ECG-R1 Differentiates Ischemic and Nonischemic T-Wave Inversion

Per a JMIR Preprint and ResearchGate listing (April 2026), researchers developed ECG-R1, a multimodal vision-language model trained with reinforcement learning to differentiate primary ischemic from secondary nonischemic T-wave inversion on ECGs. The JMIR abstract states the RL-based framework "successfully differentiates ischemic from non-ischemic TWI and demonstrates significantly better generalization than standard SFT (supervised fine-tuning)." The related arXiv preprint (2602.04279), accepted at ICML 2026 per the project GitHub repository, describes the same model as "the first reasoning ECG MLLM designed for reliable ECG interpretation," per the paper's abstract.

ECG-R1 combines Qwen3-VL-8B as the language-vision backbone with ECG-CoCa as a dedicated time-series encoder, using decoupled projectors for each modality to avoid shared-capacity bottlenecks in earlier ECG MLLMs. Training follows a two-stage pipeline: supervised fine-tuning on 30,000 protocol-guided samples from MIMIC-IV-ECG - where the five-phase interpretation protocol is derived from a clinical cardiology monograph - followed by reinforcement learning with ECG Diagnostic Evidence Rewards (EDER). Unlike general reasoning LLMs such as DeepSeek-R1, EDER rewards structured intermediate clinical reasoning steps, not only final-answer correctness. An Interleaved Modality Dropout training strategy improves robustness when ECG signal or image data is missing at inference time.

T-wave inversion requires rapid triage between ischemia and secondary causes such as bundle branch block or ventricular hypertrophy. Per the arXiv paper, existing MLLMs including GPT-5.1 and MedGemma produce "plausible but clinically incorrect" ECG analyses, with the paper providing "the first quantitative evidence that severe hallucinations are widespread" in current models across proprietary, open-source, and medical MLLM categories. Licensed cardiologist evaluation is included in the ICML paper's experiments.

Code, model weights (ECG-R1-8B-RL), and a 30,000-sample protocol-guided training dataset are publicly released on GitHub (PKUDigitalHealth/ECG-R1) and HuggingFace. A live demo is accessible at ai.heartvoice.com.cn/ECG-R1. The work is led by Shenda Hong's group at Peking University; two co-authors are Tencent employees, as disclosed in the paper.

The T-wave inversion differentiation results originate from a JMIR preprint pending peer review. The broader ECG-R1 paper is accepted at ICML 2026 for ML peer review but neither represents clinical regulatory clearance nor validated deployment in a hospital setting.