Abstract
This project looked to investigate and report the potential for detecting paroxysmal atrial
fibrillation using sinus rhythm electrocardiograms taken after an episode. Using complexity
analysis, the symmetric projector attractor reconstruction method, and restitution analysis
in conjunction with machine learning models, an understanding of the performance of such
models was obtained. Optimisation of these methods and models was performed to
maximise the performance of the models along with a review of the associated sensitivities
and specificities.
By employing a feature ranking method a formalised and reproducible method of combining
these analysis techniques was used, allowing for an improvement in performance along with
an understanding of the relationship between each analysis method and their associated
features. Finally, the optimised methods and models were applied to a new database
collected from Liverpool Heart and Chest Hospital, allowing for an understanding of the
transferability of the methods and models used. The work was reviewed in its entirety
evaluating the methods and results gained; future work to expand this area of research has
been suggested based on the work performed in this project.