Validated Artificial Intelligence Models for Automation of Standard Diagnostics in Sleep Medicine - A Systematic Review
Maha Alattar1, Alok Govind2, Shraddha Mainali1
1Neurology, Virginia Commonwealth University, 2Neurology, National Institute of Mental Health and Neurosciences
Objective:
To review and select studies from the literature that report Artificial Intelligence (AI) models for automation of current standard diagnostics in sleep medicine which were validated on clinical datasets.
Background:
The current reliance on manual labeling of overnight Polysomnogram (PSG) makes the diagnosis of sleep disorders challenging, expensive, and laborious. AI serves as valuable complement to expert technologists and clinicians, enabling personalized medicine and reaching a wider audience including under-served populations.
Design/Methods:
PubMed and IEEE databases were searched for studies reporting novel AI models for application in sleep medicine that were developed using standard diagnostics and validated using independent training, validation, and testing datasets. Data, including model architecture, size, and composition of used datasets and performance metrics including agreement with consensus manual scoring and accuracy, were extracted.
Results:
2114 articles were retrieved from the search, of which 18 met the selection criteria. Ten automated sleep staging models and seven models for detection of sleep disorders were reported, all of which were deep-learning based. Examples include a Recurrent Neural Network (RNN) for sleep staging by long sequence modeling of 200 epochs (roughly one full sleep cycle) at a time, an RNN with Long Short Term Memory (LSTM) for Obstructive Sleep Apnea (OSA) screening using RR-Intervals, an ensemble Convolutional Neural Network (CNN) designed to generate a hypnodensity graph of sleep stage probabilities for Narcolepsy Type 1 diagnosis, a CNN for detecting REM Sleep Without Atonia (RSWA) using rectified Electromyographic (EMG) signals, and an LSTM network that extracts limb movement features from anterior tibialis EMG signals to classify Periodic Limb Movements of Sleep (PLMS).
Conclusions:
A mere 0.85% of the studies that were retrieved met our criteria for validation on unseen clinical data. This highlights the need for proving the generalizability of AI models as reliable tools for integration into the clinical practice of sleep medicine.