Signal-Driven Model Order Selection for MUSIC-Based HRV Spectral Characterization

Heart rate variability (HRV) is a useful non-invasive tool for studying autonomic nervous system modulation under emotional stimulation; however, accurate estimation of dominant frequencies in HRV signals remains challenging due to their non-stationary nature and the sensitivity of some spectral methods to configuration parameters. This work presents a methodology for the spectral characterization of HRV signals derived from ECG recordings from the DREAMER database, with emphasis on optimizing the model order of the MUSIC algorithm to improve dominant frequency localization within the physiological low-frequency (LF) and high-frequency (HF) bands. The proposed methodology included ECG signal preprocessing, R-peak detection, RR interval extraction, HRV interpolation, and spectral analysis using MUSIC, while evaluating different model orders through a signal-driven composite criterion based on AIC, MDL, ESTER, eigengap, and model complexity. The criteria were normalized using min-max normalization and combined using equal predefined weights. The results showed that the signal-driven selection of the parameter p produced recording-dependent model order configurations and different dominant frequency estimates across the analyzed stimuli. The resulting LF/HF agreement was evaluated independently after model order selection and showed non-uniform correspondence across stimuli and spectral estimators. Overall, these findings indicate that model order selection can substantially influence the spectral characterization obtained with MUSIC and provide a signal-driven framework for examining this dependence in HRV recordings.