An Ensemble Learning Approach to Detect Cardiac Abnormalities in ECG Data Irrespective of Lead Availability

Tim Uhlemann1, Sebastian Wegener2, Joshua Prim1, Nils Gumpfer1, Dimitri Grün2, Jennifer Hannig1, Till Keller2, Michael Guckert1
1Cognitive Information Systems, Kompetenzzentrum für Informationstechnologie, Technische Hochschule Mittelhessen - University of Applied Sciences, 2Department of Internal Medicine I, Cardiology, Justus-Liebig-University Gießen


Aim Recent research has shown that artificial intelligence can detect heart diseases when applied to electrocardiogram data. As of now, underlying architectures have mostly been built for specific problems with restricted generalisation to other patterns, e.g. convolutional neural network (CNN)-based models able to detect local patterns do not capture abnormalities with rhythmic dependencies well. Additionally, standard deep learning approaches cannot incorporate knowledge in non-deep learning representations. Aim of this project is to overcome limitations of distinct architectures by using a hybrid ensemble of models, also incorporating expert knowledge.

Methods The core of our concept is an orchestration of architectures in which submodels working with sufficient precision for a subset of diseases are combined into a comprising computer aided diagnostic engine with comprehensible precision for a larger set of pathologies. Submodels working on subsets of available leads are combined into an ensemble which computes aggregated predictions, e.g. by voting, averaging or stacking. This approach allows for hybrid architectures in which rule-based inference engines can be incorporated.

Results In the unofficial phase of PhysioNet/CinC Challenge, our ensemble of a single CNN-based model achieved a challenge metric score of 0.43, -0.29, 0.37 and 0.33 for the 12, 6, 3 and 2 lead inputs, respectively (Team CardioIQ). We assume that the loss in performance for the 6-lead model is architecture-specific and due to missing precordial leads. Cross-validation on a local train/test-split supported the promising results achieved on the challenge data set, e.g. an AUROC score higher than 0.95 for diseases like complete right bundle branch block.

Conclusions Results support our hypothesis that single model architectures perform well on specific diseases. However, restricted generalisation power on elementary different diseases makes a combination of specifically trained expert models in an ensemble with an appropriate aggregation a promising approach to outperform singular models.