0317 Quantifying Importance Of Electroencephalography Spectral Domain Features In Automatic Diagnosis Of Chronic Insomnia

• Published in: Sleep (New York, N.Y.) N.Y.), 2019-04, Vol.42 (Supplement_1), p.A130-A130
• Main Authors: Bouchequet, Paul, Solelhac, Geoffroy, Andrillon, Thomas, Romano, Francesco, Brigham, Marco, Chennaoui, Mounir, Léger, Damien
• Format: Article
• Language: English
• Subjects: Algorithms; Electroencephalography; Eye movements; Insomnia; NREM sleep; REM sleep; Sleep; Support vector machines
• ISSN: 0161-8105; 1550-9109
• DOI: 10.1093/sleep/zsz067.316

Introduction Polysomnographies (PSG) electroencephalographic (EEG) records contains many relevant information unused in clinical processes. Algorithms commonly used in machine learning can help us identify the most important features used by models in classification problems. The objective is to compare the efficiency of EEG Rapid Eye Movement (REM) and Non-Rapid Eye Movement (NREM) sleep features from PSG in the detection of chronic insomnia between control records. Methods 299 PSG have included: 54 controls subjects and 245 chronic insomniacs. Spectral power of the EEG central derivation (C3-M2) have been extracted then divided into 0,5 Hertz (Hz) bands from 0,5 Hz to 40 Hz with Fast Fourier Transforms (FFT) for each REM and NREM 30 seconds sleep epochs. Bands powers have been normalized by dividing by the broadband (0,5-40hz) power. For each PSG, average power for each band have been computed for REM and NREM epochs. A few algorithms, including linear support vector machine (SVM) and random forests, have been trained, firstly with NREM, then with REM features to detect chronic insomnia diagnosis. Global performance have been estimated with a Cohen Kappa (CK) test on a data subset, unused during training (train/test split 0,7, Bootstrap method). Individual importance of each feature have been estimated with the area under the receiver operating characteristic curve (ROC AUC). Results SVM is better at chronic insomnia detection with REM features than with NREM features (CK > 0.89). REM bands between 3-6hz have the highest ROC AUC of all the features (>0.9). Conclusion EEG spectral domain features from REM sleep are better to diagnose chronic insomnia than spectral domain features from NREM sleep. Differences appeared in specific REM sleep brain oscillations between controls and chronic insomniacs. Support (If Any) Banque Publique d'Investissement, Dreemcare Project.