AI-Predicted ECG Age Gap as a Biomarker: Bias-Adjusted Correlation with Mortality and Cardiovascular Risk Factors

Abstract Background Artificial intelligence (AI) models can estimate a person’s age from ECG. The gap between the predicted ECG age and chronological age, predicted age deviation (PAD), has been associated with cardiovascular risk factors and mortality. However, regression bias causes PAD to correlate with chronological age itself, potentially distorting these associations. Objectives To investigate the bias introduced by age on PAD by comparing associations between PAD and a bias-corrected PAD (PADbc) with cardiovascular risk factors and survival outcomes. Methods ECG and cardiovascular risk data from Ziekenhuis Oost-Limburg (2002-2023) were linked to mortality data from the Belgian National Registry. A neural network was trained to predict age from ECGs. PADbc corresponded to the residual of PAD regressed on chronological age. Associations with risk factors were tested using chi-squared and ANOVA. Survival was analyzed with Kaplan-Meier curves and Cox proportional hazards models. Results We included 1,258,993 ECGs from 234,586 patients, split 40:10:50 into training-, validation-, and testsets by patient. In the testset (mean age 56.4 ± 16.9 years, MAE 7.9), PAD correlated with age (r = -0.54) and showed inverse associations with most risk factors, conversely, higher PADbc (r = 0.00) was associated with higher prevalence of risk factors. Kaplan–Meier revealed PADbc above its MAE was linked to lower survival, whereas for PAD showed the opposite. Multivariate Cox showed each one-year increase in both PAD or PADbc was associated with a 1.4% increased mortality hazard. Conclusion PADbc is associated with cardiovascular risk factors and mortality, offering an age-independent biomarker of biological aging.