Sensitivity and Generalization of a Neural Network for Estimating Left Atrial Fibrotic Volume Fractions from the 12-lead ECG

Features extracted from P waves of the 12-lead electrocardiogram (ECG) have proven valuable for noninvasively estimating the left atrial fibrotic volume fraction associated with the arrhythmogenesis of atrial fibrillation. However, feature extraction in the clinical context is prone to errors and oftentimes yields unreliable results in the presence of noise. This leads to inaccurate input values provided to machine learning algorithms tailored at estimating the amount of atrial fibrosis with clinical ECGs. Another important aspect for clinical translation is the network’s generalization ability regarding new ECGs. To quantify a network’s sensitivity to inaccurately extracted P wave features, we added Gaussian noise to the features extracted from 540,000 simulated ECGs consisting of P wave duration, dispersion, terminal force in lead V1, peak-to-peak amplitudes, and additionally thoracic and atrial volumes. For assessing generalization, we evaluated the network performance for train-validation-test splits divided such that ECGs simulated with the same atria or torso geometry only belonged to either the training and validation or the test set. The root mean squared error (RMSE) of the network increased the most in case of noisy torso volumes and P wave durations. Large generalization errors with a RMSE difference between training and test set of more than 2% fibrotic volume fraction only occurred if very high or low atria and torso volumes were left out during training. Our results suggest that P wave duration and thoracic volume are features that have to be measured accurately if employed for estimating atrial fibrosis with a neural network. Furthermore, our method is capable of generalizing well to ECGs simulated with anatomical models excluded during training and thus meets an important requirement for clinical translation.