Introduction and Background
Electrocardiogram (ECG) signals are essential diagnostic tools in cardiology, capturing the electrical activity of the heart over time. In addition to detecting heart-related abnormalities, recent studies have shown that ECGs also encode meta-information about individuals, such as age and gender. With the rise of deep learning, it is now possible to learn latent representations of ECGs that go beyond traditional hand-crafted features.
Representation learning aims to map high-dimensional raw data into lower-dimensional, informative representations. However, these representations are often entangled—i.e., they mix multiple explanatory factors. Disentangled representation learning seeks to separate these factors (such as age, gender, heart rate variability) into distinct components within the latent space. This approach can improve interpretability, downstream task performance, and generalization.
In this thesis, we investigate how disentangled representations can be learned from ECG data and whether such representations can help in predicting demographic variables such as age and gender, as well as uncover hidden or latent factors that may correspond to other clinical or physiological characteristics.
Research Question
Main Question:
Can disentangled representation learning be used to isolate and extract meta-information such as age, gender, and potentially other latent attributes from ECG data?
Sub-questions:
- How well can disentangled representations separate known factors (e.g., age, gender)?
- Can these representations reveal additional interpretable features without explicit supervision?
- What are the trade-offs between disentanglement quality and classification performance in ECG signal encoding?
Tasks & Goals
- Literature Review: Understand existing methods in disentangled representation learning and their applications to ECG or time-series data.
- Data Preparation: Select, preprocess, and label ECG data for training and evaluation.
- Model Development: Implement or adapt disentanglement models (e.g., β-VAE, InfoGAN) for ECG inputs.
- Evaluation: Measure how well latent variables capture age, gender, and other hidden features using classification and disentanglement metrics.
- Reporting: Analyze results, discuss findings, and document the full process in the thesis.
Expected Outcome
- A trained model that learns interpretable, disentangled representations from ECG data.
- Successful classification of age and gender using latent codes.
- Insights into whether additional latent factors relate to other physiological or clinical features.
- A compact and well-documented thesis with reproducible experiments.
Requirements
- Basic Programming Skills:
Experience with Python, including libraries such as NumPy, pandas, and matplotlib.
- Introductory Machine Learning Knowledge:
Understanding of neural networks, especially autoencoders or CNNs. Prior coursework or projects using PyTorch or TensorFlow is helpful but not mandatory.
- Interest in Biomedical Data:
Willingness to learn the basics of ECG signals and their clinical significance.