Fusion of ECG Foundation Model Embeddings to Improve Early Detection of Acute Coronary Syndromes

Zeyuan Meng, Lovely Yeswanth Panchumarthi, Saurabh Kataria, Alex Fedorov, Jessica Zègre-Hemsey, Xiao Hu, Ran XiaoarXiv 2025 · OpenReview

Abstract

Acute Coronary Syndrome (ACS) is a life-threatening cardiovascular condition where early and accurate diagnosis is critical for effective treatment and improved patient outcomes. This study explores the use of ECG foundation models, specifically ST-MEM and ECG-FM, to enhance ACS risk assessment using prehospital ECG data collected in ambulances. Both models leverage self-supervised learning (SSL), with ST-MEM using a reconstruction-based approach and ECG-FM employing contrastive learning, capturing unique spatial and temporal ECG features. We evaluate the performance of these models individually and through a fusion approach, where their embeddings are combined for enhanced prediction. Results demonstrate that both foundation models outperform a baseline ResNet-50 model, with the fusion-based approach achieving the highest performance (AUROC: 0.843 ± 0.006, AUCPR: 0.674 ± 0.012). These findings highlight the potential of ECG foundation models for early ACS detection and motivate further exploration of advanced fusion strategies to maximize complementary feature utilization.

Introduction

Acute Coronary Syndrome (ACS) is a medical emergency characterized by a sudden reduction in blood flow to the heart. In the critical "ultra-early" stage of care—within the ambulance—the electrocardiogram (ECG) is the gold standard for diagnosis. However, interpreting these signals in a prehospital environment is difficult due to noise and the need for rapid decision-making. Traditional AI approaches often rely on supervised learning, which requires massive amounts of expert-labeled data and often fails to generalize to the artifacts found in real-world EMS data. This research investigates the use of foundation models to bridge these gaps.

Methodology

The study leverages Foundation Models pretrained on large-scale unlabeled datasets using Self-Supervised Learning (SSL). Two distinct strategies were evaluated:

  1. ST-MEM (Reconstruction-based): A masked autoencoder approach that learns by reconstructing hidden patches of the ECG signal, capturing both temporal and spatial structures.
  2. ECG-FM (Contrastive-based): A model that learns by maximizing the similarity between different "views" of the same ECG recording, emphasizing lead-to-lead spatial relationships.

The researchers hypothesized that these different learning objectives capture complementary information. They developed a Fusion Model by normalizing and concatenating the embeddings from both models, followed by an XGBoost classifier for the final ACS prediction. This was compared against a ResNet-50 baseline trained from scratch on the target task.

Datasets

The models were pretrained on several large datasets, including Chapman, Ningbo, CODE-15, UHN-ECG, MIMIC-IV, and PhysioNet 2021, totaling millions of ECGs. Evaluation was performed on the MEDIC Dataset, which contains 5,813 12-lead ECGs collected by EMS providers in North Carolina from patients with suspected ACS.

Results and Discussion

The fusion approach demonstrated superior performance across all metrics compared to individual models and the supervised baseline.

Performance comparison of the baseline, individual foundation models, and the fusion approach.

As shown in the results table, the Fusion model achieved an AUROC of 0.843, a significant improvement over the ResNet-50 baseline's 0.701. This suggests that the general representations learned by foundation models provide a robust starting point for clinical diagnostic tasks.

The benefit of fusion is further supported by t-SNE visualizations of the model embeddings.

t-SNE visualizations comparing the embedding spaces of ST-MEM (left) and ECG-FM (right).

The visualizations reveal that ST-MEM produces more compact clusters, while ECG-FM results in more dispersed embeddings. This diversity confirms that the two models capture different facets of the ECG signal, allowing the fusion model to leverage a more comprehensive set of features for diagnosis.

Conclusion and Future Work

This study highlights the potential of fusing diverse ECG foundation model embeddings to improve early ACS detection. While the results are promising, the authors note that the current fusion method is relatively simple. Future research will explore more advanced integration strategies, such as attention mechanisms, and validate the models across more diverse EMS systems. Additionally, improving the interpretability of these "black box" embeddings remains a key goal for clinical implementation.

Summary updated

Abstract

Acute Coronary Syndrome (ACS) is a life-threatening cardiovascular condition where early and accurate diagnosis is critical for effective treatment and improved patient outcomes. This study explores the use of ECG foundation models, specifically ST-MEM and ECG-FM, to enhance ACS risk assessment using prehospital ECG data collected in ambulances. Both models leverage self-supervised learning (SSL), with ST-MEM using a reconstruction-based approach and ECG-FM employing contrastive learning, capturing unique spatial and temporal ECG features. We evaluate the performance of these models individually and through a fusion approach, where their embeddings are combined for enhanced prediction. Results demonstrate that both foundation models outperform a baseline ResNet-50 model, with the fusion-based approach achieving the highest performance (AUROC: 0.843 +/- 0.006, AUCPR: 0.674 +/- 0.012). These findings highlight the potential of ECG foundation models for early ACS detection and motivate further exploration of advanced fusion strategies to maximize complementary feature utilization.
BibTeX
@misc{meng2025fusion,
  author = {Zeyuan Meng and Lovely Yeswanth Panchumarthi and Saurabh Kataria and Alex Fedorov and Jessica Zègre-Hemsey and Xiao Hu and Ran Xiao},
  title = {{Fusion of ECG Foundation Model Embeddings to Improve Early Detection of Acute Coronary Syndromes}},
  year = {2025},
  eprint = {2502.17476},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2502.17476},
}
Figure