FPT Software AI Center / University of Arkansas (Nguyen et al.)
ECG foundation model designed to remain accurate when leads are missing or signals are noisy. A 1D ConvNeXt V2 encoder is trained with a dual-mode self-distillation objective (separate lead-missing and noise "teachers") alongside contrastive alignment to detailed diagnostic-criteria text reports retrieved via a lightweight, LLM-free "Cardiac Feature Retrieval" module. Consistently ranks best or second-best across PTB-XL diagnostic tasks and MIT-BIH arrhythmia classification under original, noisy, lead-missing, and combined-corruption conditions. Developed by FPT Software AI Center and the University of Arkansas.
Architecture
CNN (1D)
1D ConvNeXt V2 ECG encoder trained via dual-mode self-distillation (lead-missing + noise teachers) plus CLIP-style contrastive alignment with retrieval-augmented diagnostic text reports (Cardiac Feature Retrieval)
Framework
PyTorch
Added to catalog
2026-08-10
CC BY-NC-SA 4.0
License for model weights only. Associated code may be licensed seperately, check code source for specific terms.
800,035 12-lead ECG-report pairs from 161,352 subjects at Beth Israel Deaconess Medical Center.
48 half-hour excerpts of two-channel ambulatory ECG recordings from 47 subjects studied by the BIH Arrhythmia Laboratory; used as a cross-dataset, lead-missing/noise-robustness benchmark distinct from the pretraining data.
52% male / 48% female; age range 0-95 (median ~62). 21,837 10-second 12-lead ECG records.
PTB-XL Super-Diagnostic multi-label classification (5 diagnostic superclasses)
ECG representation embedding, robust to lead-missing and noise conditions
MIT-BIH Arrhythmia 5-class beat classification (Normal, LBBB, RBBB, PAC, PVC)