Background: The cycle length is a measure for the characterization of the electrical activity of the heart substrate, serving to guide catheter ablation. Recent research has shown that areas where catheter ablation led to an acute termination of persistent AF demonstrated rapid activity. Thus, a reliable estimation of the cycle length of atrial tissue during the mapping phase is essential. The local electrical activity is often generated by several neighboring sources with different characteristics resulting in electrograms with high variability of wave morphology and amplitude, thus the accurate analysis of CL is challenging. Intracardiac recordings are known to be influenced by added noise and far field activity during non-contact phases. Previous works are mostly based on dominant frequency estimation which is ambiguous in fractionated, poorly organized AF samples. Some methods based on amplitude thresholding were suggested recently.
Method: The input signals of 1500ms are initially preprocessed with use of wavelet high frequency noise reduction. A research of extrema with adaptive threshold is employed, obtained extrema are clustered by activations using sliding windows. A reference is chosen inside each activation and the final cycle length is estimated as the mean of intervals between reference positions. When the variability of intervals is too high, post-processing is applied. It consists in the adaptive search of missed activations in highly distant intervals and a merge of neighboring activations for the overly close ones. If the variability reduces, the post-processed segmentation is retained.
Results: A validation dataset of 40 AF, 21 AT periodic samples were annotated by two physicians. Obtained mean absolute errors are 6.86ms, 8.06ms respectively. Several examples demonstrating periodic behavior but with a cycle length difficult to identify by humans were also chosen and the method's outputs were validated by doctors. The samples where errors exceeded 15% were reviewed with physicians.