Better outcomes, greater equity, and truly personalized sleep medicine are within reach, but only if our diagnostic tools can see the full picture of sleep. Too often, patients with obstructive sleep apnea are underserved by today’s standard approaches. Women, non-obese individuals, and patients with comorbidities are especially vulnerable to underestimated disease severity or outright misdiagnosis, largely because conventional home sleep tests cannot detect accurate sleep stages or arousals — signals that are essential to understanding diverse presentations of apnea.
This is where AI can fundamentally change care. At Nox Medical, we apply AI to expand diagnostic visibility, improve accuracy across patient populations, and support more equitable access to high-quality sleep diagnostics. Our approach is grounded in scientific rigor and real-world evidence, reflecting a commitment to transparency and trust in clinical practice.
The validation dataset for Nox Medical’s most recent AI tool, DeepResp™ 2 was large and globally diverse, collected in everyday clinical environments. The validation dataset alone included approximately 5,800 sleep recordings — the largest dataset used to date in an FDA clearance for a sleep diagnostics medical device — forming the foundation of the DeepResp™ 2 FDA clearance (K252320). This scale and diversity help ensure reliable performance across a wide range of physiologies, care settings, and diagnostic challenges, bringing us closer to accurate diagnosis for every patient, not just the “typical” one.
This foundation is supported by peer-reviewed, published evidence that demonstrates the accuracy, reliability, and clinical relevance of our AI1. And unlike black-box systems that obscure their decision-making, Nox AI is designed to be transparent. That is, input data can be reviewed by the clinician so they can determine if they trust the outputs.
With this combination of large-scale validation, global data diversity, scientific publication, and transparent design, Nox AI aims to be a new benchmark for how to develop and implement AI medical devices. And Nox has been working with AI for well over a decade, quietly leading the innovation in AI, designed specifically to support conclusive testing results and to improve diagnostic confidence and throughput.
Our physiology-grounded models are helping bring greater clarity, consistency, and confidence to both PSG and home sleep testing — and our work is only accelerating.
How AI Fits into Sleep Medicine
Sleep studies produce massive amounts of physiological data — EEG, EOG (eye movement), EMG (muscle activity), respiratory flow, oxygen saturation — recorded over a full night or more. A trained sleep technologist manually scores this data to identify sleep stages, arousals, and respiratory events. This scoring is foundational for clinical interpretation but poses challenges that stand in the way of meeting the need for accuracy and scale/volume, including:
- Time constraints, requiring hours of expert review per study
- Inter-scorer variability, where different clinicians may interpret the same signals differently
- Slow or stalled workflows, particularly in under-resourced clinics
Nox Medical recognized early that AI could automate repetitive tasks — freeing clinicians to focus on interpretation and personalized patient care. At the same time, Nox realized the importance of responsibly validated AI, designed to support clinical decision-making, not eliminating clinicians from the loop. This AI-augmented approach aims to build clinician trust by respecting the complexities and nuances of physiological signal interpretation.
Foundations in AI and Machine Learning: Nox Medical’s Evolution
Automating Sleep Scoring
Nox Medical’s AI journey began in earnest in 2015, when the team undertook the challenge of automating sleep stage classification using EEG data. Sleep stages (Wake, REM, N1, N2, N3) are fundamental to almost all sleep diagnoses, but scoring them manually is slow and dependent on human interpretation.
Nox developed algorithms capable of classifying 30-second epochs into respective sleep stages — a task once thought to require highly trained human scorers. Key to making this successful was the existence of established human consensus scoring standards, which enabled reliable model training and validation.
“We started by classifying sleep stages based on EEG data because it’s well-defined and has reliable human annotations.” — Dr. Jón Skírnir Ágústsson, PhD — Vice President of AI & Data Science at Nox Medical.
Extending AI Beyond Sleep Stages
Nox Medical’s co-founder and CTO, Sveinbjörn Höskuldsson elaborates on how Nox has been at the forefront of AI in understanding sleep:
“Measuring airflow is essential for sleep diagnostics. Flow by definition is such an important parameter for diagnosing sleep apnea. Sleep apnea is a disturbance in respiration, and airflow is the core signal of disturbed respiration. The core reason we need an accurate flow measurement is to accurately diagnose sleep apnea. In addition, we have made many improvements in measuring respiratory flow using respiratory inductance plethysmography (RIP) technology from which we gain insight into respiration and can derive new descriptive parameters.”
What makes this different and valuable is that not all RIP technology is the same. Traditionally RIP has been of low quality, making it useful only to detect the absence or presence of breathing movements2. Nox RIP shows much more.
Sveinbjörn continues:
“Because of this high resolution, high accuracy and high reliability, we can actually derive something that was not possible to do with regular polysomnography as it is described today. This is very important because this gives us the insights to predict what type of treatment will suit every single patient.”
Ultimately, AI’s goal in these contexts isn’t simply automation — it’s a fundamentally reshaping of how signals are interpreted to uncover information that might otherwise be overlooked:
- Detecting respiratory events with greater sensitivity
- Estimating total sleep time more accurately, especially in systems without full EEG
- Identifying subtle patterns that correlate with clinical outcomes
The ultimate objective is supporting conclusive diagnoses. That is, reducing inconclusive or false-negative study results and minimizing the risk of underestimated OSA severity, improving the likelihood of achieving an accurate diagnosis from the first recording. The accuracy and specificity Nox AI enables moving from broad categorizations to tailored insights that reflect individual physiologies.
Driving Research in New Directions through Key Collaborations
Keeping research and innovation moving forward, Nox Medical collaborates on research projects with prestigious international universities, such as Harvard Medical School, and is an active participant in Sleep Revolution, an interdisciplinary project with nearly 40 European partners, including some of the leading academic sleep centers in Europe.
More specifically to AI, Nox collaborates on important subjects like OSA endotypes by working with luminaries in the sleep field, such as Scott Sands and others to publish groundbreaking research.
AI in Practice: Arousal Scoring Reimagined
A critical and often overlooked component of sleep diagnostic interpretation is arousal scoring — brief interruptions in sleep architecture that can signify underlying pathophysiology.
Traditionally, detecting arousals has been one of the most time-intensive and subjective tasks in sleep study analysis, and is not traditionally done in HSAT. Only hypopneas ending in desaturation are counted. Manual arousal detection requires scorers to sift through multi-channel EEG and physiological data to mark fleeting events. Nox’s FDA-cleared DeepResp offers AI in practice as a deep-learning-powered model that learns complex signal features from large datasets and maintains high accuracy compared to manual scoring standards.
The model’s performance has demonstrated strong sensitivity and specificity, and when its outputs were used as inputs for hypopnea-detection algorithms, it significantly improved AHI classification accuracy. What does this mean? HSAT-based diagnoses can avoid underestimating AHI compared with PSG, across diverse populations whose hypopneas result mainly in arousals rather than significant oxygen desaturation. More accurate AHI reduces inconclusive/false-negative HSAT and unnecessary follow-up PSG testing, enabling earlier and more accurate diagnosis and appropriate treatment
Nox employs AI technology to make the most of what AI does best — processing large volumes of data quickly and consistently — while clinicians continue to provide oversight and interpretation. When clinicians pair up with Nox sleep tests backed up by advanced AI, they can spend more time ensuring their own confidence in the data and ensuring that their patients understand and feel confident with the diagnosis they receive.
From Research to Clinic: Validation and Clinical Confidence
Central to Nox Medical’s approach is robust validation. Unlike models that are trained and tested on limited or homogeneous datasets, Nox emphasizes diverse, real-world validation to ensure performance generalizes across patients, demographics, and settings. The idea is to benchmark AI against manual scoring to build tools that contribute to clinical confidence:
- Validation on large, real-world datasets (up to 5,800 studies). Not curated samples but true clinical-use data that reflect everyday variability
- Global data sources from diverse patient populations and real-world care environments worldwide
- Peer-reviewed, published evidence that provides independent, scientific validation supporting accuracy, reliability, and clinical relevance1,2
- Transparent algorithm and outputs with clear, interpretable results that clinicians can understand, use, interact with, and rely on – not black-box scoring
AI at the Core of Clinical Sleep Diagnostics: DeepResp
While earlier AI efforts focused on automating established clinical tasks (like sleep staging and arousal scoring), DeepResp represents a leap forward — an FDA-cleared AI-enabled medical device that significantly improves AHI accuracy for home sleep tests and closely aligns with PSG AHI.
What is DeepResp?
DeepResp is a cloud-based AI medical device that uses deep learning to score 30-second epochs in sleep states (REM, NREM, wake) and arousal events directly and solely from respiratory inductance plethysmography (RIP) belt signals. By analyzing subtle changes in breathing patterns, the AI tool scores sleep states and arousals without requiring EEG, EOG, or EMG.
Technical and Regulatory Validation
Analysis of one of the largest-ever validation datasets for an FDA-cleared HSAT system demonstrates that DeepResp improves the accuracy of AHI by enabling sleep staging and arousal scoring in HSAT studies without adding the burden of EEG for patients.
DeepResp’s FDA clearance (K 252330) underscores its clinical readiness. Cleared in March 2025, the device is intended to assist in diagnosis and follow-up of sleep disorders in adults and can be used in diverse settings, including hospitals, ambulatory clinics, and home sleep testing environments.
Where Nox AI is Going
Despite progress in the efficiencies and workflows of sleep labs, AI in sleep medicine is still in its infancy. Innovations will continue to develop, helping to, as Nox’s Dr. Jón Skírnir Ágústsson states, “…shape the future of sleep medicine as we get more information from sleep studies without adding burden to the patient, allowing physicians to correctly diagnose their patients and select the most appropriate treatment for each individual.”
For Nox, this future means continuing to focus on data quality and diversity, as AI models are only as strong as the data they are trained on. AI can be used not only for efficiency gains, such as reduced scoring time, but also for providing diagnostic insights with REM/NREM OSA information and arousal scoring in HSAT as well as OSA endotypes that provide a more complete picture of sleep physiology and support better diagnoses as well as predict better therapy outcomes and treatment decisions.
Continuing on this path, Nox AI strives to score studies faster while deepening clinical insight to improve patient outcomes.
Better Sleep, Shaped by AI
Nox Medical’s approach to AI in sleep medicine combines innovation grounded in clinical reality, validated rigorously, and designed to expand and augment human expertise.
From automating routine sleep staging to redefining comprehensive scoring with DeepResp, Nox and its approach to AI show how artificial intelligence can be used responsibly to transform medical diagnostics. With this hybrid model of AI-assisted, clinician-centered practice, the future of sleep medicine is not only more efficient — it’s conclusive, more consistent, and more patient-centric than ever before.
References
1 Finnsson, E., Erlingsson, E., Hlynsson, H.D. et al. “Detecting arousals and sleep from respiratory inductance plethysmography”. Sleep and Breathing (2025).
2 Finnsson E, Arnardottir E, Montazeri K, Keenan BT, Schwab RJ, Gislason T, Pack AI, Wellman A, Islind AS, Agustsson JS, Sands SA. Respiratory Inductance Plethysmography to Quantify Changes in Ventilation in Obstructive Sleep Apnea. IEEE Trans Biomed Eng. 2025 Oct 7;PP. doi: 10.1109/TBME.2025.3618403. Epub ahead of print. PMID: 41056175.
Topic: Industry News
