Repeated heart recordings improved AI risk prediction in patients
Across three hospital cohorts, models using repeated heart recordings outperformed a single-recording model for 11 of 13 outcomes.
Our read
Biomarkers- Studied:
- Cohort study of patients, against a single-recording model (Independent replication, multicenter)
- What changed:
- Models using repeated heart recordings outperformed a single-recording model for 11 of 13 outcomes across three hospital cohorts
- The catch:
- Predictions were assessed, but health benefits from model-guided decisions were not measured
- Why it’s aging research:
- It bears on staying healthy in midlife and later life.
Editor of record: Boris Djordjevic · AgingResearch.News
Hospital records from 189,539 patients at Beth Israel Deaconess Medical Center supported an observational study that developed artificial intelligence risk models. The models used age, sex and up to ten electrocardiograms, recordings of the heart’s electrical activity. Researchers validated them in separate hospital cohorts of 1.3 million and 350,000 patients.
Compared with a model using only the latest recording, models using repeated recordings better predicted 11 of 13 outcomes, including death and stroke, across all three cohorts. Benefits increased with more past recordings and were largest in people aged 75 or older. At the development hospital, a separate model ranked ten-year cardiovascular risk better than an existing risk calculator among eligible outpatients.
Why it matters
For cardiovascular prevention as people age, the work addresses whether a patient’s history of heart recordings can add useful information beyond the latest test.
Caveats
The comparison with PREVENT used the model-development cohort, not an independent external cohort. The observational analysis assessed predictions, not health benefits from model-guided decisions, and was reported in a conference abstract.
The paper
Imperial College London
AHA Scientific Sessions 2026 Abstract, Nov 2, 2026
Presented at AHA Scientific Sessions 2026, abstract 9962
Comments
Be the first to comment.