AHA Scientific Sessions 2026 Abstract · Nov 2, 2026
Acute Changes in Machine Learning Estimated ECG Age After Hospital Admission in the MIMIC-IV Cohort
By Parker W, Vellanki B, Ahmad S et al.
Background: Delta ECG age is a metric calculated by subtracting a patient's chronological age from the age estimated by a machine learning algorithm applied to the patient’s ECG. A positive delta ECG age has been associated with an increased risk of mortality, heart failure, and stroke. However, the factors associated with acute fluctuations in delta ECG age remain unknown. We investigated whether acute events leading to hospital admission are associated with acute changes in delta ECG age. Methods: This retrospective cohort study used the MIMIC-IV ECG-linked cohort, a de-identified single-center dataset of emergency department and ICU encounters in Boston, MA, from 2008-2022. ECG age was estimated using the Resnet-based CODE model. Delta ECG age was defined as the predicted ECG age minus the actual age at the time of the ECG. Delta-delta ECG age was defined as post-admission minus pre-admission delta ECG age, using the nearest ECG 7-365 days before admission and the ECG closest to 24 hours after admission within a ±12-hour window. We applied a linear age-bias correction to delta ECG age by regressing delta ECG age on actual age and subtracting the fitted age-dependent offset. Among admissions with a single incident diagnosis (HF, STEMI, hip fracture, or gastroenteritis), we compared delta-delta ECG age across diagnosis groups using one-way ANOVA and adjusted subject-level random-intercept mixed models with hip fracture as the reference group. Results: Among 9,927 admissions from 6,270 subjects, mean±SD delta-delta ECG age was -0.52±8.97 years for HF (n=9,176), 0.47±9.01 for STEMI (n=498), -0.23±9.86 for hip fracture (n=102), and -0.02±9.57 for gastroenteritis (n=151); corresponding bias-corrected values were -0.14±8.96, 0.81±9.01, 0.10±9.85, and 0.33±9.60. Diagnosis groups did not differ by ANOVA for raw (p=0.104) or bias-corrected (p=0.131) delta-delta ECG age. In adjusted mixed models versus hip fracture, HF was not associated with delta-delta ECG age in raw or bias-corrected analyses (raw and corrected β=-0.12 years, 95% CI -1.63 to 1.39) nor was STEMI (raw and corrected β=-0.28, 95% CI -1.92 to 1.37). Conclusion: In the MIMIC-IV ECG cohort, hospitalization for HF or STEMI was not associated with clinically meaningful acute shifts in ECG-age gap, and findings were unchanged after ECG-age bias correction. These results suggest that ECG-predicted age is relatively stable across selected acute illness states.