Why the study?
Differentiating ischemic from non-ischemic T-wave inversion on ECGs is challenging and prone to high false-positive rates, while existing deep learning models are often single-modality black boxes.
Does a multi-modal Vision-Language Model trained with reinforcement learning improve diagnostic accuracy in differentiating ischemic from non-ischemic T-wave inversion compared to standard methods and clinicians?
Does a multi-modal Vision-Language Model trained with reinforcement learning improve diagnostic accuracy in differentiating ischemic from non-ischemic T-wave inversion compared to standard methods and clinicians?
A novel reinforcement learning-based vision-language model significantly improved the accuracy of differentiating ischemic from non-ischemic T-wave inversions on ECGs compared to standard AI methods and clinicians.
May aid ECG triage for TWI; leaves open prospective validation before clinical use.
BACKGROUND The differentiation of ischemic from non-ischemic T-wave inversion (TWI) on electrocardiograms (ECGs) is a critical diagnostic challenge in cardiology. The non-specific nature of TWI leads to high false-positive rates, resulting in unnecessary, costly, and risky invasive procedures for patients. Existing deep learning models are often limited by being single-modality "black boxes". OBJECTIVE The objective of this study is to develop a novel diagnostic framework designed to address the critical clinical challenge of accurately differentiating ischemic from non-ischemic TWI. By utilizing a multi-modal Vision-Language Model trained with a Reinforcement Learning (RL) paradigm, this study aims to improve diagnostic accuracy and provide interpretable reasoning. METHODS We develop ECG-R1, a multi-modal framework using the Qwen2-VL-2B Vision-Language Model to analyze both ECG waveform images and associated clinical text. Instead of SFT, the model is trained using a RL paradigm with the Group Relative Policy Optimization (GRPO) algorithm. The model is trained to generate a structured output containing an explicit reasoning trace and a final "Yes" or "No" answer. A two-component, rule-based reward function is designed to assess both format adherence and diagnostic accuracy. Performance is compared against strong Supervised Fine-Tuning (SFT) baselines. RESULTS On a multi-modal dataset of 12,917 cases with TWI, our GRPO model achieves an average accuracy of 74.07%, demonstrating strong generalization with 72.93% accuracy in cross-hospital validation. This result is an improvement of ~24 % over the ~50% diagnostic accuracy of clinicians and 8.2% higher than the best SFT baseline, using ~71% fewer parameters. CONCLUSIONS The RL-based ECG-R1 framework successfully differentiates ischemic from non-ischemic TWI and demonstrates significantly better generalization than standard SFT methods. By enhancing diagnostic accuracy and providing interpretable reasoning, this approach offers a more robust and trustworthy tool to support clinical decision-making in cardiology.
No takes yet. Share an insight, caveat, or question.
Cheng et al. (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: