Source-linked AI summary
Detection of Chagas Disease from the ECG: The George B. Moody PhysioNet Challenge 2025
Matthew A. Reyna, Zuzana Koscova, Jan Pavlus, Soheil Saghafi, James Weigle, Andoni Elola, Salman Seyedi, Kiersten Campbell, Qiao Li, Ali Bahrami Rad, Antônio H. Ribeiro, Antonio Luiz P. Ribeiro, Reza Sameni, Gari D. Clifford
TL;DR
Chagas disease is difficult to detect because symptoms may be mild early, serological testing capacity is limited, and undiagnosed cases can have serious consequences. The Challenge developed ECG-based algorithms and evaluation methods for prioritizing patients for confirmatory testing, with strongest models identifying nearly three times as many Chagas-positive patients as indiscriminate testing on ELSA-Brasil data.
Problem
Chagas disease can progress from mild or asymptomatic infection to serious cardiovascular disease, while serological testing capacity is limited and detection rates are often below 1%.
Method
The Challenge used ECG algorithms trained and evaluated across datasets with weak and strong labels, hidden data sources, and a metric incorporating local confirmatory testing capacity.
Results
The highest-performing models identified nearly three times as many Chagas-positive patients as indiscriminate testing on ELSA-Brasil data, while performance decreased on unseen data sources.
Takeaways & Limitations
ECG-based machine learning can help prioritize potential Chagas patients for confirmatory diagnosis and treatment within limited testing capacity.
Takeaways & Limitations
Model performance decreased on unseen data sources, especially ELSA-Brasil, and the relevant prevalence and testing capacity vary geographically and over time.
Abstract
from arXiv · showhide
Objective: Chagas disease is a parasitic infection that is endemic to South America, Central America, and, more recently, the U.S., primarily transmitted by insects. Chronic Chagas disease can cause cardiovascular diseases and digestive problems. Serological testing capacities for Chagas disease are limited, but Chagas cardiomyopathy often manifests in ECGs, providing an opportunity to prioritize patients for testing and treatment. Approach: The George B. Moody PhysioNet Challenge 2025 invites teams to develop algorithmic approaches for identifying Chagas disease from electrocardiograms (ECGs). Main results: This Challenge provides multiple innovations. First, we leveraged several datasets with labels from patient reports and serological testing, provided a large dataset with weak labels and smaller datasets with strong labels. Second, we augmented the data to support model robustness and generalizability to unseen data sources. Third, we applied an evaluation metric that captured the local serological testing capacity for Chagas disease to frame the machine learning problem as a triage task. Significance: Over 630 participants from 111 teams submitted over 1300 entries during the Challenge, representing diverse approaches from academia and industry worldwide.
1. Introduction
Chagas disease can progress from mild or absent early symptoms to serious cardiac complications, while limited serological testing constrains diagnosis. The Challenge therefore focuses on ECG-based screening and prioritization, building on evidence that ECG and related signal analyses can reveal disease-associated changes but may generalize poorly across cohorts.
- Disease context: Chagas disease can become lifelong and cause cardiomyopathy, heart failure, arrhythmias, thromboembolism, and increased mortality, although early disease may be mild or asymptomatic.Specific drugs can prevent progression in early infection, but serological testing capacity is limited.
- Clinical need: Detection rates are often below 10% or 1%, motivating machine-learning systems that prioritize patients for confirmatory serological testing.Serological testing remains the diagnostic gold standard.
- ECG opportunity: Chagas cardiomyopathy often manifests in ECGs, making ECG analysis a potential source of signals for disease detection and treatment planning.ECG is positioned as a practical screening signal rather than a replacement for confirmatory diagnosis.
- Prior evidence: Prior work found HRV and ECG-pattern changes before overt cardiac involvement, including spectral-index changes indicating autonomic dysfunction.These findings established HRV as a potential marker of subclinical disease progression.
- Prior evidence: Entropy-based analyses of Holter ECGs can distinguish healthy individuals from Chagas patients, including some seropositive patients without overt ECG abnormalities.Later studies combined entropy methods with machine learning to improve predictive performance.
- Open challenge: Data-driven ECG studies show promise for Chagas detection but also substantial performance drops across independent cohorts, highlighting generalization challenges.The Challenge was designed to continue ECG-based screening efforts in this context.
2. Methods
The Challenge assembled heterogeneous 12-lead ECG datasets, standardized and augmented them, and evaluated open-source algorithms on hidden data. Its scoring metric framed Chagas detection as capacity-constrained prioritization for confirmatory testing rather than conventional classification alone.
- Challenge data: 378,624 12-lead ECG recordings from 6 sources were organized into public training data and hidden validation and test sets with approximately equal Chagas prevalence.The hidden sources were withheld to assess overfitting and generalization.
- Preprocessing: Augmentation oversampled Chagas-negative records in REDS-II and SaMi-Trop 3 to approximately match the 2.04% ELSA-Brasil positivity rate and added noise and device-representative filters.These steps were intended to preserve prevalence while supporting robustness across ECG acquisition conditions.
- Challenge procedure: Teams submitted open-source training and inference code, which was run on hidden validation and test data in containerized environments.The setup supported reproducibility and evaluation under controlled runtime conditions.
- Scoring metric: The primary score was true positive rate among the top 5% of patients ranked by predicted Chagas probability, reflecting constrained serological testing capacity.The metric reframed evaluation as a constrained ranking problem rather than relying on AUROC or accuracy.
- Scoring metric: The scoring formulation defines TPR(τ) = TP(τ)/P and FPR(τ) = FP(τ)/N for a classifier threshold τ, with referral capacity represented by M.Subjects exceeding the risk threshold are referred for testing.
3. Results
The Challenge attracted broad participation and evaluated algorithms using hidden validation and test data, including a capacity-constrained metric and multiple data sources.
- 1,317 code submissions came from 111 teams with over 650 team members.
- 65 teams had entries evaluated on hidden validation and test sets, while 41 met all requirements to be ranked.
- The Challenge score measures true positive rate within the top 5% of predicted cases under a fixed testing-capacity constraint.
- Table 3 reports Challenge scores for the three highest-ranked teams across validation and test sets from REDS-II, SaMi-Trop 3, and ELSA-Brasil.
- Median Challenge scores declined from validation to test data, with drops of 1.4% for REDS-II, 15% for SaMi-Trop 3, and 64% for ELSA-Brasil.
4. Discussion
Performance generally declined on unseen datasets, especially ELSA-Brasil, although the strongest models identified nearly three times as many positive patients as indiscriminate testing.
- Models performed similarly on REDS-II validation and test data but worse on SaMi-Trop 3 and worse still on ELSA-Brasil.SaMi-Trop 3 and ELSA-Brasil used different ECG machines and collection practices; ELSA-Brasil also had a more asymptomatic population.
- ELSA-Brasil best mirrors potential national ECG-based screening data despite generally producing the lowest model performance.
- Nearly three times as many Chagas-positive patients were identified by the highest-performing models as by indiscriminate testing.The passage frames this as potentially arresting disease development while patients remain largely asymptomatic.
- Prevalence rates and serological testing capacities vary geographically and change over time, limiting the fixed assumptions used for screening evaluation.A prevalence rate of 2% and testing capacity of 5% are described as roughly accurate for Brazil at the stated time.
5. Conclusions
The Challenge assembled heterogeneous ECG datasets and used a testing-capacity-aware metric to support clinically relevant Chagas detection as a prioritization task.
- The article presents public and private 12-lead ECGs with both self-reported and serologically validated Chagas labels.
- A large weakly labeled database and smaller strongly labeled databases, evaluated on hidden sources, support assessment of generalizability to unseen populations and environments.
- The evaluation metric incorporates confirmatory testing capacity to support clinically relevant models that prioritize patients for Chagas testing.