disorders are beginning to receive attention under the AI paradigm. Digital phenotyping studies have used smartphone usage patterns, passive sensor data, accelerometry, and light exposure metrics to classify insomnia subtypes and circadian-rhythm disorders7-9. For example, early studies examine smartphone interaction and actigraphy streams to distinguish insomnia from uncomplicated short sleep10. In parallel, natural-language-processing (NLP) analyses of EHR notes have been used to identify insomnia diagnoses at scale11.
While these applications are less mature than OSA-focused AI, they reflect the broadening scope of AI in sleep medicine. Sleep clinicians and pulmonologists should remain aware of these trends, especially in integrated care models where insomnia often co-exists with OSA and other cardiopulmonary comorbidities.
PEDIATRIC SLEEP AND HEALTH-EQUITY CONSIDERATIONS
An important limitation across sleep-medicine AI literature is the relative paucity of pediatric datasets and equity-focused research. Most large training cohorts derive from adult populations in North America and Europe; children, adolescents, and historically under-represented racial/ethnic groups are frequently excluded or under-powered. This raises concerns that algorithmic performance may degrade when applied in underserved or diverse populations, potentially exacerbating disparities in sleep health5. Sleep medicine and home-respiratory providers must therefore apply extra caution when deploying AI tools in pediatric or demographically diverse cohorts. Continuous monitoring of algorithmic performance across age, race, sex, and comorbidity strata is essential.
REMOTE MONITORING, ADHERENCE PREDICTION, AND IMPLEMENTATION WORKFLOWS
Beyond diagnosis, one of the most clinically actionable areas of AI in sleep medicine concerns therapy adherence and remote monitoring. Substantial data suggest that the first few nights or weeks of PAP usage contain predictive signals for long-term adherence (e.g., leak patterns, residual apnea-hypopnea index, usage frequency, and mask engagement). Machine-learning models analyzing these early signals have achieved accuracies of 70–80% for predicting 3-month or 12-month adherence12.
Implementation studies have shown that coupling predictive analytics with telemonitoring and proactive respiratory therapist outreach leads to improved adherence outcomes compared with usual care. From an operational perspective, these tools allow sleep providers and DME partners to target limited human-resource interventions to those at highest risk of discontinuation, thereby improving cost-effectiveness. Integrating adherence-prediction models into remote monitoring platforms positions relevant organizations to deliver differentiated value to patients, payers, and referring physicians.
IMPLEMENTATION, REGULATION AND "HOW TO EVALUATE AN AI TOOL"
As providers assess vendor-offered AI tools, sleep physicians and DME stakeholders should consider three key questions: how was the algorithm trained, how was it validated, and how is performance maintained over time. The AASM’s recent position statement emphasizes the importance of transparency, dataset diversity, human-in-loop workflows, and ongoing performance monitoring.
Vendors should be able to provide:
- The size, diversity, and annotation quality of their training and validation datasets
- Independent validation results benchmarked against expert-scored PSG or clinically adjudicated outcomes, ideally stratified by age, sex, BMI, race/ethnicity, and comorbidities,
- regulatory status (FDA, CE-mark, or other) and statement of human-oversight responsibilities
- ongoing performance-monitoring protocols (e.g., drift detection) and clear documentation of how AI outputs integrate into existing clinical workflows (i.e., human review required, confidence thresholds, alert escalation).
From a practical implementation perspective, AI tools should be viewed as augmented workflow-partners, not replacements for human clinicians. Integration into existing sleep laboratory, home-sleep testing, PAP initiation, and remote monitoring pipelines requires coordinated clinical, technical, and operational planning.
CONCLUSION
AI offers meaningful opportunity to enhance detection and treatment of sleep-disordered breathing. Evidence supporting automated sleep scoring and adherence prediction is relatively robust and rapidly maturing. Earlier-stage applications, such as wearable OSA detection, mask-fitting automation, imaging-based phenotyping, insomnia/circadian disorder detection, and pediatric-equity frameworks, are promising but require further validation and real-world outcomes. For pulmonologists and sleep-medicine clinicians, the current moment calls for balanced optimism: these tools can accelerate workflows and broaden access, but physician oversight and careful evaluation remain critical.
LEARN HOW APRIA CAN HELP YOUR SLEEP PATIENTS
References
1. Chan, M., Soreca, I., Ali, M. E., Chakravorty, S., Gulla, A., Shroyer, B., & Atwood, C. (2022). 0771 Artificial Intelligence Based Mask Fit Algorithm Application in the Pittsburgh Veteran Population. Sleep, 45(Supplement_1), A335-A336.
2. Sheth, R., Sheth, S., & Audette, M. (2025). Developing Custom‐Fit CPAP Mask Prototype in Patients With Obstructive Sleep Apnea. Journal of Sleep Research, e70146.
3. Abd-Alrazaq, A., Aslam, H., AlSaad, R., Alsahli, M., Ahmed, A., Damseh, R., ... & Sheikh, J. (2024). Detection of sleep apnea using wearable AI: systematic review and meta-analysis. Journal of medical Internet research, 26, e58187.
4. Giorgi, L., Nardelli, D., Moffa, A., Iafrati, F., Di Giovanni, S., Olszewska, E., ... & Casale, M. (2025, January). Advancements in Obstructive Sleep Apnea Diagnosis and Screening Through Artificial Intelligence: A Systematic Review. In Healthcare (Vol. 13, No. 2, p. 181). MDPI.
5. Bandyopadhyay, A., Oks, M., Sun, H., Prasad, B., Rusk, S., Jefferson, F., ... & Seixas, A. (2024). Strengths, weaknesses, opportunities, and threats of using AI-enabled technology in sleep medicine: a commentary. Journal of Clinical Sleep Medicine, 20(7), 1183-1191.
6. Haghighat, S., Joghatayi, M., Issa, J., Azimian, S., Brinz, J., Ashkan, A., ... & Sangalli, L. (2025). Diagnostic accuracy of artificial intelligence for obstructive sleep apnea detection: a systematic review. BMC Medical Informatics and Decision Making, 25(1), 278.
7. Alamoudi, D., Breeze, E., Crawley, E., & Nabney, I. (2023). The feasibility of using smartphone sensors to track insomnia, depression, and anxiety in adults and young adults: narrative review. JMIR mHealth and uHealth, 11, e44123.
8. Langholm, C., Byun, A. J. S., Mullington, J., & Torous, J. (2023). Monitoring sleep using smartphone data in a population of college students. Npj mental health research, 2(1), 3.
9. Kirshenbaum, J. S., Crowley, R. N., Latham, M. D., Pagliaccio, D., Auerbach, R. P., & Allen, N. B. (2025). Comparison of Sleep Features Across Smartphone Sensors, Actigraphy, and Diaries Among Young Adults: Longitudinal Observational Study. JMIR Formative Research, 9(1), e67455.
10. BaHammam, A. S. (2024). Artificial intelligence in sleep medicine: the dawn of a new era. Nature and science of sleep, 445-450.
11. Lopez-Garcia, G., Weissenbacher, D., Stadler, M., O’Connor, K., Xu, D., Gryboski, L., ... & Gonzalez-Hernandez, G. (2025). Automated Insomnia Phenotyping from Electronic Health Records: Leveraging Large Language Models to Decode Clinical Narratives. medRxiv.
12. Scioscia, G., Tondo, P., Foschino Barbaro, M. P., Sabato, R., Gallo, C., Maci, F., & Lacedonia, D. (2022). Machine learning-based prediction of adherence to continuous positive airway pressure (CPAP) in obstructive sleep apnea (OSA). Informatics for Health and Social Care, 47(3), 274-282.
.png)
