Abstract:

Cardiovascular diseases account for 17 million deaths per year worldwide. Of these, 25% are categorized as sudden cardiac death, which can be related to ventricular tachycardia (VT). This type of arrhythmia can be caused by focal activation sources outside the sinus node. Catheter ablation of these foci is a curative treatment in order to inactivate the abnormal triggering activity. However, the localization procedure is usually time-consuming and requires an invasive procedure in the catheter lab. To facilitate and expedite the treatment, we present two novel localization support techniques based on convolutional neural networks (CNNs) that address these clinical needs. In contrast to existing methods, our approaches were designed to be independent of the patient-specific geometry and directly applicable to surface ECG signals, while also delivering a binary transmural position. Moreover, one of the method's outputs can be interpreted as several ranked solutions. The CNNs were trained on a dataset containing only simulated data and evaluated both on simulated test data and clinical data. On a novel large and open simulated dataset, the median test error was below 3 mm. The median localization error on the unseen clinical data ranged from 32 mm to 41 mm without optimizing the pre-processing and CNN to the clinical data. Interpreting the output of one of the approaches as ranked solutions, the best median error of the top-3 solutions decreased to 20 mm on the clinical data. The transmural position was correctly detected in up to 82% of all clinical cases. These results demonstrate a proof of principle to utilize CNNs to localize the activation source without the intrinsic need for patient-specific geometrical information. Furthermore, providing multiple solutions can assist physicians in identifying the true activation source amongst more than one possible location. With further optimization to clinical data, these methods have high potential to accelerate clinical interventions, replace certain steps within these procedures and consequently reduce procedural risk and improve VT patient outcomes.

Abstract:

Aims Atrial cardiomyopathy (ACM) is associated with new-onset atrial fibrillation, arrhythmia recurrence after pulmonary vein isolation (PVI) and increased risk for stroke. At present, diagnosis of ACM is feasible by endocardial contact mapping of left atrial (LA) low-voltage substrate (LVS) or late gadolinium-enhanced magnetic resonance imaging, but their complexity limits a widespread use. The aim of this study was to assess non-invasive body surface electrocardiographic imaging (ECGI) as a novel clinical tool for diagnosis of ACM compared with endocardial mapping. Methods and results Thirty-nine consecutive patients (66 ± 9 years, 85% male) presenting for their first PVI for persistent atrial fibrillation underwent ECGI in sinus rhythm using a 252-electrode-array mapping system. Subsequently, high-density LA voltage and biatrial activation maps (mean 2090 ± 488 sites) were acquired in sinus rhythm prior to PVI. Freedom from arrhythmia recurrence was assessed within 12 months follow-up. Increased duration of total atrial conduction time (TACT) in ECGI was associated with both increased atrial activation time and extent of LA-LVS in endocardial contact mapping (r = 0.77 and r = 0.66, P < 0.0001 respectively). Atrial cardiomyopathy was found in 23 (59%) patients. A TACT value of 148 ms identified ACM with 91.3% sensitivity and 93.7% specificity. Arrhythmia recurrence occurred in 15 (38%) patients during a follow-up of 389 ± 55 days. Freedom from arrhythmia was significantly higher in patients with a TACT <148 ms compared with patients with a TACT ≥148 ms (82.4% vs. 45.5%, P = 0.019). Conclusion Analysis of TACT in non-invasive ECGI allows diagnosis of patients with ACM, which is associated with a significantly increased risk for arrhythmia recurrence following PVI.

Abstract:

Ventricular ectopic beats are caused by foci that spontaneously depolarize in the ventricles and may induce life-threatening tachy-arrhythmia. The common clinical therapy is to localize the focus and then apply catheter ablation to it. Currently used localization methods are either invasive or require information about the patient-specific heart geometries that are costly to generate. Therefore, in order to localize the ectopic foci non-invasively and without the requirement of patient-specific geometries, a deep learning approach is proposed in this work. This approach is trained on a simulated dataset containing 1.8 million body surface potential maps (BSPMs) of ventricular ectopic beats. BSPMs are electrocardiogram (ECG) signals that are recorded by a matrix of electrodes on the patient’s torso. The proposed localization approach is threefold. After the prepossessing of the BSPMs, a convolutional neural network (CNN) is trained to estimate the begin and end of the activation window. Based on these predictions, the BSPMs are clipped and resampled to a fixed length. They are then passed into a localization neural network (NN) which is trained to predict the coordinates of the ectopic foci. For the localization approach, three different NNs are compared that mainly differ by their task formulation. One is formulated as a multi-task problem combining a regression with two binary classifications, the second as a fuzzy classification over barycentric coordinates and the third combines two binary classifications with a fuzzy classification over combined barycentric coordinates. The results on the simulation dataset are precise: The NN to predict the begin and end of the activation window achieves results with a mean test error smaller than 2 ms. The results of that network on the clinical dataset are plausible, but the ground truth is not known. All three localization networks achieve a mean geodesic validation error smaller than 2 mm. Without further ado, the results delivered by the localization NNs on a clinical dataset are not sufficiently accurate, but approaches to solve this in future work are proposed. Since the application of (deep learning) algorithms in medicine poses many safety require- ments, an approach for the explanation of the implemented NNs was applied: Saliency maps, which quantify the influence of the input features with respect to a specified NN output, are applied. The results conform to human intuition: the information of electrodes on the front of the torso is more influential to the NNs’ predictions than the information of electrodes on the back of the torso. Additionally, the information contained in the first time samples of the BSPMs affect the localization predictions more than of the later time samples. Therefore, this work should provide a suitable starting point for further investigations of the explainability of the proposed localization approach. To summarize, the proposed deep learning approach delivered precise results for the local- ization of ventricular ectopic foci on a large simulated dataset. The implemented NNs were only trained on simulated data and their application to a clinical dataset revealed that the current NNs have to be further optimized for clinical data. However, the algorithm does not require patient-specific heart geometries which is a novelty in this field and overcomes shortcomings of existing methods. For the first time, an explainability method was applied to an NN-based localization approach. Findings complied to human intuition: First time samples and electrodes on the front of the torso are most influential for the NNs’ prediction. The combination of NNs and explainability methods enables the development for an exact and non-invasive localization procedure that accelerates the treatment of the patient and improves the outcome.