ECG-R1: Protocol-Guided and Modality-Agnostic MLLM for Reliable ECG Interpretation

), a summary narrative, and a final diagnosis (enclosed in `). This structure forces the model to ground its reasoning in measurable features before concluding.

2. Decoupled Modality Encoding with Interleaved Modality Dropout (IMD)

To enhance robustness against missing modalities and ensure cross-modal consistency, ECG-R1 employs a decoupled architecture and a novel training strategy.

• Architecture: Unlike prior "omni-perception" models that couple time-series and image embeddings into a single token space, ECG-R1 uses a Decoupled Modalities Encoding scheme. It employs independent encoders for the ECG signal (time-series) and the ECG image, mapping them into the LLM space via separate projectors. The time-series is injected at a dedicated <ecg> token, while the image is injected at the standard <image> token.
• Interleaved Modality Dropout (IMD): During both Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), the model is trained on a mixture of environments where modalities are randomly dropped or their token block orders are swapped. Theoretically, this minimizes the mixture risk over all test-relevant environments (missing signal, missing image, swapped order), providing provable upper bounds on worst-environment risk and controlling cross-modal discrepancies.

3. Reinforcement Learning with ECG Diagnostic Evidence Rewards (EDER)

To further refine reasoning capabilities, the authors introduce a reinforcement learning phase using ECG Diagnostic Evidence Rewards (EDER).

• Reward Shaping: Unlike standard reasoning models that reward only final answer correctness, EDER incorporates a stepwise evidence coverage score. The reward function evaluates whether the model's generated reasoning trace explicitly includes key diagnostic evidence phrases extracted from the ground-truth protocol-structured reasoning.
• Optimization: The total reward combines format compliance, diagnosis accuracy (Jaccard similarity), and the stepwise evidence reward. The policy is optimized using the DAPO algorithm, encouraging the model to maintain structured, evidence-grounded intermediate reasoning steps rather than collapsing into short, potentially hallucinated answers.

Key Contributions

1. ECG-R1 Model: The first reasoning-based ECG MLLM that integrates protocol-guided data generation, decoupled modality encoding, and evidence-rewarded reinforcement learning.
2. Protocol-Guided Instruction Data Generation: A novel method for constructing a structured, six-step analysis corpus that adheres to monograph-defined thresholds and diagnostic logic, significantly reducing factual errors and hallucinations compared to prompt-based generation.
3. Interleaved Modality Dropout (IMD): A theoretically motivated training strategy that simulates modality missingness and token-order swapping, providing explicit guarantees for robustness and cross-modal consistency without requiring additional missing-modality generators.
4. Systematic Evaluation: A comprehensive benchmarking of proprietary, open-source, and medical MLLMs on ECG interpretation, revealing widespread hallucinations and modality sensitivity in current state-of-the-art models.

Experimental Results

The authors evaluated ECG-R1 against a broad spectrum of baselines, including general-purpose proprietary models (e.g., GPT-5.1, Gemini-3-Pro), open-source MLLMs, medical MLLMs (e.g., MedGemma, HuatuoGPT), and ECG-specialized models (e.g., GEM, PULSE).

• Grounded ECG Interpretation: ECG-R1 (RL) achieved a Diagnosis Accuracy of 80.29, significantly outperforming the previous best ECG-specialized model, GEM (74.70), and all general/medical MLLMs (which scored below 30.00). It also showed substantial gains in process-oriented metrics, such as Evidence-Based Reasoning (79.08 vs. 63.15 for GEM) and Clinical Diagnostic Fidelity (84.20 vs. 62.90 for GEM).
• Robustness and Consistency: Under modality missing conditions (e.g., inputting only the time-series signal), ECG-R1 maintained high performance with a relative drop of only ~2-3% in accuracy, whereas GEM suffered a 28.0% relative drop. In terms of cross-modal consistency, ECG-R1 achieved an SBERT-Score of 0.97 between image-only and signal-only interpretations, compared to 0.92 for GEM, indicating superior semantic alignment.
• Cardiologist Evaluation: Four licensed cardiologists rated ECG-R1 interpretations as significantly more reliable and useful than GEM. ECG-R1 scored 4.55/5 on reliability (vs. 4.16 for GEM) and 4.34/5 on usefulness, with specific praise for its analytical accuracy and structured reasoning quality.

Significance and Claims

The paper claims that ECG-R1 represents a substantial step toward reliable, clinical-grade ECG interpretation by AI. The authors emphasize that the widespread hallucinations found in existing models pose a critical risk for clinical deployment, where incorrect interpretations can lead to severe health outcomes. By grounding interpretations in measurable ECG features and standardized clinical protocols, and by ensuring robustness across varying data completeness scenarios, ECG-R1 addresses the primary barriers to trust in medical AI.

The work concludes that while ECG-R1 demonstrates superior performance, it is intended for research purposes and requires qualified clinical oversight for any real-world application. The authors position their contribution not as a replacement for clinicians, but as a tool to support safer, more consistent, and evidence-based ECG workflows, provided that outputs are verified by qualified professionals.