CAMEL: An ECG Language Model for Forecasting Cardiac Events

The paper introduces CAMEL, a novel ECG language model that utilizes a specialized encoder and curriculum learning to enable the first ECG-based forecasting of future cardiac events, achieving state-of-the-art performance across multiple classification and forecasting benchmarks.

The Gist

Imagine your heart is like a complex orchestra playing a continuous symphony. The Electrocardiogram (ECG) is the sheet music that records this music, showing the electrical rhythm of every beat. For decades, doctors and computers have been very good at reading this sheet music to diagnose what's happening right now (like "The violin section is out of tune" or "The drummer is skipping beats").

However, there's a huge gap: No one could reliably predict what the orchestra would play next. Could the music suddenly turn into a chaotic jazz solo (a dangerous heart rhythm) in five minutes?

Enter CAMEL (Cardiac Autoregressive Model for ECG Language-Modeling). Think of CAMEL not just as a music reader, but as a super-intelligent music critic who can also predict the future.

Here is how CAMEL works, broken down into simple concepts:

1. The Problem with Previous Models: "The Snapshot Trap"

Imagine trying to predict a storm by looking at a single photograph of the sky taken for just 10 seconds. You might see a cloud, but you can't see the wind picking up or the pressure dropping.

• Old AI models were like that. They only looked at 10-second "snippets" of heart data. They were great at saying, "This is a normal beat," but they couldn't see the slow, subtle changes happening over minutes or hours that signal a heart attack or arrhythmia is coming.

2. The CAMEL Solution: "The Long-Form Storyteller"

CAMEL is different because it can read long stories, not just snapshots. It can look at hours of heart data at once.

• The Translator: CAMEL treats heart signals like a language. It breaks the heart's electrical waves into tiny "words" (tokens).
• The Brain: It uses a massive "brain" (a Large Language Model, similar to the tech behind advanced chatbots) that has already learned how doctors speak and think.
• The Connection: CAMEL teaches this brain to understand that a specific pattern of "words" (heartbeats) means the same thing as a specific medical sentence. It bridges the gap between the raw squiggly lines of the heart and human language.

3. How It Learned: The "5-Step Boot Camp"

You can't just hand a complex medical problem to a robot and expect it to solve it immediately. The researchers trained CAMEL using a curriculum, like a student progressing through school:

• Step 1 (The Basics): CAMEL learned to recognize the raw shapes of heartbeats, like a child learning to identify letters.
• Step 2 (Multiple Choice): It practiced answering simple questions like, "Is this a normal beat or an irregular one?"
• Step 3 (The Math): It learned to do the math. It started calculating specific numbers, like "How long is the pause between beats?" or "How fast is the heart beating?" This is crucial because these numbers are the clues to the future.
• Step 4 (The Conversation): CAMEL practiced having a chat with a doctor. It learned to explain why it made a diagnosis, using the math it learned in Step 3. "I think this is dangerous because the pauses are getting shorter."
• Step 5 (The Crystal Ball): Finally, it learned to look at the current pattern and say, "Based on these specific clues, I predict a chaotic rhythm will happen in the next 3 minutes."

4. The "Magic" Trick: Lead-Aware Attention

A heart has 12 different "microphones" (leads) recording the same beat from different angles.

• Old models treated these like separate, unrelated conversations.
• CAMEL understands that all 12 microphones are listening to the same heart at the same time. It has a special "attention mask" that lets it compare all 12 views simultaneously, just like a conductor listening to the whole orchestra to hear if the violins are out of sync with the cellos.

5. Why This Matters: The "Early Warning System"

The biggest breakthrough is Forecasting.

• Before: If a patient had a heart attack, the AI might only tell you after the damage was done.
• With CAMEL: It can look at a patient's heart data, notice subtle changes (like a slight irregularity in the rhythm that a human might miss), and say: "I predict this patient will go into Atrial Fibrillation in 10 minutes."

This gives doctors a head start. Instead of reacting to a crisis, they can intervene early—perhaps giving a medication or adjusting a monitor—to prevent the crisis from ever happening.

Summary Analogy

If the heart is a car engine:

• Old AI was a mechanic who looked at the engine for 10 seconds and said, "It sounds fine right now."
• CAMEL is a mechanic who listens to the engine for an hour, notices the oil pressure is dropping very slowly, hears a faint rattle that only happens every 500 miles, and says, "This engine is going to seize up in 20 minutes. Pull over now."

CAMEL turns the heart's electrical signal from a static record into a dynamic story, allowing us to read the ending before it happens.

Technical Summary

1. Problem Statement

Electrocardiograms (ECGs) are critical for diagnosing cardiovascular conditions, but current AI approaches face significant limitations:

• Lack of Forecasting: Existing ECG Language Models (ELMs) focus primarily on classifying static 10-second snippets or generating reports. They cannot predict future cardiac events (e.g., the onset of atrial fibrillation), which is crucial for early clinical intervention.
• Context Limitations: Most foundation models are restricted to short input windows (typically 10 seconds), preventing them from capturing long-term temporal patterns and subtle prognostic markers necessary for forecasting.
• Interpretability Gap: Traditional deep learning models (CNNs, Transformers) often act as black boxes, offering predictions without the natural language reasoning or clinical justification required by physicians.

2. Methodology: The CAMEL Architecture

The authors propose CAMEL (Cardiac Autoregressive Model for ECG Language-Modeling), the first ELM designed to reason over long-duration ECG signals and forecast future events.

A. Architecture Design

• Backbone: Built upon MedGemma-4B, a pre-trained medical large language model (LLM), adapted with LoRA (Low-Rank Adaptation) to preserve medical knowledge while learning ECG-text alignment.
• ECG Encoder:
  • Uses a 3-layer CNN to encode 1-second ECG segments (at 256 Hz) into a 64-dimensional latent space.
  • A Single Linear Projection (SLP) layer maps these embeddings to the LLM's hidden token space ($h=2560$).
• Tokenization Strategy:
  • Unlike prior models that treat ECGs as fixed images or short sequences, CAMEL tokenizes 1-second segments of each lead as individual tokens.
  • This allows the model to handle variable-length inputs (up to 10,000 seconds) and interleaved text prompts.
  • Special tokens (<lead>, </lead>, <ecg>, </ecg>) structure the input, enabling the model to handle multiple leads and multiple ECG recordings within a single conversation.
• Lead-Aware Attention Masking:
  • A custom attention mechanism is introduced to handle the non-causal nature of multi-lead ECGs.
  • Tokens representing the same time step ($t$) across different leads are allowed to attend to each other bidirectionally, facilitating cross-lead learning while preserving the sequential dependencies required for autoregressive text generation.

B. Training Pipeline (5-Stage Curriculum)

CAMEL is trained using a progressive curriculum learning approach to build reasoning capabilities:

1. Autoencoder Stage: Self-supervised training on 1 billion ECG segments to learn robust signal representations (reconstructing raw waveforms).
2. Multiple Choice & Short Answer: Instruction tuning on high-level classification tasks (Harvard-Emory dataset) to establish basic diagnostic understanding.
3. Statistics Integration: Training on questions requiring the computation and interpretation of specific ECG metrics (e.g., RR intervals, QTc, heart rate) using the ECGDeli library.
4. Multi-Turn Conversations: Unifying classification and statistics into dialogue (single and multi-ECG comparisons) to simulate clinical reasoning and report generation.
5. Forecasting: The final stage trains the model to predict future rhythm transitions (e.g., Normal $\to$ AFIB) over specific time horizons. This uses the Icentia11k dataset and a rubric of electrophysiologic risk factors (e.g., RMSSD, PAC count) verified by cardiologists.

3. Key Contributions

1. CAMEL Model: The first general-purpose ELM capable of forecasting cardiac events by reasoning over long temporal contexts (up to 10,000 seconds).
2. ECGForecastBench: A new benchmark introduced to evaluate the prediction of future arrhythmias (AFIB/AFL) from baseline normal sinus rhythms, filling a gap in existing evaluation suites.
3. Curriculum Learning Pipeline: A novel 5-stage training strategy that effectively bridges the gap between raw signal processing and complex clinical reasoning.
4. Architecture Innovations: The introduction of lead-aware attention masking and 1-second tokenization enables flexible input lengths and multi-lead comparative reasoning.

4. Experimental Results

CAMEL was evaluated on 6 tasks and 9 datasets, demonstrating state-of-the-art (SOTA) performance.

• Forecasting (ECGForecastBench):
  • CAMEL outperformed fully supervised baselines (XGBoost, CNN) and zero-shot ELMs.
  • Achieved a +12.4% improvement over fully supervised models and +21.1% over zero-shot ELMs.
  • Performance scaled with input duration; longer inputs (up to 600s) significantly improved F1 scores compared to standard 10-second inputs.
• Classification:
  • Achieved SOTA results on ECGBench with a +7.0% absolute average gain.
  • Outperformed baselines on out-of-distribution datasets (e.g., CSN, Penn) and in single-lead telemetry scenarios.
• Report Generation & QA:
  • Demonstrated superior conversational ability and diagnostic accuracy on MIMIC-IV and PTB-XL datasets, scoring competitively with or exceeding models like GEM and PULSE.
  • Showed strong performance in "Grounding" tasks, achieving the lowest RMSE in ECG statistics computation (nearly half the error of GEM).
• Ablation Studies:
  • Confirmed that LoRA adapters and lead-aware attention masking are critical; removing them caused performance to collapse (e.g., 0% F1 on some tasks without LoRA).

5. Significance and Impact

• Clinical Utility: By forecasting future events rather than just classifying current states, CAMEL offers a tool for proactive intervention, potentially preventing adverse events like cardiac arrest or stroke.
• Explainability: The model generates evidence-based narratives, linking specific statistical findings (e.g., "high RMSSD") to clinical conclusions, enhancing trust and interpretability for clinicians.
• Scalability: The ability to process long-duration signals (hours vs. seconds) aligns better with real-world ambulatory monitoring (Holter monitors, wearable devices), moving beyond the constraints of short clinical snapshots.
• Foundation for Future AI: The work establishes a new paradigm for multimodal medical AI, demonstrating that curriculum learning and specialized tokenization can enable LLMs to perform complex temporal reasoning on physiological signals.

Limitations

The authors note that the 1-second tokenization strategy may occasionally truncate QRS complexes or miss subtle morphological changes. Future work aims to explore semantic tokenization based on QRS intervals or longer segment windows to optimize context window usage.