CinC2025

Detection of Chagas Disease from the ECG: The George B. Moody PhysioNet Challenge 2025

Challenge Website

:skull::skull::skull: BIG MISTAKE: Forgot to add (z-score) normalization in the preprocessing pipeline in the config!!! :skull::skull::skull:

• The Conference
• Description of the files/folders(modules)
• Key problem to solve
• Background knowledge
• Post-Conference Thoughts
  • Thoughts on the Final Results
  • Focus Issue Phase Findings

The Conference

Conference Website | Unofficial Phase Leaderboard¹ | Official Phase Leaderboard | Final Test Results

     

The Conference Poster

Click to view the details

The Conference Poster

:point_right: Back to TOC

Description of the files/folders(modules)

Files

Click to view the details

• README.md: this file, serves as the documentation of the project.
• cfg.py: the configuration file for the whole project.
• const.py: constant definitions.
• Dockerfile: docker file for building the docker image for submissions.
• requirements.txt, requirements-docker.txt, requirements-no-torch.txt: requirements files for different purposes.
• evaluate_model.py, helper_code.py, prepare_code15_data.py, run_model.py, train_model.py: scripts inherited from the official baseline and official scoring code. Modifications on these files are invalid and are immediately overwritten after being pulled by the organizers (or the submission system).
• sync_official.py: script for synchronizing data from the official baseline and official scoring code.
• team_code.py: entry file for the submissions.
• submissions: log file for the submissions, including the key hyperparameters, the scores received, commit hash, etc. The log file is updated after each submission and organized as a YAML file.

Folders(Modules)

Click to view the details

• official_baseline: the official baseline code, included as a submodule.
• official_scoring_metric: the official scoring code, included as a submodule.
• models: folder for model definitions, typically we used a CRNN model. Some custom loss functions are also defined in this module.
• utils: various utility functions, including custom scoring functions, and some training-validation split files.
• results: folder containing some typical experiment log files, for reproducibility.

:point_right: Back to TOC

Key problem to solve

The data is highly imbalanced, with only approximately 2% of the data being positive. Dealing with the imbalanced data is the key problem to solve in this challenge. Possible solutions include:

• Upsampling the positive data
• Downsampling the negative data
• Using Focal Loss, Asymmetric Loss, etc.
• Using class weights
• Using data augmentation, including Mixup, Cutmix, etc.

Background knowledge

Chagas Disease and ECG

According to a review paper about ECG abnormalities in Chagas Disease, the most common ECG abnormalities are:

• Prevalence of overall ECG abnormalities was higher in participants with CD (40.1%; 95%CIs=39.2-41.0) compared to non-CD (24.1%; 95%CIs=23.5-24.7) (OR=2.78; 95%CIs=2.37-3.26).

• Among specific ECG abnormalities, prevalence of

  • complete right bundle branch block (RBBB) (OR=4.60; 95%CIs=2.97-7.11),
  • left anterior fascicular block (LAFB) (OR=1.60; 95%CIs=1.21-2.13),
  • combination of complete RBBB/LAFB (OR=3.34; 95%CIs=1.76-6.35),
  • first-degree atrioventricular block (A-V B) (OR=1.71; 95%CIs=1.25-2.33),
  • atrial fibrillation (AF) or flutter (OR=2.11; 95%CIs=1.40-3.19),
  • ventricular extrasystoles (VE) (OR=1.62; 95%CIs=1.14-2.30)

  was higher in CD compared to non-CD participants

Post-Conference Thoughts

Thoughts on the Final Results

• High-performing teams often accepted very low accuracy in exchange for better recall/risk ranking.
• Foundation/self-supervised ECG encoders (ViT/Transformer backbones, distilled/foundation models pretrained on large ECG corpora) are widely used.

Focus Issue Phase Findings

Empirical results from the Focus Issue Phase submissions reveal a key insight about handling the imbalanced data:

• Upsampling is harmful for this task. Heavy upsampling of positive samples (e.g., SaMi-Trop ×12) causes the model to overfit to the few positive instances, severely hurting generalization across datasets (especially the hidden SaMi-Trop III and ELSA-Brasil test sets).
• Quality > Quantity. Replacing upsampling with Weighted BCE + Soft Labels (reliability-aware label smoothing) improved the SaMi-Trop III score from 0.148 to 0.295 using the same ResNet backbone — nearly doubling performance.
• This held even without a Foundation Model backbone: the CRNN-ResNet with the correct training strategy (no upsampling, Weighted BCE, soft labels) outperformed all prior submissions across all datasets except ELSA-Brasil (slight drop, within noise).

Miscellaneous

CinC2020 | CinC2021 | CinC2022 | CinC2023 | CinC2024

Footnotes

1. As clarified by the organizers, the validation set for the official phase was updated, hence the unofficial and official phase leaderboards are not comparable. ↩