Reliable discrimination of FF from NF BVE from the pulmonary veins during catheter ablation of AF is of high clinical importance. Especially for single-shot devices such as the CB catheter, correct electrogram interpretation can be challenging. The main findings of our study are as follows: [1] For the discrimination of farfield from nearfield signals using a single feature, the absolute power in high-frequency showed the best overall accuracy with 79.4%, followed by the proportion with a slew rate > 0.15 V/s (slew-rate), voltage amplitude, absolute power in low-frequencies (PLF), relative high-frequency power (PHF-rel) and PHF_Neighbor. [2] With multiple features in the prediction model, the combination of a frequency domain (PHF) and time domain analysis (Vmax) feature yielded the best overall accuracy of 82.7% to predict NF-PV signal from a single-beat BVE. ##With the (relatively) high specificity of 89% and a sensitivity of 77%, the implementation as a diagnostic test to identify if additional ablation for PVI is needed is reasonable. [3] On a vein-selective analysis, the overall accuracy was highest for the right inferior PV (96.6%) and lowest for the left superior PV (76.9%). Implementing the information on the distance between LSPV and the left atrial appendage, the accuracy of the algorithm could be improved for the LSPV. [4] The algorithm showed numerically comparable accuracy to the classification by five experienced EP specialists. However, further external testing on larger datasets is required to confirm the results.

Intracardiac uni- and bipolar voltage electrograms are the fundamental basis for any invasive electrophysiological study. The BVE measured between two electrodes of an EP catheter is used to characterize the underlying propagation of the depolarization of the myocardial cells based on temporal activation, electrogram morphology and amplitude. Despite reflecting more local myocardium depolarization than for unipolar voltage electrogram, the BVE is still influenced to some extent by distant or farfield depolarization of the myocardial tissue. In addition to the underlying tissue characteristics, the shape and size of the EGM [6] are strongly influenced by the electrode size, inter-electrode distance and relative orientation of the bipolar electrodes in relation to the propagating wavefront of the cellular depolarization [7].

PV potentials are defined and colloquially described as “sharp” nearfield BVEs following a farfield BVE from the LA. Dependent on the position of the CMC within the PV, the PV signal is more or less easily discernible and separable from the atrial-FF BVE and varies between the veins [8]. This “sharpness” of the local PV BVE is mentioned throughout the publications, but to the best of our knowledge, a reproducible, quantitative measure has never been published. The sharpness of an electrogram can be defined in the time domain based on the signal width and the slew rate of the deflection. For the sensing of atrial signals in cardiovascular implantable electronic devices, for instance, an interpolated slew rate in the range of 0.5V/s is recommended as a characteristic for nearfield atrial signal detection [9]. In our study, with a “sharpness criteria” of a cumulative threshold of a slew rate above 0.15V/s for the local BVEs, this feature, however, was not identified as a reliable predictor to identify nearfield BVE. However, the frequency domain–based high power frequency spectrum at a cut-off of 150 Hz might address this lack of definition for nearfield characterization.

Farfield (interference) elimination was aimed at using a novel dipole source model to describe the impact of the nearfield and farfield source on the measured voltage signal [10, 11]. With this approach, an improvement in the spatial resolution of the electrogram from approximately 10 to 2.5 mm was expected. The assumed spatial resolution of 10 mm with the standard voltage calculation is in line with our observation on the impact of the LAA on the LSPV BVE. With the inclusion of the distance between LSPV and LAA, we observed that our frequency-dependent algorithm works highly reliably at distances between LAA and LSPV above 10 mm with 0% false positive LA NF detection. However, with distances below 5 mm, the farfield BVE of the LAA shows similar frequency characteristics (especially the high-frequency power spectrum) with the nearfield PV signal. Using this dipole density modelling approach on our contact-based BVE might help to further improve the accuracy of our algorithm, especially for the LSPV with the LAA in close proximity to the CMC.

Another established approach to eliminate the impact of farfield on the BVE is to use dedicated catheter designs. In general, catheters with closely spaced electrodes are recommended. Computational simulations showed that smaller electrodes with narrow spacing produce sharper BVE with higher electrogram amplitude [5, 7]. However, when the option of selecting a specific catheter design is not available (as for CB PVI), the above-described frequency analysis with a dedicated high-frequency cut-off for the power spectrum might be a powerful alternative to eliminate the farfield impact of the local BVE.

A simple way to discriminate between FF and NF BVE is to observe the temporal evolution of the signal during ablation. When a local PV-NF is detectable, this signal shows a temporal delay with the advancement of the lesion, allowing the definition of a local PV signal based on the progression of the PV entrance block. However, no observable delay must not inevitably imply that the vein is already isolated, since BVE might still be hidden in the LA farfield component. Refraining from further ablation, the endpoint of PVI will not be reached. On the other side, with an already isolated vein, additional unnecessary ablation might result in complications, such as phrenic nerve palsy reported for CB PVI of the RSPV.

Numerous pacing strategies have been established in clinical practice, including decremental pacing, differential pacing, perivenous pacing or intravenous pacing [3]. However, when pacing from the distal CS, overlapping of the two BVEs (PV-NF and atrial-FF) was still observed in 65% of the patients.

Another approach to differentiate PV nearfield from LA farfield has been described using a multi-electrode mapping catheter in combination with an automated software implemented in the mapping system (Rhythmia, Boston Scientific, USA) [12]. With this software (Lumipoint), areas with a simultaneous electrical activation were highlighted, allowing for an identification of a farfield effect from a surrounding structure, such as for instance the LAA. However, this approach requires a detailed electroanatomical mapping after ablation. Time domain bipolar voltage electrogram characteristics were used to define a library of characteristic PV electrograms [13]. Besides the typology (including the amplitude and the number of peaks), the minimal and maximal slope of the BVE, its peak angle and amplitude were used to characterize the BVE. Using this library, a 2-step algorithm showing an accuracy of 93% was developed. In a subsequent study, the library-dependent classification algorithm on BVE was expanded and validated to the herein-used octapolar CMC for CB PVI [14]. In contrast to this strategy, our algorithm is not dependent on a training library and has the potential to be implemented as fully automatic approach, since only two simple features are required. A frequency-based analysis of the PV BVE from a decapolar CMC with 15 mm, 20 mm or 25 mm diameter (Lasso, Biosense Webster) was already performed 10 years ago [15]. After FFT, the bimodal amplitude spectrum (similar to that shown by our study in Fig. 2) was characterized by the full-width half maximum (FWHM) for the first and maximum peak in the spectrum, similar to the low-frequency power feature in our manuscript. Furthermore, the frequency of the maximum second peak times the amplitude divided by the peak amplitude and the frequency cut-off that divides the FFT area in two equal halves were computed, similar to our frequency cut-off. Our model showed a comparable accuracy with the inclusion of only two features, making the interpretation and automatic implementation easier.