An InceptionTime-Inspired Convolutional Neural Network to Detect Cardiac Abnormalities in Reduced-Lead ECG Data

Harry Crocker and Aaron Costall
University of Bath


Abstract

Cardiovascular disease is the leading cause of death worldwide. The twelve-lead electrocardiogram (ECG) is a common tool for diagnosing cardiac abnormalities, but its interpretation requires a trained cardiologist. Thus there is growing interest in automated ECG diagnosis, especially using fewer leads. Hence the PhysioNet-CinC Challenge 2021: Will two (leads) do? The University of Bath team (UoB_HBC) developed InceptionTime-inspired deep convolutional neural networks, using parallel 1D convolutions of varying length. Twelve-, six-, three-, and two-lead models achieved unofficial cross-validation Challenge metric scores of 0.542, 0.470, 0.521 and 0.452, and 0.50, 0.40, 0.49 and 0.42 on the hidden validation set. The small difference between cross-validation and validation scores, and consistency in reduced-lead model rankings, suggests the approach generalizes well. At the official phase, however, Challenge metric scores dropped to 0.402, 0.385, 0.401, and 0.391, respectively. Though the twelve-lead model performs best, three-lead performance was lower by just 2–3 % at the unofficial phase and by only 0.25 % on the official validation set, suggesting potential for reliable reduced-lead diagnoses. Furthermore, the three-lead model performed consistently better than the six-lead, highlighting the importance of selection of type of lead, not just their number.