TL;DR
MyoECG is a cross-modal deep learning framework that "teaches" standard 12-lead ECGs to recognize the structural and functional signatures typically only visible on Cardiac MRI (CMR). By distilling CMR-derived phenotypes into ECG representations, the model achieves state-of-the-art cardiomyopathy detection and subtype classification using only low-cost, widely available electrical signals.
What Problem Does This Paper Address?
Cardiomyopathies—diseases of the heart muscle like dilated (DCM) or hypertrophic (HCM) cardiomyopathy—are leading causes of heart failure and sudden death, yet they often remain undiagnosed until a major clinical event occurs.
The clinical "gold standard" for diagnosis is Cardiac MRI (CMR), which provides exquisite detail on heart structure and function. However, CMR is expensive, time-consuming, and inaccessible for population-scale screening. In contrast, the 12-lead electrocardiogram (ECG) is ubiquitous and inexpensive but primarily reflects electrical activity. While AI has previously been used to detect heart disease from ECGs, these models often act as "black boxes" that rely on dataset-specific shortcuts rather than grounded biological features, limiting their clinical trust and generalizability.
Methods
The authors developed MyoECG using a teacher-student representation learning strategy to bridge the gap between electrical signals and physical structure.
- CMR Teacher Construction: A CMR encoder was trained to predict 27 quantitative phenotypes (e.g., ventricular volumes, wall thickness, ejection fraction) from four-chamber cine MRI sequences.
- Cross-Modal Alignment (Pretraining): Using paired ECG-CMR data, an ECG encoder (based on the ST-MEM transformer architecture) was trained to align its latent representation with the CMR teacher. This involved three loss functions:
- Semantic Distillation: Forcing ECG embeddings to be close to CMR embeddings in latent space.
- Phenotype Regression: Training the ECG encoder to directly predict the 27 CMR-derived phenotypes.
- Relational Alignment: Ensuring that if two patients have similar heart structures on MRI, their ECG embeddings are also similar.
- Supervised Fine-Tuning: The "CMR-informed" ECG encoder was then fine-tuned on clinical labels to detect cardiomyopathy and classify it into three subtypes: DCM, HCM, and restrictive cardiomyopathy (RCM).
Datasets
The study utilized a massive multi-centre dataset comprising over 1.8 million ECGs:
- Paired Training: 45,241 paired ECG-CMR examinations from the UK Biobank (UKB) and Fuwai Central China Cardiovascular Hospital (FCCCH).
- External Validation: Six independent datasets including the Harvard–Emory ECG Database (HEEDB), MIMIC-IV-ECG, and several large Chinese hospital cohorts (SRRSH, FAHZU, SAHZU).
- Reader Study: A dedicated dataset (FCCCH-RS) of 403 ECGs used to evaluate human-AI collaboration.
Key Insights and Ideas
- Biologically Grounded "Vision": Unlike standard ECG-AI, MyoECG is explicitly "informed" by anatomy. Saliency maps (Figure 4b) show the model focuses on the QRS complex and T-wave—regions known to reflect ventricular depolarization and repolarization—rather than random noise.
- The "Genetic Ablation" Proof: To prove the model isn't just finding shortcuts, the authors performed a GWAS on the MyoECG score. They found strong genetic correlations with cardiomyopathy. Crucially, when they adjusted the analysis for the 27 CMR traits, this correlation vanished. This confirms that the model’s predictive power comes specifically from the structural/functional information it "inherited" from the MRI teacher.
- Superior Generalizability: By learning fundamental biological representations rather than just fitting labels, MyoECG maintained high performance across seven different datasets with vastly different disease prevalences (0.2% to 54.7%).
Key Results
- Detection Performance: MyoECG achieved AUROCs ranging from 0.800 to 0.974 for cardiomyopathy detection, consistently outperforming five state-of-the-art ECG foundation models (like ECGFounder and ST-MEM).
- Subtype Classification: For distinguishing between DCM, HCM, and RCM, the model achieved a macro-AUROC of 0.941 in internal testing and remained robust in external cohorts.
- Clinical Utility: In a reader study with six cardiologists, AI assistance significantly boosted diagnostic accuracy. Mean sensitivity increased by 0.109 and specificity by 0.056.
- Biological Links: The MyoECG score was associated with 69 circulating proteins (including the heart stress marker NT-proBNP) and identified four independent genomic loci (including NFIA and KCNQ1) linked to heart disease.
Limitations and Future Work
- Retrospective Nature: All evaluations were performed on historical data. Prospective clinical trials are needed to see if this actually improves patient outcomes in the real world.
- Limited MRI Input: The CMR teacher only used "cine" (motion) sequences. It did not incorporate late gadolinium enhancement (LGE), which is used to detect heart scarring (fibrosis). Adding this could further improve detection of specific disease stages.
- Low PPV in General Population: While MyoECG is much better than previous models, the Positive Predictive Value (PPV) remains low in very low-prevalence settings (like the general public). The authors suggest it is best used as a "triage" tool to prioritize patients for follow-up imaging rather than a definitive diagnostic test.
Summary updated