MEETI: A Multimodal ECG Dataset from MIMIC-IV-ECG with Signals, Images, Features and Interpretations.

Area of Science:

• Cardiovascular Medicine
• Artificial Intelligence
• Biomedical Informatics

Background:

• Electrocardiograms (ECGs) are crucial for diagnosing cardiac conditions like arrhythmias and ischemia.
• Current machine learning for ECG interpretation faces limitations due to a lack of comprehensive, multimodal public datasets.
• Existing datasets often lack integration of raw signals, diagnostic images, and textual interpretations, hindering clinical AI deployment.

Purpose of the Study:

• To introduce MEETI (MIMIC-IV-Ext ECG-Text-Image), the first large-scale dataset synchronizing raw ECG waveforms, plotted images, and LLM-generated interpretations.
• To address the gap in multimodal ECG datasets for developing clinically applicable AI systems.
• To provide a benchmark for next-generation cardiovascular AI research.

Main Methods:

• Developed MEETI by integrating approximately 800,000 recordings from the MIMIC-IV-ECG database.
• Aligned four data components: raw ECG signals, high-resolution plotted images, per-beat quantitative parameters, and textual interpretations.
• Utilized unique identifiers for seamless data synchronization across all components.

Main Results:

• MEETI is the first dataset to synchronize raw ECG signals, plotted images, beat-level parameters, and detailed textual interpretations.
• The dataset enables fine-grained analysis and enhances model interpretability through beat-level quantitative parameters.
• Facilitates multimodal transformer learning for integrated cardiovascular AI analysis.

Conclusions:

• MEETI provides a robust foundation for developing advanced, multimodal AI in cardiovascular research.
• The dataset supports explainable and integrated analysis, overcoming limitations of single-modality approaches.
• MEETI is poised to accelerate the development of clinically deployable AI for ECG interpretation.