Fig. 4: Top-performing features in sleep-wake classification by importance. | npj Biosensing

Fig. 4: Top-performing features in sleep-wake classification by importance.

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

Fig. 4

(1) Unbalanced dataset: PPG time-domain and nonlinear dynamics features (e.g., PPG_skew, PPG_TM25, PPG_LC) and time-domain PPI characteristics (e.g., PPI_RMS, PPI_SVD) were among the most influential. (2) ADASYN-balanced dataset: Frequency-domain PPI metrics (e.g., PPI_LF_HF_power, PPI_VLF_LF_power) became more dominant, suggesting their increased role in distinguishing wake from sleep after balancing. For both (1) and (2), (a) shows the top 20 features for all participants, (b) for NFLE patients, and (c) for RBD patients. NFLE nocturnal frontal lobe epilepsy, RBD REM behavior disorder.

Back to article page