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).
Electrocardiograms (ECG) are multi-dimensional recordings of the heart's electrical activity and serve as a primary tool for diagnosing and triaging conditions such as heart attacks, arrhythmias, and other cardiac abnormalities (Kaplan Berkaya et al., 2018;Savonitto et al., 1999). From symbolic algorithms and statistical models to CNNs, automated ECG classification has moved from extensive academic study to widespread deployment in both ambulatory and in-hospital environments, as seen in systems like GE's Marquette 12SL (GE Healthcare, 2019). More recently, foundation models have emerged for jointly processing ECG and text, which we call ECG Language Models (ELMs). ELMs combine ECG representation learning with natural language generation to produce interpretable classifications and reports (Liu et al., 2024b,a;Wang et al., 2025;Lan et al., 2025). Despite their promise, existing ELMs only target classification and do not predict a patient's future state, thus offering limited support for early intervention.
Forecasting cardiac events from ECG signals is a key challenge for AI in cardiac care. Unlike classification, forecasting requires detecting subtle, prognostic patterns in ECGs to anticipate future adverse events. Such early warning of cardiac events, such as ventricular tachycardia, could allow clinicians to intervene to improve patient outcomes (Pollack et al., 2016;Soar et al., 2021). While classical ML models and CNNs have been applied to this task (Kenet et al., 2023;Rooney et al., 2023), they rely on fully supervised training for fixed-length inputs and offer interpretability only through post-hoc explanations, limiting their ability to generalize across tasks and 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. clinical contexts. In contrast, ELMs contain an LLM backbone trained on clinical knowledge, allowing them to generalize across tasks and generate natural language explanations along with their predictions.
To meet this challenge, we propose CAMEL (Cardiac Autoregressive Model for ECG Language-Modeling), the first general-purpose ELM designed to support long temporal context windows of ECG signals. While existing benchmarks are largely restricted to the classification of 10-second snippets, we introduce ECGForecastBench, a new benchmark for predicting future arrhythmias from baseline normal sinus rhythms as input. Our model generates forecasting reports by leveraging ECG signal statistics with established clinical associations (Zhang et al., 2025). These statistics provide physiologically grounded explanations for the risk of a future cardiac event.
Like prior ELMs, CAMEL builds on a pre-trained large language model backbone, namely MedGemma-4B (Sellergren et al., 2025), to support reasoning and natural language generation. The core insight that allows CAMEL to reason over long temporal contexts of ECG signals is how the integration of signal embeddings with text embeddings operates at the token level. By encoding each one-second segment of each lead in an ECG as an individual token, CAMEL can interleave multiple signal sequences of any duration with textual prompts. This design supports flexibility in both input length and lead configuration, enabling CAMEL to reason over long ECG contexts and variable, potentially incomplete sets of leads common in real-world settings. This is in contrast to prior ELMs, whose contexts are generally restricted to 10-second, 12-lead ECGs (Table 1).
To train CAMEL, we introduce a 5-stage curriculum that gradually builds the model’s reasoning and forecasting capabilities. Training starts with an autoencoder stage to learn robust ECG representations. Subsequent stages teach the model multiple-choice and short-answer tasks, understanding of ECG statistics, multi-turn conversational reasoning, and finally the generation of forecasting reports. This curriculum enables CAMEL to compute ECG statistics from long contexts and use them as evidence for clinically grounded forecasts (Fig. 1).
In summary, the main contributions of this paper are as follows. First, we introduce the architecture for CAMEL, which enables its unique capability to reason over long-duration ECG signals and identify predictive markers of future events. Next, we present a large-scale data generation pipeline that supports curriculum learning for ECG comprehension, including the development of a novel benchmark, ECGForecastBench, for evaluating forecasting of future adverse cardiac events. We then present a staged training algorithm that progressively builds CAMEL’s ECG grounding, reasoning, and forecasting capabilities. Finally, we demonstrate that CAMEL achieves strong zero-shot perfo
This content is AI-processed based on open access ArXiv data.