Comparing hidden Markov model and hidden semi-Markov model based detectors of apnea-bradycardia episodes in preterm infants

In this paper, we propose, evaluate and compare two detectors of apnea-bradycardia episodes, based on hidden Markov models (HMM) and hidden semi-Markov models (HSMM). Evaluation is performed on a database of 233 apnea-bradycardia episodes manually annotated. The acquired ECG signals are processed to obtain RR series. The proposed detectors, applied to these RR series, are composed of two HMM or HSMM models, each one representing two distinct physiopathological states: absence and presence of apnea-bradycardia. A learning phase is firstly applied to each model in order to estimate their parameters from a learning dataset. Then, using a sliding window, the models are applied to a set of new observations, to compute the log-likelihood of each model for each time instant. Detection of the events of interest is based on the comparison of log-likelihoods with respect to a threshold. The optimal detection configuration was obtained in terms of sensitivity, specificity and detection delay. Results show that the analysis of the dynamics of RR series, through the HSMM, allows for a significant improvement of sensitivity (90.38% vs 88.42%) and specificity (92.23% vs 89.67%), with a reduction of the detection delay (0.92±3.56s vs 1.60±3.72s).