Performance of an arrhythmia analysis system based on hidden Markov models

An arrhythmia analysis system based on hidden Markov models is evaluated to investigate noise tolerance. Although detection performance of the system using model parameters estimated from clean data was unimpressive, QRS detection performance was significantly improved using a simple noise compensation strategy which consisted of adding the noise variance to output function variance parameters. Unfortunately, this strategy appears to cause the model to lose its ability to discriminate between normal/abnormal QRS complexes. The results presented are preliminary, but they suggest that alternate noise compensation algorithms should be investigated