Frequency Tracking of Atrial Fibrillation using Hidden Markov Models

Abstract

A Hidden Markov Model (HMM) is used to improve the robustness to noise when tracking the atrial fibrillation (AF) frequency in the ECG. Each frequency interval corresponds to a state in the HMM. Following QRST cancellation, a sequence of observed states is obtained from the residual ECG, using the short time Fourier transform. Based on the observed state sequence, the Viterbi algorithm, which uses a state transition matrix, an observation matrix and an initial state vector, is employed to obtain the optimal state sequence. The state transition matrix incorporates knowledge of intrinsic AF characteristics, e.g., frequency variability, while the observation matrix incorporates knowledge of the frequency estimation method and SNRs. An evaluation is performed using simulated AF signals where noise obtained from ECG recordings have been added at different SNR. The results show that the use of HMM considerably reduces the average RMS error associated with the frequency tracking: at 5 dB SNR the RMS error drops from 1.2 Hz to 0.2 Hz