Interpretable Multimodal Zero-Shot ECG Diagnosis via Structured Clinical Knowledge Alignment

TL;DR

ZETA tackles the opacity and limited generalization of automated ECG diagnosis by aligning 12-lead ECG signals with structured positive/negative clinical observations via a zero-shot, multimodal framework. A human-in-the-loop pipeline uses LLMs to generate observations, which are expert-validated and organized into disease-specific P/N sets; a frozen multimodal model then computes similarities between ECG embeddings and observation embeddings, aggregating evidence with a softmax-based reasoning to produce a clinically grounded likelihood. Across six datasets, ZETA achieves competitive zero-shot ROC AUC (average $77.2$) and strong cross-dataset generalization (e.g., PTBXL-Rhythm $90.2$, transfers up to $83.98$), while enabling interpretable predictions grounded in concrete diagnostic features. A physician study demonstrates that accompanying ZETA's explanations improves decision support, highlighting practical potential for transparent AI in clinical workflows. The work also provides a curated observation dataset and code, offering a reproducible platform for advancing interpretable, generalizable ECG AI.

Abstract

Electrocardiogram (ECG) interpretation is essential for cardiovascular disease diagnosis, but current automated systems often struggle with transparency and generalization to unseen conditions. To address this, we introduce ZETA, a zero-shot multimodal framework designed for interpretable ECG diagnosis aligned with clinical workflows. ZETA uniquely compares ECG signals against structured positive and negative clinical observations, which are curated through an LLM-assisted, expert-validated process, thereby mimicking differential diagnosis. Our approach leverages a pre-trained multimodal model to align ECG and text embeddings without disease-specific fine-tuning. Empirical evaluations demonstrate ZETA's competitive zero-shot classification performance and, importantly, provide qualitative and quantitative evidence of enhanced interpretability, grounding predictions in specific, clinically relevant positive and negative diagnostic features. ZETA underscores the potential of aligning ECG analysis with structured clinical knowledge for building more transparent, generalizable, and trustworthy AI diagnostic systems. We will release the curated observation dataset and code to facilitate future research.

Figures (4)

• Figure 1: Overview of our ZETA framework. (1) ECG Observation Generation & Curation: LLMs generate candidate positive/negative observations (e.g., for 1AVB), which are validated and refined by clinical experts. (2) Observation Comparison: A pre-trained multimodal model compares input ECG embeddings against expert-reviewed observation embeddings, yielding aggregated similarity scores for a final likelihood score. This process provides interpretable insights by linking predictions back to structured, clinically relevant diagnostic features.
• Figure 2: Workflow for preprocessing LLM-generated clinical observations. Candidate observations from multiple LLMs are selected and organized into dataset-specific and cross-dataset sets for subsequent expert review.
• Figure 3: ZETA's inference architecture: encoding ECG and P/N observations, computing similarities, and reasoning via observation scores to predict condition likelihood. Background color intensity represents ECG–observation similarity, with darker colors indicating higher similarity.
• Figure 4: Qualitative analysis of ZETA's interpretable predictions for First Degree Atrioventricular Block (1AVB), Axis Left Shift (ALS), and Wolf-Parkinson-White syndrome (WPW). Scores indicate the alignment between the ECG (sample leads shown) and expert-reviewed P/N clinical observations, providing a transparent rationale for the final diagnosis likelihood.