Fig. 2.

The order of feature importance for (a) automated rapid-eye movement (REM) classification (sleep staging) using electrocardiogram (ECG), electrooculogram (EOG), and electromyogram (EMG) features (top 25) and (b) REM sleep behaviour (RBD) detection (top 13). From (a) the EOG features appear most important for REM classification, specifically permutation entropy, max peak, and coastline features. Followed by elapsed recording time and EMG features, such as the 75th percentile, entropy, relative power, and motor activity. These additional features supplement the EOG features to provide a boost in REM classification performance, shown in Fig. 1 (see sensitivity of B1 and C2). From (b) the feature importance for RBD detection clearly illustrated that EMG metrics outperform ECG metrics. Of the ECG metrics, irregular evidence (origin count and irregular index) during non-REM (NREM) appeared the most effective for RBD detection. These are followed by frequency-based ECG metrics, such as low frequency peak, high frequency peak, and the low frequency to high frequency (LFHF) index.