Approaches to age bias correction

This commentary refers to ‘Sex-specific body fat distribution predicts cardiovascular ageing’, by V. Losev et al., https://doi.org/10.1093/eurheartj/ehaf553 and the discussion piece ‘When bias correction becomes bias: lessons for cardiovascular age models’, by M. Lebowitz et al., https://doi.org/10.1093/eurheartj/ehaf800. We thank the correspondents for their interest in our work and the underlying methodology. We are also pleased that providing open-source code enables transparent reporting of the analysis. The model was developed on healthy adults using a conventional partitioning of the data for training, tuning, and evaluation. This was then used to assess deviation from healthy aging in the remaining population. Using these methods, there is an inevitable symmetrical bias with over-estimation of age for younger individuals and under-estimation in older individuals. Different approaches have been developed to correct this bias—but whatever method is used, the results are almost identical whether using an independent dataset or not for fitting the parameters.1 Also in our paper, we report effect sizes for linear regression models which include age and age2 as covariates to adjust for linear and quadratic relationships with chronological age in the population of interest. There is ongoing research in this field to derive reliable estimates of age deltas that correct for different biases during model training.