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

Figures

• 

  Figure 1. Circumplex model of affect based on Russell’s framework [10], illustrating the valence–arousal space and the distribution of representative emotional states.

• 

  Figure 2. Valence–arousal representation of the DREAMER stimuli. Colors indicate the a priori expected LF/HF tendency. Dashed lines at 3.0 define the high/low thresholds. Numbers identify the corresponding stimuli listed in Table 1; some labels may be partially or completely obscured by overlapping markers. Stimuli 2, 9, and 17 are boundary cases. Assignments are descriptive hypotheses, not direct measures of autonomic activity.

• 

  Figure 3. Block diagram of the proposed HRV analysis methodology using ECG signals from the DREAMER database. The framework includes ECG preprocessing, R-peak detection, HRV interpolation, filtering of the interpolated HRV signal, window-based spectral analysis, MUSIC model order optimization, CWT analysis, and autonomic interpretation based on LF/HF band analysis.

• 

  Figure 4. ECG preprocessing stages for Subject 9 under Stimulus 11 (calmness): raw ECG, band-pass filtered ECG, and detected R-peaks. The horizontal axis represents time (s), and the vertical axis represents ECG amplitude (mV). The recording segment shown corresponds to the analyzed 60 s stimulus interval. Red circles indicate the detected R-peaks. (a) Raw ECG signal for Subject 9 under Stimulus 11 (calmness). (b) Band-pass filtered ECG signal for Subject 9 under Stimulus 11 (calmness). (c) Detected R-peaks from the filtered ECG signal for Subject 9 under Stimulus 11 (calmness).

• 

  Figure 5. HRV generation and preprocessing stages for Subject 9 under Stimulus 11 (calmness): RR-derived HRV, PCHIP-interpolated HRV, and filtered interpolated HRV. The horizontal axis represents time (s), and the vertical axis represents RR interval (s). The interpolated HRV was sampled at 8 Hz, and the final signal was filtered in the 0.03–0.40 Hz range. The recording segment shown corresponds to the analyzed 60 s stimulus interval. (a) HRV signal obtained from RR intervals for Subject 9 under Stimulus 11 (calmness). (b) Interpolated HRV signal used for spectral analysis for Subject 9 under Stimulus 11 (calmness). (c) Filtered interpolated HRV signal used for spectral analysis for Subject 9 under Stimulus 11 (calmness).

• 

  Figure 6. Normalized MUSIC pseudo-spectrum for Subject 9 under Stimulus 11 (calmness), obtained from the final 60 s HRV segment using the signal-driven model order selection procedure. The horizontal axis represents frequency (Hz) and the vertical axis represents the normalized MUSIC pseudo-spectrum. The LF and HF regions are indicated using the 0.04–0.15 Hz and 0.15–0.40 Hz frequency ranges, respectively, with the LF/HF boundary marked at 0.15 Hz.

• 

  Figure 7. Time–frequency representation of the HRV signal for Subject 9 under Stimulus 11 (calmness) obtained using the Continuous Wavelet Transform (CWT). The horizontal axis represents time (s), and the vertical axis represents frequency (Hz), covering the analyzed 0.04–0.40 Hz frequency range. The dashed white line at 0.15 Hz indicates the LF/HF boundary, while the upper dashed line marks the dominant frequency estimated from the CWT representation. The color scale represents the normalized CWT magnitude.

• 

  Figure 8. Distribution of the selected MUSIC model orders obtained with the fixed p=4, AIC, MDL, ESTER, eigengap, and proposed signal-driven criteria across the 414 DREAMER participant–stimulus recordings. AIC and MDL consistently selected the upper candidate order (p=10), whereas the proposed criterion selected recording-dependent orders within p=3–5. No physiological or emotional information was used during model order selection.

• 

  Figure 9. Temporal sequence of the signal-driven MUSIC model order p∗(w) obtained for Subject 9 under Stimulus 11 (calmness). The HRV signal was analyzed using 30 s windows with 50% overlap. The dashed horizontal line indicates the most frequently selected window-level model order. The analysis is presented only as an exploratory assessment of temporal model order stability and does not involve LF/HF classification.