The clinical stakes are considerable. Obstructive sleep apnea is associated with fragmented sleep, intermittent hypoxemia, excessive daytime sleepiness, neurocognitive impairment, and increased cardiometabolic risk, yet many affected individuals remain undiagnosed or wait years for a diagnosis. The gold standard, in-laboratory polysomnography, captures a rich set of physiological signals including EEG, ECG, airflow, respiratory effort, oxygen saturation, and body position, but it is expensive, labor-intensive, and difficult to scale to population-level screening. Home sleep apnea testing has improved access, but it creates new demands for automated, robust, and interpretable analysis. The review argues that this is precisely where machine learning and deep learning can transform the field, provided the technology is validated rigorously enough for real-world deployment.

To bring order to a rapidly expanding literature, the authors propose a multidimensional taxonomy organized around four complementary dimensions: acquisition platforms, data modalities, AI methods, and clinical applications. Six primary research domains emerge from this framework: polysomnography and home sleep apnea testing signal analysis; wearable, smartphone, and contactless screening; audio and reduced-channel physiological-signal diagnosis; imaging and craniofacial anatomical risk prediction; phenotyping, endotyping, and risk stratification; and treatment management, CPAP adherence, and remote monitoring. A seventh, cross-cutting theme covers explainable, trustworthy, and clinically deployable AI, reflecting the growing recognition that interpretability, robustness, and deployment feasibility matter across every modality rather than belonging to a single one.

The historical arc the review traces is striking. The earliest phase focused on automating diagnosis from conventional physiological recordings, relying on handcrafted signal-processing features and classical machine learning classifiers for apnea detection, severity classification, and apnea-hypopnea index estimation. A second phase expanded into reduced-channel and home-oriented screening, exploring ECG and heart rate variability signals and wearable multimodal sensing as lower-burden alternatives to the sleep laboratory. The third phase marked the deep learning transition, with convolutional, recurrent, transformer-style, and multimodal architectures becoming common for PSG, HSAT, ECG, PPG, and audio signals. The fourth phase broadened the field into wearable, contactless, and multimodal assessment, incorporating wrist-worn consumer devices, sleep-sound analysis, radar-based sensing, and bed-integrated tactile sensors. Most recently, the field has moved toward trustworthy, clinically deployable AI and, at the frontier, large language models that support patient education, report summarization, clinical documentation, and decision-support workflows.

Technically, the reviewed corpus reveals a methodological landscape in transition. Deep learning, convolutional neural networks, multimodal fusion, transformer-based models, and attention mechanisms dominate recent work, while classical machine learning, feature engineering, support vector machines, and random forests remain widely used, particularly for clinical-variable and risk-stratification studies. A visualization-based synthesis of the 218 core papers, using term co-occurrence mapping and tag-based evidence maps, confirms that the literature remains strongly diagnosis-centered: OSA detection, respiratory-event detection, severity classification, screening, and AHI estimation are the dominant tasks. By contrast, treatment management, CPAP adherence support, and remote monitoring are strikingly underrepresented, with only six multi-label assignments in the taxonomy tables, suggesting a major opportunity for future work beyond diagnostic automation.

The data landscape is equally heterogeneous. PSG and HSAT studies center on airflow, EEG, SpO2, and respiratory-effort signals; wearable and contactless studies rely on PPG, actigraphy, sleep audio, accelerometry, and radar or radio-frequency sensing; imaging studies use craniofacial photographs, MRI radiomics, endoscopy, and airway morphology; and phenotyping studies depend on clinical variables, PSG-derived summaries, and comorbidity features. Representative examples span the full spectrum: pediatric OSA detection from airflow and oximetry, apnea and hypopnea localization in PSG using multi-scale object detection, transformer-based audio segmentation for fine-grained non-invasive apnea detection, radar-based apnea-hypopnea detection with hybrid CNN-LSTM networks, contactless breathing-vibration monitoring, and multimodal diagnostic prediction models combining two-dimensional facial images with clinical data.

Yet the review’s most sobering findings concern validation. Only 31 of 218 studies, or 14.2 percent, evaluated a locked model in a strict external or cross-dataset test, and only four used a prospective validation cohort. The unit of data partitioning emerged as a critical source of optimistic bias: when epochs, windows, or events from the same participant are randomly allocated to both training and test sets, performance estimates can be inflated. One audio-based study illustrates the point dramatically, achieving 88.32 percent accuracy with a subject-dependent segment split but only 68.55 percent accuracy with a subject-independent split. A cross-database analysis of heart rate variability-based apnea detection found that external-validation accuracy, sensitivity, and specificity were 20 to 40 percent lower than epoch-wise cross-validation results. These findings support participant-level or recording-level partitioning, explicit reporting of allocation units, and external testing whenever the intended use involves previously unseen patients or devices.

Performance reporting also depends heavily on the clinical endpoint and evaluation metric. A single-effort-belt model trained on MGH data yielded an external SHHS AUROC of 0.92 but an AUPRC of only 0.56, showing how class prevalence shapes interpretation. A model trained and validated on SHHS and tested on unseen MESA and MrOS cohorts reported an event-level AUROC of 0.87, an AUPRC of 0.90, and an F1-score of roughly 0.72 for AHI-severity classification. A prospectively validated commercial smartwatch algorithm achieved an AUC of 0.931 for moderate-or-greater OSA against physician-scored PSG in 53 participants, while a questionnaire-based model retained an AUC of 0.91 in a later prospective temporal cohort of 189 participants. Within-study comparisons, such as ANN and random-forest models achieving AUCs of 0.70 in a prospective SAGIC testing cohort, similar to the STOP-BANG questionnaire at 0.71, show that comparative conclusions are most reliable when models are evaluated on the same cohort, target, reference standard, and participant-separated test set.

The challenges the review identifies extend well beyond statistics. Data heterogeneity across devices, sleep laboratories, scoring protocols, and patient populations makes generalization difficult. Interpretability is essential because sleep clinicians must understand which respiratory events, oxygen desaturation patterns, sleep stages, or anatomical features influenced a prediction before acting on it; explainable ECG models and airflow-based models using Grad-CAM, LIME, SHAP, relevance maps, and attention mechanisms demonstrate how this can be achieved. Privacy and data governance concerns are amplified when data flow through home devices, mobile applications, and cloud platforms, making lightweight, on-device, energy-efficient AI particularly relevant. Large language models introduce their own risks, including hallucination, guideline non-adherence, and unclear accountability, requiring source grounding, clinical validation, and clinician oversight before deployment.

The opportunities, however, are equally compelling. Scalable screening through wearables, smartphones, and contactless sensors could identify at-risk individuals before formal sleep testing and triage patients for PSG, HSAT, or specialist evaluation. Precision sleep medicine, powered by multimodal AI that integrates physiological signals, symptoms, comorbidities, demographics, imaging, and treatment data, could move the field beyond the apnea-hypopnea index alone toward treatable traits and personalized management. Remote monitoring and adaptive management, drawing on CPAP device logs, residual events, usage patterns, and patient feedback, could reduce clinic burden while improving continuity of care. The authors conclude that future progress will depend on rigorous external validation, transparent reporting, reproducible benchmarking, privacy-preserving data governance, fairness across demographic groups, and safe integration into sleep-medicine workflows, positioning trustworthy AI and carefully validated language models as key enablers of more accessible, scalable, and personalized sleep apnea care.

Subject of Research: Artificial intelligence applications for the diagnosis and management of sleep-disordered breathing and obstructive sleep apnea

Article Title: Artificial intelligence for sleep-disordered breathing: a taxonomy of methods, data modalities, and clinical applications

Article References: Dinh, T., Lisik, D., Fournier-Viger, P., Tran, D., Yu, P. S., Van Hong, H., & Zou, D. (2026). Artificial intelligence for sleep-disordered breathing: a taxonomy of methods, data modalities, and clinical applications. Frontiers in Sleep, 5, Article 1927453. https://doi.org/10.3389/frsle.2026.1927453

Image Credits: AI Generated

DOI: 10.3389/frsle.2026.1927453

Keywords: obstructive sleep apnea, sleep-disordered breathing, artificial intelligence, machine learning, deep learning, polysomnography, wearable sensors, large language models, precision sleep medicine, explainable AI, CPAP adherence, external validation