A Machine Learning Approach to Classify Myocardial Infarction Using Vectorcardiographic Ventricular Depolarization and Repolarization

Filip Karisik1 and Martin Baumert2
1University of Adelaide, 2The University of Adelaide


Aim: QT interval beat-to-beat variability has indicated diagnostic/prognostic abilities in myocardial infarction. Furthermore, research has suggested that the vectorcardiogram has superior diagnostic abilities compared to the standard electrocardiogram in myocardial infarction. This study aimed to assess the ability of ventricular depolarization and repolarization to classify myocardial infarction patients versus control subjects. Method: vectorcardiogram data was obtained from the publicly available PTB Database. 147 recordings (78 MI vs. 69 Control) were utilized. For each recording, 60 QRS-complex and T-wave vectorcardiogram beats were extracted using the 2DSW algorithm. Templates for the QRS-loop and T-loop were respectively constructed using the average of the 60 beats. An inhomogeneous three-dimensional template adaptation scheme was applied on each QRS-loop and T-loop to capture subtle beat-to-beat morphological changes. The results of the adapted templates were then fed into a long short-term memory network as raw three-dimensional data. A regularized three-layer network was implemented. The network was trained to produce a binary output, classifying 0 as control subjects and 1 as myocardial infarction patients. Results: The classifier produced test set classification results with an overall 89.1% accuracy, 89.1% sensitivity and 90.0% specificity. The validation and performance of the system were analysed on 10% and 20% of the extracted dataset respectively. Conclusion: In conclusion, high classification accuracy has been achieved on a relatively small subset of the PTB database. Future work will look to improve the classification results by increasing the number of beats fed into the neural network and extending the analysis across all recordings in the PTB database.