A comparative study of a hidden Markov model detector for atrial fibrillation

Abstract

A comparative study of several atrial fibrillation (AF) detection algorithms was done to determine the algorithm best suited for use in real clinical environments to detect AF in ambulatory ECGs. The algorithms that were investigated for this paper are based on the Hidden Markov Model (HMM), measures of variance, linear predictive coding, and measurement of approximate entropy (AE). Based on the results from the test data set, the HMM algorithm performed best for this application. In general, there is little difference between the performance of the HMM and AE algorithms. However, the implementation of the HMM algorithm is more computationally efficient. Because of the large amount of data that must be analyzed in ambulatory ECG recordings, the computational efficiency must be considered as an issue of practicality. Review of the data illuminated some of the strengths and weaknesses of the various algorithms. Variance measures performed with either high sensitivity or high positive predictivity, but were not able to achieve a desirable operating point that had both acceptable sensitivity and positive predictivity. Although AE and LPC had acceptable sensitivity and positive predictivity, the HMM performed even better than both of these in terms of overall error rate. It would seem that an observational model such as the HMM, fits the data better than parametric models such as AE and LPC. Finally, as the computing power of medical systems increases, more sophisticated algorithms may be exploited in ways that leads to more accurate computerized ECG interpretation.