Decoding obstructive sleep apnea phenotypes through integrated multimodal sleep EEG

• Obstructive sleep apnea (OSA): repeated collapse of the upper airway during sleep, causing breathing pauses.

• Phenotype/subtype: a subgroup of patients who share a recognizable pattern of characteristics.

• EEG (electroencephalography): recording of the brain's electrical activity through scalp electrodes.

• PSG (polysomnography): the full overnight sleep study that records brain activity, breathing, and other body signals.

• AHI (apnea-hypopnea index): breathing interruptions per hour of sleep, the standard severity measure. This study required AHI ≥ 5 events/hour.

• Sleep macrostructure: the overall architecture of sleep, such as time spent in each stage and sleep efficiency.

• Sleep stages (N1, N2, N3, REM): N1 is light sleep, N2 is intermediate sleep, N3 is deep "slow-wave" sleep, and REM is the dreaming stage.

• Spectral power / PSD (power spectral density): how much of the EEG signal falls in each frequency band.

• Frequency bands (delta, theta, alpha, beta, gamma): EEG rhythms from slow (delta) to fast (gamma).

• Multimodal data: several types of data combined, here demographics, PSG measures, EEG, and blood biomarkers.

• Transformer: a deep-learning model that uses "attention" to weigh which inputs matter most.

• Unsupervised learning / clustering: computer methods that find groups in data without being told the groups in advance.

• Gaussian mixture model (GMM): a clustering method that models data as overlapping bell-shaped groups.

• Bayesian information criterion (BIC): a score used to choose the model that fits well without being needlessly complex.

• Consensus clustering: repeating clustering with different methods and keeping the groupings that appear consistently.

• Permutation feature importance: shuffling one variable at a time to see how much the grouping gets worse, which shows how much that variable matters.

• False discovery rate (FDR): a statistical correction that limits false positives when many tests are run.

• Neuro-respiratory control: the brain's regulation of breathing.

1. Can OSA subtypes be identified from objective brain signals (sleep EEG), instead of symptom reports or the AHI alone? The authors argue that current subtyping relies on subjective reports or coarse physiological measures. These cannot explain the underlying neurophysiological mechanisms.

2. Does combining EEG with demographics, sleep-study measures, and blood biomarkers produce stable and clinically meaningful subtypes?

3. Which EEG features best separate the subtypes, and how do the subtypes differ clinically and physiologically?

Sample and approach

• The source was the Wuhan database of 903 people. Of these, 505 had OSA (AHI ≥ 5), and 196 were excluded for poor EEG channels or more than 50% wakefulness. The final cohort was 309 patients (Figure 1).

• Four data types were combined: demographics (7 features), sleep-study macrostructure (6), blood biomarkers (4), and EEG spectral features (226 after filtering). That gave 243 features in total.

• A cascaded multimodal transformer encoded each data type and fused them by cross-modal attention. Several clustering algorithms were compared. The Gaussian mixture model with three components was chosen after internal validation metrics and BIC minimization, with consensus clustering used to check robustness.

Three subtypes emerged (Table 1; Figure 2A shows them separating visually):

• Group of 52 (16.8%), labeled Phenotype 0 in the figures. The Table 1 column header for this group was lost in extraction, so this label was inferred from the figures. Mean age was about 53, the youngest group.

• Phenotype 1, 131 patients (42.4%). This group was the oldest (about 61 years) and had almost no deep sleep (N3 about 0.9% of sleep). It also had the lowest sleep efficiency (0.71).

• Phenotype 2, 126 patients (40.8%). Mean age was about 56. It had the highest proportion of deep sleep (N3 about 9.5%), the least N2 sleep, and the highest sleep efficiency (0.77).

Where the subtypes differed in Table 1

• Age (p = 0.0002), N2 and N3 sleep percentages (both p < 0.0001), and sleep efficiency (p = 0.0184) differed.

• One further row, whose label was lost, also differed significantly (p = 0.0019).

• Body measures (neck, waist, hip, BMI), snoring, arousal index, N1 and REM percentages, and blood cell counts did not differ significantly. Daytime sleepiness (ESS) and the comorbidity rows that were visible also showed no significant differences. The rows for the apnea index and oxygen measures were cut off, so these results could not be assessed.

EEG features drove the grouping

• The top-ranked features (Figure 2B) were mainly REM-related or REM-to-N3 ratio measures. Examples include REM-over-N3 frontal theta, REM frontal beta, N2 frontal SWI, N3 frontal-over-occipital gamma, and REM-over-N3 occipital alpha. Figure 4 shows how the groups differ on these features.

• Figure 3 compares band power across sleep stages and brain regions (frontal, central, occipital). Several differences between groups are marked, especially in deep sleep (N3) and REM. The blue group appears to have notably higher beta and gamma power in N3. The figure does not provide enough detail to state this as a firm result.

• Figure 5 correlates EEG features with clinical and blood variables, using FDR correction. Only a few associations were marked significant.

The authors present their approach as a data-driven framework that goes beyond AHI-centered and symptom-based classification of OSA.

By combining multimodal data with a transformer and unsupervised learning, they report three stable, clinically distinct subtypes defined mainly by sleep neurophysiology.

They say EEG features were the main drivers of the clustering. They point to the weight of the EEG modality in the fusion stage and to the top ten discriminating features, which included REM-derived and cross-stage REM/N3 measures.

They link this to the view that OSA is a sleep-state-dependent disorder of neuro-respiratory control, not just a matter of how often breathing events occur. Their introduction adds that EEG's high resolution may help reveal the mechanisms behind OSA's variability.

The authors' stated limitations, clinical implications, and recommendations for future work were not in the readable text.

Limitations

• This is a single-cohort study of 309 patients, and no external validation was visible in the readable text.

• The subtypes differ in age, so some of the apparent brain-signal differences could partly reflect age. The visible text does not show whether the authors adjusted for this.

• Figure 3 comparisons were explicitly not corrected for multiple testing, as the Methods state. They are exploratory.