MEETI: Multimodal ECG Analysis Dataset

• MEETI is a multimodal ECG dataset integrating raw waveforms, quantitative beat-level features, high-resolution plotted images, and LLM-generated narrative interpretations.
• The dataset comprises 784,680 12-lead ECG records from 160,597 adult ICU patients, supporting tasks like arrhythmia classification and parameter regression.
• MEETI employs advanced signal processing and wavelet analysis to ensure precise feature extraction and reliable mapping between clinical signals and narrative diagnostics.

MEETI is a large-scale, multimodal dataset constructed from the MIMIC-IV-ECG collection and specifically designed to support advanced research in explainable, transformer-ready ECG analysis, retrieval, and interpretation. Each MEETI record is precisely aligned across four modalities—raw ECG waveforms, high-resolution plotted images, quantitative beat-level parameters, and detailed narrative interpretations—enabling rigorous multimodal tasks and providing a unified resource for fine-grained and interpretable cardiovascular AI. All records are systematically linked with unique study identifiers and a master index, supporting seamless data integration and retrieval. MEETI covers 784,680 12-lead, ten-second clinical ECGs from 160,597 adult ICU patients and introduces LLM-generated structured text outputs for each trace. Benchmark tasks include arrhythmia classification, parameter regression, interpretability assessment, and multimodal retrieval. The dataset addresses the prevailing limitations of prior ECG resources, which to date had been constrained to single- or dual-modality coverage, and it provides a foundation for clinically relevant, multimodal explainable machine learning in arrhythmia and cardiac health analysis (Zhang et al., 21 Jul 2025).

1. Dataset Structure and Modalities

MEETI unifies four key data types for each 12-lead ECG recording:

• Raw Waveforms: Stored in WFDB format, with 12 channels sampled at 500 Hz for 10 seconds per record. All 784,680 records include raw traces, enabling time-series analysis at the lead and beat levels.
• Quantitative Features: Beat-by-beat parameters (e.g., HR, RR intervals, amplitudes, durations) computed per lead using the open-source FeatureDB toolkit. Features are provided for every record and stored in MATLAB .mat files under per-lead fields.
• Plotted Images: High-resolution PNGs render the full 12-lead ECG trace in a 12×1 lead layout at 300 dpi, conforming to widely used clinical visualization standards. Due to file-size considerations, approximately 10,000 records include these images; the remainder may require regeneration from signals using the provided scripts.
• Interpretive Text: LLM-generated (GPT-4o) clinical narratives, each conditioned on the original report and full quantitative feature table, provide structured explanations mapping waveform- and feature-level findings to clinical constructs.

All modalities are anchored to a unique study identifier and cataloged within a central CSV registry, providing direct mapping across signals, features, reports, and images.

2. Signal Processing and Quantitative Feature Extraction

Signal preprocessing and feature extraction are performed via the FeatureDB pipeline, which consists of:

• Peak Detection: Adaptive algorithms locate the onset and offset of P-waves, QRS complexes, and T-waves in each lead.
• Wavelet Analysis: Multi-scale discrete wavelet transforms refine boundary delineation of each waveform component and effect baseline wander removal.
• Parameter Calculation: For each beat, the following are computed and stored:
  • Heart Rate: $HR = \frac{60000}{(RR_1 + RR_2) \times 500}$
  • R-R Intervals: $RR_1$, $RR_2$
  • Amplitudes and Durations: $P_{amplitude}$, $QRS_{amplitude}$, $T_{amplitude}$; durations and intervals (PR, QRS, QT, QTc, ST)
  • Morphological Categoricals: ST form (horizontal, upsloping, downsloping)

Measured parameters are serialized in .mat files with a uniform schema, enabling per-beat, per-lead statistical analysis and direct interpretability mapping.

3. Textual Interpretation Generation and Quality Control

Interpretative narratives are generated for each ECG record using GPT-4o. The text generation process employs "role-based prompts" that integrate the original clinician report and all computed per-beat features, with instructions to explicitly link quantitative findings to diagnostic statements (e.g., “QRS durations range 110–130 ms, suggesting intraventricular conduction delay”). Output consistency is verified by:

• Conditioning prompts on ground-truth diagnoses.
• Applying plausibility filtering against known diagnostic labels to flag and remove inconsistent or implausible generations.

Each narrative provides a concise but detailed account suitable for both clinical and interpretability-focused research tasks.

4. Data Formats, Directory Layout, and Access

All data are systematically organized to facilitate multimodal research workflows:

Modality	File Type & Structure	Availability
Raw waveform	WFDB (.dat/.hea)	All 784,680 records
Quantitative features	MATLAB .mat (per-lead arrays, interpretations)	All 784,680 records
Plotted images	PNG (300 dpi, 12×1 layout)	~10,000 records
Master index	CSV (study and subject IDs, file pointers)	All records

The directory tree is divided by patient and study IDs (e.g., MEETI/pNNNN/pXXXXXXXX/sYYYYYYYY/), with record-level mappings in the CSV file. Each .mat record includes fields such as id, report, LLM_Interpretation, and per-lead feature dictionaries.

5. Preprocessing, Quality Assurance, and Limitations

Preprocessing enforces:

• Signal integrity (all 12 leads, correct sampling rate, artifact exclusion).
• Standardized image rendering (paper speed: 25 mm/s; amplitude: 10 mm/mV; grid: 0.5 mV × 0.2 s blocks).
• Internal baseline wander removal within the FeatureDB wavelet stage.

Notable limitations include:

• All cases are derived from Beth Israel Deaconess adult ICU patients, potentially introducing demographic biases.
• LLM interpretations may hallucinate or omit rare features; critical results require external verification.
• Only ~10,000 images are provided due to file-size constraints. Full-scale image-based studies necessitate image regeneration using released scripts.
• All data are de-identified under HIPAA Safe Harbor, and secondary users must adhere to MIMIC-IV-ECG data-use agreements.

6. Benchmark Tasks and Downstream Applications

MEETI supports a range of clinically and computationally relevant tasks:

• Arrhythmia Classification: Leveraging any, or all, of the four modalities.
• ECG Parameter Regression: Predicting beat-level or lead-level physiological intervals and amplitudes.
• Interpretability Studies: Mapping text explanations to corresponding quantitative features and signal events.
• Multimodal Retrieval: Querying by image, text, or signal to locate semantically or morphologically similar ECGs.

Evaluation metrics include accuracy, $F_1$, AUC for classification, MSE for regression, and BLEU/ROUGE for generated text quality. Distributional statistics and representative LLM outputs are included for baseline reference; pretrained models are not bundled with the dataset.

7. Example Record and Interpretive Output

For illustration, a single MEETI record (study ID 40000369) contains:

• Raw waveform snippet: [0.00, 0.12, 0.25, 0.10, –0.10, …] (Lead II, 500 Hz)
• Image: PNG export in 12x1 format showing all 12 leads on a standardized clinical grid
• Quantitative parameters (lead II, first 3 beats):
  • HR = 76 bpm, RR₁ = 395 ms, RR₂ = 410 ms
  • P_amplitude = 0.15 mV, P_duration = 90 ms
  • PR_interval = 140 ms, QRS_amplitude = 0.80 mV, QRS_duration = 88 ms
  • QT_interval = 360 ms, QTc = 404 ms
• LLM-generated interpretation:

"Normal sinus rhythm at ~76 bpm. PR intervals are consistently 140 ms, within normal limits. QRS complexes are narrow (88 ms) with no abnormal conduction. QTc of 404 ms is normal. No evidence of ischemia or arrhythmia."

A plausible implication is that this multimodal packaging supports both evaluation of signal-processing pipelines and benchmarking of transformer-based models in explainable ECG analysis (Zhang et al., 21 Jul 2025).