Multi-Task Learning for ECG Analysis with Self- and Semi-Supervised Representation Learning

📋 Type MA thesis
⚡ Status running
📅 Duration Aug 1, 2026 – Feb 1, 2027
👤 Primary supervisor Tomás Arias Vergara
👥 Co-supervisor Sheethal Bhat
🎓 Student Mehran Pashaei Medical Engineering MSc

Abstract

General objective. This thesis aims to design, implement, and evaluate a unified ECG representation learning
framework for multiple clinical ECG tasks, and to assess whether self- and semi-supervised learning improve data
efficiency and robustness compared to single-task approaches.
Specific objectives. The study will: (i) develop a shared 1D encoder with task-specific heads for AF detection, signal
quality estimation, waveform delineation, and event detection; (ii) pretrain the encoder using masked signal reconstruction;
(iii) apply semi-supervised learning for annotation-scarce delineation; (iv) compare CNN, CNN-Transformer, and Mamba
encoder variants; and (v) analyse task interactions through ablations and gradient analysis

General objective. This thesis aims to design, implement, and evaluate a unified ECG representation learning
framework for multiple clinical ECG tasks, and to assess whether self- and semi-supervised learning improve data
efficiency and robustness compared to single-task approaches.
Specific objectives. The study will: (i) develop a shared 1D encoder with task-specific heads for AF detection, signal
quality estimation, waveform delineation, and event detection; (ii) pretrain the encoder using masked signal reconstruction;
(iii) apply semi-supervised learning for annotation-scarce delineation; (iv) compare CNN, CNN-Transformer, and Mamba
encoder variants; and (v) analyse task interactions through ablations and gradient analysis