ECG-MoE: Mixture-of-Expert Electrocardiogram Foundation Model

Yuhao Xu; Xiaoda Wang; Yi Wu; Wei Jin; Xiao Hu; Carl Yang

cs.AIMar 4, 2026

ECG-MoE: Mixture-of-Expert Electrocardiogram Foundation Model

Authors: Yuhao Xu, Xiaoda Wang, Yi Wu, Wei Jin, Xiao Hu, Carl Yang

Organizations: Department of Computer Science, Emory University, Atlanta, GA 30307 · School of Computer Science, University of Oklahoma, Norman, OK 73019 · Center of Data Science, Emory University, Atlanta, GA 30307

Abstract

Electrocardiography (ECG) analysis is crucial for cardiac diagnosis, yet existing foundation models often fail to capture the periodicity and diverse features required for varied clinical tasks. We propose ECG-MoE, a hybrid architecture that integrates multi-model temporal features with a cardiac period-aware expert module. Our approach uses a dual-path Mixture-of-Experts to separately model beat-level morphology and rhythm, combined with a hierarchical fusion network using LoRA for efficient inference. Evaluated on five public clinical tasks, ECG-MoE achieves state-of-the-art performance with 40% faster inference than multi-task baselines.

Figures & tables

Figure 1: Existing foundation models have limitations because different models capture distinct ECG features, hindering multi-task performance within a single model. Furthermore, despite the strong periodicity inherent in ECG signals, this characteristic is often overlooked by existing ECG foundation models. To address these issues, we propose an ensemble learning method based on Period MoE. The effectiveness of our approach is validated across five common downstream tasks.

Figure 2: ECG-MoE framework: ① Multi-model feature extraction with adaptive downsampling, ② Period-aware gating weight generation via specialized MoE, ③ Multi-Task fusion using LoRA.

		TimesNet	DLinear	MOMENT	TEMPO	ECG-FM
Regression (MAE $\downarrow$ )	RR	$817.0\pm 2.5$	$816.4\pm 2.9$	$816.6\pm 2.1$	$816.3\pm 1.9$	$816.3\pm 1.9$
Regression (MAE $\downarrow$ )	Age	$62.28\pm 0.36$	$62.63\pm 0.50$	$62.61\pm 0.37$	$62.33\pm 0.38$	$62.27\pm 0.38$
Binary Class (F1 $\uparrow$ )	Sex	$0.60\pm 0.00$	$0.51\pm 0.08$	$0.34\pm 0.00$	$0.42\pm 0.00$	$0.33\pm 0.00$
Binary Class (F1 $\uparrow$ )	Ka	$0.06\pm 0.00$	$0.05\pm 0.00$	$0.02\pm 0.00$	$0.18\pm 0.00$	$0.35\pm 0.00$
15 Class (ACC $\uparrow$ )	AD	$0.06\pm 0.01$	$0.03\pm 0.02$	$0.03\pm 0.02$	$0.02\pm 0.00$	$0.07\pm 0.00$

Table 1: Zero-shot performance of baseline models on ECG data. Highlighted are the top first and second results. RR Interval Estimation (RR), Age Estimation (Age), Sex Classification (Sex), Potassium Abnormality Prediction (Ka), Arrhythmia Detection (AD).

		TimesNet	DLinear	MOMENT	TEMPO	ECG-FM	ECG-MoE
Regression (MAE $\downarrow$ )	RR	$304.3\pm 4.3$	$786.0\pm 5.4$	$146.9\pm 1.3$	$141.5\pm 2.1$	$147.3\pm 1.3$	76.37 $\pm$ 4.7
Regression (MAE $\downarrow$ )	Age	$24.89\pm 0.07$	$28.46\pm 0.74$	$13.41\pm 0.45$	$13.52\pm 0.31$	$13.49\pm 0.17$	12.83 $\pm$ 0.42
Binary Class (F1 $\uparrow$ )	Sex	$0.51\pm 0.05$	$0.57\pm 0.01$	0.69 $\pm$ 0.02	$0.54\pm 0.01$	$0.52\pm 0.05$	0.69 $\pm$ 0.01
Binary Class (F1 $\uparrow$ )	Ka	$0.41\pm 0.01$	$0.43\pm 0.01$	$0.49\pm 0.00$	$0.50\pm 0.00$	$0.49\pm 0.00$	0.57 $\pm$ 0.00
15 Class (ACC $\uparrow$ )	AD	$0.17\pm 0.00$	$0.48\pm 0.02$	$0.66\pm 0.03$	$0.54\pm 0.14$	$0.49\pm 0.03$	0.73 $\pm$ 0.01

Table 2: Performance of fine-tuned baseline models on five ECG downstream tasks. Bold values represent the best performance.

		MoE	Time MoE	Self-Attn	Cross-Attn	ECG-MoE
Regression (MAE $\downarrow$ )	RR	$103.47\pm 2.4$	$78.64\pm 1.4$	$117.46\pm 3.2$	$86.71\pm 5.8$	$\mathbf{76.37\pm 4.7}$
Regression (MAE $\downarrow$ )	Age	$16.49\pm 0.87$	$13.04\pm 0.14$	$15.47\pm 0.85$	$14.71\pm 0.51$	$\mathbf{12.83\pm 0.42}$
Binary Class (F1 $\uparrow$ )	Sex	$0.53\pm 0.13$	$0.61\pm 0.10$	$0.54\pm 0.01$	$0.61\pm 0.02$	$\mathbf{0.69\pm 0.01}$
Binary Class (F1 $\uparrow$ )	Ka	$0.39\pm 0.17$	$0.50\pm 0.10$	$0.48\pm 0.00$	$0.48\pm 0.00$	$\mathbf{0.57\pm 0.00}$
15 Class (ACC $\uparrow$ )	AD	$0.62\pm 0.01$	$\mathbf{0.74\pm 0.00}$	$0.67\pm 0.00$	$0.69\pm 0.01$	$0.73\pm 0.00$

Table 3: Comprehensive performance comparison of various MoE and attention mechanisms against the proposed ECG-MoE on ECG tasks.

Panel of Figure 3 — Figure 3: Comparative Attention Heatmaps for Different Attention Mechanisms: An Analysis Based on the RR Interval Task in the MIMIC-IV Dataset.

Electrocardiogram Foundation Models

Electrocardiogram foundation models learn reusable representations from large ECG datasets for diagnosis, risk prediction, and transfer across clinical settings.

See all 28 papers

CAMEL: An ECG Language Model for Forecasting Cardiac Events
Feb 17, 2026Neelay Velingker, Alaia Solko-Breslin, Mayank Keoliya +9Electrocardiogram Foundation Models Electrocardiogram Interpretation
University of Pennsylvania
Foundation Model for Cardiac Time Series via Masked Latent Attention
Mar 27, 2026Moritz Vandenhirtz, Samuel Ruipérez-Campillo, Simon Böhi +6Electrocardiogram Representation Learning Electrocardiogram Foundation Models
Department of Computer Science, ETH Zurich, Switzerland · Department of Biomedical Engineering, University of Basel, Switzerland
MS-ECG-FM: Towards a More Universal Electrocardiogram Foundation Model for Health Monitoring using Multi-source Contrastive Learning
Oct 6, 2026Robert A. Lewis, I-Min Chiu, Kyle Verrier +7Electrocardiogram Foundation Models Electrocardiogram Classification
Massachusetts Institute of Technology (work done while at Apple). · Apple, Inc. · Division of Cardiology, Johns Hopkins Medicine. +1

Feb 17, 2026J/K move · Enter open · S save

CAMEL: An ECG Language Model for Forecasting Cardiac Events

Neelay Velingker, Alaia Solko-Breslin, Mayank Keoliya, Seewon Choi+8

University of Pennsylvania

Electrocardiogram Foundation Models Electrocardiogram Interpretation Clinical Outcome Prediction Electrocardiogram Classification Forecasting Benchmarks

Open paper AskPDF arXiv

Electrocardiograms (ECG) are electrical recordings of the heart that are critical for diagnosing cardiovascular conditions. ECG language models (ELMs) have recently emerged as a promising framework for ECG classification accompanied by report generation. However, current models cannot forecast future cardiac events despite the immense clinical value for planning earlier intervention. To address this gap, we propose CAMEL, the first ELM that is capable of inference over longer signal durations which enables its forecasting capability. Our key insight is a specialized ECG encoder which enables cross-understanding of ECG signals with text. We train CAMEL using established LLM training procedures, combining LoRA adaptation with a curriculum learning pipeline. Our curriculum includes ECG classification, metrics calculations, and multi-turn conversations to elicit reasoning. CAMEL demonstrates strong zero-shot performance across 6 tasks and 9 datasets, including ECGForecastBench, a new benchmark that we introduce for forecasting arrhythmias. CAMEL is on par with or surpasses ELMs and fully supervised baselines both in- and out-of-distribution, achieving SOTA results on ECGBench (+7.0% absolute average gain) as well as ECGForecastBench (+12.4% over fully supervised models and +21.1% over zero-shot ELMs).

Abstract

Electrocardiogram Foundation Models

Explore similar work

CAMEL: An ECG Language Model for Forecasting Cardiac Events

Foundation Model for Cardiac Time Series via Masked Latent Attention

MS-ECG-FM: Towards a More Universal Electrocardiogram Foundation Model for Health Monitoring using Multi-source Contrastive Learning