Cha-Cha-Chagas: Auxiliary Pretraining and Fine-Tuning Across Heterogeneous Datasets for ECG-Based Chagas Disease Detection

This repository contains the Python implementation for our entry in the George B. Moody PhysioNet Challenge 2025, titled "Auxiliary Pretraining and Fine-Tuning Across Heterogeneous Datasets for ECG-Based Chagas Disease Detection".

Our submission was developed by the Cha-Cha-Chagas team and builds upon the official PhysioNet Challenge 2025 Python example repository, expanding it with deep learning–based architectures, auxiliary pretraining, and fine-tuning strategies for improved ECG-based Chagas disease detection.

---

🧠 Overview

Chagas disease (American trypanosomiasis) is a parasitic infection caused by Trypanosoma cruzi, soemtimes leading to chronic Chagas cardiomyopathy (CCC). Detection from ECG signals remains challenging due to the scarcity of high-quality labeled data.

Our approach investigates whether auxiliary pretraining on weakly labeled ECG data (from the large CODE-15% dataset) can improve downstream Chagas detection when fine-tuned on datasets with stronger labels such as SaMi-Trop (serologically confirmed positives) and PTB-XL (assumed negatives).

Despite the hypothesis, experiments revealed that this pretraining strategy did not outperform conventional supervised training, underscoring the importance of dataset balance, label reliability, and domain similarity in multi-dataset ECG modeling.

---

📁 Repository Structure

• train_model.py — Wrapper for training.

• run_model.py — Wrapper for inference.

• evaluate_model.py — Used for local validation with the official PhysioNet evaluation code.

• team_code.py — Main implementation containing:

  • Deep neural network architecture (Net1D).
  • Dataset handling (ECGDataset).
  • Auxiliary pretraining and fine-tuning logic.
  • Model saving/loading utilities.

As instructed by the Organizers, we did not modify the official PhysioNet scripts (train_model.py, run_model.py, helper_code.py). Only team_code.py.

---

🧩 Method Summary

Data Sources

We used the following datasets:

• CODE-15%: Large Brazilian ECG dataset with weak self-reported Chagas labels.
• SaMi-Trop: Serologically verified Chagas cohort (positive cases only).
• PTB-XL: German ECG dataset (assumed Chagas-negative controls).

Data Processing

• Resampling to 400 Hz.
• Fixed window of 7 seconds (shortest common duration).
• All 12 standard ECG leads used.
• Resampling and zero-padding handled dynamically during dataset loading.
• Signals and metadata read using WFDB package.

Model Architecture

We implemented a 1D convolutional neural network (Net1D) inspired by the InceptionTime architecture:

• Residual connections for stable training.
• Variable receptive fields for multi-scale temporal feature extraction.
• Integrated Squeeze-and-Excitation blocks.
• Final linear classifier for binary Chagas prediction.

Training Strategy

1. Auxiliary Pretraining

• Model pretrained on CODE-15%.
• Auxiliary tasks included demographic and rhythm-related ECG features.

2. Fine-Tuning

• Model fine-tuned using SaMi-Trop (positive) and PTB-XL (negative) data.
• Used balanced mini-batches via a WeightedRandomSampler.
• Optimization: Adam (lr=1e-4), weight decay=1e-5.
• Loss: Binary cross-entropy (BCEWithLogitsLoss).
• Training for 5 epochs per phase, with early stopping based on validation AUROC.

---

🧪 Results Summary

Data used for pretraining	Data used for fine-tune or conventional training	AUROC on internal validation data	AUPRC on internal validation data	Challenge score on hidden validation data	REDS-II (hidden test)	SaMi-Trop (hidden test)	ELSA-Brasil (hidden test)
CODE-15%	SaMi-Trop and PTB-XL	0.69	0.22	0.040	–	–	–
–	SaMi-Trop and PTB-XL	–	–	0.066	–	–	–
–	CODE-15%	–	–	0.143	–	–	–
–	SaMi-Trop and CODE-15%	0.81	0.41	0.316	0.296	0.247	0.136

Key Finding: Auxiliary pretraining using weak CODE-15% labels did not generalize better than direct supervised training on the development data nor the hidden validation. Therefore the best performing model, using conventional training on CODE-15% and SaMi-Trop, were applied on the final hidden test set (only one attempt). A possible cause of the unexpectedly poor results may be domain shift between datasets (e.g., Brazilian vs. German populations), which likely introduced dataset bias during fine-tuning.

---

🚀 Running the Code

1. Installation

Create a virtual environment or use Docker. Then install dependencies:

pip install -r requirements.txt

2. Training

python train_model.py -d training_data -m model -v

3. Inference

python run_model.py -d holdout_data -m model -o holdout_outputs -v

4. Evaluation

Clone and run the official evaluation repo:

python evaluate_model.py -d holdout_data -o holdout_outputs -s scores.csv

5. Docker (Recommended)

To ensure reproducibility:

docker build -t chagas_image .
docker run -it -v ~/example/training_data:/challenge/training_data \
               -v ~/example/model:/challenge/model \
               -v ~/example/holdout_data:/challenge/holdout_data \
               -v ~/example/holdout_outputs:/challenge/holdout_outputs chagas_image bash

---

⚙️ Model Loading and Saving

• Save: The model is stored as model.keras.h5 inside the model folder.
• Load: Automatically reloaded via load_model() in team_code.py.

---

📊 Limitations and Future Work

• Dataset Bias: Differences in acquisition equipment, population, and temporal context introduce confounders.
• Label Quality: CODE-15% self-reported labels remain unreliable for fine-grained learning.
• Future Directions: Domain adaptation, adversarial debiasing, or contrastive pretraining could enhance generalization across diverse ECG datasets.

---

🧩 References

This implementation accompanies the paper:

Auxiliary Pretraining and Fine-Tuning Across Heterogeneous Datasets for ECG-Based Chagas Disease Detection Bjørn-Jostein Singstad, Nikolai Olsen Eidheim, Amila Ruwan Guruge, Ola Marius Lysaker, Vimala Nunavath. University of South-Eastern Norway & Akershus University Hospital, 2025. GitHub Repository

---

🧑‍💻 Authors

Cha-Cha-Chagas Team:

• Bjørn-Jostein Singstad
• Nikolai Olsen Eidheim
• Amila Ruwan Guruge
• Ola Marius Lysaker
• Vimala Nunavath

---

Useful Links

• PhysioNet Challenge 2025
• Evaluation Code Repository
• Official Discussion Forum