Key result
ECG transformer detects MI with ~91% accuracy, but retinal model underperforms standard risk factors.
Why the study?
Claims of near-perfect accuracy from multimodal deep learning on retinal fundus images and ECGs for cardiovascular risk assessment are frequently confounded by data leakage and evaluation on datasets lacking the stated modalities.
Does a transformer-based multimodal framework using retinal fundus images and ECG signals improve cardiovascular risk assessment compared to demographic and machine-learning baselines under a leakage-free protocol?
Observational
Yes
Does a transformer-based multimodal framework using retinal fundus images and ECG signals improve cardiovascular risk assessment compared to demographic and machine-learning baselines under a leakage-free protocol?
Answers are grounded in this paper and cite their sources.
Under a strict leakage-free evaluation protocol, deep learning on 12-lead ECGs accurately detects myocardial infarction, whereas retinal fundus imaging fails to outperform basic demographic variables (age and sex) for predicting carotid intima-media thickness risk.
Background: Deep learning on retinal fundus images and on the electrocardiogram (ECG) has each shown promise for cardiovascular risk assessment, and multimodal fusion is often claimed to reach near-perfect accuracy. Such claims are frequently confounded by data leakage and by evaluation on data that does not actually contain the stated modalities. Methods: We re-implemented a transformer-based multimodal framework and evaluated it under a strict leakage-free protocol on two public datasets: PTB-XL (21,799 twelve-lead ECGs; myocardial-infarction vs. normal, 1513 held-out test ECGs on the official patient-respecting fold) and China-Fundus-CIMT (2,903 patients, carotid intima-media-thickness-defined risk; three-fold patient-level cross-validation). A spectro-temporal ECG transformer and a retinal vision transformer were trained independently and combined by decision-level fusion; all preprocessing, model selection and thresholds were confined to training folds. Each modality was compared with strong baselines. Results: The ECG transformer reached 90.75% accuracy and an AUC of 0.970 (sensitivity 88.91%, specificity 91.80%), exceeding logistic-regression, random-forest and gradient-boosting baselines (AUC 0.835-0.869). The retinal transformer reached 71.34+/-4.57% accuracy and an AUC of 0.764+/-0.016; notably, an age-plus-sex logistic baseline (AUC 0.854) outperformed the image model, indicating limited added value of the fundus images for this label. Because the two cohorts share no patients, fusion could only be studied by simulation and did not constitute per-patient validation. Conclusions: Under rigorous evaluation, ECG-based MI detection is strong and reproducible, whereas retinal CIMT-risk prediction is modest and does not beat demographics, and genuine multimodal fusion will be investigated in future research.
No takes yet. Share an insight, caveat, or question.
Jinkathoti et al. (2026) conducted an observational in Cardiovascular risk (myocardial infarction and carotid intima-media-thickness-defined risk). Spectro-temporal ECG transformer and retinal vision transformer vs. Logistic-regression, random-forest, gradient-boosting, and age-plus-sex logistic baselines was evaluated on Diagnostic accuracy and AUC for myocardial infarction detection (ECG) and carotid intima-media-thickness-defined risk prediction (retinal images) (AUC 0.970 (ECG); AUC 0.764 (retinal)). The ECG transformer achieved 90.75% accuracy for MI detection, whereas the retinal transformer for CIMT risk underperformed an age-plus-sex baseline.
Synapse has enriched 10 related papers. Consider them for comparative context: