Une approche multivariée pour la détection d'épisodes d'apnée-bradycardie par modèles semi-Markoviens cachés.
Résumé
In this work, hidden Markov models (HMM) and hidden semi-Markov models (HSMM) are adapted for the classification, according to the maximum likelihood, of the dynamics of multivariate time series, obtained before and after apnée-bradycardie events in perterm infants. A phase of preprocessing of the observations, including signal quantization and the integration of delayed versions of each data source, is proposed. Results highlight the importance of considering the dynamics of the signals, show that HSMM are better adapted than HMM to our problem and emphasize that, with a suitable preprocessing, such as the quantification of observations and the introduction of an optimal delay between the observables, a significant gain in performance can be obtained.