AI heart-test predictions varied within hospitalized patients
Disagreement rates across repeated heart recordings ranged from 14.3% to 24.5% for disease prediction tasks.
Our read
Biomarkers- Studied:
- Cohort study of hospitalized patients, 5-6 days (Multicenter)
- What changed:
- Disagreement rates across repeated heart recordings ranged from 14.3% to 24.5% for disease prediction tasks
- The catch:
- Only hospital stays with at least five recordings were analyzed
Editor of record: Boris Djordjevic · AgingResearch.News
Hospitalized patients with at least five electrocardiograms—recordings of the heart’s electrical activity—were studied in an observational analysis of nine artificial intelligence prediction models. The analysis included 9,498 groups of recordings at Beth Israel Deaconess Medical Center, 187,092 at Massachusetts General Hospital and 13,035 at Emory University Hospital. Each group came from one hospital stay, with no more than five days between successive recordings.
The models predicted age, sex, several diseases and risk of death. Age predictions were the most reproducible; diabetes and 30-day mortality predictions were the least reproducible. Greater variation in predicted age was linked to higher five-year mortality across all three hospitals. Greater variation in predicted sex was linked to lower mortality. These links remained after accounting for age, sex, other illnesses and each group’s average prediction.
Why it matters
For disease screening, the authors emphasize caution when relying on a single recording. They recommend reporting reproducibility when evaluating these models for clinical use.
Caveats
Only hospital stays with at least five recordings were analyzed; reliability during routine outpatient screening was not tested. The results were 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 9994
Comments
Be the first to comment.