Defesa de doutorado do discente Guilherme Augusto, dia 20/03/2026 às 13:00.

Defesa de doutorado do discente Guilherme Augusto, dia 20/03/2026 às 13:00.

Título: A Multi-Context Evaluation of ECGWavePuzzle for ECG Learning: From Multitask Supervision to Quantized Personalization and MoE

Banca: Prof. Dr. Eduardo José da Silva Luz (Orientador); Prof. Dr. Jadson Gertrudes Castro (UFOP); Prof. Dr. Luiz Torres Bambirra (UFOP); Prof. Dr. Alexei Manso Correa Machado (PUCMinas); Prof. Dr. Raphael Teixeira (UFPA)

Abstract:

Cardiovascular diseases are the leading cause of death globally, and reliable arrhythmia detection from electrocardiogram (ECG) signals is essential for timely diagnosis and monitoring. Although deep learning has advanced automatic ECG interpretation, many studies still overlook deployment constraints such as memory, latency, and reduced numerical precision, and they often struggle with patient variability and scarce labeled data.

This thesis evaluates the self-supervised pretext task ECGWavePuzzle across multiple scenarios to test its viability under low-data and resource-constrained settings aligned with E3C. In a compact multitask framework that couples binary classification with RR-interval regression using lightweight backbones on PTB-XL, MIT-BIH, and IEGM, ECGWavePuzzle improves stability and discrimination without larger models, for example increasing VANet accuracy from 0.650 to 0.761 on PTB-XL and from 0.436 to 0.632 on MIT-BIH, while raising AUC from 0.808 to 0.846 and from 0.686 to 0.714. In a human-in-the-loop personalization pipeline with mixed precision and selective INT8 quantization-aware training, the ECGWavePuzzle-based approach remains stable under INT8/QAT, reaching accuracy around 0.90 and macro-F1 around 0.53. In a multi-lead, order-aware pretraining setting, ECGWavePuzzle achieves exam-level transfer with macro-AUROC close to 0.885 on PTB-XL and consistent trends under low-label fractions. Finally, as Stage I initialization in Puzzle-MoE, ECGWavePuzzle improves macro-AUROC from 0.8567 to 0.8784 in a controlled setting and supports routing-based specialization when load balancing prevents expert collapse.

Therefore, ECGWavePuzzle emerges as a practical morphology-aware signal that improves data efficiency across diverse ECG learning pipelines and remains compatible with deployment-oriented constraints, while highlighting open challenges for efficient, reliable on-device adaptation.