Fig. 5: Overlap and variability in top features across datasets and balancing methods. | npj Biosensing

Fig. 5: Overlap and variability in top features across datasets and balancing methods.

From: Interpretable feature-based machine learning for automatic sleep detection using photoplethysmography

Fig. 5

a Top 20 features across different dataset configurations: showing substantial overlap in feature selection across groups. b Top 5 features across datasets: highlights greater variability in the most critical predictors depending on dataset composition and sleep disorder type. Yellow indicates that a feature was included in the top-ranked features for a given dataset, while purple indicates that it was not. While some features, such as PPG_skew and PPG_TM25, were consistently among the top-performing features in all unbalanced datasets, and PPI frequency-domain features in all ADASYN-balanced datasets, the highest-ranked features differed more substantially between NFLE and RBD patients. This suggests that dataset composition and disorder-specific physiological influences impact feature importance.

Back to article page