CAMEL: An ECG Language Model for Forecasting Cardiac Events

Neelay Velingker; Alaia Solko-Breslin; Mayank Keoliya; Seewon Choi; Jiayi Xin; Anika Marathe; Alireza Oraii; Rajat Deo

cs.LGFeb 17, 2026

CAMEL: An ECG Language Model for Forecasting Cardiac Events

Authors: Neelay Velingker, Alaia Solko-Breslin, Mayank Keoliya, Seewon Choi, Jiayi Xin, Anika Marathe, Alireza Oraii, Rajat Deo,

Organizations: University of Pennsylvania

Abstract

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).

Figures & tables

Figure 1: Example of CAMEL’s forecasting capability. In the top example, CAMEL takes as input normal sinus rhythm ECG at time $T$ and correctly forecasts AFIB at $T+3$ minutes by reasoning over the RMSSD, RR-interval, and PAC count (reasoning highlighted). In the bottom example, CAMEL correctly predicts a normal outcome based on accurately extracted statistics.

	Tasks			Inputs
Model	Input Classification	Report Generation	Forecasting	Modality	# ECGs	Duration (s)
ECGFounder ( 2025b )	✓	✗	✗	waveform	single	10
ECG-FM ( 2025 )	✓	✗	✗	waveform	single	10
MERL ( 2024a )	✓	✗	✗	waveform	single	10
ECG-Byte ( 2025 )	✓	✗	✗	waveform	single	10
MELP ( 2025 )	✓	✓	✗	waveform	single	10

Table 1: Comparison of ECG foundation models across capabilities. The duration field indicates the ECG lengths supported during training and evaluation for each model. CAMEL is the first ELM to support forecasting, multi-ECG inputs, and long-context signals.

Figure 2: An overview of the CAMEL architecture. 1-second, single-lead ECG segments from two patients are encoded and combined with text token embeddings. The resulting sequence is processed by an LLM backbone (MedGemma-4B with LoRA adapters) to generate a clinical report. Fixed models are shown in blue, and trainable models are shown in green.

Panel of Figure 3 — Figure 3: Forecasting performance (Macro-F1) in predicting AFib, AFlutter or Sinus Rhythm across input window $w$ $(s)$ and horizon $h$ $(s)$ . We report zero-shot results from GPT-5.2 (with Code Interpreter and high effort), PULSE, GEM, and CAMEL, supervised training results for XGB and CNN, and linear probing results for CAMEL Probe. PULSE and GEM only support 10-second ECG inputs ( $w=10$ ). CAMEL outperforms all baselines, with increased performance with higher input windows, highlighting the importance of supporting longer ECG recordings.

	Classification							ECG-QA		Stat	Report
	PTB-XL	Code15	CPSC	CSN	HEEDB	Icentia	Penn	Single	Multi	CPSC	PTB-XL	MIMIC
Method	F1							Accuracy		RMSE	LLM
MERL	19.22	19.72	24.86	6.38	16.33	15.80	2.53	-	-	-	-	-
MELP	18.62	26.07	20.75	5.00	19.03	10.80	2.53	-	-	-	-	-
PULSE	28.06	75.90	57.08	12.62	18.62	25.15	34.12	79.07	64.41	144148	18.40	22.05
GEM	29.82	80.83	61.06	8.19	19.40	18.70	33.33	80.80	65.89	304	20.45	44.65

Table 2: Comprehensive evaluation results across three tasks: Forecasting, Classification, and Report Generation. Best results are highlighted in “-” indicates method not evaluated on task/dataset.

Methods	PTBXL-Rhythm			PTBXL-Sub			PTBXL-Form			PTBXL-Super			CPSC2018			CSN
Training Ratio	1%	10%	100%	1%	10%	100%	1%	10%	100%	1%	10%	100%	1%	10%	100%	1%	10%	100%
MERL ( 2024a )	53.33	82.88	88.34	64.90	80.56	84.72	58.70	72.43	79.65	82.39	86.27	88.67	70.33	85.22	90.57	66.60	82.74	87.95
MELP ( 2025 )	88.83	94.65	96.91	79.22	84.40	87.46	63.44	76.71	83.30	85.82	87.61	87.87	88.54	91.75	94.32	78.25	84.83	90.17
PULSE ( 2024b )	98.38	99.07	99.25	89.21	92.44	93.88	82.38	90.19	91.97	89.88	92.32	93.08	67.81	73.90	82.48	87.25	89.49	91.87
GEM ( 2025 )	97.55	98.78	99.12	89.48	92.57	93.19	83.32	88.07	90.77	87.61	91.88	93.03	74.58	82.41	91.38	87.12	90.33	91.69
CAMEL (ECGInstruct)	97.74	98.98	99.07	90.07	92.38	93.75	81.28	89.47	92.17	87.59	91.58	92.33	78.54	93.37	98.52	90.00	95.05	97.11

Table 3: Linear probing performance (AUC [%]) of ECG classification across multiple datasets, 12-leads. Results are reported for different training data proportions (1%, 10%, 100%) CAMEL (ECGInstruct) refers to CAMEL trained only on ECGInstruct data Liu et al. (2024b) , whereas CAMEL refers to the model trained on our curriculum specified in

\lx@sectionsign

3 .

Method	PTBXL-Super	PTBXL-Sub	PTBXL-Rhythm	PTBXL-Form	Code15	CPSC-2018	CSN
No LoRA + Lead-aware masking	0.00	0.00	0.00	0.00	27.35	1.18	0.00
LoRA + Full masking	54.39	17.22	21.34	16.98	75.13	46.69	10.15
LoRA + Causal masking	69.15	22.13	21.08	14.72	77.34	42.30	10.02
LoRA + Lead-aware masking	75.91	26.49	34.27	19.03	81.54	51.10	14.30

Table 4: Architecture ablations (F1 [%]). Training without LoRA adapters leads to poor ECG representations, as evidenced by the zero F-1 scores on PTB-XL tasks. Lead-aware masking, which enables bidirectionality in attention computation for ECG leads, proves be better, or at-par with full masking. Both LoRA adapters and lead-aware masking are necessary for achieving SOTA accuracy.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Supervision	Dataset Name	Source/Collection	Records	Duration	Leads
Signal-only	MITDB ( 2001 )	Beth Israel	48	30 m	2
	NTUH ( 2021 )	National Taiwan University Hospital	1,000	10 s	12
	PTB ( 2009 )	University Clinic Benjamin Franklin	549	30–120 s	12
	QTDB ( 1997 )	Beth Israel, Multi-center Europe	105	15 m	2
	SPH ( 2022 )	Shandong Provincial Hospital	25,770	10–60 s	12
	G12 ( 2020 )	Emory University	10,344	10 s	12

Appendix

Table 5: Overview of the ECG Datasets.

Stage	Learning Rate	Epochs	Batch Size	Warmup steps
Stage 2	1e-4	3	8	500
Stage 3	1e-4	1	8	500
Stage 4	1e-4	9	4	256
Stage 5	1e-4	3	2	256

Appendix

Table 6: Hyperparameters for Training Stages 2-5

Hyperparameter	Value
Total Gradient Accumulation Target	4
Gradient Clipping	1.0
Optimizer	AdamW
Learning rate scheduler	linear
LoRA Rank	128
Mask strategy	semantic

Appendix

Table 7: Global Hyperparameters

	$h=60$						$h=180$
Method	10	30	60	120	300	600	10	30	60	120	300	600
XGB	59.55	55.58	56.89	56.09	59.64	55.20	56.85	56.60	58.18	57.09	58.07	56.17
CNN	46.89	54.62	53.59	50.79	52.35	53.10	49.59	54.72	57.08	54.40	52.76	50.21
CAMEL Probe	69.83	72.52	69.43	74.30	72.56	75.14	69.81	67.38	69.91	69.85	75.46	74.64
GPT-5.2	43.26	47.54	48.00	48.24	54.36	51.28	37.67	46.71	51.22	54.43	49.29	52.01
PULSE	43.21	-	-	-	-	-	40.50	-	-	-	-	-

Appendix

Table 8: Forecasting results (F1 [%]) on Icentia. The task is forecasting atrial fibrillation or flutter (abnormal) or normal sinus rhythm at varying horizons

h

(s) and input-windows

w

(s). We report supervised training results for XGB and CNN, zero-shot results for GPT-5.2, PULSE, GEM, and CAMEL, and linear probing result for CAMEL Probe.

Method	PTBXL Super	PTBXL Sub	PTBXL Rhythm	PTBXL Form	Code15 Binary	Code15 Multi
MERL	41.39	12.36	10.84	12.31	3.09	36.35
MELP	40.45	10.45	11.43	12.16	20.50	31.63
PULSE	73.47	8.11	18.07	12.59	62.28	89.52
GEM	75.77	8.01	20.87	14.64	71.12	90.53
CAMEL	67.75	28.89	39.29	25.54	69.45	86.95

Appendix

Table 9: Full classification results (F1 [%]).

Method	10	30	60	120	300	600	900
CAMEL	46.19	41.32	40.45	45.10	58.28	64.95	64.15

Appendix

Table 10: AFIB classification results (F1 [%]) with varying ECG durations (s)

Table 11: Classification results (F1 [%]) with fewer leads at test time.

Table 15

Method	Single Choose	Single Query	Single Verify	Multi CQ	Multi CV	Multi IQ	Multi IV
PULSE	82.22	78.07	76.91	67.96	53.75	84.71	51.21
GEM	83.67	78.19	80.54	67.94	57.36	84.61	53.64
CAMEL	70.69	72.38	75.89	67.83	56.85	84.17	68.06

Appendix

Table 13: Full ECG-QA results (Accuracy [%]) on PTB-XL dataset.

Table 14: Full report generation result.

Panel of Figure 4 — Figure 4: Example report generated by CAMEL for a sample from the PTB-XL dataset. The ground truth label is Sinus rhythm. Incomplete right bundle branch block. PR interval is at the upper limit. Otherwise, normal ECG.

Panel of Figure 5 — Figure 5: Example of a report generated by CAMEL for a sample from the MIMIC-IV dataset.

Figure 6: Stage 3 conversation examples across the 5 task types.

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

ELF: A Family of Encoder-Free ECG-Language Models
Jan 5, 2026William Han, Tony Chen, Chaojing Duan +6Healthcare Electrocardiography
Carnegie Mellon University, USA · Allegheny Health Network, USA · University of California, Los Angeles, USA +1
ECG-MoE: Mixture-of-Expert Electrocardiogram Foundation Model
Mar 4, 2026Yuhao Xu, Xiaoda Wang, Yi Wu +3Electrocardiogram Foundation Models Electrocardiography
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
ECGQuest: Benchmarking and Fine-Tuning Language Models for Electrocardiography
Aug 31, 2026Mohammadsina Hassannia, Matthew A. Reyna, Reza SameniElectrocardiography LLM Fine-Tuning
Department of Biomedical Informatics Emory University Atlanta, GA 30322, USA · Department of Biomedical Engineering Georgia Institute of Technology and Department of Biomedical Informatics Emory University Atlanta, GA 30332, USA

Jan 5, 2026J/K move · Enter open · S save

ELF: A Family of Encoder-Free ECG-Language Models

William Han, Tony Chen, Chaojing Duan, Xiaoyu Song+5

Carnegie Mellon University, USA · Allegheny Health Network, USA · University of California, Los Angeles, USA+1

Healthcare Electrocardiography Multimodal Large Language Models Electrocardiogram Interpretation Multimodal Learning

Open paper AskPDF arXiv

ECG-Language Models (ELMs) extend recent advances in Multimodal Large Language Models (MLLMs) to automated ECG interpretation. However, most existing ELMs inherit Vision-Language Model (VLM) design choices and rely on pretrained ECG encoders, introducing substantial architectural and training complexity. Inspired by encoder-free VLMs, we introduce ELF, a family of three encoder-free ELMs that remain competitive with, and often outperform, prior state-of-the-art ELMs across two datasets despite substantially simpler architectures and training pipelines. All code and data are available at github.com/ELM-Research/ECG-Language-Models.

Abstract

Electrocardiogram Foundation Models

Explore similar work

ELF: A Family of Encoder-Free ECG-Language Models

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

ECGQuest: Benchmarking and Fine-Tuning Language Models for Electrocardiography