Heart recordings predict future blocks in electrical pathways
One-year prediction scores distinguishing future cases from noncases were 0.910 for left-sided block and 0.859 for right-sided block in a separate database.
Our read
Biomarkers- Studied:
- Cohort study of patients, up to 5 years
- What changed:
- AUROCs for 0-1-year prediction were 0.910 for LBBB and 0.859 for RBBB in external testing
Editor of record: Boris Djordjevic · AgingResearch.News
Patients’ digital, 10-second heart recordings were analyzed in a machine-learning prediction study of bundle branch block, a delay or blockage in the heart’s electrical pathways. The models used standard 12-lead electrocardiograms, which record the heart’s electrical activity. They excluded recordings with incomplete, ambiguous, current or previous complete block.
The team trained separate models on 1.77 million recordings for left-sided block and 1.72 million for right-sided block. Testing in another database used 81,945 and 81,812 recordings, respectively. The models predicted newly documented complete block up to five years ahead. Adding clinical information and routine recording measurements did not materially improve prediction. Among patients without previous or same-admission heart failure, five-year heart failure incidence was 48.6% in the highest predicted left-sided block risk third versus 5.5% in the lowest.
Why it matters
Heart failure is an age-related disease for which earlier risk assessment may matter. The authors suggest these electrical warning signs could help guide earlier evaluation of heart structure and planning for worsening conduction.
Caveats
Database prediction results do not show whether using the models would improve care. The conference abstract does not give patient ages or the number of unique patients, limiting assessment of whom the results apply to.
- Deep learning
- Bundle-branch block
- Heart failure
- Incident complete left bundle branch block
- Incident complete right bundle branch block
- Humans
The paper
Machine Learning-Derived 12-Lead ECG Prediction of Incident Bundle Branch Block
University of North Carolina at Chapel Hill
AHA Scientific Sessions 2026 Abstract, Nov 2, 2026
Presented at AHA Scientific Sessions 2026, abstract 18733
Comments
Be the first to comment.