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

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Signal Processing

Background:

• Automated electrocardiogram (ECG) interpretation faces challenges in distinguishing waveform morphology from rhythm dynamics.
• Existing multimodal models often conflate these attributes, leading to semantic ambiguity in ECG analysis.

Purpose of the Study:

• To develop a framework (ECG-MTDA) that explicitly decouples waveform morphology and rhythm dynamics for improved ECG interpretation.
• To create a semantically grounded embedding space by aligning ECG morphology with text descriptions.

Main Methods:

• ECG-MTDA employs a PQRST-guided masked autoencoder for morphology representation and continuous wavelet transform for temporal dynamics.
• A contrastive objective aligns learned morphology with label-conditioned text descriptions generated by a large language model (LLM).

Main Results:

• ECG-MTDA achieved superior performance on PTB-XL and CPSC 2018 benchmarks, outperforming a strong multimodal baseline.
• The model demonstrated high accuracy (AUC 0.97) in predicting short-term paroxysmal atrial fibrillation progression on an in-house cohort.
• Ablation studies confirmed the benefits of the decoupled design and morphology-text alignment.

Conclusions:

• The clinically-inspired decoupling strategy in ECG-MTDA yields more precise and robust multimodal representations for complex ECG analysis.
• This framework enhances both diagnostic classification and near-term risk stratification for cardiac conditions.