• Kaj Printz Madsen
  • Morten Brodersen Jensen
  • Carsten Vestergaard Risager
4. term, Software, Master (Master Programme)
We propose an automatic and interpretive heartbeat classification approach for identifying heart arrhythmias based on learned shapelets from annotated electrocardiograms (ECG).

The heartbeat classification approach consists of ECG signal preprocessing, shapelet extraction, shapelet transformation, and classification.
The preprocessing step involves removal of baseline wander, noise filtering and dimension reduction of the multi-lead ECG signals. The binary shapelet transform is used to extract discriminative subsequences from the ECG, which are used to transform the ECG data into feature vectors. Finally, we train a heterogeneous ensemble of classifiers on the shapelet transformed data set.

We evaluate the performance of the approach on two data sets. The MIT-BIH data set for comparison with previous work within ECG classification following the inter-patient scheme and the AAMI recommendations. As well as AAU-ECG, a real-world multi-labeled ECG data set consisting of 413,151 ECG records where the performance is tested against the industry-leading knowledge-based Marquette 12SL ECG analysis program (Marquette).

The MIT-BIH experiments show that shapelets improves the recall metric of normal and ventricular ectopic heartbeats as well as the precision of fusion beats. In addition, our approach achieves the highest global performance for the four classes with an overall accuracy of 94.3%. For the AAU-ECG data set, the knowledge-based Marquette, in general, surpasses the performance of the learn-based shapelet approach. However, our approach has good discrimination power for right and left bundle branch block, associated with significant cardiovascular mortality, and outperforms Marquette on four diagnoses related to left ventricular hypertrophy.
Publication date8 Jun 2018
Number of pages94
ID: 280595476