Background: Ventricular tachycardia (VT) is a major cause of sudden cardiac deaths, treated with catheter ablation that targets and destroys the tissue responsible for initiating or sustaining VT. Pace-mapping is a key practice for identifying ablation targets by matching stimulated sites' ECG to the observed clinical VT. An AI-assisted pace-mapping approach can help clinicians identify the most likely target sites, reducing procedure time and increasing the success rate.
Objective: We have previously shown that active learning (AL) can progressively guide clinicians to optimal pace-mapping sites by refining the prediction of VT targets with collected ECG data. However, this approach required the AL process to start from scratch for each VT target, even within the same patient. To allow knowledge transfer across pace-mapping procedures, both across or within patients, will further improve the efficiency of pace-mapping. This study aims to achieve this with a novel AL framework combining neural surrogates with continual meta-learning.
Method: We train a neural surrogate fNN to map a pacing site r to its ECG signal 𝜙(r) = fNN(r), using predictions and their uncertainty to identify the next best pace-mapping site. Rather than training fNN from scratch each time, we use meta-learning to learn a fNN that can be quickly fine-tuned to fNN' using AL on current task while maintaining knowledge from past pace-mapping processes
Results: We evaluated our method on two heart geometries, each with four infarcted and one healthy condition, totalling 3701 and 3357 labeled pacing sites - simulated ECG pairs. Results showed an average reduction of 85.26% and 51.79% pace-mapping sites needed in comparison to pace-mapping without AL and with standard AL, respectively.
Conclusion: We propose a meta-learning active learning framework for reducing the number of pace-mapping sites by enabling knowledge transfer across tasks.