Impact of contrastive learning methods for training foundation models in ECG analysis
Dive into an experimental study to discover the best methods for applying contrastive learning to ECG data, including algorithm selection, data pairing techniques, and optimal data formats for deep learning.
Unsupervised contrastive learning is getting more attention as it does not need any supervised training to understand the domain. In this regard, serveral ECG foundation models are there. However, it is ongoing research to find the best contrastive learning method for ECG analysis. For this we have to setup an experimental study to find the best contrastive learning method. For example: what is the best algorithm? What is the best method to make data pairs used in contrastive learning? What is the best data format for ECG signals to process via DL? (ex: signal, Image or spectrum diagram etc.)
Goal
To identify the most effective contrastive learning method for ECG analysis by conducting an experimental study that evaluates various algorithms, data pairing techniques, and data formats (such as signals, images, or spectrum diagrams) suitable for deep learning models.
Learning outcome
- ECG analysis using AI
- Machine learning / AI: unsupervised contrastive learning
- Building proper research methodology for testing the effectiveness of contrastive learning method to final output
- Writing a scientific paper and how to publish it
Qualifications
- Hard working, motivated
- Fluent in Python programming
- Interested in learning
- (The rest can be learned during the thesis work)
Supervisors
- Vajira Thambawita
- Pål Halvorsen
- Michael Riegler
References
- Foundation Models for Electrocardiograms
- Large-scale Training of Foundation Models for Wearable Biosignals
