When Not to Diagnose: Evaluating Evidence-Sufficiency Recognition in ECG Multimodal Language Models with ECGTrustBench

Mobile electrocardiogram (ECG) capture routes photographed or scanned 12-lead images to an AI model or a clinician, yet missing leads, calibration loss, or photographic degradation can remove diagnosis-critical evidence while the tracing remains visually readable. A condition-aware evidence-sufficiency check is therefore needed to flag unusable captures for recapture or clinician review before diagnosis. Existing ECG multimodal large language model (MLLM) evaluations cannot support this, as they assume every question is answerable. We introduce ECGTrustBench, which evaluates evidence-sufficiency recognition as answerable-versus-insufficient QA over 3,300 image-question pairs, with corruption criteria unanimously approved by three licensed clinicians. Across ECG-specific, medical-domain, general-domain, and proprietary MLLMs, we find a consistent gap between answering and abstaining: ECG-specific models reach up to 77.3% clean accuracy yet abstain on at most 10% of evidence-deficient cases, while proprietary models achieve recognition up to 82%. Overall, current models respond more to visible degradation than to missing diagnostic evidence, highlighting a critical gap for evidence-sufficiency gating.