Key result
CARDIO-ML-ED dataset demonstrates high diagnostic agreement between clinicians with a Cohen's kappa of ~0.82.
Why the study?
Publicly available cardiovascular databases that integrate ECG imagery with comprehensive clinical metadata such as laboratory results, vital signs, and medical history were in critical shortage.
Observational (n=582)
Yes
Answers are grounded in this paper and cite their sources.
The CARDIO-ML-ED dataset provides a publicly available, multimodal resource pairing 408 digitized 12-lead ECG images with 82 clinical metadata features across 582 subjects to support machine learning and multimodal fusion research in cardiovascular care.
Objectives: The CARDIO-ML-ED dataset was collected to address the critical shortage of publicly available cardiovascular databases that integrate electrocardiogram (ECG) imagery with comprehensive clinical metadata. Existing ECG repositories lack detailed patient-level information such as laboratory results, vital signs, and medical history—factors essential for developing clinically meaningful machine learning models. We aimed to create a curated, well-labelled, multimodal resource that supports research in automated ECG interpretation, multimodal fusion, and exploratory clinical phenotyping, while promoting reproducibility and innovation in artificial intelligence (AI) driven cardiovascular care. Data description: The CARDIO-ML-ED dataset comprises two complementary components: (1) structured clinical metadata for 582 subjects, and (2) a curated subset of 408 high-resolution ECG images derived from paper-based records digitized through standardized scanning, normalization, cropping, quality filtering, and noise reduction. The dataset includes 82 clinical features per subject—covering demographics, vital signs, laboratory tests (complete blood count, cardiac enzymes, electrolytes, inflammatory markers, arterial blood gases), and medical history. Diagnostic labels were independently assigned by a cardiologist and an emergency medicine physician, achieving substantial interrater agreement (Cohen's kappa = 0.82). This resource offers unique value through its dense integration of ECG data with clinical covariates, enabling multimodal fusion research, unsupervised clustering, imbalanced learning, and predictive analytics.
No takes yet. Share an insight, caveat, or question.
moattar et al. (2026) conducted an observational in Cardiovascular and non-cardiac emergency department presentations (n=582). Multimodal clinical metadata and 12-lead ECG image curation was evaluated on Interrater agreement for ECG diagnostic labeling between a cardiologist and an emergency medicine physician (Cohen's kappa = 0.82).
Synapse has enriched 10 related papers. Consider them for comparative context: