Scaler–Sampler–Cost Interactions in Classical Sleep Staging: Utility-Based Trade-offs for N1 on Sleep-EDF

Abstract

Automated sleep staging is still difficult. Because of its transitional nature and overlap with adjacent stages, stage N1 shows the lowest reliability. In this work, we examine whether N1 performance under leakage-safe evaluation can be enhanced by cost-sensitive learning. We also look at how it affects SMOTE-family resampling and feature scaling. We use a multilayer perceptron with five stages and 30-s epochs from Sleep-EDF Expanded dataset. 28 handcrafted features from Fpz-Cz EEG channel are used to train the model. Three scalers— z-score, robust, and min-max—are assessed. Additionally, we assess three cost formulations (Inverse-frequency, Effective-number, and Log-scaled) and three resampling techniques (SVMSMOTE, SMOTETOMEK, and SMOTE with AllKNN). Every experiment is carried out using subject-wise 5-fold cross-validation. The best-performing global option, according to the results, is inverse-frequency sample weighting without resampling. Compared to SMOTE-based methods, it achieves competitive N1 improvements. On the other hand, SVM-based resampling with effective-number weighting is most advantageous for min-max scaling. Additionally, a within-cell added-value analysis indicates that cost-sensitive learning yields its greatest benefits when resampling is not used. Once SMOTE-family sampling is used, it provides little additional benefit. Reducing misclassifications into W/N2/REM is the main source of N1 improvements, according to confusion-matrix inspection. All things considered, the study offers useful advice for choosing N1-oriented configurations. Under realistic, leakage-safe evaluation, it also takes into consideration trade-offs in non-target stages.