Clinically Interpretable Zero-Shot ECG Classification via Multimodal Learning and Expert-Aligned Descriptors

Luiz Facury de Souza1, Jose Geraldo Fernandes1, Pedro Robles Dutenhefner1, Turi Vasconcelos Rezende1, Gisele Pappa2, Gabriela Paixão3, Antonio Luiz Ribeiro1, Wagner Meira Jr1
1Universidade Federal de Minas Gerais, 2UFMG, 3Doctor


Abstract

The widespread adoption of artificial intelligence inclinical cardiology is hampered by a critical factor:the lack of transparency in automated electrocardiogram(ECG) interpretation systems. While deep learning mod-els can accurately classify cardiac abnormalities, their"black-box" nature prevents clinicians from verifying thediagnostic reasoning, undermining clinical trust. To ad-dress this, we developed a multimodal diagnostic frame-work that emulates clinical reasoning by directly linkingECG signal features to their corresponding textual de-scriptions from clinical reports. Our system learns torecognize abnormalities like Left Bundle Branch Block(LBBB) not merely as a classification label, but by identify-ing and associating it with established diagnostic criteria,such as a 'widened QRS complex (> 120 ms)'. Trainedon a large dataset of paired ECG signals and narrativereports, our model achieves diagnostic accuracies compa-rable to conventional supervised models without requiringexplicit training on classification labels. By grounding itspredictions in clinically relevant subfeatures, the systemprovides transparent, verifiable evidence for its conclu-sions. This approach represents a paradigm shift towardAI systems that augment clinical decision-making with in-telligible, evidence-based insights, fostering greater trustand facilitating integration into the diagnostic workflow