Identification of individual sleep apnea events from the ECG using neural networks and a dynamic Markovian state model

A two step method is introduced that uses heart beat related information from the ECG to identify individual sleep apnea events. In the first step a set of features based on the RR interval, QRS dynamic, T-wave morphology and high frequency noise is used as input of a neural network to assign each heart beat to one of four consecutive apnea states. In the second step the output of the neural network is proceeded by a dynamic Markovian state model which determines the most probable state trajectory. The 35 records of the learning set from challenge 2000 are used for model adjustment. For that purpose the authors examined about each 10^th minute of the learning records and manually annotated the location of apnea states. Although only a part of the information is used we find 89% agreement with the original annotations within the learning set and 84% agreement if applied to the 35 records of the test set from challenge 2000.