Ledezma Rondon, Carlos Alberto; (2020) Populations of models and machine learning for the assessment of cardiac electrophysiology. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Main.pdf - Accepted Version Download (43MB) | Preview

Abstract

The purpose of this thesis was to develop a novel methodology that would enable the multi-scale analysis of cardiac electrophysiology signals and the development of methods for detecting cardiac disease. This goal was achieved through new techniques combining the use of populations of mathematical models and machine learning. This resulted in two novel methodologies. First, an approach that underpins variations in cellular ion-channels that lead to pathological observations in organ-level acquisitions, using populations of models, was developed. It was tested in an ex-vivo beating heart experiment and the results obtained by applying the method developed in this work were consistent with other experimental data acquired simultaneously. The presented algorithm, in contrast to the acquisitons that are currently made in the experiments, is applicable in real time and gives timely and predictive information. Furthermore, the method is modular, allowing to change its components according to the available models, computational capabilities and experimental acquisitions. Second, the previously experimentally validated method was used to create a database, which included the effects of inter-subject variability, to enable the study of ischaemia using machine learning techniques. Together, the populations contained over 20000 signals, which were either healthy or ischaemic and included the spatial and temporal evolution of ischaemic events. Neural networks were then trained to classify the different signals contained in the database and the relevance of their input features was analysed to elucidate the effect of inter-subject variability in the detection of ischaemic events. The results showed that variability precludes the detection of ischaemia by simple visual inspection, that non-linear multi-feature models are required for correct classification and that neural networks are appropriate to assess ECG signals when looking for ischaemic traits. The use of computational models to bridge scales in cardiac electrophysiology and the use of virtual signals to elucidate the effects of inter-subject variability in the detection of ischaemia have been, hitherto, unexplored.

Download activity - last month

Export as JSONXMLCSV

Download activity - last 12 months

Export as JSONCSVXML

Downloads by country - last 12 months

Export as JSONXMLCSV

Archive Staff Only

View Item