Exploring Heart Rate Variability Feature Importance for REM Sleep Behavior Disorder Classification: A Comprehensive Multi Dataset Study

Heart rate variability (HRV) is a promising biomarker for detecting subtle autonomic alterations in sleep disorders such as REM Sleep Behavior Disorder (RBD). However, the relevance of HRV-derived features depends heavily on methodological choices, including data segmentation, preprocessing, and dataset composition. This study systematically investigates how four key factors (epoch duration, inclusion of wake segments, dataset heterogeneity, and the presence of comorbidities such as Obstructive Sleep Apnea Syndrome (OSAS)), affect HRV feature discriminative power. Using multiple polysomnography datasets, we extract a comprehensive suite of 528 features per subject and evaluate their importance using a combination of statistical and machine learning-based techniques. Results show that longer epochs typically yield more robust features, though shorter windows enhance the prominence of non-linear metrics. Including wake segments improves feature relevance in heterogeneous datasets, supporting their use in real-world conditions. Furthermore, dataset aggregation enhances generalizability for full-night recordings, but may reduce performance when analyzing sleep-only data. Finally, including OSAS subjects had minimal effect on top-ranked features, indicating the robustness of HRV-based RBD markers across varying clinical contexts. These findings highlight the importance of thoughtful experimental design in HRV-based classification models in sleep research.