Combining HRV features for automatic arousal detection

Arousals are vital for sleep as they ensure its reversibility. However, an increased amount of arousals might indicate sleep disturbances or disorders. Since arousal events are similar to wake states but much shorter than the standard annotation epoch length of 30 s, they degrade sleep staging classification performance. Arousals are also related to physiological activities, such as cardiac activation, thus making the detection in a less disturbing way than with polysomnographies in sleep laboratories possible. Therefore, we analyzed 72 features derived from the heart rate variability (HRV) of 15 whole-night polysomnographic ECG recordings to quantify cardiac activation during sleep. After calculating the Mahalanobis distance (MD), ranking the best uncorrelated features and performing MANOVA, we show that combining multiple features increases the discriminative power (MD=1.56, χ²=33117) to detect arousals during the night compared to the best single feature (MD=1.16, χ²=16633). A linear mixed model is used to show between-subject effects and to validate the significance of each feature based on Wald test statistics.