A fully automated framework for detecting phasic REM sleep based on hybrid adaptive segmentation of a single EOG channel yielded 92.9% correct detections and physiologically consistent REM microstructure when validated on clinical polysomnography recordings.
Key Findings
Results
The hybrid adaptive segmentation framework achieved 92.9% correct event-level detections on EyeCon datasets, with 5.3% fragmentation and 1.8% missed events.
Event-level analysis was performed on controlled EyeCon datasets with precise ground-truth event markers.
All segmentation and classification hyperparameters were optimized exclusively on EyeCon datasets before being applied to clinical PSG recordings.
The framework used a single EOG channel for detection.
Results
Phasic REM accounted for 31.8 ± 3.5% of total REM duration, and tonic REM accounted for 68.2 ± 3.5% when the model was applied to 21 full-night clinical PSG recordings.
The model was applied without modification to 21 full-night clinical polysomnography recordings.
Phasic REM proportion: 31.8 ± 3.5% of REM duration.
Tonic REM proportion: 68.2 ± 3.5% of REM duration.
The authors describe these results as 'physiologically consistent' REM microstructure.
Results
EEG analysis confirmed increased beta and gamma power during phasic REM compared to tonic REM, supporting the physiological validity of the automated classifications.
Additional EEG power spectral analysis was conducted to validate the physiological plausibility of detected phasic and tonic REM segments.
Increased beta and gamma frequency band power was observed during phasic REM periods.
The authors state this finding provides 'physiological validity' for the proposed framework.
This EEG analysis was performed on the 21 clinical PSG recordings.
Methods
The segmentation algorithm fuses median absolute deviation (MAD) amplitude-change detection with a morphology score derived from a custom saccade kernel built from manually verified EyeCon recordings.
Segment boundaries were refined using local derivative extrema to improve temporal alignment.
A supervised support vector machine (SVM) classifier further refined segment labels using features based on saccade morphology.
Features included correlations with custom log-sigmoid templates and a morphology similarity measure.
The custom saccade kernel was built from manually verified EyeCon recordings rather than generic wavelet families.
Background
The framework was designed to address limitations of existing EOG methods that rely on fixed thresholds or generic wavelet families that do not accurately capture real saccade morphology in clinical polysomnography.
Existing methods were characterized as relying on 'fixed thresholds or generic wavelet families.'
The authors note that 'automated quantification of phasic versus tonic REM remains uncommon.'
Phasic REM is defined by brief clusters of saccadic eye movements and transient cortical activation.
Phasic REM has been linked to emotional memory consolidation, sensorimotor integration, and autonomic modulation.
Methods
The framework uses a single-channel EOG design and was externally validated on clinical PSG recordings without modification of hyperparameters optimized on EyeCon data.
The single-channel design was identified as making the tool 'suitable for both retrospective analyses and future clinical applications.'
Hyperparameters were optimized exclusively on controlled EyeCon datasets and the final model was applied without modification to 21 full-night clinical PSG recordings.
The authors describe the framework as 'interpretable, morphology-aware, and computationally efficient.'
External validation on clinical PSG represents a separate dataset from the optimization dataset.
What This Means
This research suggests that REM sleep — the stage of sleep associated with dreaming — is not uniform but actually consists of two distinct sub-states: 'phasic' REM, characterized by rapid bursts of eye movements and brief brain activation, and 'tonic' REM, which is quieter. Scientists have long known about these sub-states, but automatically detecting and measuring them in overnight sleep recordings has been difficult because existing computer methods use overly simplistic rules that don't match how real eye movements actually look in medical sleep studies. This study developed a new automated computer system that learns the shape of real saccadic (rapid) eye movements from carefully verified recordings and uses this knowledge, along with a machine learning classifier, to distinguish phasic from tonic REM using just a single eye-movement recording channel.
When tested on controlled recordings with known ground truth, the system correctly identified 92.9% of eye movement events, missed only 1.8%, and slightly over-split another 5.3%. When applied to 21 full overnight clinical sleep recordings from real patients, the system found that phasic REM made up about 31.8% of total REM sleep time and tonic REM made up 68.2% — proportions that align with what sleep researchers expect physiologically. Brain wave (EEG) data also confirmed that the periods the system labeled as phasic REM had higher activity in the beta and gamma frequency ranges, which is what neuroscience predicts, providing additional evidence that the automated labels are accurate.
This research suggests that it is now feasible to automatically and accurately measure the fine structure of REM sleep — how much time is spent in phasic versus tonic sub-states — from standard clinical sleep recordings without special equipment or manual scoring. This could be valuable for large-scale research into how REM sleep sub-states relate to memory, emotional health, and conditions such as depression or PTSD, and may eventually support clinical tools for assessing sleep quality in greater detail than current methods allow.
Nagy T, Piorecký M, Janků K, Piorecká V. (2026). Hybrid Adaptive Segmentation and Morphology-Based Classification of EOG for Automated Detection of Phasic and Tonic REM Sleep.. Sensors (Basel, Switzerland). https://doi.org/10.3390/s26041389