Summary
We carried out a systematic investigation of supervised learning techniques for biological age modeling. The biological aging acceleration is associated with the remaining health- and life-span. Artificial Deep Neural Networks (DNN) could be used to reduce the error of chronological age predictors, though often at the expense of the ability to distinguish health conditions. Mortality and morbidity hazards models based on survival follow-up data showed the best performance. Alternatively, logistic regression trained to identify chronic diseases was shown to be a good approximation of hazards models when data on survival follow-up times were unavailable. In all models, the biological aging acceleration was associated with disease burden in persons with diagnosed chronic age-related conditions. For healthy individuals, the same quantity was associated with molecular markers of inflammation (such as C-reactive protein), smoking, current physical, and mental health (including sleeping troubles, feeling tired or little interest in doing things). The biological age thus emerged as a universal biomarker of age, frailty and stress for applications involving large scale studies of the effects of longevity drugs on risks of diseases and quality of life.
To be published as Chapter 2 in “Biomarkers of aging”, ed. A. Moskalev, Springer, 2019.
