Automated quantification and identification of decremental response after extrastimulus: validation of a novel software

Abstract Funding Acknowledgements Type of funding sources: Public grant(s) – National budget only. Main funding source(s): Junta de Andalucia Background Automatic identification of myocardial electrograms (EGMs) with decremental conduction (DC) is a promising field in catheter ablation procedures, allowing the identification of the critical isthmus of tachycardias with re-entry mechanisms. In this work, a novel artificial intelligence (AI)-based software was developed for the automatic analysis of delays between evoked responses in the coronary sinus (CS), allowing signal quantification. Methods Atrioventricular (AV) and ventriculoatrial (VA) conduction was analysed in coronary sinus signals obtained from patients undergoing electrophysiological study. A simple pacing (S=400-600ms) and extra-stimulus protocol (S2=ERP + 20-60ms) were used in all cases. An AI-based algorithm was developed for EGM delineation with the objective of localising the onsets and offsets of each independent local and far field activations. The algorithm consisted in several steps. Firstly, an AI-based QRS detector was employed (Figure 1A). Secondly, the located QRS’ were used alongside another AI model to delineate the EGMs corresponding to all bipolar electrodes in the CS (Figure 1B). Thirdly, majority voting is performed on the delineations to obtain an overall prediction for all CS electrodes (Figure 1C). Fourthly, the predictions are cleaned and a rule-based algorithm is employed to automatically exclude faulty registries (e.g. loss of capture or AV/VA block after S2). Finally, the determination of decremental response is performed by computing the time difference between the last paced beat and the S2 extrastimulus. This was measured using the onsets of morphologically similar activations in the trace, correcting for misalignments. A decremental response was defined as a difference >10 ms between the AV/VA interval during the S1 drive and the same interval immediately after the S2. Four experts validated results of the automated analysis. Results 321 tracings from 50 patients were analysed. 81 tracings were automatically excluded due to loss of capture or AV/VA block after S2. The AV/VA interval was automatically annotated in 236 EGMs. 192 tracings exhibited decremental conduction properties, whereas 48 tracings showed non-decremental AV or VA conduction due to accessory pathways. Automatic analysis accurately classified 191 tracings as decremental (93.75% sensitivity, 91.47% specificity) and 45 tracings as non-decremental (85.42% sensitivity, 98.53% specificity). Conclusions AI-based algorithms are able to produce objective quantifications over real-world data, discriminating between local and far field signals and accurately identifying EGMs with decremental evoked response. Automated data analysis can accelerate research and reduce clinical workload. Moreover, a quantification-based approach can be employed in many other applications. Validation of this software on ventricular signals could facilitate ventricular tachycardia substrate ablation procedures.