A Machine Learning Model for Obstructive Sleep Apnea Detection Using Ensemble Learning and Single-Lead EEG Signal Data

Sleep is crucial for cognitive and physical functions, and sleep disorders like Obstructive Sleep Apnea (OSA) can significantly affect a person's health. Polysomnography is the gold standard for diagnosing OSA, but despite its effectiveness, it is time-consuming and prone to human errors. To address this issue, this paper proposes an Ensemble Expert System for Obstructive Sleep Apnea Detection - II (EESOSAD-II) that leverages the single channel (C4-A1) Electroencephalography (EEG) signal and an ensemble learning model. The proposed model employs Discrete Wavelet Transform (DWT) with db8 for efficient EEG sub-band separation and statistical feature extraction. To enhance the data quality, the proposed model incorporates a Gaussian filter for feature smoothing and an Isolation Forest for outlier treatment. To further enhance the pre-processing pipeline, Recursive Feature Elimination (RFE) is used for sub-optimal feature set selection, and the Extra Tree classifier is employed for efficient classification of apnea and non-apnea events. The performance of the proposed model is evaluated using multiple evaluation metrics like - Precision, Recall, Accuracy, F1-Score and ROC_AUC curve for detailed analytical and benchmark comparison. The verification result shows that the proposed model achieved an average accuracy of 86% in comparatively optimized computational time than the state-of-the-art feature selection techniques. Furthermore, the EESOSAD-II outperformed the benchmark OSA detection model with optimal performance margin and achieved efficient performance results.