Time-frequency decompositions: Bayesian model-based approaches

Summary form only given. A range of developments in Bayesian time series modelling in prevoius years has focussed on issues of identifying latent structure in non-stationary time series, particularly driven by applications in which time-varying spectral structure of time series is an inherent and prime feature. This article reviews some of these developments, including the theoretical and methodological basis of decomposition methods in state-space models. The resulting methods can be viewed as providing a time-domain representation of changing spectral characteristics. Examples are drawn from problems in clinical EEG studies, where the assessment of changes over time in the frequency structure of components of EEG signals is key to characterising brain seizures under various treatments.