Machine-learning-based classification of obstructive sleep apnea using 19-channel sleep EEG data.

OBJECTIVE

This study aimed to investigate the neurophysiological effects of obstructive sleep apnea (OSA) using multi-channel sleep electroencephalography (EEG) through machine learning methods encompassing various analysis methodologies including power spectral analysis, network analysis, and microstate analysis.

METHODS

Twenty participants with apnea-hypopnea index (AHI) ≥ 15 and 18 participants with AHI <15 were recruited. Overnight polysomnography was conducted concurrently with 19-channel EEG. Preprocessed EEG data underwent computation of relative spectral power. A weighted network based on graph theory was generated; and indices of strength, path length, eigenvector centrality, and clustering coefficient were calculated. Microstate analysis was conducted to derive four topographic maps. Machine learning techniques were employed to assess EEG features capable of differentiating two groups.

RESULTS

Among 71 features that showed significant differences between the two groups, seven exhibited good classification performance, achieving 88.3 % accuracy, 92 % sensitivity, and 84 % specificity. These features were power at C4 theta, P3 theta, P4 theta, and F8 gamma during NREM1 sleep and at Pz gamma during REM sleep from power spectral analysis; eigenvector centrality at F7 gamma during REM sleep from network analysis; and duration of microstate 4 during NREM2 sleep from microstate analysis. These seven EEG features were significantly correlated with polysomnographic parameters reflecting the severity of OSA.

CONCLUSIONS

The application of machine learning techniques and various EEG analytical methods resulted in a model that showed good performance in classifying moderate to severe OSA and highlights the potential of EEG to serve as a biomarker of functional changes in OSA.