ECG-MTDA: A Framework for Morphology-Temporal Decoupling and LLM-Guided Text Alignment.

Abstract

Effective automated electrocardiogram (ECG) interpretation hinges on disentangling waveform morphology from rhythm dynamics, a challenge for existing multimodal models that often conflate these heterogeneous attributes and introduce semantic ambiguity. We introduce ECG-MTDA, a framework that explicitly decouples these components. It learns morphology-oriented representations via a PQRST-guided masked autoencoder, while separately modeling temporal dynamics using continuous wavelet transform. Crucially, we align the learned morphology with concise, label-conditioned textual descriptions generated by a large language model (LLM) using a contrastive objective, creating a semantically grounded embedding space. ECG-MTDA demonstrates superior performance on the PTB-XL and CPSC 2018 benchmarks (e.g., AUC 93.16 on PTB-XL Superclass), with statistically significant gains over a strong multimodal baseline. Furthermore, on a challenging in-house cohort (n=620) for short-term paroxysmal atrial fibrillation (pAF) progression prediction, the model achieves an AUC of 0.97 $\pm$ 0.02 with high sensitivity (0.80 $\pm$ 0.04) and specificity (0.98 $\pm$ 0.01). Ablation studies and qualitative analyses confirm the benefits of our decoupled design and morphology-text alignment. Our results demonstrate that this clinically-inspired decoupling strategy yields more precise and robust multimodal representations for complex ECG analysis, enhancing both diagnostic classification and near-term risk stratification.