Explanation-visualised deep learning model for accessory pathway localisation using 12-lead electrocardiography

Abstract

Funding Acknowledgements Type of funding sources: Public Institution(s). Main funding source(s): British Heart Foundation Imperial Centre of Research Excellence Background/Introduction ECG algorithms for identifying accessory pathway (AP) locations are inaccurate and difficult to use. Human expert interpretation is poorly reproducible. Artificial intelligence (AI) techniques such as machine learning can improve accuracy in classification tasks by eschewing theory-driven predictions. More reproducible and accurate AP localisation could shorten procedure time and personalise ablation consent. Purpose We developed a neural network to perform AP localisation using 12-lead ECGs. Its decision-making process was analysed to enable explainability of the neural network. Methods A convolutional neural network was trained on raw, digital, intra-procedural 12-lead ECGs of patients with manifest APs who underwent successful ablation. ECGs were labelled with AP locations as left-sided, septal or right-sided using procedure reports, fluoroscopy and electro-anatomical maps. Accuracy of the neural network was assessed via 4-fold cross-validation and was compared to the Arruda algorithm. Five cardiologists were also assessed for their accuracy in determining locations in sub-groups of cases. The neural network was retrospectively analysed to identify areas of ECGs most influential to its predictions using importance mapping. Results In 156 cases, accuracy of the neural network (92.9%) was significantly higher than the Arruda algorithm (76.9%; p < 0.0001) and all five cardiologists (37.5% to 65.9%; p = 0.0001 to 0.0290). Importance mapping demonstrated that the QRS complexes of leads aVL and V1 were perceived as most influential, indicating interrogation of the lateral and anterior-posterior axes respectively. The figure shows (A) architecture of the neural network, (B) accuracy of the neural network, Arruda algorithm and five cardiologists, (*, p = 0.05 – 0.01; **, p = 0.01 – 0.001; ***, p = 0.001 - 0.0001; ****, p < 0.0001; as compared to the neural network) and (C) example importance maps for 12-lead ECGs of left-sided, septal and right-sided APs (in order from left to right), with darker regions corresponding to greater relative importance. Conclusion AI ECG interpretation allows accurate, reproducible prediction of AP locations, superior to conventional algorithms and human interpretation. Although AI decision-making is thought of as a ‘black box’, explanation visualisation techniques such as importance mapping allow humans to understand aspects of how a neural network make decisions. A prospectively validated neural network could be integrated into clinical practice to improve pre-procedural AP localisation. Abstract Figure. Summary of results