CLAP: Closed-Loop Agentic Perception Unlocks Latent ECG Diagnostic Capability in Multimodal Large Language Models

The electrocardiogram (ECG) is the most accessible tools for diagnosing cardiovascular disease. Recent advances in multimodal large language models (MLLMs) offer a promising path toward automated ECG diagnosis, yet a gap remains between current performance and clinical reliability. What fundamentally limits their diagnostic capability? In this paper, we reveal a perception–reasoning gap: both ECG-specific and general-purpose MLLMs struggle to reliably ground diagnoses in ECG evidence, while exhibiting stronger diagnostic ability when the relevant evidence is reliably perceived. To unlock this capability, we introduce Closed-Loop Agentic Perception (CLAP), an inference-time framework that turns ECG perception from a passive prediction into an active, verifiable, and iterative process. CLAP combines Perceptual Backprojection, which re-instantiates intermediate perceptual hypotheses on the source ECG for verification and refinement, with ECG Perceptual Scaffolding, which organizes interdependent perceptual tasks and coordinates grounded evidence across them. Experimental results show that, without additional training, CLAP consistently improves diagnostic performance across three general-purpose MLLMs, increasing patient-wise F1 by at least $1.44\times$ and enabling them to approach or surpass SOTA ECG-specific MLLMs. These gains, accompanied by consistent improvements in perceptual fidelity, highlight the broader potential of agentic intelligence to close the perception–reasoning gap toward clinically reliable ECG intelligence.