Universidade Federal do Ceará
Mestrado e Doutorado em Ciências da Computação

Título: An Attention-based Machine Learning Solution for Predicting Low Left Ventricular Ejection Fraction in Chagas Disease Patients

Data: 18/07/2024
Horário: 16h
Local: Sala de Seminários – Bloco 952

Resumo:

Chagas Disease (CD) is an endemic infection with a generally chronic evolution, discovered at the beginning of the 20th century. CD is an independent risk factor for ischemic stroke (IS), representing a significant thromboembolic manifestation of the disease, with left ventricular ejection fraction (LVEF) being one of its indicators. Although not the only factor, a low ejection fraction, such as 40%, associated with heart failure, is related to a high mortality rate, functioning as a marker for identifying and treating more sensitive patients. As a way to identify a low LVEF, the electrocardiogram is a more economical and accessible technique, which consumes less time than other exams. In order to enhance the ECG’s capacity for these analyses, more and more machine learning mechanisms have been incorporated to increase precision in data processing. However, deep learning mechanisms are generally considered black box models and are difficult to interpret. In this context, the present work uses the IMLE-NET model, which incorporates attention mechanisms in its architecture, the same ones that are widely used in Natural Language Processing (NLP) in Transformer architectures, with the aim of providing a certain degree of interpretability regarding regions of interest in ECG signals. Using weights pre-trained on the CinC2017 and ShareeDB datasets, both from the Physionet database (https://physionet.org/), it was possible to apply the transfer learning technique to evaluate the behavior of the model on a set of 4-hour ECG signals from patients with Chagas disease, containing 219 samples, for the problem of classifying ejection fraction below 50%. The experiments were carried out using the cross-validation technique with a holdout set of 15%, which allowed obtaining average recall of 0.82 (maximum of 0.93) in architectures with different topologies and in both datasets used in pre-workout. The F1-score presented average values greater 0.78. In terms of interpretability, models generated with both datasets were able to highlight regions between the T wave and the P-R interval that are important for identifying the patients’ ejection fraction status.

Banca examinadora:

• Prof. Dr. João Paulo do Vale Madeiro (MDCC/UFC – Orientador)
• Prof. Dr. César Lincoln Cavalcante Mattos (MDCC/UFC)
• Prof. Dr. Luís Otávio Rigo Júnior (UNILAB)