Intracardiac bipolar electrograms (EGM) comprise nearfield (NF) and farfield (FF) signals. For the treatment of atrial fibrillation, the discrimination between left atrial (LA) FF and pulmonary vein (PV) NF is of fundamental importance to confirm pulmonary vein isolation (PVI). This study aims at developing an algorithm to discriminate, from a single heartbeat, PV-NF from LA-FF signals within the PVs during PVI. This algorithm may provide a real-time confirmation of complete PVI (absence of PV-NF).
We retrospectively analysed EGM using the cryoballoon technology and a decapolar circular catheter (Achieve, Medtronic) of 42 patients. Signals were manually classified based on the disappearance of the PV-NF signal during PVI as local PV-NF or LA-FF. Features were calculated for the two groups using a 60-ms window. The powers in different frequency bands were calculated using FFT for the lead with the highest power in the high frequency band 150-350 Hz suggesting the closest NF signal. Supervised machine learning models were trained with 4-fold cross-validation. Overall predictive accuracy, and NF false discovery rate (leading to unsafe excessive ablation), were used to select the best classifier.
We analysed 261 signals (73 from the left-inferior PV, 104 from left-superior, 55 from right-superior and 29 from the right-inferior). The high-est predictive accuracy of 79% was obtained with a SVM (quadratic) model including the signals from all veins. The two best features were the power in the high frequency band and the maximal bipolar voltage of the signal. Including more features like wavelet decomposition did not improve the accuracy.
The classifier accuracy depends on the location of the vein. Classification is more accurate for the inferior veins (positive prediction value of 86%) and lower for the left superior vein (75%) due to its close proximity of the LA appendage, with NF false discovery rate of 5% and 27% respectively.