Deep biomarkers of aging are population-dependent

just published an excellent article showing how machine learning methods (specifically deep neural networks, DNNs) can be used to quantify the aging process using a set of 41 standard clinical biomarkers, most of which are not specifically recognized as biomarkers of aging. DNNs provide a method to obtain a predictive algorithm from raw data (the biomarkers in this case) with minimal to no a priori assumptions (see Mamoshina et al. 2016 [2] for details). This is an important finding because (a) it confirms that aging is not a single specific process, but rather a suite of changes that are felt across multiple physiological systems, probably within a complex systems framework , and (b) it suggests that measurement of the aging process is feasible with simple, standard measures. Both of these agree with recent findings from our lab showing that similar sets of biomarkers perform well for measurement of physiological dysregulation [3-7]. The difference is that our models are geared toward understanding the biology, and Putin et al. [1]'s toward prediction (i.e., estimation of biological age, though they do not use the term). Their model substantially outperforms ours for age prediction, but because the underlying algorithm is sufficiently complex as to remain a black box, it can provide relatively little insight into mechanisms. The two approaches are thus complementary. There is, however, a substantial caveat to Putin et al. [1]'s approach that was not mentioned in their article. Their algorithm was developed based on clinical data from a single source covering Eastern Europe (90% Russia), and the applicability to data from other settings or to population subsets was not verified. There are a number of reasons to suspect that their algorithm would need to be adjusted for application in other settings: (1) Aging rates may differ across countries; (2) Genetic and environmental determinants of physiology may differ across countries/cultures, independent of aging; and (3) There may be specific biases in how clinical lab samples are taken and analyzed that differ substantially across health systems. These distinctions are not trivial: a universal measure of biological age has very different practical and biological implications than one that is highly contextual. They also represent a more general challenge for machine learning in the health domain: traditional applications of such techniques (e.g. facial Letters to the Editor recognition, sentence completion [2]) are not generally subject to bias or anything related to the epidemiological concept of …