A heart-test model’s false alarms linked to later heart failure
In Alberta’s full test cohort, the model’s false positives had an adjusted hazard ratio of 4.19 for subsequent heart failure versus correct negatives.
159,608 participants, 5 years, independent replication
Limit: The study did not measure whether using the model's predictions improved care
Edited by Boris Djordjevic · AgingResearch.News
In an observational cohort study, researchers evaluated a heart failure model using 297,773 electrocardiograms—recordings of the heart’s electrical activity—from 97,631 patients in Alberta, Canada. False positives were patients the model flagged despite no concurrent physician-confirmed heart failure diagnosis. Researchers tracked subsequent heart failure and death.
In the full Alberta cohort, false positives had an age- and sex-adjusted hazard ratio of 4.19 for subsequent heart failure versus correct negatives, a relative comparison of event rates over time. Among patients with no prior or current heart failure, the ratio was 3.88. Their absolute excess risk versus correct negatives was 5.35 percentage points at one year, 9.17 at three years and 11.89 at five years. In the independent MIMIC-IV dataset of 159,608 patients, the adjusted hazard ratio was 2.61 among those without prior or current heart failure.
Why it matters
For heart failure prediction, the researchers suggested that examining diagnostic errors may improve how cardiovascular models are interpreted and trusted.
Caveats
This observational study assessed associations, not whether using the model’s predictions improved care. The findings were reported in a conference abstract.
The paper
Rethinking false positives: can ECG-based diagnostic ML models predict future heart failure risk?
University of Alberta; Canadian Vigour Center
ESC Congress 2026 Abstract, Oct 26, 2026
Presented at ESC Congress 2026
