Towards broader application of deep learning methods to the automated analysis of electrocardiograms

  • Rob Brisk

Student thesis: Doctoral Thesis


Introduction: Automated analysis of the electrocardiogram (ECG) is one of the most impactful forms of computerised clinical decision support. Recent advances in deep learning (DL) techniques have shown the potential to overcome historical limitations of manually engineered ECG analysis. However, DL-based ECG analysis is a nascent technology and questions remain about its limitations and usability.

Methods: To investigate the limits of DL’s capacity for detecting acute myocardial infarction (AMI), a DL classifier was trained to detect early onset of coronary artery occlusion. To investigate DL’s ability to broaden access to ECG analysis through high-quality interpretation of ECG images, a DL classifier was trained to detect atrial fibrillation (AF) from images of ambulatory ECG recordings. To evaluate a novel approach to reducing the volume of training data required to train DL-based ECG analysers, a system was developed to simulate ECG signals and corresponding wave segmentation masks. DL models were pretrained using this synthetic data, fine-tuned to detect AMI and AF from real ECGs, then compared again non-pretrained models.

Results: A DL model was unable to detect hyperacute coronary occlusion better than a random chance classifier. Performance appeared better in an earlier iteration of the experiment, but this appeared to be due to data leakage. A DL model detected AF from ECG images with equivalent accuracy to raw ECG samples. Pretraining with synthetic ECG data reduced the need for training on real ECGs to achieve comparable accuracy and provided a potential mechanism for clinician confidence calibration.

Conclusion: DL requires large volumes of training data and suffers from a “black box” effect. DL can broaden access to automated ECG analysis through high quality interpretation of ECG image data. Wave segmentation pretraining reduces the need for training data and provides a potential mechanism for confidence calibration. This may ameliorate the black box phenomenon.
Date of AwardFeb 2023
Original languageEnglish
SponsorsEastern Corridor Medical Engineering Centre
SupervisorRaymond Bond (Supervisor), Jim McLaughlin (Supervisor), Dewar Finlay (Supervisor) & David J. McEneaney (Supervisor)


  • Artificial intelligence
  • ECG

Cite this