Background: Freedom from AF following ablation remains at around 60% even with the addition of electrogram guided procedures alongside pulmonary vein isolation. Gap junctional (GJ) connexin proteins have been established as being deterministic in the development of atrial fibrillation (AF), therefore identifying areas of GJ abnormality (cellular uncoupling) may provide a target to guide ablation strategy. The contact electrogram (EGM) contains electrophysiological information beyond what is currently interpreted clinically. More effective utilisation of subtle changes in EGM morphology caused by cellular uncoupling could lead to improved treatments. The Langendorff system enables EGMs from intact human and porcine hearts to be recorded ex-vivo, where cellular uncoupling can be pharmacologically induced in a controlled manner. We aimed to train a machine learning model that would be capable of identifying the presence of GJ uncoupling in the myocardium from the EGM morphology.
Methods: Unipolar EGMs were recorded on the left atrial and left ventricular epicardium of ex vivo human (n=3) and porcine (n=8) hearts using a Langendorff system and high-density grid catheter (Abbot Medical). All hearts were paced at cycle lengths (CL) between 300-1500ms and administered 1mM carbenoxolone (CBX) via bolus to induce cellular uncoupling. EGM recordings were sequentially mapped at the same sites before and after CBX administration. Nineteen morphological features were extracted from each EGM using automated algorithms.
A random forest machine learning algorithm was trained on 80% of the EGMs (n=378132), selected at random by site and CL. Prediction accuracy of classifying between baseline (BL) and CBX recordings was assessed using the remaining unseen 20% of the EGMs (n=94533)
Results: The average prediction accuracy on the validation dataset was found to be 92%, with BL accuracy at 94% and CBX at 88%. Precision was 92% BL, 95% CBX and recall was 96% BL, 81% CBX. Of the features chosen by the machine learning algorithm, the group percentage changes from BL to CBX were RS interval -60.2%, R point Amplitude -5.9%, QS Interval -56.5%, EGM Duration -52.6%, Amplitude +10.0%, QR Interval -49.8%, Q Point Amplitude -23.7%, S Point Amplitude -24.1%, RS Gradient -120%, QR Gradient -5.9%, S-Endpoint Gradient +33.8%, Fractionation Index +60.5%, RS Ratio +10.5%, stimulus to (-dV/dt)max Interval +20.3%.
Conclusion: Machine learning can be used to accurately and automatically detect reduced cellular coupling from the EGM morphology of intact ex vivo hearts. The methodology used enables interpretation of the EGMs beyond the current clinical binary classification of simple/complex or early/late. Using the Langendorff to pharmacologically induce other channelopathies in the ex vivo hearts would enable a machine learning model to be trained, which could predict a more diverse array of abnormalities from the EGMs, that, if translated to the clinic, could be of further benefit to guide ablation procedures.