ECG-R1: Protocol-Guided and Modality-Agnostic MLLM for Reliable ECG Interpretation
Researchers have developed ECG-R1, a new multimodal large language model (MLLM) specifically designed for reliable electrocardiogram (ECG) interpretation. The model incorporates protocol-guided instruction generation, a modality-decoupled architecture to handle missing data, and reinforcement learning with diagnostic evidence rewards. Evaluations indicate that many existing MLLMs, including proprietary and open-source versions, suffer from widespread hallucinations when interpreting ECGs, underscoring the need for independent verification of their outputs. AI
IMPACT Introduces a more reliable method for AI-driven medical diagnostics, highlighting the risks of current models in critical applications.