<![CDATA[Class imbalance poses a challenge in EEG-based sleep stage classification, particularly for the N1 class, which has the fewest samples. This study compares three class imbalance handling strategies—SMOTE, Class-Weighted Loss, and Focal Loss—against a baseline using an identical 1D-CNN architecture. The Sleep-EDF Expanded dataset (30 subjects) was used, and evaluation was conducted via 3-fold subject-level cross-validation, ensuring subject-level separation between training, validation, and test sets. Results show that while the baseline achieved the highest overall accuracy (91.0%), it yielded the lowest balanced accuracy (72.9%) and N1 recall (25.1%). SMOTE improved N1 recall to 38.8%, whereas Class-Weighted Loss and Focal Loss achieved the highest N1 recall rates at 56.2% and 50.6%, respectively, with balanced accuracies of 77.6% and 77.8%. These findings indicate that, in this experiment, algorithm-level approaches were more effective than SMOTE at enhancing N1 class sensitivity and balancing performance across classes. However, this advantage did not extend to all metrics, as SMOTE produced higher accuracy and N1 F1-scores compared to the two algorithm-level approaches. Therefore, selecting a strategy to address class imbalance requires weighing the trade-off between minority class sensitivity and overall classification performance.]]>
Copyrights © 2026
