Comprehensive Summary
Yang et al. developed a deep-learning model that estimates hemoglobin levels directly from standard 12-lead ECGs (“ECG-Hb”) and examined whether these ECG-derived estimates are associated with long-term outcomes. The model was trained on a very large retrospective dataset (388,166 ECGs from 187,202 patients) and evaluated in both internal and external validation cohorts. ECG-Hb showed a moderate correlation with laboratory hemoglobin values (r: 0.56 internally and 0.53 externally). Further, the model performed reasonably well for identifying moderate-to-severe anemia, with AUCs of 0.8545 in the internal validation cohort and 0.8243 in the external validation cohort, and demonstrated improved discrimination for severe anemia (AUC: 0.9038 internally and 0.8766 externally). Model interpretability and analyses suggested the heart-rate variability metrics and repolarization features, particularly QT-interval-related measures, contributed most to hemoglobin estimation. Over a follow-up of up to 8 years, patients with severely low ECG-Hb had substantially higher risks of all-cause mortality and new-onset heart failure, even after adjustment for age, sex, and measured hemoglobin.
Outcomes and Implications
This work suggests that ECG-derived hemoglobin estimates capture clinically meaningful physiologic information and may function as a low-cost screening or risk-stratification signal, especially in settings where blood testing is limited. Importantly, ECG-Hb should not be viewed as a replacement for laboratory hemoglobin: correlations are modest, positive predictive values are low, and abnormal ECG-Hb likely reflects broader cardiovascular stress rather than anemia alone. The retrospective design, reliance on hospital-based cohorts, and lack of prospective clinical decision testing limit immediate clinical adoption. Future studies will need to clarify whether ECG-Hb primarily identifies anemia or cardiovascular vulnerability as well as whether acting on these signals improves patient outcomes.