<![CDATA[Sleep stage classification is essential for diagnosing sleep disorders such as insomnia and sleep apnea. However, manual scoring of polysomnography (PSG) is time consuming and subjective. Automatic systems based on single channel EEG are promising for home based monitoring, but they face challenges due to class imbalance and inter subject variability. This study proposes a hybrid model that combines 15 handcrafted features (statistical and spectral) with 128 dimensional features extracted by a one dimensional Convolutional Neural Network (1D CNN), followed by a stacking ensemble (Random Forest and Support Vector Machine as base learners, Logistic Regression as meta learner). Using 40 subjects from the Sleep EDF Expanded dataset, a strict subject independent split (80% train / 20% test) was applied to avoid data leakage. The dataset contained 107,258 epochs with extreme imbalance (Wake 67.6%, N1 2.95%). After SMOTE oversampling on the training set, the model achieved an accuracy of 67.5%, macro F1 score of 31.4%, and Cohen’s Kappa of 0.34. An ablation study showed that CNN features alone (72.2% accuracy) outperformed handcrafted features (70.4%) and hybrid features (67.5%). The confusion matrix revealed that minority stages (especially N1, N3, REM) were poorly recognized. These results highlight that cross subject generalization remains a major challenge in EEG based sleep staging, and proper subject independent validation is critical to avoid overoptimistic claims.]]>
