Source-linked AI summary
Leveraging Cardiac Imaging to Improve ECG-Based Detection of Chagas Disease in Resource-Constrained Settings
Laura Alvarez-Florez, Daniel Uyterlinde, Samuel Ruipérez-Campillo, Lukas P. A. Arts, Folkert W. Asselbergs, Fleur V. Y. Tjong
TL;DR
Chagas disease detection is needed where CMR is scarce, while ECG is widely available but does not directly capture cardiac structure. This paper transfers CMR-derived structural knowledge to ECG through contrastive pre-training, improving Chagas detection and performance on shifted cohorts without using Chagas cases during pre-training.
Problem
CMR provides structural cardiac information but requires scarce resources, whereas ECG is widely available yet represents cardiac structure only indirectly.
Method
The method contrastively aligns ECG representations with a clinically grounded CMR target space using paired ECG-CMR examinations, then deploys the resulting ECG encoder without CMR during inference.
Results
The representation transfers to Chagas detection despite no Chagas cases during pre-training and achieves the highest AUROC on SaMi-Trop-3 and the best ELSA-Brasil Challenge score among the compared top-performing methods.
Takeaways & Limitations
Imaging-supervised ECG representations can transfer structural cardiac information to Chagas detection across diseases, populations, and healthcare settings beyond the imaging cohort.
Takeaways & Limitations
Pre-training requires large paired ECG-CMR datasets and substantial computational resources, and the analysis does not identify which structural phenotypes contribute to Chagas prediction.
Abstract
from arXiv · showhide
Chagas disease is a major cause of cardiomyopathy in Latin America. Cardiac magnetic resonance (CMR) imaging can characterize its structural abnormalities, but scanners and expert readers remain scarce in endemic regions. Electrocardiography (ECG) is inexpensive and widely available, yet structural disease must be inferred indirectly from electrical signals. We propose to transfer CMR-derived structural knowledge to ECG through contrastive pre-training. Using 63,193 paired ECG-CMR examinations from the UK Biobank, we align an ECG encoder with a clinically grounded CMR embedding space using an asymmetric InfoNCE objective. Despite seeing no Chagas cases during pre-training, the resulting representation improves ECG-based Chagas detection. Across CODE-15% and SaMi-Trop, a frozen linear probe achieves an AUROC of 0.851 and sensitivity at the top 5% of predicted risk (Top5%-TPR) of 0.427 in five-fold cross-validation, compared with 0.827 and 0.377 for an unaligned ECG-FM baseline. On the PhysioNet/CinC 2025 Challenge test set, our model obtains the highest AUROC on SaMi-Trop-3 and the best ELSA-Brasil challenge score among the three top-performing methods, indicating that imaging-supervised ECG representations can generalize to populations and settings beyond the pre-training distribution.
1 Introduction
Chagas cardiomyopathy is structurally characterized but often assessed with widely available ECG because CMR access is limited. The paper proposes transferring CMR-derived structural knowledge to ECG through contrastive pre-training for diagnosis without imaging at deployment.
- Motivation: CMR directly characterizes cardiac structure, but cost, infrastructure, and specialist expertise limit its routine use in many regions.ECG is inexpensive, portable, and broadly available, but records electrical activity rather than structure.
- Motivation: Chagas cardiomyopathy involves myocardial fibrosis, ventricular remodeling, and progressive ventricular dysfunction that CMR can visualize.Chagas disease is a leading cause of non-ischemic cardiomyopathy in Latin America.
- Motivation: Prior multimodal studies align paired ECG and CMR representations, improving downstream cardiovascular prediction over ECG-only self-supervised approaches.These methods aim to make ECG representations capture structure and function otherwise directly observable through imaging.
- Contribution: The proposed approach pre-trains an ECG encoder against paired UK Biobank ECG-CMR examinations, then adapts it for Chagas detection using ECG alone.Chagas disease is absent from the paired ECG-CMR training cohorts, testing transfer to an unseen disease.
- Contribution: The work reports transfer of structural information to Chagas detection and validation on the PhysioNet/Computing in Cardiology 2025 benchmark against ECG-only baselines.The stated contribution is transfer across diseases, populations, and healthcare settings beyond pre-training.
2 Related Work
ECG-based Chagas models have generally remained unimodal, whereas recent ECG-CMR contrastive methods introduce structural supervision. However, transfer to diseases and populations absent from pre-training remains unresolved.
- Existing approaches: Self-supervised ECG foundation models learn transferable features for cardiovascular tasks, while Chagas predictors commonly use disease-specific cohorts or unimodal ECG pre-training.These models infer Chagas-related structural patterns from electrical signals without structural supervision.
- Open gap: The unresolved question is whether structural representations retain diagnostic value under domain shift when the disease and its imaging manifestations are absent from pre-training.This motivates testing transfer to Chagas disease and distinct populations.
- Multimodal learning: Recent multimodal methods align ECGs with CMR embeddings and deploy the ECG encoder alone after pre-training.They have achieved state-of-the-art ECG performance for structural phenotype regression.
3 Methodology
The method aligns ECG representations with a clinically grounded CMR target space, using asymmetric contrastive learning to transfer structural and functional information. The aligned ECG encoder is then deployed unimodally for Chagas classification.
- Framework: The framework first aligns paired ECG-CMR examinations and then adapts the structurally aware ECG encoder to Chagas classification in an unseen imaging-scarce cohort.The deployment phase uses ECG without requiring CMR.
- Framework: ECG-FM and CineMA provide initialized ECG and CMR representations, isolating the effect of cross-modal transfer.ECG-FM was pre-trained on 1.4 million 12-lead ECGs, while CineMA represents multi-view cine CMR sequences.
- Clinically grounded target space: Raw CineMA embeddings are nearly indistinguishable between end-diastole and end-systole, with average cosine similarity above 0.99.This near-degeneracy means static anatomy dominates and obscures subtle functional dynamics.
- Clinically grounded target space: End-diastolic and end-systolic CMR representations are concatenated and passed through a supervised MLP to construct the target space.The MLP is trained on structural and functional cardiac phenotypes.
- Asymmetric contrastive pre-training: An asymmetric InfoNCE objective aligns normalized ECG projections with normalized CMR targets while updating only the ECG encoder and projection head.The fixed target space organizes ECG embeddings according to the clinical geometry of CMR.
- Deployment: After pre-training, the CMR branch is discarded and the aligned ECG encoder is evaluated as a frozen feature extractor with a linear Chagas classification head.Because Chagas patients and Chagas-specific remodeling were absent from pre-training, this evaluates transfer to an unseen disease.
4 Experiments
The experiments pre-train on paired UK Biobank ECG-CMR data and evaluate Chagas detection on CODE-15% and SaMi-Trop under comparisons with ECG-only and end-to-end baselines. The protocol includes frozen linear probing and challenge-oriented submissions.
- Data: 63,193 matched UK Biobank ECG-CMR examinations are used, with 47,683 for alignment training and 15,510 for validation and held-out evaluation.Splits are stratified by LVEF, LV mass, sex, and AFib diagnosis.
- Data: The Chagas evaluation combines CODE-15% and SaMi-Trop, comprising approximately 345,000 ECGs and 8,192 Chagas-positive cases.The combined prevalence is approximately 2.4%; PTB-XL is added for the challenge-protocol comparison.
- Comparisons: The ablation compares the proposed model with unaligned ECG-FM, an end-to-end ensemble, and the PhysioNet 2025 winning approach.These comparisons separate cross-modal alignment from ECG-only pre-training and alternative supervised systems.
- Evaluation: The downstream analysis uses layer 9 as a frozen feature extractor with a linear head trained in five-fold cross-validation.Layer selection is independent of Chagas labels and is based on performance across clinically relevant cardiac phenotypes.
- Evaluation: The challenge submission ensembles the frozen linear probe with an end-to-end fine-tuned model initialized from the contrastive checkpoint.Class imbalance uses positive-class reweighting, and checkpoint selection uses AUPRC.
5 Results
Cross-modal ECG-CMR pre-training improved Chagas detection over ECG-only baselines in cross-validation and performed competitively on the official challenge test sets.
- Value of cross-modal alignment: 5.0 percentage points in Top5%-TPR separated ECG-CMR contrastive pre-training from the unaligned ECG-FM baseline on CODE-15% and SaMi-Trop.The comparison used five-fold cross-validation with frozen encoders and linear probes.
- Comparison with existing methods for Chagas prediction: The CMR-aligned encoder surpassed the AUROC of 0.80 reported by Jidling et al., despite never encountering Chagas patients or Chagas-related CMR during pre-training.Jidling et al.'s model was trained end-to-end on Chagas data.
- Challenge protocol: 5.0 percentage points in Top5%-TPR separated the frozen linear probe from the challenge winner's frozen-backbone ablation under the full challenge protocol.Both methods used frozen encoders and linear probes.
- Official challenge test set: 0.269 was the overall Challenge score, placing the submission just below the top three teams' 0.280–0.323 range.The overall score was reported on the official hidden test sets.
- Official challenge test set: 0.773 was the highest AUROC among compared methods on SaMi-Trop-3, versus 0.767 for the winner.On ELSA-Brasil, the submission achieved the best Challenge score, 0.132.
6 Discussion
The discussion presents imaging-supervised ECG pre-training as a way to transfer structural cardiac knowledge across diseases and populations without CMR at inference. It also identifies unresolved mechanisms and practical deployment boundaries.
- 6 Discussion: The framework distills structural cardiac knowledge from UK Biobank ECG-CMR pairs into a deployable ECG encoder without CMR during inference.The CMR branch is used for alignment during pre-training and is not required for downstream deployment.
- 6 Discussion: Contrastive ECG-CMR pre-training provided diagnostic benefit beyond the diseases and populations represented in the imaging cohort.The reported transfer includes Chagas cardiomyopathy despite its imaging manifestations being absent during pre-training.
- 6 Discussion: The CMR-aligned encoder outperformed ECG-only self-supervised baselines and models trained end-to-end on Chagas-labeled data without disease-specific supervision.On the official benchmark, overall performance was just below the top-ranked teams, while ELSA-Brasil results exceeded those of the top three winners.
- 6 Discussion: The analysis does not identify which specific structural phenotypes contribute to Chagas prediction.The authors propose relating learned ECG representations to Chagas-relevant cardiac manifestations in future work.
- 6 Discussion: Large paired ECG-CMR datasets and substantial computational resources remain necessary for representation development, although deployment uses only a standard 12-lead ECG.The authors call for prospective evaluation across Latin American populations and assessment of deployment requirements.
7 Impact in RCS
The framework enables ECG-based diagnosis where advanced cardiac imaging infrastructure is unavailable by transferring structural cardiac knowledge into an ECG encoder.
- 7 Impact in RCS: The framework enables diagnosis in settings where advanced imaging infrastructure is not available.Structural cardiac knowledge is distilled into an ECG encoder for deployment.