Flutter waves (f waves) in atrial flutter (AFL) can be used as a diagnostic marker for estimating characteristics of the pathology. During AFL, the high synchronicity between f waves and the QRST complex create a dependency link, which invalidates many QRST cancellation algorithm. Recently a new technique was developed to correct f waves overlapped with T waves. This technique requires the segmentation of f waves overlapped with T waves: a non-trivial task. To address this problem, we developed a new detector algorithm based on statistical principles.
Several detector models were developed based on the principle of statistical hypothesis testing, consisting in maximizing the ratio of the likelihood function between two hypotheses. The problem was formulated as a weak signal detection problem in different noise type (Gaussian white and Laplace noise). The formulation is parameterized by 2 variables which represent amplitude modulation of the reference and the noise variance, and was modified to account for multiple leads and/or T wave overlap. Overall, there were 32 different detectors.
25 records of 12-lead ECG data from AFL patients were used for testing. f wave annotations were available over the total signal length. A single f wave reference was segmented by hand and serves as the reference. The test value was calculated at each time instant, and the subsequent peaks corresponding to an f wave were detected. To determine a suitable threshold for automatic detection using the limited number of annotations available, each model was tested on the ensemble of annotations of all the records using a leave-one-out scheme. Receiver-operating characteristic curves were obtained and the best model was selected.
The results of the test show that the single-lead Laplace likelihood ratio detector had the best performance (87% sensitivity and 77% specificity, AUC = 0.89) for an error of 20ms from the true annotations.