SLEEP 2024, the premier clinical and scientific conference in the sleep field, was held in Houston on June 1-5[1], highlighting the adoption of clinical artificial intelligence in sleep disorders.
How do AI systems assist physicians in developing better therapies? What can machine learning do to enhance the processing of sleep data, evaluate sleep patterns, and identify sleep abnormalities? Which sleep illnesses exactly can AI help manage? What are major companies’ responses to AI trends? Read on to uncover AI’s revolutionary role in sleep medicine.
Applications of Artificial Intelligence in Medicine Examples
1. AI Enhancing Diagnostic Enhancements
Clinical artificial intelligence is promising to revolutionize diagnostics in sleep medicine with automated sleep study scoring and early detection of sleep disorders. Traditional manual scoring of sleep studies is labor-intensive and subject to variability among scorers. To accurately identify and classify various sleep stages and events requires detailed, continuous analysis of extensive sleep data, often spanning several hours. As part of regular diagnostic procedures in sleep medicine, this process can be tedious and monotonous, making it prone to errors due to the repetitive nature of the work[2]. Additionally, there is a recognized high level of variability in sleep scoring results among different expert scorers. This inter-scorer variability in sleep staging is largely attributed to equivocal epochs that contain features of more than one stage, leading to differing interpretations[3].
However, AI auto-scoring systems are developing to rationalize this process for accuracy levels comparable to human experts with Cohen's kappa values of up to 0.80.[4] For FDA-cleared autoscoring software decreases the time and costs of manual scoring while standardizing results across studies. Moreover, AI models are adopted to detect OSA using smart oximeters monitoring SpO2 levels. They surpass conventional methods for earlier and more accessible detection of OSA. Further, AI predictive models utilize demographic and physiological data such as age, sex, and BMI to identify individuals at risk of developing sleep disorders. Thus, it is likely to facilitate timely intervention and better patient outcomes.
2. AI Redefining Personalized Treatment
Firstly, AI is apt to alter personalized treatment in sleep medicine with predicted treatment responses with high accuracy. Machine Learning (ML) systems analyze extensive datasets to classify correlations between patient traits, sleep physiology, and treatment outcomes. It facilitates precise phenotyping and endotyping of OSA. Phenotyping refers to the clinical manifestations of OSA, such as partial or complete upper airway collapse, while endotyping focuses on the underlying physiological mechanisms, like anatomical compromise, unstable ventilatory control, and low arousal threshold[5]. With a more comprehensive analysis of the clinical features and physiological mechanisms of patients, clinicians can modify treatment details as per the characteristics of each patient, such as upper airway collapsibility, muscle responsiveness, arousal thresholds, and ventilatory control.
Secondly, AI is transforming the landscape of OSA treatment with simplified workflow. To provide OSA patients with targeted treatment, former practices to evaluate upper-airway collapsibility demand at least three techniques. These include the application of negative pressure pulses (NPPs), the resistance change observed in response to externally applied inspiratory resistive loading (IRL), and the determination of the critical pressure (Pcrit) needed to induce a cessation of airflow[6].
However, apart from more accurate OSA classification, with predictive modeling, AI algorithms can predict the likelihood and severity of airway collapse by analyzing patient data that require less effort to access, such as anatomical features, sleep patterns, and respiratory metrics, prompting personalized treatment plans with minimal workload.
Another AI application in personalized treatment is its ability to predict CPAP therapy, enabling early involvement to improve patient compliance. The AI monitoring system utilizes compliance, mask leaks, and residual respiratory events data to predict expected compliance and provide feedback and interventions. A former study has proven that the patients with this intervention had a mean of 1.14 h/day higher adjusted CPAP compliance than the control without the system[7].
3. AI Boosting Research and Big Data
AI and big data are taking sleep medicine research to new heights by utilizing publicly available datasets for analysis. The National Sleep Research Resource (NSRR) [8] and the UK Biobank[9] offer vast repositories of sleep data for researchers to discover new insights into sleep disorders. The NSRR is a comprehensive system for sharing sleep data, containing information on more than 26,808 subjects and 31,166 signal files. It provides access to analysis-ready physiological signals and various clinical data to support sleep research[8]. Besides, the Human Sleep Project (HSP) gathers real-world sleep data from ~19K patients evaluated at the Massachusetts General Hospital and will grow over the coming years to include data from >200K patients[10]. It addresses fundamental questions about sleep duration, quality, and genetic and environmental impacts. To handle extensive data resources, AI techniques are trained to disregard overfitting and underfitting with large, diverse, and representative datasets for model accuracy and generalizability, generating more robust algorithms and eventually contributing to more reliable and effective treatment outcomes.
4. AI Inspiring OSA Complication Prediction
ML models analyze EEG signals to predict neurological diseases or sleep disorders. They can determine phenoconversion timing in idiopathic Rapid Eye Movement sleep behavior disorder. It is a precursor to neurodegenerative diseases.
Apart from that, Margaux et al have developed an AI model, for patients suspected of OSA, which allows the clinician to evaluate the risk of cardiovascular morbidity and all-cause mortality (MMCV) while diagnosing sleep disorders. The model is interpretable with the smallest combination of easily accessible clinical and sleep features derived from sleep recordings. Strengths of the model are, in addition to estimating an MMCV risk, the ability to find other risk factors and to calculate risk thresholds for these factors[11].
Furthermore, AI models help predict the risks of developing major depressive disorder (MDD) by analyzing sleep data. Extensive experiments with the XGBoost machine learning model have shown that severe insomnia, poor sleep quality, and occasional night terrors significantly increase the likelihood of developing MDD. By focusing on sleep-wake disorder symptoms[12], this model can raise awareness of the risks linking sleep-wake disorders and MDD in adolescents, thereby enhancing primary care and prevention efforts.
Updates about AI Technologies in Sleep Medicine
EnsoData and Aeroflow Partnered to Enhance Sleep Apnea Care Using AI Tools[13]
EnsoData’s EnsoTherapy AI model’s task generation and prioritization help Aeroflow’s sleep coaches streamline processes to provide personalized support and improve patient adherence to PAP therapy. The AI predicts patient compliance within days of starting treatment, which allows targeted interventions. Announced in May 2024, this partnership aims to streamline the CPAP therapy onboarding process, resulting in better health outcomes and higher compliance rates. Early results have shown significant improvements in patient adherence within the first 30 days.
IHH Invests in Belun Technology for AI Sleep Diagnostic Technology[14]
It was announced in May 2024 that IHH Healthcare's investment in Belun Technology utilizes Belun's FDA-cleared wearable ring device for sleep apnea detection. It helps assess sleep stages to back personalized treatment plans. IHH collaborates with Belun to reconnoiter applications for other sleep-related health conditions for diagnostic accuracy and patient outcomes in clinical settings.
Mount Sinai to Develop Sleep Apnea Outcome Risk Prediction Models[15]
Mount Sinai scholars were reported in Mar 2024 to have received a $3 million NIH grant for AI models predicting cardiovascular risks in OSA patients. They will use data from the SAVE trial and MESA cohort to recognize predictors of atherosclerosis progression. Plus, the models will be legalized with Mount Sinai's EHR data. It shows clinical artificial intelligence's implications in custom medicine.
AcuPebble Sleep Apnoea Device Used in the University of Warwick Trial[16]
The University of Warwick's FOUND Project uses the AcuPebble device to update OSA diagnosis starting from Feb 2024. AcuPebble (a sensor on the neck) collects sleep data and uploads it via a mobile app for analysis. The trial recruits 1,426 patients and compares AcuPebble's efficiency with traditional methods. The innovation shortens diagnosis times from months to days.
HoneyNaps' AI Sleep Analysis Algorithm Receives FDA Approval[17]
HoneyNaps' SOMNUM technology uses deep learning to analyze multi-channel, time-series biosignals. The FDA approval in Sep 2023 shows its authority over other video image systems. It endured clinical trials with 400 subjects, including direct US participants, over three years. It expands diagnostic functions to contain cardiovascular and neuromuscular diseases.
Sleep Illnesses and AI Technologies
OSA
In OSA diagnosis, firstly, AI algorithms analyze PSG data with hopingly higher precision, recognizing apnea events and AHI more accurately than traditional methods. Secondly, AI algorithms revolutionize the classification of upper airway collapse sites and thus accelerate OSA diagnosis by analyzing snoring sounds, which is clinically significant for understanding the pathogenesis and selecting optimal treatments. This method is simpler and more patient-friendly compared to the widely accepted drug-induced sleep endoscopy (DISE)[18], as it is non-invasive and requires fewer equipment and personnel. By innovatively analyzing comprehensive snore sound parameters such as formant frequencies (FF), Mel frequency cepstral coefficients (MFCC), non-Gaussianity score (NGS), loudness, and pitch, ML models are proven to estimate the severity of obstructive sleep apnea (OSA) with a proved accuracy of 88% to 91%[19], which, if popularized, will significantly streamline the clinical diagnosis process. Thirdly, AI genetic algorithms are able to categorize biomarkers and quicken the process of identifying acetylation-related genes in OSA patients[20], aiding in preemptive treatment strategies.
In OSA treatment, AI, in conjunction with ubiquitous wearable devices, significantly improves the accuracy of detecting issues with PAP adherence and mask fit, triggering self-management interventions that empower patients to optimize treatment adherence. Secondly, with AI algorithms accelerating data processing, clinicians are considering cooperating with more factors in addition to PSG data, like BMI, pharyngeal critical closing pressure (Pcrit), upper airway dilator muscle recruitment[21], etc., to help determine precise treatment. AI algorithms greatly enhance this process by learning from large, multifaceted datasets and constantly developing layered mathematical models, which improves therapy accuracy with more data points[21]. Moreover, AI systems excel at analyzing CPAP usage data in real time, giving rise to enhanced treatment adherence and effectiveness by providing prompt feedback to clinicians and patients.
Hypersomnia
AI technologies are proven capable of expediting the differentiation of many types of hypersomnia with progressive data analysis techniques. Through ML procedures, clinical artificial intelligence processes polysomnographic data to identify subtle sleep architecture and stability variations. For instance, an AI algorithm was employed to obtain automatic hypnograms and hypodensity graphs of patients with different types of hypersomnia, and then random forest classifiers were trained and tested in a 5-fold-cross-validation scheme to distinguish between narcolepsy type 1 (NT1), narcolepsy type 2 (NT2), and idiopathic hypersomnia (IH) with high accuracy[22]. It is verified by F1 scores of 0.74 for NT1 vs. NT2 and 0.89 for NT1 vs. IH[23].
Furthermore, an unsupervised ML approach, agglomerative hierarchical clustering, is adopted in a study to identify clusters of clinically distinct central hypersomnolence disorders[24]. AI models analyze hypodensity graphs depicting sleep stage probabilities to sense specific biomarkers of each hypersomnia subtype. Cluster-guided classification will result in a more solid hypersomnolence classification system that is less vulnerable to the instability of single features. It helps shape appropriate treatment plans for patient outcomes.
Insomnia
Clinical artificial intelligence exerts great potential to transfigure the diagnosis and treatment of insomnia through sleep micro-event recognition. AI algorithms analyze EEG, EOG, and EMG signals to discover K-complexes, spindles, and rapid eye movements with high precision. For example, convolutional neural networks (CNNs) and recurrent neural networks (RNNs) are two types of artificial neural networks used respectively for image recognition and sequential data processing. They attain high accuracy in identifying sleep spindles and K-complexes to diagnose insomnia and learn its etiology. As a result, AI supplements valuable insights into sleep quality and disturbances that cause insomnia. In addition, AI models stand a good chance to predict the effectiveness of insomnia treatment while examining changes in micro-event patterns for clinicians to mold interventions for patients. It helps comprehend insomnia's pathophysiology, which prompts more practical treatment strategies.
Circadian Rhythm Sleep-Wake Disorders
AI technologies also manage circadian rhythm sleep-wake disorders while identifying circadian time. Clinical artificial intelligence models predict circadian gene expression patterns using transcriptomic data with lower sampling points to optimize experimental designs. For example, LightGBM classifies circadian transcripts using DNA sequence features for F1 scores of 0.766[25]. AI also predicts internal circadian time from a single transcriptomic timepoint and recognizes marker transcripts showing the clock phase. It lets clinicians assess circadian disruptions due to environmental factors or diseases. Similarly, AI models help design personalized treatment schedules for better synchronization with external cues and circadian disorder treatment efficacy.
REM Sleep Behavior Disorder
Clinical artificial intelligence is making progress to potentially improve the diagnosis of REM Sleep Behavior Disorder (REM RBD). AI algorithms investigate PSG data to detect REM sleep without atonia (RSWA) and other characteristic markers of RBD. E.g., semi-automatic methods preprocess and filter EMG and ECG signals to find muscle activity during REM sleep. Meanwhile, CNNs classify REM sleep stages and detect abnormal muscle tone with high sensitivity and specificity. They differentiate RBD from other parasomnias while analyzing the frequency as well as the intensity of muscle cramps and dream enactment behaviors. Also, AI video analysis of sleep movements provides a non-invasive diagnostic tool for the reliability of RBD diagnosis.
Sleep-Related Movement Disorders
AI-assisted medical diagnosis helps in sleep-related movement disorders, including restless legs syndrome (RLS) and periodic limb movement disorder (PLMD). Clinical artificial intelligence models analyze multi-channel PSG data to isolate movement patterns and sleep stages. For example, easy and accurate ML models have been made to achieve the automatic detection of PLMS events by processing electromyography (EMG) signals from PSG data[26]. AI electroencephalogram (EEG) and EMG data analysis identifies biomarkers and specific movement patterns of such disorders for precise diagnosis. Moreover, AI models foresee the severity of movement disorders while quantifying movement frequency, duration, and impact on sleep quality. Along these lines, it helps clinicians develop directed treatment plans, monitor treatment efficacy, and adjust therapies in real time.
Challenges of AI Technologies in Sleep Medicine in the Near Future
How Can AI be Integrated into Sleep Medicine?
For AI technologies to be effective, they must seamlessly integrate into existing clinical workflows. For example, clinical artificial intelligence in sleep medicine must integrate heterogeneous data from wearables, polysomnography, and EHRs. This requires not only technical compatibility but also training for healthcare providers to use these tools effectively.
How Accurate is AI Sleep Detection?
AI models need rigorous validation to ensure their accuracy and reliability across diverse populations and settings. However, AI deployment is complicated by the absence of standard techniques and the requirement for constant model validation in varied populations. Therefore, extensive work is being done to ensure that AI models generalize well to different patient demographics and clinical scenarios, just as they do in controlled research environments.
How to Make AI More Transparent?
It is difficult to understand how AI models, particularly deep learning algorithms, arrive at specific conclusions. Conveying this internal state and dependencies in a humanly comprehensible way is extremely challenging[27]. Although these models often function as “black boxes,” continuous efforts are focused on making AI systems more interpretable and their decision-making processes transparent. This is crucial for gaining the trust of healthcare providers and patients and for the effective use of AI in clinical settings.
Wrapping Up
In conclusion, clinical artificial intelligence (AI) in sleep medicine holds significant promise for enhancing the diagnosis and treatment of various sleep disorders. By integrating heterogeneous data from wearables, polysomnography, and electronic health records (EHRs), AI can provide a comprehensive view of a patient's sleep patterns and health status.
As the sleep monitoring market innovates, wearable devices have become the preferred tools for AI-enhanced sleep monitoring. Viatom's devices are delicately engineered to track an extensive array of vital signs, catering to diverse usage scenarios. As the leading supplier in the sleep monitoring market, we stay open-minded and actively seek industry collaborations to develop cutting-edge solutions that improve sleep medicine and overall well-being.
References:
[1] SLEEP 2024 | APSS Annual Meeting | AASM | SRS (2024)
[2] Multicentre sleep‐stage scoring agreement in the Sleep Revolution project (2024)
[3] An approach for determining the reliability of manual and digital scoring of sleep stages (2023)
[4] Sleep scoring moving from visual scoring towards automated scoring (2022)
[5] Endotypes and phenotypes in obstructive sleep apnea (2020)
[6] Upper-Airway Collapsibility: Measurements and Sleep Effects (2015)
[7] Management and treatment of patients with obstructive sleep apnea using an intelligent monitoring system based on machine learning aiming to improve continuous positive airway pressure treatment compliance: randomized controlled trial (2021)
[8] The National Sleep Research Resource: towards a sleep data commons (2018)
[9] UK Biobank: An Open Access Resource for Identifying the Causes of a Wide Range of Complex Diseases of Middle and Old Age (2015)
[10] The Human Sleep Project - Brain Data Science Platform (2023)
[11] Cardiovascular risk and mortality prediction in patients suspected of sleep apnea: a model based on an artificial intelligence system (2021)
[12] Major Depressive Disorder Prediction Based on Sleep-Wake Disorders Symptoms in US Adolescents: A Machine Learning Approach from National Sleep Research Resource (2024)
[13] Aeroflow Sleep leverages EnsoData’s novel PAP therapy prediction AI model, EnsoTherapy, to improve treatment compliance (2024)
[14] IHH invests in Belun Technology for AI-powered sleep diagnostic technology (2024)
[15] Mount Sinai to develop sleep apnea outcome risk prediction models (2024)
[16] AcuPebble sleep apnoea device used in University of Warwick trial (2024)
[17] FDA Clears HoneyNaps’ AI-Powered Sleep-Scoring Software (2023)
[18] LwPTL: a novel classification for upper airway collapse in sleep endoscopies (2019)
[19] Snore Sound Analysis Can Detect the Presence of Obstructive Sleep Apnea Specific to NREM or REM Sleep (2018)
[20] The theranostic value of acetylation gene signatures in obstructive sleep apnea derived by machine learning (2023)
[21] An introduction to artificial intelligence in sleep medicine (2021)
[22] Multiple sleep latency test and polysomnography in patients with central disorders of hypersomnolence (2020)
[23] Differentiation of central disorders of hypersomnolence with manual and artificial-intelligence-derived polysomnographic measures (2024)
[24] Data-Driven Phenotyping of Central Disorders of Hypersomnolence With Unsupervised Clustering (2022)
[25] Interpreting machine learning models to investigate circadian regulation and facilitate exploration of clock function (2024)
[26] Deep Learning for Automatic Detection of Periodic Limb Movement Disorder Based on Electrocardiogram Signals (2022)
[27] Oxford Handbook of Digital Ethics (2022)