ECG segmentation in a body sensor network using adaptive hidden Markov models

In this paper, a novel approach for segmenting ECG signal in a body sensor network is presented. Hidden Markov Modeling (HMM) technique is employed. The parameter adaptation in traditional HMM methods is conservative and slow to respond to these beat interval changes. Since people's heart rates vary a lot, the corresponding characteristic waveform intervals and durations change with time as well. Moreover, for patients with cardiac diseases, such as arrhythmia, the heart beat interval may even change abruptly and irregularly. Therefore inadequate and slow parameter adaptation is largely responsible for the low positive predictivity rate (+P). To solve the problem, we introduce an active HMM parameter adaptation and ECG segmentation algorithm, which includes three parts: the pre-segmentation and classification, the HMM model training, and the detailed segmentation. Body sensor networks are used to pre-segment the raw ECG data by performing QRS detection. Then the R-R interval information that directly reflects the beat interval variation is extracted and used to classify the raw ECG data into several groups. One specific HMM is trained for each of the groups. Hence, instead of one single generic HMM, multiple individualized HMMs are set up. In the detailed segmentation, each HMM is only responsible for extracting the characteristic waveforms of the ECG signals with similar temporal features from the same group, so that the temporal parameter adaptation can be naturally achieved.