---
language: en
license: mit
tags:
  - electrocardiogram
  - multimodal-learning
  - zero-shot-learning
  - medical-ai
  - interpretability
---

# Interpretable Multimodal Zero-Shot ECG Diagnosis (ZETA)

[![📄 Paper](https://img.shields.io/badge/Paper-Nature-blue)](https://www.nature.com/articles/s44325-025-00099-x)
[![💻 GitHub](https://img.shields.io/badge/Code-GitHub-black)](https://github.com/Tang-Jia-Lu/Zeta)

## 🧠 Overview

We propose **ZETA**, a zero-shot multimodal framework for ECG diagnosis that aligns signals with **structured clinical observations**.

Instead of directly predicting diseases, ZETA **compares ECG signals with positive and negative clinical evidence**, mimicking differential diagnosis.


## 🖼️ Framework

<p align="left">
  <img src="zeta.png" width="80%">
</p>


## ⚙️ Method

- **Structured observations**: LLM-generated + expert-validated  
- **Multimodal alignment**: pretrained ECG-text model  
- **Inference**:
  - ✅ match with positive observations  
  - ❌ match with negative observations  

Prediction is based on **relative evidence strength**.



## 📊 Results

<p align="left">
  <img src="zeta_performance.png" width="80%">
</p>


## 📦 Checkpoint

```bash
ZETA/checkpoints/best.pt