TLDR: MEETI is the first large-scale ECG dataset that integrates raw waveform signals, high-resolution plotted images, extracted quantitative parameters, and detailed AI-generated textual interpretations. Built on MIMIC-IV-ECG, it addresses the critical need for comprehensive multimodal data to develop more advanced, interpretable, and clinically deployable AI systems for cardiovascular diagnosis and research.
Electrocardiograms (ECGs) are fundamental in diagnosing heart conditions like arrhythmias and myocardial ischemia. While artificial intelligence (AI) has shown great promise in interpreting ECGs, a significant hurdle has been the lack of publicly available datasets that combine all necessary data types: raw signals, diagnostic images, and detailed interpretation text. Most existing datasets offer only single or dual modalities, making it challenging to develop AI systems that can truly understand and integrate diverse ECG information as clinicians do in real-world settings.
To bridge this critical gap, researchers have introduced MEETI (MIMIC-IV-Ext ECG-Text-Image), a groundbreaking large-scale ECG dataset. MEETI is the first of its kind to synchronize raw waveform data, high-resolution plotted images, and comprehensive textual interpretations generated by advanced large language models (LLMs). Beyond these, MEETI also includes beat-level quantitative ECG parameters extracted from each lead, providing structured data for fine-grained analysis and enhancing model interpretability.
What Makes MEETI Unique?
Built upon MIMIC-IV-ECG, one of the largest open-access collections of 12-lead clinical ECG recordings with over 800,000 ten-second recordings paired with expert reports, MEETI significantly expands this foundation with three key additions:
- High-resolution ECG images: These are rendered from the raw waveforms using standardized plotting parameters, mimicking clinical paper recordings.
- Beat-level quantitative ECG parameters: Extracted using the FeatureDB toolkit, these include precise measurements of P waves, QRS complexes, T waves, and intervals like PR, QRS duration, QT, and QTc for every heartbeat.
- Detailed textual interpretations: Generated by GPT-4o, these interpretations are highly granular and explicitly link diagnoses to measurable ECG parameters, going beyond the often brief human-written reports.
Each record in MEETI is meticulously aligned across these four components—raw ECG waveform, corresponding plotted image, extracted feature parameters, and detailed interpretation text—using consistent, unique identifiers. This unified structure is designed to support advanced transformer-based multimodal learning, enabling more interpretable reasoning about cardiac health.
How Was MEETI Created?
The raw ECG signals originate from routine clinical care at the Beth Israel Deaconess Medical Center (BIDMC), part of the MIMIC-IV-ECG dataset. For parameter extraction, the open-source FeatureDB toolkit was integrated, which precisely localizes waveform components and measures key temporal intervals. High-fidelity ECG images were generated using the `ecg_plot` Python library, ensuring a standardized graphical format. For the textual interpretations, role-based prompts combined with the extracted ECG parameters allowed GPT-4o to generate clinically detailed and parameter-grounded insights, effectively leveraging its medical knowledge.
Also Read:
- Glucose-ML: A New Resource for Advancing AI in Diabetes Management
- Acoustic Index: A New AI Tool for Heart Health Assessment
Impact and Accessibility
MEETI represents a significant step forward for AI in cardiology. By integrating visual, temporal, and semantic modalities, it facilitates cross-modal alignment and supports the development of models capable of performing ECG analysis from multiple perspectives. This resource encourages a shift from traditional ‘black-box’ diagnostic pipelines towards inference-based reasoning, enhancing the ability of multimodal LLMs to generate meaningful, context-aware interpretations of cardiac health.
The dataset is publicly accessible, aiming to lower the barrier to entry for multimodal ECG research. Researchers can seamlessly integrate waveform analysis, image-based deep learning, feature engineering, and natural language processing within a single, comprehensive resource. While the initial release includes about 10,000 ECG images due to size constraints, users can generate additional samples from the underlying MIMIC-IV-ECG data using provided tools. For more details, you can refer to the original research paper.
