25 February 2025
09:00 Master's Defense Room 85 of IC2
Topic
Exploring Representation Learning in Electrocardiogram Signals
Student
João Victor Palhares Barbosa Avanzini
Advisor / Teacher
Marcos Medeiros Raimundo
Brief summary
Electrocardiogram (ECG) signals give vital details about cardiovascular health, and effective Machine Learning (ML) algorithms have permitted the extraction of accurate representations from this data. While architectures such as Convolutional Neural Network (CNN) Autoencoder (AE), Multi-layer Perceptron (MLP) AE, and Transformer-Based Autoencoder With Series Embedding (TAE) have shown promise in capturing the temporal and spatial complexities of ECG signals, the role of Latent Space (LS) dimensionality in improving reconstruction accuracy and disease classification still lacks careful study. This work studies a variety of AE architectures with multiple LS dimensions to evaluate reconstruction errors. Similarly, classification experiments show that disease separability in the LS is dependent on the physiological impact recorded in ECG signals, with disorders such as Varicose Veins and Hepatocellular Carcinoma giving high Area under the Receiver Operating Characteristic Curve (ROC-AUC) values when classified wrt other conditions such as Thyroid Nodules. As a result, this study carefully assesses LS dimensionalities across various architectures, indicating that balanced configurations, such as a CNN AE with 200-dimensional LS, produce the best reconstruction and classification performance.
Examination Board
Headlines:
| Marcos Medeiros Raimundo | IC / UNICAMP |
| Petra Maria Bartmeyer | IMECC / UNICAMP |
| Anderson de Rezende Rocha | IC / UNICAMP |
Substitutes:
| Sandra Eliza Fontes de Avila | IC / UNICAMP |
| Cristiano Torezzan | FCA/UNICAMP |