BrainAGE revisited: Keeping up the pace with deep learning methods

Abstract

AbstractSince the introduction of the BrainAGE method (Franke et al., 2010), novel machine learning methods of brain age prediction have continued to emerge. The idea of estimating the chronological age from (different modalities of) magnetic resonance images proved to be an interesting field of research due to the simplicity of its interpretation and its potential use as a biomarker of brain health. With the advancement of deep learning methods that are starting to dominate the field, it may seem like standard machine learning methods are not competitive anymore.Here we present our revised BrainAGE framework that combines multiple weighted BrainAGE models. We employed the analyses on a sample of 36840 T1-weighted images from the UK Biobank. The most accurate prediction with a mean absolute error (MAE) of years was obtained by relevance vector regression weighted model combination. Our results are comparable to those of deep learning methods. We additionally tested the relation of different BrainAGE models and white matter hyperintensities as markers of ageing brain and found no differences in association to biomedical data between models of differing accuracies.HighlightsStandard machine learning (BrainAGE) algorithms can outperform deep learning methods.Combining machine learning models can help increase accuracy and robustness of predictions.Higher accuracy of BrainAGE prediction does not necessarily imply higher/lower association with biomedical data.