this paper aims at investigating different methods for the detection of the start and end of bradycardias in heart rate signal of premature babies. We present two methods based on a disturbance detector and on a decision tree that are compared to classical thresholding approaches. Decision tree obtained the best detection results (Se=78.2%, PP=68.7%) against the disturbance detector (Se=90.2%, PP=61.3%) and the best thresholding method (Se=92.5%, PP=46.5%). Moreover, the decision tree exhibits better performance for the boundaries estimation (median delay = 7-5 seconds) than the disturbance detector (median delay = 8-5 seconds) with a better stability (STD=8.5 to 8.7s vs. STD=35.3 to 19.9s). These methods will be integrated to the BabyTalk project which aims at summarizing neonatal clinical data as text in order to improve data management in Neonatal ICU.
展开▼