Spectra in low-rank localized layers (SpeLLL) for interpretable time-frequency analysis.

Area of Science:

• Biomedical Signal Processing
• Time-Frequency Analysis
• Statistical Modeling

Background:

• Biomedical time series often exhibit time-varying frequency characteristics crucial for scientific understanding.
• The high dimensionality of time-varying power spectra hinders direct application by researchers and clinicians.
• Elucidating complex biological mechanisms is challenging with traditional spectral analysis methods.

Purpose of the Study:

• To introduce a novel approach for time-frequency analysis of biomedical signals.
• To provide a parsimonious representation of the time-varying power spectrum.
• To facilitate the elucidation of complex biological mechanisms by applied researchers and clinicians.

Main Methods:

• Decomposition of the time-varying power spectrum into orthogonal rank-one layers in time and frequency.
• Application in fully nonparametric or semiparametric analyses, incorporating exogenous information and time-varying covariates.
• Estimation via a penalized reduced-rank regression framework for interpretable layer estimates.

Main Results:

• The proposed method yields interpretable layers representing power localized in specific time blocks and frequency bands.
• Simulation studies demonstrate the empirical properties of the estimation procedure.
• The approach effectively analyzes relationships between power at different times and frequencies.

Conclusions:

• The new time-frequency analysis method offers a simplified and interpretable representation of complex biomedical data.
• This approach enhances the utility of spectral analysis for understanding biological mechanisms.
• Practical application is demonstrated in heart rate variability analysis during sleep.