ECG classification of canine arrhythmias using deep neural networks

University of Technology Sydney. Faculty of Engineering and Information Technology.With an ever-increasing canine population comes a rising number of cardiovascular diseases. Of the multitude of issues surrounding congenital heart defects veterinarians have identified that canine arrythmia classification is often inaccurate. Thus, current arrhythmia classification methods are unreliable when compared to human arrhythmia classification techniques. This study investigates the capabilities of current classification standards, their effectiveness on canine sinus rhythms and develops classification models tailored for canines. Veterinarians are obtaining unreliable results from current animal ECG classification software. It is hypothesised that classification software has been developed using human sinus rhythms, which does not accurately identify canine arrythmias. This research determines whether training a neural network using canine sinus rhythms will provide a higher degree of accuracy than models which were trained using human sinus rhythms. To test this hypothesis, two sets of neural networks were developed for the purpose of identifying canine arrythmias. The first set was trained using human sinus rhythms, obtained from the PhysioNet MIT-BIH Arrhythmia database to replicate the current academic models used for arrythmia classification. The second set was trained using canine sinus rhythms obtained from the Small Animal Specialist Hospital, in Sydney Australia. Relationally, it tests whether arrythmia classification can be improved by using a neural network tailored for canines. All datasets followed the AAMI ECG standards, and the neural networks used in this research included CNN, LSTM, and MLP architectures. The findings of this research display that all models trained using human sinus rhythms provided an accuracy score of less than 50% when tested on canine sinus rhythms. Similar results were achieved when testing human sinus rhythms on models that were trained using canine sinus rhythms. Alternatively, once each neural network was trained and tested using data from a single species, human or canine, the models achieved an accuracy score of over 90%. In conclusion this project showcases a need for classification models to be trained on patient specific data and provides the groundwork for developing methods to ensure accurate arrhythmia classification