ECG-GPT

ECG-GPT generates free-text, expert-level diagnostic interpretations directly from 12-lead ECG images using a vision encoder-decoder architecture.

Key Features:

• Format-independent image analysis: The vision encoder-decoder model processes ECG images regardless of layout or format, eliminating the need for signal data or device integration.
• Large-scale training and validation: The model was developed using 2.9 million ECGs and externally validated across 3.8 million additional ECGs from 7 geographically distinct health settings including US health systems, Brazil (Minas Gerais), UK Biobank, Germany (PTB-XL), and a Missouri community hospital.
• Comprehensive diagnostic output: The model generates full-text interpretation statements that identify the complete context of diagnosis statements with allied conditions.
• Multi-condition diagnostic performance: Across 26 extracted diagnostic labels, the model achieved AUROCs of 0.80-0.95 and AUPRCs of 0.50-0.86 for atrial fibrillation, sinus tachycardia, sinus bradycardia, premature atrial contractions, and premature ventricular contractions; AUROCs of 0.88-0.96 and AUPRCs of 0.23-0.86 for left bundle branch block, right bundle branch block, first degree atrioventricular block, left anterior fascicular block, and left posterior fascicular block; and diagnostic accuracy ranging from 0.93-0.99 across all conditions.
• Semantic similarity validation: The model achieved a median pairwise cosine similarity of 0.90 (IQR 0.83-0.97) for generated interpretations compared to expert statements.

Scientific Applications:

• Automated ECG interpretation: ECG-GPT enables automated, expert-level diagnostic reporting for clinical triage and management of patients with cardiac conditions.
• Low-resource settings expansion: The image-based approach provides accessible ECG analysis without requiring proprietary signal acquisition systems or rule-based algorithms.
• Cross-population clinical validation: The model supports reliable ECG interpretation across demographically and geographically diverse patient populations and care settings.

Methodology:

ECG-GPT uses a vision-text transformer model with an encoder-decoder architecture trained on 12-lead ECG images and corresponding diagnosis statements from the Yale-New Haven Health System (2000-2022). Validation employed structured clinical assessment, semantic similarity metrics, and conventional natural language generation metrics across 7 geographically distinct health settings.