Pediatric/Adult Congenital Heart Disease -> Pediatric Cardiology D-PO04 - Poster Session IV (ID 15) Poster

D-PO04-035 - Using Machine Learning To Localize The Accessory Pathway In Patients With Wpw (ID 1184)

Disclosure
 B.J. Lipman: Nothing relevant to disclose.

Abstract

Background: The application of conventional algorithms to identify accessory pathway (AP) locations in pediatric patients with Wolff-Parkinson-White syndrome (WPW) is limited. Machine learning (ML) technology may offer superior results.
Objective: To assess ML algorithms sensitivity and positive predictive value (PPV) in localizing the site of the AP in pediatric patients with WPW.
Methods: ECGs of 245 unique WPW patients (mean age 12 ± 4years) were evaluated, with AP location (left, right, septal)identified at electrophysiological study. A random forest model was trained on 320 automated ECG measurements followed by a Convolutional Neural Network (CNN) comprising four convolutional and two fully connected layers trained on numerical data of calibrated and filtered one dimensional (1D) ECG images. Training with differing lead combination and recording intervals were tested. Finally, the CNN was trained upon two-dimensional (2D) ECG images. Accuracy was compared for each model.
Results: Random forest precision (Positive predictive value) and recall (sensitivity) was calculated for each location, with a random forest precision (PPV) ranging between 0.6-1.00, and recall (sensitivity) ranging between 0.43-1.00. (Table). Right sided pathways had the highest positive predictive value while septal pathways the highest sensitivity. For 1D CNNs, the optimal lead combination of 3 orthogonal leads and one beat had an accuracy of 0.76, with no improvement with more leads or addition of 2D CNN.
Conclusion: ML offers a high sensitivity for detection of septal pathways and may be useful clinically. Development of more sophisticated ML techniques may further improve AP localization.
Collapse