Back to the roots: Using the old inverse Dower transformation as a dimensionality reduction tool for a 12-lead ECG arrhythmia classifier

Fabian Rabe1 and Fabian Stieler2
1University of Augsburg, Germany, 2University of Augsburg


A 12-lead ECG is an important diagnostic tool for detecting arrhythmias in a human heart. In the past, Deep Neural Networks have shown potential in assisting physicians diagnosing these ECGs. A recent study observed that a Deep Learning model trained on a single augmented ECG lead had a comparable performance to a model trained on all 12 leads.

Overall the 12 leads are a highly correlated representation of the difference in electrical potential originating in the 3D physical space of the human heart. The inverse Dower transformation is a technique to transform the 12 dimensional ECG into a 3D VectorCardioGram (VCG). Thus the VCG is a much conciser representation of the information contained in a 12-lead ECG.

Due to correlated features being an undesired trait in input data for machine learning, different techniques for feature reduction have been developed. However the human's ability to interpret these reduced features is lost in most of the transformations.

In the context of the PhysioNet/CinC Challenge 2020, we examine the utility of the inverse Dower transformation as a tool for reducing the dimensions of the input data for a machine learning model, whilst retaining a signal interpretable by a human. Our classifier, which did not yet employ the inverse Dower transformation, achieved a F2-Score of 0.134 and a G2-Score of 0.048 on the hidden test set. In a 5-fold cross validation on the training data we achieved a F2-Score of 0.653 and a G2-Score of 0.465.