NUS Introduces MEETI: Multimodal ECG Dataset Revolutionizing Cardiovascular AI from MIMIC-IV

NUS Pioneers Multimodal ECG Innovation with MEETI Dataset Release

Researchers from the National University of Singapore (NUS) have launched the MEETI Multimodal ECG Dataset, a transformative resource derived from the vast MIMIC-IV-ECG database. This dataset marks a significant advancement in cardiovascular artificial intelligence, synchronizing raw electrocardiogram (ECG) signals, high-resolution images, quantitative features, and AI-generated interpretations for the first time at scale. Led by NUS faculty and alumni, MEETI empowers developers to build more accurate, explainable AI models for heart disease diagnosis, potentially saving lives through faster, more reliable clinical decisions.

The release, published today in Scientific Data, addresses a longstanding gap in public ECG resources. Traditional datasets like PTB-XL or Chapman-Shaoxing offer limited modalities, hindering multimodal learning where models process signals, visuals, and text simultaneously. MEETI changes that, providing 784,680 aligned records from 160,597 patients, enabling transformer-based systems that mimic human cardiologists.

Background: The MIMIC-IV-ECG Foundation and NUS Expertise

MIMIC-IV-ECG, hosted on PhysioNet, contains nearly 800,000 de-identified 12-lead ECGs collected at Beth Israel Deaconess Medical Center over a decade. Each 10-second recording, sampled at 500 Hz, comes with machine measurements and links to clinician reports. NUS's Saw Swee Hock School of Public Health, through Prof. Mengling Feng's AI for Public Health (AI4PH) program, specializes in leveraging such data for real-world health AI.

Xiang Lan, a recent NUS PhD graduate under Prof. Feng and now at Yale, co-led the effort alongside collaborators from Peking University and HeartVoice Medical Technology. Their work builds on NUS's strengths in data science, where Feng directs research extracting actionable insights from electronic health records to enhance care quality.

This Singapore-led initiative aligns with the nation's Research, Innovation and Enterprise 2025 (RIE2025) plan, investing S$25 billion in health tech, positioning NUS as a global leader in biomedical AI.

Decoding MEETI's Four Core Components

MEETI's power lies in its perfect alignment across modalities, using unique study IDs for seamless integration:

• Raw ECG Waveforms: Original 12-lead signals in WFDB format, preserving full temporal fidelity for arrhythmia detection.
• High-Resolution Images: Plotted ECGs in clinical layouts (e.g., 3x4 grid) at 300 DPI via ecg_plot tool, ideal for CNN-based vision models.
• Beat-Level Features: Extracted via FeatureDB—PR/QRS/QT intervals, amplitudes, heart rate per lead and beat, enabling precise electrophysiological analysis.
• LLM Interpretations: GPT-4o-generated texts, prompted as a cardiologist using features and reports, e.g., "Irregular RR intervals and absent P waves suggest atrial fibrillation, confirmed by QRS widening in V1-V3."

This structure supports end-to-end multimodal training, where models learn cross-modal reasoning like linking ST elevation in images to ischemia texts.

Technical Construction: From MIMIC to MEETI Step-by-Step

Building MEETI involved rigorous processing:

1. Filter MIMIC-IV-ECG for valid 12-lead, 10s records, yielding 784,680.
2. Extract features with adaptive peak detection and wavelet transforms for interval/amplitude accuracy.
3. Generate images matching clinical standards (25 mm/s speed, 10 mm/mV gain).
4. Craft role-based prompts for GPT-4o: input parameters + clinician notes → detailed, evidence-based reports.
5. Align via IDs, validate distributions (e.g., median QRS 90 ms).

Open-source tools on GitHub ensure reproducibility. Data on Zenodo, pending PhysioNet integration.

NUS's Strategic Role in Global Health Data Science

Prof. Feng's group at NUS exemplifies Singapore's push for AI-health fusion. His prior works on LLMs in healthcare and time-series contrastive learning underpin MEETI. Xiang Lan's thesis advanced ECG multimodal fusion, now scaling to MIMIC-IV.

This positions NUS graduates for top roles; explore research jobs in Singapore's booming biomedical sector. NUS's Institute of Data Science fosters such innovations, attracting global talent.

Applications: Transforming ECG AI from Single to Multimodal

MEETI unlocks:

• Diagnosis: Fuse signal-text for superior arrhythmia/myocardial infarction detection vs. unimodal baselines.
• Prognostics: Predict outcomes using features + interpretations.
• Explainability: Ground decisions in parameters, e.g., QTc prolongation risks.
• Education: Train clinicians with aligned visuals/texts.

Early users report 10-20% accuracy gains in multimodal transformers over signal-only models on held-out MIMIC subsets.

Challenges Addressed and Dataset Statistics

MEETI tackles data silos: prior datasets like PTB-XL (21k ECGs, signals+labels) or ECG-Text (image-text pairs) lack full multimodality. Stats:

• Patients: 160,597 (diverse demographics).
• Recordings: 784,680 (avg 4.9/patient).
• Parameters: 20+ per beat/lead (e.g., median HR 74 bpm).
• Texts: ~200-500 words/report, clinically precise.

Demographics mirror MIMIC: adults, emergency settings, common pathologies like sinus rhythm (60%), AFib (10%).

Future Implications for Singapore's Healthcare Ecosystem

In Singapore, where heart disease is the top killer (NHG data: 20% deaths), MEETI accelerates AI deployment. Integrate with NHCP's ECG AI pilots at National Heart Centre Singapore. NUS's efforts align with Smart Nation, fostering startups via career advice for AI-health pros.

Global impact: Benchmarks show multimodal models outperform singles by 5-15% F1 on tasks like MI detection. Expect citations surge, tools adoption.

Accessing MEETI: Tools and Community Impact

Download from GitHub (10 stars already). Requires PhysioNet credentialing for MIMIC base. Tools:

• FeatureDB: Extract features.
• ecg_plot: Generate images.

Join NUS AI4PH for collaborations; check higher-ed-jobs for postdocs in data science.

Photo by Andy Wang on Unsplash

Conclusion: NUS Leading the Charge in Explainable Health AI

MEETI exemplifies NUS's role in bridging academia-industry for societal good. As Singapore invests in precision medicine, datasets like this propel AI from lab to bedside. Researchers, clinicians: dive in today. For career growth in this field, visit Rate My Professor, Higher Ed Jobs, Career Advice, University Jobs, and post opportunities at Recruitment.