Benchmarking Supervised and Self-Supervised Mortality Risk Prediction Models on ECG Data
The project systematically compare supervised and self-supervised pretrained models for ECG-based mortality prediction, to examine whether self-supervised pretraining improves performance in deep survival analysis for all-cause mortality prediction, and how it compares with purely supervised models trained directly on the target task. Specifically, we adapted two state-of-the-art SSL backbones: ResNet and HuBERT-ECG, and three survival modeling approaches: DeepSurv, DeepHit, and Logistic Hazard, and also investigated the impact of different fine-tuning strategies on model performance.
Contributions:
- Literature review and experimental design
- Implementation of data loading, training pipelines, and fine-tuning strategies
- Enhancement of hyperparameter optimization using Optuna and TensorBoard
- Comparative evaluation and results analysis
References
This project builds upon existing open-source resources, including:
- Pretrained benchmark models adapted from cavalab/ecg-survival-benchmark
- HuBERT-ECG architecture based on self-supervised pretraining research HuBERT-ECG as a Self-Supervised Foundation Model for Broad and Scalable Cardiac Application
- PyCox survival analysis framework