The extracellular contact electrogram (EGM) is used in the cardiac catheter laboratory to record the electrical activity and architecture of the local myocardium. Whilst changes in the electro-architecture can be reflected in the EGM morphology, interpretation of the EGM is currently limited. The hypothesis of this thesis was that features of the EGM using supervised machine learning could be used to predict the functional abnormalities in intact hearts.
Addressing the hypothesis involved the use of intact porcine and human hearts to obtain EGM data. The hearts (n=34) were perfused and restarted ex vivo using the Langendorff apparatus, and pharmacological modulators were administered to explore the effects of functional abnormalities, an example of which would be changes in ion channel activity and gap junction uncoupling. The computational methods used to analyse the data included first signal processing the EGMs, then development of an automated feature extraction process. The morphology of these EGMs at baseline and following modulation were investigated, then characterised. The dataset was then used to develop and train supervised machine learning prediction models.
The results of this thesis found that the functional determinants of electrogram morphology can be accurately predicted using supervised machine learning. When classifying Baseline (BL) against all modulated (Mod) EGMs grouped together, BL EGMs were correctly predicted 90% of the time, and Mod EGMs were correctly predicted 92% of the time. When classifying Mod EGMs separately, BL EGMs were classified with a 96% accuracy, 76% accuracy for Carbenoxolone (CBX), 91% accuracy for Lidocaine (Lido), and 78% accuracy for Pinacidil (Pin). The modulators used to induce abnormalities are responsible for various cardiac arrhythmias, so the effective prediction of these modulations indicate that leveraging machine learning could contribute to improvement in diagnostics and treatment of these conditions.