Controllable ECG Synthesis Through Neural Network Input Modelling

• Autor(a) principal: Herrera, Ana Laura Batista
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10362/162739
• topic: ECG synthesis; controllable synthesis; GRU; data augmentation; arrhythmia synthesis; Domínio/Área Científica::Engenharia e Tecnologia::Outras Engenharias e Tecnologias

Artificial Intelligence (AI) has brought promising innovations to healthcare, particularly in patient diagnosis and monitoring. However, challenges such as data scarcity, quality, and privacy concerns often hinder its effective implementation. Electrocardiogram (ECG) signals are crucial for cardiac disease diagnosis, but collecting these signals can be challenging due to patient privacy concerns and the rarity of some cardiac conditions. This thesis presents a novel method for synthesising ECG signals in a controllable manner, with the aim of augmenting existing datasets and, by extension, promoting the development of more ethically responsible AI algorithms in healthcare. The methods used in this thesis include the usage of a non-linear modified sawtooth signal as an input to train a Gated Recurrent Unit (GRU) generator. Then, by modifying this input signal, with alterations that result in parameterised changes to the synthetic signal, it was possible to generate different ECG traces, such as normal sinus rhythms, tachycardia, bradycardia, and first and second-degree atrioventricular (AV) blocks. The model successfully retained the specific ECG characteristics of individual subjects across these rhythms and arrhythmias. In addition, a Random Forest (RF) model was trained and then tested to classify the synthetic arrhythmias. Once this was successfully achieved, the synthetic arrhythmias were used to augment the previous training dataset. The results showed improved performance on the same test set, highlighting the utility and effectiveness of the synthetic signals. This thesis contributes to the field of AI in healthcare by developing new methods to generate data for underrepresented groups and providing evaluation metrics to assess the quality of synthetic data. Ultimately, this work serves as a foundation for the equitable use of synthetic signals in healthcare AI, promoting unbiased and ethically responsible applications.